diff --git "a/multimodal/L2048-w8/CuaS1Decoder_part1.mlmodelc/model.mil" "b/multimodal/L2048-w8/CuaS1Decoder_part1.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/multimodal/L2048-w8/CuaS1Decoder_part1.mlmodelc/model.mil" @@ -0,0 +1,11973 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.2"}})] +{ + func main(tensor cos, tensor hidden, tensor sin) { + int32 var_34 = const()[name = string("op_34"), val = int32(-2)]; + int32 var_63 = const()[name = string("op_63"), val = int32(-1)]; + fp16 var_62_promoted_to_fp16 = const()[name = string("op_62_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_72_cast_fp16 = pow(x = hidden, y = var_62_promoted_to_fp16)[name = string("op_72_cast_fp16")]; + tensor var_74_axes_0 = const()[name = string("op_74_axes_0"), val = tensor([-1])]; + bool var_74_keep_dims_0 = const()[name = string("op_74_keep_dims_0"), val = bool(true)]; + tensor var_74_cast_fp16 = reduce_mean(axes = var_74_axes_0, keep_dims = var_74_keep_dims_0, x = var_72_cast_fp16)[name = string("op_74_cast_fp16")]; + fp16 var_75_to_fp16 = const()[name = string("op_75_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_76_cast_fp16 = add(x = var_74_cast_fp16, y = var_75_to_fp16)[name = string("op_76_cast_fp16")]; + fp32 var_77_epsilon_0 = const()[name = string("op_77_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_77_cast_fp16 = rsqrt(epsilon = var_77_epsilon_0, x = var_76_cast_fp16)[name = string("op_77_cast_fp16")]; + tensor x_1_cast_fp16 = mul(x = hidden, y = var_77_cast_fp16)[name = string("x_1_cast_fp16")]; + tensor var_80_to_fp16 = const()[name = string("op_80_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor x_3_cast_fp16 = mul(x = x_1_cast_fp16, y = var_80_to_fp16)[name = string("x_3_cast_fp16")]; + tensor layers_0_linear_attn_in_proj_qkv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20976832))))[name = string("layers_0_linear_attn_in_proj_qkv_weight_to_fp16_quantized")]; + tensor linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20993280)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_linear_attn_in_proj_qkv_weight_to_fp16_quantized, x = x_3_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 3, 0])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_1_to_fp16 = const()[name = string("const_1_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = linear_0_cast_fp16)[name = string("transpose_55")]; + tensor input_3_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")]; + int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(8192)]; + tensor input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor([1])]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0])]; + tensor input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor([1])]; + tensor layers_0_linear_attn_conv1d_weight_to_fp16 = const()[name = string("layers_0_linear_attn_conv1d_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21009728)))]; + tensor input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_0_linear_attn_conv1d_weight_to_fp16, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor var_115_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_115_cast_fp16")]; + tensor qkv_1_perm_0 = const()[name = string("qkv_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_117_begin_0 = const()[name = string("op_117_begin_0"), val = tensor([0, 0, 0])]; + tensor var_117_end_0 = const()[name = string("op_117_end_0"), val = tensor([1, 2048, 2048])]; + tensor var_117_end_mask_0 = const()[name = string("op_117_end_mask_0"), val = tensor([true, true, false])]; + tensor qkv_1_cast_fp16 = transpose(perm = qkv_1_perm_0, x = var_115_cast_fp16)[name = string("transpose_54")]; + tensor var_117_cast_fp16 = slice_by_index(begin = var_117_begin_0, end = var_117_end_0, end_mask = var_117_end_mask_0, x = qkv_1_cast_fp16)[name = string("op_117_cast_fp16")]; + tensor var_118 = const()[name = string("op_118"), val = tensor([1, 2048, 16, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_118, x = var_117_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor var_120_begin_0 = const()[name = string("op_120_begin_0"), val = tensor([0, 0, 2048])]; + tensor var_120_end_0 = const()[name = string("op_120_end_0"), val = tensor([1, 2048, 4096])]; + tensor var_120_end_mask_0 = const()[name = string("op_120_end_mask_0"), val = tensor([true, true, false])]; + tensor var_120_cast_fp16 = slice_by_index(begin = var_120_begin_0, end = var_120_end_0, end_mask = var_120_end_mask_0, x = qkv_1_cast_fp16)[name = string("op_120_cast_fp16")]; + tensor var_121 = const()[name = string("op_121"), val = tensor([1, 2048, 16, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_121, x = var_120_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor var_123_begin_0 = const()[name = string("op_123_begin_0"), val = tensor([0, 0, 4096])]; + tensor var_123_end_0 = const()[name = string("op_123_end_0"), val = tensor([1, 2048, 8192])]; + tensor var_123_end_mask_0 = const()[name = string("op_123_end_mask_0"), val = tensor([true, true, true])]; + tensor var_123_cast_fp16 = slice_by_index(begin = var_123_begin_0, end = var_123_end_0, end_mask = var_123_end_mask_0, x = qkv_1_cast_fp16)[name = string("op_123_cast_fp16")]; + tensor var_124 = const()[name = string("op_124"), val = tensor([1, 2048, 32, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_124, x = var_123_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor layers_0_linear_attn_in_proj_z_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21075328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31561152))))[name = string("layers_0_linear_attn_in_proj_z_weight_to_fp16_quantized")]; + tensor linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31569408)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_linear_attn_in_proj_z_weight_to_fp16_quantized, x = x_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_128 = const()[name = string("op_128"), val = tensor([1, 2048, 32, 128])]; + tensor z_1_cast_fp16 = reshape(shape = var_128, x = linear_1_cast_fp16)[name = string("z_1_cast_fp16")]; + tensor layers_0_linear_attn_in_proj_b_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31577664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31659648))))[name = string("layers_0_linear_attn_in_proj_b_weight_to_fp16_quantized")]; + tensor linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31659776)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_linear_attn_in_proj_b_weight_to_fp16_quantized, x = x_3_cast_fp16)[name = string("linear_2_cast_fp16")]; + tensor beta_1_cast_fp16 = sigmoid(x = linear_2_cast_fp16)[name = string("beta_1_cast_fp16")]; + tensor var_137_weight_0_to_fp16 = const()[name = string("op_137_weight_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31659904)))]; + tensor var_137_bias_0_to_fp16 = const()[name = string("op_137_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31823808)))]; + tensor var_137_cast_fp16 = linear(bias = var_137_bias_0_to_fp16, weight = var_137_weight_0_to_fp16, x = x_3_cast_fp16)[name = string("op_137_cast_fp16")]; + tensor var_138_cast_fp16 = softplus(x = var_137_cast_fp16)[name = string("op_138_cast_fp16")]; + tensor var_134_to_fp16 = const()[name = string("op_134_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31823936)))]; + tensor g_1_cast_fp16 = mul(x = var_134_to_fp16, y = var_138_cast_fp16)[name = string("g_1_cast_fp16")]; + tensor var_140_cast_fp16 = mul(x = q_1_cast_fp16, y = q_1_cast_fp16)[name = string("op_140_cast_fp16")]; + tensor var_142_axes_0 = const()[name = string("op_142_axes_0"), val = tensor([-1])]; + bool var_142_keep_dims_0 = const()[name = string("op_142_keep_dims_0"), val = bool(true)]; + tensor var_142_cast_fp16 = reduce_sum(axes = var_142_axes_0, keep_dims = var_142_keep_dims_0, x = var_140_cast_fp16)[name = string("op_142_cast_fp16")]; + fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_144_cast_fp16 = add(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")]; + fp32 var_145_epsilon_0 = const()[name = string("op_145_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_145_cast_fp16 = rsqrt(epsilon = var_145_epsilon_0, x = var_144_cast_fp16)[name = string("op_145_cast_fp16")]; + tensor q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_145_cast_fp16)[name = string("q_3_cast_fp16")]; + tensor var_147_cast_fp16 = mul(x = k_1_cast_fp16, y = k_1_cast_fp16)[name = string("op_147_cast_fp16")]; + tensor var_149_axes_0 = const()[name = string("op_149_axes_0"), val = tensor([-1])]; + bool var_149_keep_dims_0 = const()[name = string("op_149_keep_dims_0"), val = bool(true)]; + tensor var_149_cast_fp16 = reduce_sum(axes = var_149_axes_0, keep_dims = var_149_keep_dims_0, x = var_147_cast_fp16)[name = string("op_149_cast_fp16")]; + fp16 var_150_to_fp16 = const()[name = string("op_150_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_151_cast_fp16 = add(x = var_149_cast_fp16, y = var_150_to_fp16)[name = string("op_151_cast_fp16")]; + fp32 var_152_epsilon_0 = const()[name = string("op_152_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_152_cast_fp16 = rsqrt(epsilon = var_152_epsilon_0, x = var_151_cast_fp16)[name = string("op_152_cast_fp16")]; + tensor k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_152_cast_fp16)[name = string("k_3_cast_fp16")]; + tensor var_157_axes_0 = const()[name = string("op_157_axes_0"), val = tensor([3])]; + tensor var_157_cast_fp16 = expand_dims(axes = var_157_axes_0, x = q_3_cast_fp16)[name = string("op_157_cast_fp16")]; + tensor var_159_reps_0 = const()[name = string("op_159_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_159_cast_fp16 = tile(reps = var_159_reps_0, x = var_157_cast_fp16)[name = string("op_159_cast_fp16")]; + tensor var_160 = const()[name = string("op_160"), val = tensor([1, 2048, 32, 128])]; + tensor q_5_cast_fp16 = reshape(shape = var_160, x = var_159_cast_fp16)[name = string("q_5_cast_fp16")]; + tensor var_165_axes_0 = const()[name = string("op_165_axes_0"), val = tensor([3])]; + tensor var_165_cast_fp16 = expand_dims(axes = var_165_axes_0, x = k_3_cast_fp16)[name = string("op_165_cast_fp16")]; + tensor var_167_reps_0 = const()[name = string("op_167_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_167_cast_fp16 = tile(reps = var_167_reps_0, x = var_165_cast_fp16)[name = string("op_167_cast_fp16")]; + tensor var_168 = const()[name = string("op_168"), val = tensor([1, 2048, 32, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_168, x = var_167_cast_fp16)[name = string("k_5_cast_fp16")]; + fp16 var_170_to_fp16 = const()[name = string("op_170_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor q_7_cast_fp16 = mul(x = q_5_cast_fp16, y = var_170_to_fp16)[name = string("q_7_cast_fp16")]; + tensor var_172_begin_0 = const()[name = string("op_172_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_172_end_0 = const()[name = string("op_172_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_172_end_mask_0 = const()[name = string("op_172_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_172_squeeze_mask_0 = const()[name = string("op_172_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_172_cast_fp16 = slice_by_index(begin = var_172_begin_0, end = var_172_end_0, end_mask = var_172_end_mask_0, squeeze_mask = var_172_squeeze_mask_0, x = q_7_cast_fp16)[name = string("op_172_cast_fp16")]; + tensor var_173 = const()[name = string("op_173"), val = tensor([1, 0, 2])]; + tensor var_175 = const()[name = string("op_175"), val = tensor([32, 32, 64, 128])]; + tensor var_174_cast_fp16 = transpose(perm = var_173, x = var_172_cast_fp16)[name = string("transpose_53")]; + tensor q_9_cast_fp16 = reshape(shape = var_175, x = var_174_cast_fp16)[name = string("q_9_cast_fp16")]; + tensor var_177_begin_0 = const()[name = string("op_177_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_177_end_0 = const()[name = string("op_177_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_177_end_mask_0 = const()[name = string("op_177_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_177_squeeze_mask_0 = const()[name = string("op_177_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_177_cast_fp16 = slice_by_index(begin = var_177_begin_0, end = var_177_end_0, end_mask = var_177_end_mask_0, squeeze_mask = var_177_squeeze_mask_0, x = k_5_cast_fp16)[name = string("op_177_cast_fp16")]; + tensor var_178 = const()[name = string("op_178"), val = tensor([1, 0, 2])]; + tensor var_180 = const()[name = string("op_180"), val = tensor([32, 32, 64, 128])]; + tensor var_179_cast_fp16 = transpose(perm = var_178, x = var_177_cast_fp16)[name = string("transpose_52")]; + tensor k_7_cast_fp16 = reshape(shape = var_180, x = var_179_cast_fp16)[name = string("k_7_cast_fp16")]; + tensor var_182_begin_0 = const()[name = string("op_182_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_182_end_0 = const()[name = string("op_182_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_182_end_mask_0 = const()[name = string("op_182_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_182_squeeze_mask_0 = const()[name = string("op_182_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_182_cast_fp16 = slice_by_index(begin = var_182_begin_0, end = var_182_end_0, end_mask = var_182_end_mask_0, squeeze_mask = var_182_squeeze_mask_0, x = v_1_cast_fp16)[name = string("op_182_cast_fp16")]; + tensor var_183 = const()[name = string("op_183"), val = tensor([1, 0, 2])]; + tensor var_185 = const()[name = string("op_185"), val = tensor([32, 32, 64, 128])]; + tensor var_184_cast_fp16 = transpose(perm = var_183, x = var_182_cast_fp16)[name = string("transpose_51")]; + tensor v_3_cast_fp16 = reshape(shape = var_185, x = var_184_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor var_187_begin_0 = const()[name = string("op_187_begin_0"), val = tensor([0, 0, 0])]; + tensor var_187_end_0 = const()[name = string("op_187_end_0"), val = tensor([1, 2048, 32])]; + tensor var_187_end_mask_0 = const()[name = string("op_187_end_mask_0"), val = tensor([false, true, true])]; + tensor var_187_squeeze_mask_0 = const()[name = string("op_187_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_187_cast_fp16 = slice_by_index(begin = var_187_begin_0, end = var_187_end_0, end_mask = var_187_end_mask_0, squeeze_mask = var_187_squeeze_mask_0, x = beta_1_cast_fp16)[name = string("op_187_cast_fp16")]; + tensor var_188_perm_0 = const()[name = string("op_188_perm_0"), val = tensor([1, 0])]; + tensor var_189 = const()[name = string("op_189"), val = tensor([32, 32, 64, 1])]; + tensor var_188_cast_fp16 = transpose(perm = var_188_perm_0, x = var_187_cast_fp16)[name = string("transpose_50")]; + tensor beta_3_cast_fp16 = reshape(shape = var_189, x = var_188_cast_fp16)[name = string("beta_3_cast_fp16")]; + tensor var_191_begin_0 = const()[name = string("op_191_begin_0"), val = tensor([0, 0, 0])]; + tensor var_191_end_0 = const()[name = string("op_191_end_0"), val = tensor([1, 2048, 32])]; + tensor var_191_end_mask_0 = const()[name = string("op_191_end_mask_0"), val = tensor([false, true, true])]; + tensor var_191_squeeze_mask_0 = const()[name = string("op_191_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_191_cast_fp16 = slice_by_index(begin = var_191_begin_0, end = var_191_end_0, end_mask = var_191_end_mask_0, squeeze_mask = var_191_squeeze_mask_0, x = g_1_cast_fp16)[name = string("op_191_cast_fp16")]; + tensor var_192_perm_0 = const()[name = string("op_192_perm_0"), val = tensor([1, 0])]; + tensor var_193 = const()[name = string("op_193"), val = tensor([32, 32, 64])]; + tensor var_192_cast_fp16 = transpose(perm = var_192_perm_0, x = var_191_cast_fp16)[name = string("transpose_49")]; + tensor g_3_cast_fp16 = reshape(shape = var_193, x = var_192_cast_fp16)[name = string("g_3_cast_fp16")]; + tensor g_col_1_axes_0 = const()[name = string("g_col_1_axes_0"), val = tensor([-1])]; + tensor g_col_1_cast_fp16 = expand_dims(axes = g_col_1_axes_0, x = g_3_cast_fp16)[name = string("g_col_1_cast_fp16")]; + bool var_196_transpose_x_0 = const()[name = string("op_196_transpose_x_0"), val = bool(false)]; + bool var_196_transpose_y_0 = const()[name = string("op_196_transpose_y_0"), val = bool(false)]; + tensor layers_0_linear_attn_lower_incl_to_fp16 = const()[name = string("layers_0_linear_attn_lower_incl_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31824064)))]; + tensor var_196_cast_fp16 = matmul(transpose_x = var_196_transpose_x_0, transpose_y = var_196_transpose_y_0, x = layers_0_linear_attn_lower_incl_to_fp16, y = g_col_1_cast_fp16)[name = string("op_196_cast_fp16")]; + tensor cum_1_axes_0 = const()[name = string("cum_1_axes_0"), val = tensor([-1])]; + tensor cum_1_cast_fp16 = squeeze(axes = cum_1_axes_0, x = var_196_cast_fp16)[name = string("cum_1_cast_fp16")]; + bool var_198_transpose_x_0 = const()[name = string("op_198_transpose_x_0"), val = bool(false)]; + bool var_198_transpose_y_0 = const()[name = string("op_198_transpose_y_0"), val = bool(false)]; + tensor layers_0_linear_attn_suffix_sum_t_to_fp16 = const()[name = string("layers_0_linear_attn_suffix_sum_t_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31832320)))]; + tensor var_198_cast_fp16 = matmul(transpose_x = var_198_transpose_x_0, transpose_y = var_198_transpose_y_0, x = layers_0_linear_attn_suffix_sum_t_to_fp16, y = g_col_1_cast_fp16)[name = string("op_198_cast_fp16")]; + tensor to_end_1_axes_0 = const()[name = string("to_end_1_axes_0"), val = tensor([-1])]; + tensor to_end_1_cast_fp16 = squeeze(axes = to_end_1_axes_0, x = var_198_cast_fp16)[name = string("to_end_1_cast_fp16")]; + bool var_200_transpose_x_0 = const()[name = string("op_200_transpose_x_0"), val = bool(false)]; + bool var_200_transpose_y_0 = const()[name = string("op_200_transpose_y_0"), val = bool(false)]; + tensor layers_0_linear_attn_pair_sum_t_to_fp16 = const()[name = string("layers_0_linear_attn_pair_sum_t_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31840576)))]; + tensor var_200_cast_fp16 = matmul(transpose_x = var_200_transpose_x_0, transpose_y = var_200_transpose_y_0, x = layers_0_linear_attn_pair_sum_t_to_fp16, y = g_col_1_cast_fp16)[name = string("op_200_cast_fp16")]; + tensor var_201 = const()[name = string("op_201"), val = tensor([32, 32, 64, 64])]; + tensor pair_1_cast_fp16 = reshape(shape = var_201, x = var_200_cast_fp16)[name = string("pair_1_cast_fp16")]; + tensor var_203_cast_fp16 = exp(x = pair_1_cast_fp16)[name = string("op_203_cast_fp16")]; + tensor pair_decay_1_cast_fp16 = mul(x = var_203_cast_fp16, y = layers_0_linear_attn_lower_incl_to_fp16)[name = string("pair_decay_1_cast_fp16")]; + tensor k_beta_1_cast_fp16 = mul(x = k_7_cast_fp16, y = beta_3_cast_fp16)[name = string("k_beta_1_cast_fp16")]; + tensor v_beta_1_cast_fp16 = mul(x = v_3_cast_fp16, y = beta_3_cast_fp16)[name = string("v_beta_1_cast_fp16")]; + bool var_208_transpose_x_1 = const()[name = string("op_208_transpose_x_1"), val = bool(false)]; + bool var_208_transpose_y_1 = const()[name = string("op_208_transpose_y_1"), val = bool(true)]; + tensor var_208_cast_fp16 = matmul(transpose_x = var_208_transpose_x_1, transpose_y = var_208_transpose_y_1, x = k_beta_1_cast_fp16, y = k_7_cast_fp16)[name = string("op_208_cast_fp16")]; + tensor kkt_1_cast_fp16 = mul(x = var_208_cast_fp16, y = pair_decay_1_cast_fp16)[name = string("kkt_1_cast_fp16")]; + bool var_211_transpose_x_1 = const()[name = string("op_211_transpose_x_1"), val = bool(false)]; + bool var_211_transpose_y_1 = const()[name = string("op_211_transpose_y_1"), val = bool(true)]; + tensor var_211_cast_fp16 = matmul(transpose_x = var_211_transpose_x_1, transpose_y = var_211_transpose_y_1, x = q_9_cast_fp16, y = k_7_cast_fp16)[name = string("op_211_cast_fp16")]; + tensor attn_intra_1_cast_fp16 = mul(x = var_211_cast_fp16, y = pair_decay_1_cast_fp16)[name = string("attn_intra_1_cast_fp16")]; + tensor layers_0_linear_attn_strict_lower_to_fp16 = const()[name = string("layers_0_linear_attn_strict_lower_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32364928)))]; + tensor var_213_cast_fp16 = mul(x = kkt_1_cast_fp16, y = layers_0_linear_attn_strict_lower_to_fp16)[name = string("op_213_cast_fp16")]; + tensor layers_0_linear_attn_eye_to_fp16 = const()[name = string("layers_0_linear_attn_eye_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32373184)))]; + tensor t_1_cast_fp16 = add(x = layers_0_linear_attn_eye_to_fp16, y = var_213_cast_fp16)[name = string("t_1_cast_fp16")]; + tensor var_217 = const()[name = string("op_217"), val = tensor([-1, 64, 64])]; + tensor t_3_cast_fp16 = reshape(shape = var_217, x = t_1_cast_fp16)[name = string("t_3_cast_fp16")]; + tensor var_222 = const()[name = string("op_222"), val = tensor([1024, 32, 2, 32, 2])]; + tensor var_223_cast_fp16 = reshape(shape = var_222, x = t_3_cast_fp16)[name = string("op_223_cast_fp16")]; + tensor var_227_to_fp16 = const()[name = string("op_227_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32381440)))]; + tensor var_228_cast_fp16 = mul(x = var_223_cast_fp16, y = var_227_to_fp16)[name = string("op_228_cast_fp16")]; + tensor blocks_1_axes_0 = const()[name = string("blocks_1_axes_0"), val = tensor([-2])]; + bool blocks_1_keep_dims_0 = const()[name = string("blocks_1_keep_dims_0"), val = bool(false)]; + tensor blocks_1_cast_fp16 = reduce_sum(axes = blocks_1_axes_0, keep_dims = blocks_1_keep_dims_0, x = var_228_cast_fp16)[name = string("blocks_1_cast_fp16")]; + tensor var_231_begin_0 = const()[name = string("op_231_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_231_end_0 = const()[name = string("op_231_end_0"), val = tensor([1024, 32, 2, 2])]; + tensor var_231_end_mask_0 = const()[name = string("op_231_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_231_cast_fp16 = slice_by_index(begin = var_231_begin_0, end = var_231_end_0, end_mask = var_231_end_mask_0, x = blocks_1_cast_fp16)[name = string("op_231_cast_fp16")]; + tensor x_5_begin_0 = const()[name = string("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = string("x_5_end_0"), val = tensor([1024, 32, 1, 1])]; + tensor x_5_end_mask_0 = const()[name = string("x_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, x = var_231_cast_fp16)[name = string("x_5_cast_fp16")]; + bool var_241_transpose_x_0 = const()[name = string("op_241_transpose_x_0"), val = bool(false)]; + bool var_241_transpose_y_0 = const()[name = string("op_241_transpose_y_0"), val = bool(false)]; + tensor var_235_to_fp16 = const()[name = string("op_235_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32383552)))]; + tensor var_241_cast_fp16 = matmul(transpose_x = var_241_transpose_x_0, transpose_y = var_241_transpose_y_0, x = x_5_cast_fp16, y = var_235_to_fp16)[name = string("op_241_cast_fp16")]; + bool var_242_transpose_x_0 = const()[name = string("op_242_transpose_x_0"), val = bool(false)]; + bool var_242_transpose_y_0 = const()[name = string("op_242_transpose_y_0"), val = bool(false)]; + tensor var_242_cast_fp16 = matmul(transpose_x = var_242_transpose_x_0, transpose_y = var_242_transpose_y_0, x = var_235_to_fp16, y = var_241_cast_fp16)[name = string("op_242_cast_fp16")]; + fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_1_cast_fp16 = mul(x = var_242_cast_fp16, y = const_6_promoted_to_fp16)[name = string("lower_1_cast_fp16")]; + bool bottom_1_interleave_0 = const()[name = string("bottom_1_interleave_0"), val = bool(false)]; + tensor bottom_1_cast_fp16 = concat(axis = var_63, interleave = bottom_1_interleave_0, values = (lower_1_cast_fp16, var_235_to_fp16))[name = string("bottom_1_cast_fp16")]; + bool inv_3_interleave_0 = const()[name = string("inv_3_interleave_0"), val = bool(false)]; + tensor top_1_to_fp16 = const()[name = string("top_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32449152)))]; + tensor inv_3_cast_fp16 = concat(axis = var_34, interleave = inv_3_interleave_0, values = (top_1_to_fp16, bottom_1_cast_fp16))[name = string("inv_3_cast_fp16")]; + tensor var_251 = const()[name = string("op_251"), val = tensor([1024, 16, 4, 16, 4])]; + tensor var_252_cast_fp16 = reshape(shape = var_251, x = t_3_cast_fp16)[name = string("op_252_cast_fp16")]; + tensor var_256_to_fp16 = const()[name = string("op_256_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32580288)))]; + tensor var_257_cast_fp16 = mul(x = var_252_cast_fp16, y = var_256_to_fp16)[name = string("op_257_cast_fp16")]; + tensor blocks_3_axes_0 = const()[name = string("blocks_3_axes_0"), val = tensor([-2])]; + bool blocks_3_keep_dims_0 = const()[name = string("blocks_3_keep_dims_0"), val = bool(false)]; + tensor blocks_3_cast_fp16 = reduce_sum(axes = blocks_3_axes_0, keep_dims = blocks_3_keep_dims_0, x = var_257_cast_fp16)[name = string("blocks_3_cast_fp16")]; + tensor var_260_begin_0 = const()[name = string("op_260_begin_0"), val = tensor([0, 0, 2, 0])]; + tensor var_260_end_0 = const()[name = string("op_260_end_0"), val = tensor([1024, 16, 4, 4])]; + tensor var_260_end_mask_0 = const()[name = string("op_260_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_260_cast_fp16 = slice_by_index(begin = var_260_begin_0, end = var_260_end_0, end_mask = var_260_end_mask_0, x = blocks_3_cast_fp16)[name = string("op_260_cast_fp16")]; + tensor x_7_begin_0 = const()[name = string("x_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_7_end_0 = const()[name = string("x_7_end_0"), val = tensor([1024, 16, 2, 2])]; + tensor x_7_end_mask_0 = const()[name = string("x_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, x = var_260_cast_fp16)[name = string("x_7_cast_fp16")]; + tensor var_262 = const()[name = string("op_262"), val = tensor([1024, 16, 2, 2, 2])]; + tensor pair_5_cast_fp16 = reshape(shape = var_262, x = inv_3_cast_fp16)[name = string("pair_5_cast_fp16")]; + tensor var_264_begin_0 = const()[name = string("op_264_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_264_end_0 = const()[name = string("op_264_end_0"), val = tensor([1024, 16, 1, 2, 2])]; + tensor var_264_end_mask_0 = const()[name = string("op_264_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_264_squeeze_mask_0 = const()[name = string("op_264_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_264_cast_fp16 = slice_by_index(begin = var_264_begin_0, end = var_264_end_0, end_mask = var_264_end_mask_0, squeeze_mask = var_264_squeeze_mask_0, x = pair_5_cast_fp16)[name = string("op_264_cast_fp16")]; + tensor var_267_begin_0 = const()[name = string("op_267_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_267_end_0 = const()[name = string("op_267_end_0"), val = tensor([1024, 16, 2, 2, 2])]; + tensor var_267_end_mask_0 = const()[name = string("op_267_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_267_squeeze_mask_0 = const()[name = string("op_267_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_267_cast_fp16 = slice_by_index(begin = var_267_begin_0, end = var_267_end_0, end_mask = var_267_end_mask_0, squeeze_mask = var_267_squeeze_mask_0, x = pair_5_cast_fp16)[name = string("op_267_cast_fp16")]; + bool var_270_transpose_x_0 = const()[name = string("op_270_transpose_x_0"), val = bool(false)]; + bool var_270_transpose_y_0 = const()[name = string("op_270_transpose_y_0"), val = bool(false)]; + tensor var_270_cast_fp16 = matmul(transpose_x = var_270_transpose_x_0, transpose_y = var_270_transpose_y_0, x = x_7_cast_fp16, y = var_264_cast_fp16)[name = string("op_270_cast_fp16")]; + bool var_271_transpose_x_0 = const()[name = string("op_271_transpose_x_0"), val = bool(false)]; + bool var_271_transpose_y_0 = const()[name = string("op_271_transpose_y_0"), val = bool(false)]; + tensor var_271_cast_fp16 = matmul(transpose_x = var_271_transpose_x_0, transpose_y = var_271_transpose_y_0, x = var_267_cast_fp16, y = var_270_cast_fp16)[name = string("op_271_cast_fp16")]; + fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_3_cast_fp16 = mul(x = var_271_cast_fp16, y = const_7_promoted_to_fp16)[name = string("lower_3_cast_fp16")]; + bool top_3_interleave_0 = const()[name = string("top_3_interleave_0"), val = bool(false)]; + tensor var_273_to_fp16 = const()[name = string("op_273_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32580864)))]; + tensor top_3_cast_fp16 = concat(axis = var_63, interleave = top_3_interleave_0, values = (var_264_cast_fp16, var_273_to_fp16))[name = string("top_3_cast_fp16")]; + bool bottom_3_interleave_0 = const()[name = string("bottom_3_interleave_0"), val = bool(false)]; + tensor bottom_3_cast_fp16 = concat(axis = var_63, interleave = bottom_3_interleave_0, values = (lower_3_cast_fp16, var_267_cast_fp16))[name = string("bottom_3_cast_fp16")]; + bool inv_5_interleave_0 = const()[name = string("inv_5_interleave_0"), val = bool(false)]; + tensor inv_5_cast_fp16 = concat(axis = var_34, interleave = inv_5_interleave_0, values = (top_3_cast_fp16, bottom_3_cast_fp16))[name = string("inv_5_cast_fp16")]; + tensor var_280 = const()[name = string("op_280"), val = tensor([1024, 8, 8, 8, 8])]; + tensor var_281_cast_fp16 = reshape(shape = var_280, x = t_3_cast_fp16)[name = string("op_281_cast_fp16")]; + tensor var_285_to_fp16 = const()[name = string("op_285_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32712000)))]; + tensor var_286_cast_fp16 = mul(x = var_281_cast_fp16, y = var_285_to_fp16)[name = string("op_286_cast_fp16")]; + tensor blocks_5_axes_0 = const()[name = string("blocks_5_axes_0"), val = tensor([-2])]; + bool blocks_5_keep_dims_0 = const()[name = string("blocks_5_keep_dims_0"), val = bool(false)]; + tensor blocks_5_cast_fp16 = reduce_sum(axes = blocks_5_axes_0, keep_dims = blocks_5_keep_dims_0, x = var_286_cast_fp16)[name = string("blocks_5_cast_fp16")]; + tensor var_289_begin_0 = const()[name = string("op_289_begin_0"), val = tensor([0, 0, 4, 0])]; + tensor var_289_end_0 = const()[name = string("op_289_end_0"), val = tensor([1024, 8, 8, 8])]; + tensor var_289_end_mask_0 = const()[name = string("op_289_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_289_cast_fp16 = slice_by_index(begin = var_289_begin_0, end = var_289_end_0, end_mask = var_289_end_mask_0, x = blocks_5_cast_fp16)[name = string("op_289_cast_fp16")]; + tensor x_9_begin_0 = const()[name = string("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = string("x_9_end_0"), val = tensor([1024, 8, 4, 4])]; + tensor x_9_end_mask_0 = const()[name = string("x_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, x = var_289_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_291 = const()[name = string("op_291"), val = tensor([1024, 8, 2, 4, 4])]; + tensor pair_7_cast_fp16 = reshape(shape = var_291, x = inv_5_cast_fp16)[name = string("pair_7_cast_fp16")]; + tensor var_293_begin_0 = const()[name = string("op_293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_293_end_0 = const()[name = string("op_293_end_0"), val = tensor([1024, 8, 1, 4, 4])]; + tensor var_293_end_mask_0 = const()[name = string("op_293_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_293_squeeze_mask_0 = const()[name = string("op_293_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_293_cast_fp16 = slice_by_index(begin = var_293_begin_0, end = var_293_end_0, end_mask = var_293_end_mask_0, squeeze_mask = var_293_squeeze_mask_0, x = pair_7_cast_fp16)[name = string("op_293_cast_fp16")]; + tensor var_296_begin_0 = const()[name = string("op_296_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_296_end_0 = const()[name = string("op_296_end_0"), val = tensor([1024, 8, 2, 4, 4])]; + tensor var_296_end_mask_0 = const()[name = string("op_296_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_296_squeeze_mask_0 = const()[name = string("op_296_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_296_cast_fp16 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, squeeze_mask = var_296_squeeze_mask_0, x = pair_7_cast_fp16)[name = string("op_296_cast_fp16")]; + bool var_299_transpose_x_0 = const()[name = string("op_299_transpose_x_0"), val = bool(false)]; + bool var_299_transpose_y_0 = const()[name = string("op_299_transpose_y_0"), val = bool(false)]; + tensor var_299_cast_fp16 = matmul(transpose_x = var_299_transpose_x_0, transpose_y = var_299_transpose_y_0, x = x_9_cast_fp16, y = var_293_cast_fp16)[name = string("op_299_cast_fp16")]; + bool var_300_transpose_x_0 = const()[name = string("op_300_transpose_x_0"), val = bool(false)]; + bool var_300_transpose_y_0 = const()[name = string("op_300_transpose_y_0"), val = bool(false)]; + tensor var_300_cast_fp16 = matmul(transpose_x = var_300_transpose_x_0, transpose_y = var_300_transpose_y_0, x = var_296_cast_fp16, y = var_299_cast_fp16)[name = string("op_300_cast_fp16")]; + fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_5_cast_fp16 = mul(x = var_300_cast_fp16, y = const_8_promoted_to_fp16)[name = string("lower_5_cast_fp16")]; + bool top_5_interleave_0 = const()[name = string("top_5_interleave_0"), val = bool(false)]; + tensor var_302_to_fp16 = const()[name = string("op_302_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32712192)))]; + tensor top_5_cast_fp16 = concat(axis = var_63, interleave = top_5_interleave_0, values = (var_293_cast_fp16, var_302_to_fp16))[name = string("top_5_cast_fp16")]; + bool bottom_5_interleave_0 = const()[name = string("bottom_5_interleave_0"), val = bool(false)]; + tensor bottom_5_cast_fp16 = concat(axis = var_63, interleave = bottom_5_interleave_0, values = (lower_5_cast_fp16, var_296_cast_fp16))[name = string("bottom_5_cast_fp16")]; + bool inv_7_interleave_0 = const()[name = string("inv_7_interleave_0"), val = bool(false)]; + tensor inv_7_cast_fp16 = concat(axis = var_34, interleave = inv_7_interleave_0, values = (top_5_cast_fp16, bottom_5_cast_fp16))[name = string("inv_7_cast_fp16")]; + tensor var_309 = const()[name = string("op_309"), val = tensor([1024, 4, 16, 4, 16])]; + tensor var_310_cast_fp16 = reshape(shape = var_309, x = t_3_cast_fp16)[name = string("op_310_cast_fp16")]; + tensor var_314_to_fp16 = const()[name = string("op_314_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32974400)))]; + tensor var_315_cast_fp16 = mul(x = var_310_cast_fp16, y = var_314_to_fp16)[name = string("op_315_cast_fp16")]; + tensor blocks_7_axes_0 = const()[name = string("blocks_7_axes_0"), val = tensor([-2])]; + bool blocks_7_keep_dims_0 = const()[name = string("blocks_7_keep_dims_0"), val = bool(false)]; + tensor blocks_7_cast_fp16 = reduce_sum(axes = blocks_7_axes_0, keep_dims = blocks_7_keep_dims_0, x = var_315_cast_fp16)[name = string("blocks_7_cast_fp16")]; + tensor var_318_begin_0 = const()[name = string("op_318_begin_0"), val = tensor([0, 0, 8, 0])]; + tensor var_318_end_0 = const()[name = string("op_318_end_0"), val = tensor([1024, 4, 16, 16])]; + tensor var_318_end_mask_0 = const()[name = string("op_318_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, x = blocks_7_cast_fp16)[name = string("op_318_cast_fp16")]; + tensor x_11_begin_0 = const()[name = string("x_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_11_end_0 = const()[name = string("x_11_end_0"), val = tensor([1024, 4, 8, 8])]; + tensor x_11_end_mask_0 = const()[name = string("x_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, x = var_318_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_320 = const()[name = string("op_320"), val = tensor([1024, 4, 2, 8, 8])]; + tensor pair_9_cast_fp16 = reshape(shape = var_320, x = inv_7_cast_fp16)[name = string("pair_9_cast_fp16")]; + tensor var_322_begin_0 = const()[name = string("op_322_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_322_end_0 = const()[name = string("op_322_end_0"), val = tensor([1024, 4, 1, 8, 8])]; + tensor var_322_end_mask_0 = const()[name = string("op_322_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_322_squeeze_mask_0 = const()[name = string("op_322_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_322_cast_fp16 = slice_by_index(begin = var_322_begin_0, end = var_322_end_0, end_mask = var_322_end_mask_0, squeeze_mask = var_322_squeeze_mask_0, x = pair_9_cast_fp16)[name = string("op_322_cast_fp16")]; + tensor var_325_begin_0 = const()[name = string("op_325_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_325_end_0 = const()[name = string("op_325_end_0"), val = tensor([1024, 4, 2, 8, 8])]; + tensor var_325_end_mask_0 = const()[name = string("op_325_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_325_squeeze_mask_0 = const()[name = string("op_325_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_325_cast_fp16 = slice_by_index(begin = var_325_begin_0, end = var_325_end_0, end_mask = var_325_end_mask_0, squeeze_mask = var_325_squeeze_mask_0, x = pair_9_cast_fp16)[name = string("op_325_cast_fp16")]; + bool var_328_transpose_x_0 = const()[name = string("op_328_transpose_x_0"), val = bool(false)]; + bool var_328_transpose_y_0 = const()[name = string("op_328_transpose_y_0"), val = bool(false)]; + tensor var_328_cast_fp16 = matmul(transpose_x = var_328_transpose_x_0, transpose_y = var_328_transpose_y_0, x = x_11_cast_fp16, y = var_322_cast_fp16)[name = string("op_328_cast_fp16")]; + bool var_329_transpose_x_0 = const()[name = string("op_329_transpose_x_0"), val = bool(false)]; + bool var_329_transpose_y_0 = const()[name = string("op_329_transpose_y_0"), val = bool(false)]; + tensor var_329_cast_fp16 = matmul(transpose_x = var_329_transpose_x_0, transpose_y = var_329_transpose_y_0, x = var_325_cast_fp16, y = var_328_cast_fp16)[name = string("op_329_cast_fp16")]; + fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_7_cast_fp16 = mul(x = var_329_cast_fp16, y = const_9_promoted_to_fp16)[name = string("lower_7_cast_fp16")]; + bool top_7_interleave_0 = const()[name = string("top_7_interleave_0"), val = bool(false)]; + tensor var_331_to_fp16 = const()[name = string("op_331_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32974528)))]; + tensor top_7_cast_fp16 = concat(axis = var_63, interleave = top_7_interleave_0, values = (var_322_cast_fp16, var_331_to_fp16))[name = string("top_7_cast_fp16")]; + bool bottom_7_interleave_0 = const()[name = string("bottom_7_interleave_0"), val = bool(false)]; + tensor bottom_7_cast_fp16 = concat(axis = var_63, interleave = bottom_7_interleave_0, values = (lower_7_cast_fp16, var_325_cast_fp16))[name = string("bottom_7_cast_fp16")]; + bool inv_9_interleave_0 = const()[name = string("inv_9_interleave_0"), val = bool(false)]; + tensor inv_9_cast_fp16 = concat(axis = var_34, interleave = inv_9_interleave_0, values = (top_7_cast_fp16, bottom_7_cast_fp16))[name = string("inv_9_cast_fp16")]; + tensor var_338 = const()[name = string("op_338"), val = tensor([1024, 2, 32, 2, 32])]; + tensor var_339_cast_fp16 = reshape(shape = var_338, x = t_3_cast_fp16)[name = string("op_339_cast_fp16")]; + tensor var_343_to_fp16 = const()[name = string("op_343_to_fp16"), val = tensor([[[[0x1p+0], [0x0p+0]]], [[[0x0p+0], [0x1p+0]]]])]; + tensor var_344_cast_fp16 = mul(x = var_339_cast_fp16, y = var_343_to_fp16)[name = string("op_344_cast_fp16")]; + tensor blocks_9_axes_0 = const()[name = string("blocks_9_axes_0"), val = tensor([-2])]; + bool blocks_9_keep_dims_0 = const()[name = string("blocks_9_keep_dims_0"), val = bool(false)]; + tensor blocks_9_cast_fp16 = reduce_sum(axes = blocks_9_axes_0, keep_dims = blocks_9_keep_dims_0, x = var_344_cast_fp16)[name = string("blocks_9_cast_fp16")]; + tensor var_347_begin_0 = const()[name = string("op_347_begin_0"), val = tensor([0, 0, 16, 0])]; + tensor var_347_end_0 = const()[name = string("op_347_end_0"), val = tensor([1024, 2, 32, 32])]; + tensor var_347_end_mask_0 = const()[name = string("op_347_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_347_cast_fp16 = slice_by_index(begin = var_347_begin_0, end = var_347_end_0, end_mask = var_347_end_mask_0, x = blocks_9_cast_fp16)[name = string("op_347_cast_fp16")]; + tensor x_13_begin_0 = const()[name = string("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = string("x_13_end_0"), val = tensor([1024, 2, 16, 16])]; + tensor x_13_end_mask_0 = const()[name = string("x_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, x = var_347_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_349 = const()[name = string("op_349"), val = tensor([1024, 2, 2, 16, 16])]; + tensor pair_11_cast_fp16 = reshape(shape = var_349, x = inv_9_cast_fp16)[name = string("pair_11_cast_fp16")]; + tensor var_351_begin_0 = const()[name = string("op_351_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_351_end_0 = const()[name = string("op_351_end_0"), val = tensor([1024, 2, 1, 16, 16])]; + tensor var_351_end_mask_0 = const()[name = string("op_351_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_351_squeeze_mask_0 = const()[name = string("op_351_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_351_cast_fp16 = slice_by_index(begin = var_351_begin_0, end = var_351_end_0, end_mask = var_351_end_mask_0, squeeze_mask = var_351_squeeze_mask_0, x = pair_11_cast_fp16)[name = string("op_351_cast_fp16")]; + tensor var_354_begin_0 = const()[name = string("op_354_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_354_end_0 = const()[name = string("op_354_end_0"), val = tensor([1024, 2, 2, 16, 16])]; + tensor var_354_end_mask_0 = const()[name = string("op_354_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_354_squeeze_mask_0 = const()[name = string("op_354_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_354_cast_fp16 = slice_by_index(begin = var_354_begin_0, end = var_354_end_0, end_mask = var_354_end_mask_0, squeeze_mask = var_354_squeeze_mask_0, x = pair_11_cast_fp16)[name = string("op_354_cast_fp16")]; + bool var_357_transpose_x_0 = const()[name = string("op_357_transpose_x_0"), val = bool(false)]; + bool var_357_transpose_y_0 = const()[name = string("op_357_transpose_y_0"), val = bool(false)]; + tensor var_357_cast_fp16 = matmul(transpose_x = var_357_transpose_x_0, transpose_y = var_357_transpose_y_0, x = x_13_cast_fp16, y = var_351_cast_fp16)[name = string("op_357_cast_fp16")]; + bool var_358_transpose_x_0 = const()[name = string("op_358_transpose_x_0"), val = bool(false)]; + bool var_358_transpose_y_0 = const()[name = string("op_358_transpose_y_0"), val = bool(false)]; + tensor var_358_cast_fp16 = matmul(transpose_x = var_358_transpose_x_0, transpose_y = var_358_transpose_y_0, x = var_354_cast_fp16, y = var_357_cast_fp16)[name = string("op_358_cast_fp16")]; + fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_9_cast_fp16 = mul(x = var_358_cast_fp16, y = const_10_promoted_to_fp16)[name = string("lower_9_cast_fp16")]; + bool top_9_interleave_0 = const()[name = string("top_9_interleave_0"), val = bool(false)]; + tensor var_360_to_fp16 = const()[name = string("op_360_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33498880)))]; + tensor top_9_cast_fp16 = concat(axis = var_63, interleave = top_9_interleave_0, values = (var_351_cast_fp16, var_360_to_fp16))[name = string("top_9_cast_fp16")]; + bool bottom_9_interleave_0 = const()[name = string("bottom_9_interleave_0"), val = bool(false)]; + tensor bottom_9_cast_fp16 = concat(axis = var_63, interleave = bottom_9_interleave_0, values = (lower_9_cast_fp16, var_354_cast_fp16))[name = string("bottom_9_cast_fp16")]; + bool inv_11_interleave_0 = const()[name = string("inv_11_interleave_0"), val = bool(false)]; + tensor inv_11_cast_fp16 = concat(axis = var_34, interleave = inv_11_interleave_0, values = (top_9_cast_fp16, bottom_9_cast_fp16))[name = string("inv_11_cast_fp16")]; + tensor var_367 = const()[name = string("op_367"), val = tensor([1024, 1, 64, 1, 64])]; + tensor var_368_cast_fp16 = reshape(shape = var_367, x = t_3_cast_fp16)[name = string("op_368_cast_fp16")]; + tensor blocks_11_axes_0 = const()[name = string("blocks_11_axes_0"), val = tensor([-2])]; + bool blocks_11_keep_dims_0 = const()[name = string("blocks_11_keep_dims_0"), val = bool(false)]; + tensor blocks_11_cast_fp16 = reduce_sum(axes = blocks_11_axes_0, keep_dims = blocks_11_keep_dims_0, x = var_368_cast_fp16)[name = string("blocks_11_cast_fp16")]; + tensor var_376_begin_0 = const()[name = string("op_376_begin_0"), val = tensor([0, 0, 32, 0])]; + tensor var_376_end_0 = const()[name = string("op_376_end_0"), val = tensor([1024, 1, 64, 64])]; + tensor var_376_end_mask_0 = const()[name = string("op_376_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_376_cast_fp16 = slice_by_index(begin = var_376_begin_0, end = var_376_end_0, end_mask = var_376_end_mask_0, x = blocks_11_cast_fp16)[name = string("op_376_cast_fp16")]; + tensor x_15_begin_0 = const()[name = string("x_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_15_end_0 = const()[name = string("x_15_end_0"), val = tensor([1024, 1, 32, 32])]; + tensor x_15_end_mask_0 = const()[name = string("x_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, x = var_376_cast_fp16)[name = string("x_15_cast_fp16")]; + tensor var_378 = const()[name = string("op_378"), val = tensor([1024, 1, 2, 32, 32])]; + tensor pair_13_cast_fp16 = reshape(shape = var_378, x = inv_11_cast_fp16)[name = string("pair_13_cast_fp16")]; + tensor var_380_begin_0 = const()[name = string("op_380_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_380_end_0 = const()[name = string("op_380_end_0"), val = tensor([1024, 1, 1, 32, 32])]; + tensor var_380_end_mask_0 = const()[name = string("op_380_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_380_squeeze_mask_0 = const()[name = string("op_380_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_380_cast_fp16 = slice_by_index(begin = var_380_begin_0, end = var_380_end_0, end_mask = var_380_end_mask_0, squeeze_mask = var_380_squeeze_mask_0, x = pair_13_cast_fp16)[name = string("op_380_cast_fp16")]; + tensor var_383_begin_0 = const()[name = string("op_383_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_383_end_0 = const()[name = string("op_383_end_0"), val = tensor([1024, 1, 2, 32, 32])]; + tensor var_383_end_mask_0 = const()[name = string("op_383_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_383_squeeze_mask_0 = const()[name = string("op_383_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_383_cast_fp16 = slice_by_index(begin = var_383_begin_0, end = var_383_end_0, end_mask = var_383_end_mask_0, squeeze_mask = var_383_squeeze_mask_0, x = pair_13_cast_fp16)[name = string("op_383_cast_fp16")]; + bool var_386_transpose_x_0 = const()[name = string("op_386_transpose_x_0"), val = bool(false)]; + bool var_386_transpose_y_0 = const()[name = string("op_386_transpose_y_0"), val = bool(false)]; + tensor var_386_cast_fp16 = matmul(transpose_x = var_386_transpose_x_0, transpose_y = var_386_transpose_y_0, x = x_15_cast_fp16, y = var_380_cast_fp16)[name = string("op_386_cast_fp16")]; + bool var_387_transpose_x_0 = const()[name = string("op_387_transpose_x_0"), val = bool(false)]; + bool var_387_transpose_y_0 = const()[name = string("op_387_transpose_y_0"), val = bool(false)]; + tensor var_387_cast_fp16 = matmul(transpose_x = var_387_transpose_x_0, transpose_y = var_387_transpose_y_0, x = var_383_cast_fp16, y = var_386_cast_fp16)[name = string("op_387_cast_fp16")]; + fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_11_cast_fp16 = mul(x = var_387_cast_fp16, y = const_11_promoted_to_fp16)[name = string("lower_11_cast_fp16")]; + bool top_11_interleave_0 = const()[name = string("top_11_interleave_0"), val = bool(false)]; + tensor var_389_to_fp16 = const()[name = string("op_389_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34547520)))]; + tensor top_11_cast_fp16 = concat(axis = var_63, interleave = top_11_interleave_0, values = (var_380_cast_fp16, var_389_to_fp16))[name = string("top_11_cast_fp16")]; + bool bottom_11_interleave_0 = const()[name = string("bottom_11_interleave_0"), val = bool(false)]; + tensor bottom_11_cast_fp16 = concat(axis = var_63, interleave = bottom_11_interleave_0, values = (lower_11_cast_fp16, var_383_cast_fp16))[name = string("bottom_11_cast_fp16")]; + bool inv_13_interleave_0 = const()[name = string("inv_13_interleave_0"), val = bool(false)]; + tensor inv_13_cast_fp16 = concat(axis = var_34, interleave = inv_13_interleave_0, values = (top_11_cast_fp16, bottom_11_cast_fp16))[name = string("inv_13_cast_fp16")]; + tensor var_396 = const()[name = string("op_396"), val = tensor([32, 32, 64, 64])]; + tensor inv_15_cast_fp16 = reshape(shape = var_396, x = inv_13_cast_fp16)[name = string("inv_15_cast_fp16")]; + bool new_v_1_transpose_x_0 = const()[name = string("new_v_1_transpose_x_0"), val = bool(false)]; + bool new_v_1_transpose_y_0 = const()[name = string("new_v_1_transpose_y_0"), val = bool(false)]; + tensor new_v_1_cast_fp16 = matmul(transpose_x = new_v_1_transpose_x_0, transpose_y = new_v_1_transpose_y_0, x = inv_15_cast_fp16, y = v_beta_1_cast_fp16)[name = string("new_v_1_cast_fp16")]; + tensor var_399_cast_fp16 = exp(x = cum_1_cast_fp16)[name = string("op_399_cast_fp16")]; + tensor var_400_axes_0 = const()[name = string("op_400_axes_0"), val = tensor([3])]; + tensor var_400_cast_fp16 = expand_dims(axes = var_400_axes_0, x = var_399_cast_fp16)[name = string("op_400_cast_fp16")]; + tensor var_401_cast_fp16 = mul(x = k_beta_1_cast_fp16, y = var_400_cast_fp16)[name = string("op_401_cast_fp16")]; + bool k_cumdecay_1_transpose_x_0 = const()[name = string("k_cumdecay_1_transpose_x_0"), val = bool(false)]; + bool k_cumdecay_1_transpose_y_0 = const()[name = string("k_cumdecay_1_transpose_y_0"), val = bool(false)]; + tensor k_cumdecay_1_cast_fp16 = matmul(transpose_x = k_cumdecay_1_transpose_x_0, transpose_y = k_cumdecay_1_transpose_y_0, x = inv_15_cast_fp16, y = var_401_cast_fp16)[name = string("k_cumdecay_1_cast_fp16")]; + tensor q_dec_1_cast_fp16 = mul(x = q_9_cast_fp16, y = var_400_cast_fp16)[name = string("q_dec_1_cast_fp16")]; + tensor var_406_cast_fp16 = exp(x = to_end_1_cast_fp16)[name = string("op_406_cast_fp16")]; + tensor var_407_axes_0 = const()[name = string("op_407_axes_0"), val = tensor([3])]; + tensor var_407_cast_fp16 = expand_dims(axes = var_407_axes_0, x = var_406_cast_fp16)[name = string("op_407_cast_fp16")]; + tensor k_dec_1_cast_fp16 = mul(x = k_7_cast_fp16, y = var_407_cast_fp16)[name = string("k_dec_1_cast_fp16")]; + tensor var_409_begin_0 = const()[name = string("op_409_begin_0"), val = tensor([0, 0, -1])]; + tensor var_409_end_0 = const()[name = string("op_409_end_0"), val = tensor([32, 32, 64])]; + tensor var_409_end_mask_0 = const()[name = string("op_409_end_mask_0"), val = tensor([true, true, true])]; + tensor var_409_cast_fp16 = slice_by_index(begin = var_409_begin_0, end = var_409_end_0, end_mask = var_409_end_mask_0, x = cum_1_cast_fp16)[name = string("op_409_cast_fp16")]; + tensor var_410_cast_fp16 = exp(x = var_409_cast_fp16)[name = string("op_410_cast_fp16")]; + tensor chunk_decay_1_axes_0 = const()[name = string("chunk_decay_1_axes_0"), val = tensor([3])]; + tensor chunk_decay_1_cast_fp16 = expand_dims(axes = chunk_decay_1_axes_0, x = var_410_cast_fp16)[name = string("chunk_decay_1_cast_fp16")]; + tensor var_415_begin_0 = const()[name = string("op_415_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_415_end_0 = const()[name = string("op_415_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_415_end_mask_0 = const()[name = string("op_415_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_415_squeeze_mask_0 = const()[name = string("op_415_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_415_cast_fp16 = slice_by_index(begin = var_415_begin_0, end = var_415_end_0, end_mask = var_415_end_mask_0, squeeze_mask = var_415_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_415_cast_fp16")]; + tensor var_417_begin_0 = const()[name = string("op_417_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_417_end_0 = const()[name = string("op_417_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_417_end_mask_0 = const()[name = string("op_417_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_417_squeeze_mask_0 = const()[name = string("op_417_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_417_cast_fp16 = slice_by_index(begin = var_417_begin_0, end = var_417_end_0, end_mask = var_417_end_mask_0, squeeze_mask = var_417_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_417_cast_fp16")]; + bool var_418_transpose_x_0 = const()[name = string("op_418_transpose_x_0"), val = bool(false)]; + bool var_418_transpose_y_0 = const()[name = string("op_418_transpose_y_0"), val = bool(false)]; + tensor state_1_to_fp16 = const()[name = string("state_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36644736)))]; + tensor var_418_cast_fp16 = matmul(transpose_x = var_418_transpose_x_0, transpose_y = var_418_transpose_y_0, x = var_417_cast_fp16, y = state_1_to_fp16)[name = string("op_418_cast_fp16")]; + tensor v_new_1_cast_fp16 = sub(x = var_415_cast_fp16, y = var_418_cast_fp16)[name = string("v_new_1_cast_fp16")]; + tensor var_421_begin_0 = const()[name = string("op_421_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_421_end_0 = const()[name = string("op_421_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_421_end_mask_0 = const()[name = string("op_421_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_421_squeeze_mask_0 = const()[name = string("op_421_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_421_cast_fp16 = slice_by_index(begin = var_421_begin_0, end = var_421_end_0, end_mask = var_421_end_mask_0, squeeze_mask = var_421_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_421_cast_fp16")]; + bool var_422_transpose_x_0 = const()[name = string("op_422_transpose_x_0"), val = bool(false)]; + bool var_422_transpose_y_0 = const()[name = string("op_422_transpose_y_0"), val = bool(false)]; + tensor var_422_cast_fp16 = matmul(transpose_x = var_422_transpose_x_0, transpose_y = var_422_transpose_y_0, x = var_421_cast_fp16, y = state_1_to_fp16)[name = string("op_422_cast_fp16")]; + tensor var_424_begin_0 = const()[name = string("op_424_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_424_end_0 = const()[name = string("op_424_end_0"), val = tensor([32, 1, 64, 64])]; + tensor var_424_end_mask_0 = const()[name = string("op_424_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_424_squeeze_mask_0 = const()[name = string("op_424_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_424_cast_fp16 = slice_by_index(begin = var_424_begin_0, end = var_424_end_0, end_mask = var_424_end_mask_0, squeeze_mask = var_424_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_424_cast_fp16")]; + bool var_425_transpose_x_0 = const()[name = string("op_425_transpose_x_0"), val = bool(false)]; + bool var_425_transpose_y_0 = const()[name = string("op_425_transpose_y_0"), val = bool(false)]; + tensor var_425_cast_fp16 = matmul(transpose_x = var_425_transpose_x_0, transpose_y = var_425_transpose_y_0, x = var_424_cast_fp16, y = v_new_1_cast_fp16)[name = string("op_425_cast_fp16")]; + tensor var_426_cast_fp16 = add(x = var_422_cast_fp16, y = var_425_cast_fp16)[name = string("op_426_cast_fp16")]; + tensor var_428_begin_0 = const()[name = string("op_428_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_428_end_0 = const()[name = string("op_428_end_0"), val = tensor([32, 1, 1, 1])]; + tensor var_428_end_mask_0 = const()[name = string("op_428_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_428_squeeze_mask_0 = const()[name = string("op_428_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_428_cast_fp16 = slice_by_index(begin = var_428_begin_0, end = var_428_end_0, end_mask = var_428_end_mask_0, squeeze_mask = var_428_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_428_cast_fp16")]; + tensor var_429_cast_fp16 = mul(x = state_1_to_fp16, y = var_428_cast_fp16)[name = string("op_429_cast_fp16")]; + tensor var_431_begin_0 = const()[name = string("op_431_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_431_end_0 = const()[name = string("op_431_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_431_end_mask_0 = const()[name = string("op_431_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_431_squeeze_mask_0 = const()[name = string("op_431_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_431_cast_fp16 = slice_by_index(begin = var_431_begin_0, end = var_431_end_0, end_mask = var_431_end_mask_0, squeeze_mask = var_431_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_431_cast_fp16")]; + bool var_433_transpose_x_1 = const()[name = string("op_433_transpose_x_1"), val = bool(true)]; + bool var_433_transpose_y_1 = const()[name = string("op_433_transpose_y_1"), val = bool(false)]; + tensor var_433_cast_fp16 = matmul(transpose_x = var_433_transpose_x_1, transpose_y = var_433_transpose_y_1, x = var_431_cast_fp16, y = v_new_1_cast_fp16)[name = string("op_433_cast_fp16")]; + tensor state_3_cast_fp16 = add(x = var_429_cast_fp16, y = var_433_cast_fp16)[name = string("state_3_cast_fp16")]; + tensor var_436_begin_0 = const()[name = string("op_436_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_436_end_0 = const()[name = string("op_436_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_436_end_mask_0 = const()[name = string("op_436_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_436_squeeze_mask_0 = const()[name = string("op_436_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_436_cast_fp16 = slice_by_index(begin = var_436_begin_0, end = var_436_end_0, end_mask = var_436_end_mask_0, squeeze_mask = var_436_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_436_cast_fp16")]; + tensor var_438_begin_0 = const()[name = string("op_438_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_438_end_0 = const()[name = string("op_438_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_438_end_mask_0 = const()[name = string("op_438_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_438_squeeze_mask_0 = const()[name = string("op_438_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_438_cast_fp16 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, squeeze_mask = var_438_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_438_cast_fp16")]; + bool var_439_transpose_x_0 = const()[name = string("op_439_transpose_x_0"), val = bool(false)]; + bool var_439_transpose_y_0 = const()[name = string("op_439_transpose_y_0"), val = bool(false)]; + tensor var_439_cast_fp16 = matmul(transpose_x = var_439_transpose_x_0, transpose_y = var_439_transpose_y_0, x = var_438_cast_fp16, y = state_3_cast_fp16)[name = string("op_439_cast_fp16")]; + tensor v_new_3_cast_fp16 = sub(x = var_436_cast_fp16, y = var_439_cast_fp16)[name = string("v_new_3_cast_fp16")]; + tensor var_442_begin_0 = const()[name = string("op_442_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_442_end_0 = const()[name = string("op_442_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_442_end_mask_0 = const()[name = string("op_442_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_442_squeeze_mask_0 = const()[name = string("op_442_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_442_cast_fp16 = slice_by_index(begin = var_442_begin_0, end = var_442_end_0, end_mask = var_442_end_mask_0, squeeze_mask = var_442_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_442_cast_fp16")]; + bool var_443_transpose_x_0 = const()[name = string("op_443_transpose_x_0"), val = bool(false)]; + bool var_443_transpose_y_0 = const()[name = string("op_443_transpose_y_0"), val = bool(false)]; + tensor var_443_cast_fp16 = matmul(transpose_x = var_443_transpose_x_0, transpose_y = var_443_transpose_y_0, x = var_442_cast_fp16, y = state_3_cast_fp16)[name = string("op_443_cast_fp16")]; + tensor var_445_begin_0 = const()[name = string("op_445_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_445_end_0 = const()[name = string("op_445_end_0"), val = tensor([32, 2, 64, 64])]; + tensor var_445_end_mask_0 = const()[name = string("op_445_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_445_squeeze_mask_0 = const()[name = string("op_445_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_445_cast_fp16 = slice_by_index(begin = var_445_begin_0, end = var_445_end_0, end_mask = var_445_end_mask_0, squeeze_mask = var_445_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_445_cast_fp16")]; + bool var_446_transpose_x_0 = const()[name = string("op_446_transpose_x_0"), val = bool(false)]; + bool var_446_transpose_y_0 = const()[name = string("op_446_transpose_y_0"), val = bool(false)]; + tensor var_446_cast_fp16 = matmul(transpose_x = var_446_transpose_x_0, transpose_y = var_446_transpose_y_0, x = var_445_cast_fp16, y = v_new_3_cast_fp16)[name = string("op_446_cast_fp16")]; + tensor var_447_cast_fp16 = add(x = var_443_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")]; + tensor var_449_begin_0 = const()[name = string("op_449_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_449_end_0 = const()[name = string("op_449_end_0"), val = tensor([32, 2, 1, 1])]; + tensor var_449_end_mask_0 = const()[name = string("op_449_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_449_squeeze_mask_0 = const()[name = string("op_449_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_449_cast_fp16 = slice_by_index(begin = var_449_begin_0, end = var_449_end_0, end_mask = var_449_end_mask_0, squeeze_mask = var_449_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_449_cast_fp16")]; + tensor var_450_cast_fp16 = mul(x = state_3_cast_fp16, y = var_449_cast_fp16)[name = string("op_450_cast_fp16")]; + tensor var_452_begin_0 = const()[name = string("op_452_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_452_end_0 = const()[name = string("op_452_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_452_end_mask_0 = const()[name = string("op_452_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_452_squeeze_mask_0 = const()[name = string("op_452_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_452_cast_fp16 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, squeeze_mask = var_452_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_452_cast_fp16")]; + bool var_454_transpose_x_1 = const()[name = string("op_454_transpose_x_1"), val = bool(true)]; + bool var_454_transpose_y_1 = const()[name = string("op_454_transpose_y_1"), val = bool(false)]; + tensor var_454_cast_fp16 = matmul(transpose_x = var_454_transpose_x_1, transpose_y = var_454_transpose_y_1, x = var_452_cast_fp16, y = v_new_3_cast_fp16)[name = string("op_454_cast_fp16")]; + tensor state_5_cast_fp16 = add(x = var_450_cast_fp16, y = var_454_cast_fp16)[name = string("state_5_cast_fp16")]; + tensor var_457_begin_0 = const()[name = string("op_457_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_457_end_0 = const()[name = string("op_457_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_457_end_mask_0 = const()[name = string("op_457_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_457_squeeze_mask_0 = const()[name = string("op_457_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_457_cast_fp16 = slice_by_index(begin = var_457_begin_0, end = var_457_end_0, end_mask = var_457_end_mask_0, squeeze_mask = var_457_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_457_cast_fp16")]; + tensor var_459_begin_0 = const()[name = string("op_459_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_459_end_0 = const()[name = string("op_459_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_459_end_mask_0 = const()[name = string("op_459_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_459_squeeze_mask_0 = const()[name = string("op_459_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_459_cast_fp16 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, squeeze_mask = var_459_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_459_cast_fp16")]; + bool var_460_transpose_x_0 = const()[name = string("op_460_transpose_x_0"), val = bool(false)]; + bool var_460_transpose_y_0 = const()[name = string("op_460_transpose_y_0"), val = bool(false)]; + tensor var_460_cast_fp16 = matmul(transpose_x = var_460_transpose_x_0, transpose_y = var_460_transpose_y_0, x = var_459_cast_fp16, y = state_5_cast_fp16)[name = string("op_460_cast_fp16")]; + tensor v_new_5_cast_fp16 = sub(x = var_457_cast_fp16, y = var_460_cast_fp16)[name = string("v_new_5_cast_fp16")]; + tensor var_463_begin_0 = const()[name = string("op_463_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_463_end_0 = const()[name = string("op_463_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_463_end_mask_0 = const()[name = string("op_463_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_463_squeeze_mask_0 = const()[name = string("op_463_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_463_cast_fp16 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, squeeze_mask = var_463_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_463_cast_fp16")]; + bool var_464_transpose_x_0 = const()[name = string("op_464_transpose_x_0"), val = bool(false)]; + bool var_464_transpose_y_0 = const()[name = string("op_464_transpose_y_0"), val = bool(false)]; + tensor var_464_cast_fp16 = matmul(transpose_x = var_464_transpose_x_0, transpose_y = var_464_transpose_y_0, x = var_463_cast_fp16, y = state_5_cast_fp16)[name = string("op_464_cast_fp16")]; + tensor var_466_begin_0 = const()[name = string("op_466_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_466_end_0 = const()[name = string("op_466_end_0"), val = tensor([32, 3, 64, 64])]; + tensor var_466_end_mask_0 = const()[name = string("op_466_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_466_squeeze_mask_0 = const()[name = string("op_466_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_466_cast_fp16 = slice_by_index(begin = var_466_begin_0, end = var_466_end_0, end_mask = var_466_end_mask_0, squeeze_mask = var_466_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_466_cast_fp16")]; + bool var_467_transpose_x_0 = const()[name = string("op_467_transpose_x_0"), val = bool(false)]; + bool var_467_transpose_y_0 = const()[name = string("op_467_transpose_y_0"), val = bool(false)]; + tensor var_467_cast_fp16 = matmul(transpose_x = var_467_transpose_x_0, transpose_y = var_467_transpose_y_0, x = var_466_cast_fp16, y = v_new_5_cast_fp16)[name = string("op_467_cast_fp16")]; + tensor var_468_cast_fp16 = add(x = var_464_cast_fp16, y = var_467_cast_fp16)[name = string("op_468_cast_fp16")]; + tensor var_470_begin_0 = const()[name = string("op_470_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_470_end_0 = const()[name = string("op_470_end_0"), val = tensor([32, 3, 1, 1])]; + tensor var_470_end_mask_0 = const()[name = string("op_470_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_470_squeeze_mask_0 = const()[name = string("op_470_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_470_cast_fp16 = slice_by_index(begin = var_470_begin_0, end = var_470_end_0, end_mask = var_470_end_mask_0, squeeze_mask = var_470_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = state_5_cast_fp16, y = var_470_cast_fp16)[name = string("op_471_cast_fp16")]; + tensor var_473_begin_0 = const()[name = string("op_473_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_473_end_0 = const()[name = string("op_473_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_473_end_mask_0 = const()[name = string("op_473_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_473_squeeze_mask_0 = const()[name = string("op_473_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_473_cast_fp16 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, squeeze_mask = var_473_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_473_cast_fp16")]; + bool var_475_transpose_x_1 = const()[name = string("op_475_transpose_x_1"), val = bool(true)]; + bool var_475_transpose_y_1 = const()[name = string("op_475_transpose_y_1"), val = bool(false)]; + tensor var_475_cast_fp16 = matmul(transpose_x = var_475_transpose_x_1, transpose_y = var_475_transpose_y_1, x = var_473_cast_fp16, y = v_new_5_cast_fp16)[name = string("op_475_cast_fp16")]; + tensor state_7_cast_fp16 = add(x = var_471_cast_fp16, y = var_475_cast_fp16)[name = string("state_7_cast_fp16")]; + tensor var_478_begin_0 = const()[name = string("op_478_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_478_end_0 = const()[name = string("op_478_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_478_end_mask_0 = const()[name = string("op_478_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_478_squeeze_mask_0 = const()[name = string("op_478_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, squeeze_mask = var_478_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_478_cast_fp16")]; + tensor var_480_begin_0 = const()[name = string("op_480_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_480_end_0 = const()[name = string("op_480_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_480_end_mask_0 = const()[name = string("op_480_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_480_squeeze_mask_0 = const()[name = string("op_480_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_480_cast_fp16 = slice_by_index(begin = var_480_begin_0, end = var_480_end_0, end_mask = var_480_end_mask_0, squeeze_mask = var_480_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_480_cast_fp16")]; + bool var_481_transpose_x_0 = const()[name = string("op_481_transpose_x_0"), val = bool(false)]; + bool var_481_transpose_y_0 = const()[name = string("op_481_transpose_y_0"), val = bool(false)]; + tensor var_481_cast_fp16 = matmul(transpose_x = var_481_transpose_x_0, transpose_y = var_481_transpose_y_0, x = var_480_cast_fp16, y = state_7_cast_fp16)[name = string("op_481_cast_fp16")]; + tensor v_new_7_cast_fp16 = sub(x = var_478_cast_fp16, y = var_481_cast_fp16)[name = string("v_new_7_cast_fp16")]; + tensor var_484_begin_0 = const()[name = string("op_484_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_484_end_0 = const()[name = string("op_484_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_484_end_mask_0 = const()[name = string("op_484_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_484_squeeze_mask_0 = const()[name = string("op_484_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_484_cast_fp16 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, squeeze_mask = var_484_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_484_cast_fp16")]; + bool var_485_transpose_x_0 = const()[name = string("op_485_transpose_x_0"), val = bool(false)]; + bool var_485_transpose_y_0 = const()[name = string("op_485_transpose_y_0"), val = bool(false)]; + tensor var_485_cast_fp16 = matmul(transpose_x = var_485_transpose_x_0, transpose_y = var_485_transpose_y_0, x = var_484_cast_fp16, y = state_7_cast_fp16)[name = string("op_485_cast_fp16")]; + tensor var_487_begin_0 = const()[name = string("op_487_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_487_end_0 = const()[name = string("op_487_end_0"), val = tensor([32, 4, 64, 64])]; + tensor var_487_end_mask_0 = const()[name = string("op_487_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_487_squeeze_mask_0 = const()[name = string("op_487_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_487_cast_fp16 = slice_by_index(begin = var_487_begin_0, end = var_487_end_0, end_mask = var_487_end_mask_0, squeeze_mask = var_487_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_487_cast_fp16")]; + bool var_488_transpose_x_0 = const()[name = string("op_488_transpose_x_0"), val = bool(false)]; + bool var_488_transpose_y_0 = const()[name = string("op_488_transpose_y_0"), val = bool(false)]; + tensor var_488_cast_fp16 = matmul(transpose_x = var_488_transpose_x_0, transpose_y = var_488_transpose_y_0, x = var_487_cast_fp16, y = v_new_7_cast_fp16)[name = string("op_488_cast_fp16")]; + tensor var_489_cast_fp16 = add(x = var_485_cast_fp16, y = var_488_cast_fp16)[name = string("op_489_cast_fp16")]; + tensor var_491_begin_0 = const()[name = string("op_491_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_491_end_0 = const()[name = string("op_491_end_0"), val = tensor([32, 4, 1, 1])]; + tensor var_491_end_mask_0 = const()[name = string("op_491_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_491_squeeze_mask_0 = const()[name = string("op_491_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_491_cast_fp16 = slice_by_index(begin = var_491_begin_0, end = var_491_end_0, end_mask = var_491_end_mask_0, squeeze_mask = var_491_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_491_cast_fp16")]; + tensor var_492_cast_fp16 = mul(x = state_7_cast_fp16, y = var_491_cast_fp16)[name = string("op_492_cast_fp16")]; + tensor var_494_begin_0 = const()[name = string("op_494_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_494_end_0 = const()[name = string("op_494_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_494_end_mask_0 = const()[name = string("op_494_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_494_squeeze_mask_0 = const()[name = string("op_494_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_494_cast_fp16 = slice_by_index(begin = var_494_begin_0, end = var_494_end_0, end_mask = var_494_end_mask_0, squeeze_mask = var_494_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_494_cast_fp16")]; + bool var_496_transpose_x_1 = const()[name = string("op_496_transpose_x_1"), val = bool(true)]; + bool var_496_transpose_y_1 = const()[name = string("op_496_transpose_y_1"), val = bool(false)]; + tensor var_496_cast_fp16 = matmul(transpose_x = var_496_transpose_x_1, transpose_y = var_496_transpose_y_1, x = var_494_cast_fp16, y = v_new_7_cast_fp16)[name = string("op_496_cast_fp16")]; + tensor state_9_cast_fp16 = add(x = var_492_cast_fp16, y = var_496_cast_fp16)[name = string("state_9_cast_fp16")]; + tensor var_499_begin_0 = const()[name = string("op_499_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_499_end_0 = const()[name = string("op_499_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_499_end_mask_0 = const()[name = string("op_499_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_499_squeeze_mask_0 = const()[name = string("op_499_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_499_cast_fp16 = slice_by_index(begin = var_499_begin_0, end = var_499_end_0, end_mask = var_499_end_mask_0, squeeze_mask = var_499_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_499_cast_fp16")]; + tensor var_501_begin_0 = const()[name = string("op_501_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_501_end_0 = const()[name = string("op_501_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_501_end_mask_0 = const()[name = string("op_501_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_501_squeeze_mask_0 = const()[name = string("op_501_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_501_cast_fp16 = slice_by_index(begin = var_501_begin_0, end = var_501_end_0, end_mask = var_501_end_mask_0, squeeze_mask = var_501_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_501_cast_fp16")]; + bool var_502_transpose_x_0 = const()[name = string("op_502_transpose_x_0"), val = bool(false)]; + bool var_502_transpose_y_0 = const()[name = string("op_502_transpose_y_0"), val = bool(false)]; + tensor var_502_cast_fp16 = matmul(transpose_x = var_502_transpose_x_0, transpose_y = var_502_transpose_y_0, x = var_501_cast_fp16, y = state_9_cast_fp16)[name = string("op_502_cast_fp16")]; + tensor v_new_9_cast_fp16 = sub(x = var_499_cast_fp16, y = var_502_cast_fp16)[name = string("v_new_9_cast_fp16")]; + tensor var_505_begin_0 = const()[name = string("op_505_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_505_end_0 = const()[name = string("op_505_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_505_end_mask_0 = const()[name = string("op_505_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_505_squeeze_mask_0 = const()[name = string("op_505_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_505_cast_fp16 = slice_by_index(begin = var_505_begin_0, end = var_505_end_0, end_mask = var_505_end_mask_0, squeeze_mask = var_505_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_505_cast_fp16")]; + bool var_506_transpose_x_0 = const()[name = string("op_506_transpose_x_0"), val = bool(false)]; + bool var_506_transpose_y_0 = const()[name = string("op_506_transpose_y_0"), val = bool(false)]; + tensor var_506_cast_fp16 = matmul(transpose_x = var_506_transpose_x_0, transpose_y = var_506_transpose_y_0, x = var_505_cast_fp16, y = state_9_cast_fp16)[name = string("op_506_cast_fp16")]; + tensor var_508_begin_0 = const()[name = string("op_508_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_508_end_0 = const()[name = string("op_508_end_0"), val = tensor([32, 5, 64, 64])]; + tensor var_508_end_mask_0 = const()[name = string("op_508_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_508_squeeze_mask_0 = const()[name = string("op_508_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_508_cast_fp16 = slice_by_index(begin = var_508_begin_0, end = var_508_end_0, end_mask = var_508_end_mask_0, squeeze_mask = var_508_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_508_cast_fp16")]; + bool var_509_transpose_x_0 = const()[name = string("op_509_transpose_x_0"), val = bool(false)]; + bool var_509_transpose_y_0 = const()[name = string("op_509_transpose_y_0"), val = bool(false)]; + tensor var_509_cast_fp16 = matmul(transpose_x = var_509_transpose_x_0, transpose_y = var_509_transpose_y_0, x = var_508_cast_fp16, y = v_new_9_cast_fp16)[name = string("op_509_cast_fp16")]; + tensor var_510_cast_fp16 = add(x = var_506_cast_fp16, y = var_509_cast_fp16)[name = string("op_510_cast_fp16")]; + tensor var_512_begin_0 = const()[name = string("op_512_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_512_end_0 = const()[name = string("op_512_end_0"), val = tensor([32, 5, 1, 1])]; + tensor var_512_end_mask_0 = const()[name = string("op_512_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_512_squeeze_mask_0 = const()[name = string("op_512_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_512_cast_fp16 = slice_by_index(begin = var_512_begin_0, end = var_512_end_0, end_mask = var_512_end_mask_0, squeeze_mask = var_512_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_512_cast_fp16")]; + tensor var_513_cast_fp16 = mul(x = state_9_cast_fp16, y = var_512_cast_fp16)[name = string("op_513_cast_fp16")]; + tensor var_515_begin_0 = const()[name = string("op_515_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_515_end_0 = const()[name = string("op_515_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_515_end_mask_0 = const()[name = string("op_515_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_515_squeeze_mask_0 = const()[name = string("op_515_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_515_cast_fp16 = slice_by_index(begin = var_515_begin_0, end = var_515_end_0, end_mask = var_515_end_mask_0, squeeze_mask = var_515_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_515_cast_fp16")]; + bool var_517_transpose_x_1 = const()[name = string("op_517_transpose_x_1"), val = bool(true)]; + bool var_517_transpose_y_1 = const()[name = string("op_517_transpose_y_1"), val = bool(false)]; + tensor var_517_cast_fp16 = matmul(transpose_x = var_517_transpose_x_1, transpose_y = var_517_transpose_y_1, x = var_515_cast_fp16, y = v_new_9_cast_fp16)[name = string("op_517_cast_fp16")]; + tensor state_11_cast_fp16 = add(x = var_513_cast_fp16, y = var_517_cast_fp16)[name = string("state_11_cast_fp16")]; + tensor var_520_begin_0 = const()[name = string("op_520_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_520_end_0 = const()[name = string("op_520_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_520_end_mask_0 = const()[name = string("op_520_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_520_squeeze_mask_0 = const()[name = string("op_520_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_520_cast_fp16 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, squeeze_mask = var_520_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_520_cast_fp16")]; + tensor var_522_begin_0 = const()[name = string("op_522_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_522_end_0 = const()[name = string("op_522_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_522_end_mask_0 = const()[name = string("op_522_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_522_squeeze_mask_0 = const()[name = string("op_522_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_522_cast_fp16 = slice_by_index(begin = var_522_begin_0, end = var_522_end_0, end_mask = var_522_end_mask_0, squeeze_mask = var_522_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_522_cast_fp16")]; + bool var_523_transpose_x_0 = const()[name = string("op_523_transpose_x_0"), val = bool(false)]; + bool var_523_transpose_y_0 = const()[name = string("op_523_transpose_y_0"), val = bool(false)]; + tensor var_523_cast_fp16 = matmul(transpose_x = var_523_transpose_x_0, transpose_y = var_523_transpose_y_0, x = var_522_cast_fp16, y = state_11_cast_fp16)[name = string("op_523_cast_fp16")]; + tensor v_new_11_cast_fp16 = sub(x = var_520_cast_fp16, y = var_523_cast_fp16)[name = string("v_new_11_cast_fp16")]; + tensor var_526_begin_0 = const()[name = string("op_526_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_526_end_0 = const()[name = string("op_526_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_526_end_mask_0 = const()[name = string("op_526_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_526_squeeze_mask_0 = const()[name = string("op_526_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_526_cast_fp16 = slice_by_index(begin = var_526_begin_0, end = var_526_end_0, end_mask = var_526_end_mask_0, squeeze_mask = var_526_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_526_cast_fp16")]; + bool var_527_transpose_x_0 = const()[name = string("op_527_transpose_x_0"), val = bool(false)]; + bool var_527_transpose_y_0 = const()[name = string("op_527_transpose_y_0"), val = bool(false)]; + tensor var_527_cast_fp16 = matmul(transpose_x = var_527_transpose_x_0, transpose_y = var_527_transpose_y_0, x = var_526_cast_fp16, y = state_11_cast_fp16)[name = string("op_527_cast_fp16")]; + tensor var_529_begin_0 = const()[name = string("op_529_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_529_end_0 = const()[name = string("op_529_end_0"), val = tensor([32, 6, 64, 64])]; + tensor var_529_end_mask_0 = const()[name = string("op_529_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_529_squeeze_mask_0 = const()[name = string("op_529_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_529_cast_fp16 = slice_by_index(begin = var_529_begin_0, end = var_529_end_0, end_mask = var_529_end_mask_0, squeeze_mask = var_529_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_529_cast_fp16")]; + bool var_530_transpose_x_0 = const()[name = string("op_530_transpose_x_0"), val = bool(false)]; + bool var_530_transpose_y_0 = const()[name = string("op_530_transpose_y_0"), val = bool(false)]; + tensor var_530_cast_fp16 = matmul(transpose_x = var_530_transpose_x_0, transpose_y = var_530_transpose_y_0, x = var_529_cast_fp16, y = v_new_11_cast_fp16)[name = string("op_530_cast_fp16")]; + tensor var_531_cast_fp16 = add(x = var_527_cast_fp16, y = var_530_cast_fp16)[name = string("op_531_cast_fp16")]; + tensor var_533_begin_0 = const()[name = string("op_533_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_533_end_0 = const()[name = string("op_533_end_0"), val = tensor([32, 6, 1, 1])]; + tensor var_533_end_mask_0 = const()[name = string("op_533_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_533_squeeze_mask_0 = const()[name = string("op_533_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_533_cast_fp16 = slice_by_index(begin = var_533_begin_0, end = var_533_end_0, end_mask = var_533_end_mask_0, squeeze_mask = var_533_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_533_cast_fp16")]; + tensor var_534_cast_fp16 = mul(x = state_11_cast_fp16, y = var_533_cast_fp16)[name = string("op_534_cast_fp16")]; + tensor var_536_begin_0 = const()[name = string("op_536_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_536_end_0 = const()[name = string("op_536_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_536_end_mask_0 = const()[name = string("op_536_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_536_squeeze_mask_0 = const()[name = string("op_536_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_536_cast_fp16 = slice_by_index(begin = var_536_begin_0, end = var_536_end_0, end_mask = var_536_end_mask_0, squeeze_mask = var_536_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_536_cast_fp16")]; + bool var_538_transpose_x_1 = const()[name = string("op_538_transpose_x_1"), val = bool(true)]; + bool var_538_transpose_y_1 = const()[name = string("op_538_transpose_y_1"), val = bool(false)]; + tensor var_538_cast_fp16 = matmul(transpose_x = var_538_transpose_x_1, transpose_y = var_538_transpose_y_1, x = var_536_cast_fp16, y = v_new_11_cast_fp16)[name = string("op_538_cast_fp16")]; + tensor state_13_cast_fp16 = add(x = var_534_cast_fp16, y = var_538_cast_fp16)[name = string("state_13_cast_fp16")]; + tensor var_541_begin_0 = const()[name = string("op_541_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_541_end_0 = const()[name = string("op_541_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_541_end_mask_0 = const()[name = string("op_541_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_541_squeeze_mask_0 = const()[name = string("op_541_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_541_cast_fp16 = slice_by_index(begin = var_541_begin_0, end = var_541_end_0, end_mask = var_541_end_mask_0, squeeze_mask = var_541_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_541_cast_fp16")]; + tensor var_543_begin_0 = const()[name = string("op_543_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_543_end_0 = const()[name = string("op_543_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_543_end_mask_0 = const()[name = string("op_543_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_543_squeeze_mask_0 = const()[name = string("op_543_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_543_cast_fp16 = slice_by_index(begin = var_543_begin_0, end = var_543_end_0, end_mask = var_543_end_mask_0, squeeze_mask = var_543_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_543_cast_fp16")]; + bool var_544_transpose_x_0 = const()[name = string("op_544_transpose_x_0"), val = bool(false)]; + bool var_544_transpose_y_0 = const()[name = string("op_544_transpose_y_0"), val = bool(false)]; + tensor var_544_cast_fp16 = matmul(transpose_x = var_544_transpose_x_0, transpose_y = var_544_transpose_y_0, x = var_543_cast_fp16, y = state_13_cast_fp16)[name = string("op_544_cast_fp16")]; + tensor v_new_13_cast_fp16 = sub(x = var_541_cast_fp16, y = var_544_cast_fp16)[name = string("v_new_13_cast_fp16")]; + tensor var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_547_squeeze_mask_0 = const()[name = string("op_547_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, squeeze_mask = var_547_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_547_cast_fp16")]; + bool var_548_transpose_x_0 = const()[name = string("op_548_transpose_x_0"), val = bool(false)]; + bool var_548_transpose_y_0 = const()[name = string("op_548_transpose_y_0"), val = bool(false)]; + tensor var_548_cast_fp16 = matmul(transpose_x = var_548_transpose_x_0, transpose_y = var_548_transpose_y_0, x = var_547_cast_fp16, y = state_13_cast_fp16)[name = string("op_548_cast_fp16")]; + tensor var_550_begin_0 = const()[name = string("op_550_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_550_end_0 = const()[name = string("op_550_end_0"), val = tensor([32, 7, 64, 64])]; + tensor var_550_end_mask_0 = const()[name = string("op_550_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_550_squeeze_mask_0 = const()[name = string("op_550_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_550_cast_fp16 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, squeeze_mask = var_550_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_550_cast_fp16")]; + bool var_551_transpose_x_0 = const()[name = string("op_551_transpose_x_0"), val = bool(false)]; + bool var_551_transpose_y_0 = const()[name = string("op_551_transpose_y_0"), val = bool(false)]; + tensor var_551_cast_fp16 = matmul(transpose_x = var_551_transpose_x_0, transpose_y = var_551_transpose_y_0, x = var_550_cast_fp16, y = v_new_13_cast_fp16)[name = string("op_551_cast_fp16")]; + tensor var_552_cast_fp16 = add(x = var_548_cast_fp16, y = var_551_cast_fp16)[name = string("op_552_cast_fp16")]; + tensor var_554_begin_0 = const()[name = string("op_554_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_554_end_0 = const()[name = string("op_554_end_0"), val = tensor([32, 7, 1, 1])]; + tensor var_554_end_mask_0 = const()[name = string("op_554_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_554_squeeze_mask_0 = const()[name = string("op_554_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_554_cast_fp16 = slice_by_index(begin = var_554_begin_0, end = var_554_end_0, end_mask = var_554_end_mask_0, squeeze_mask = var_554_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_554_cast_fp16")]; + tensor var_555_cast_fp16 = mul(x = state_13_cast_fp16, y = var_554_cast_fp16)[name = string("op_555_cast_fp16")]; + tensor var_557_begin_0 = const()[name = string("op_557_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_557_end_0 = const()[name = string("op_557_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_557_end_mask_0 = const()[name = string("op_557_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_557_squeeze_mask_0 = const()[name = string("op_557_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_557_cast_fp16 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, squeeze_mask = var_557_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_557_cast_fp16")]; + bool var_559_transpose_x_1 = const()[name = string("op_559_transpose_x_1"), val = bool(true)]; + bool var_559_transpose_y_1 = const()[name = string("op_559_transpose_y_1"), val = bool(false)]; + tensor var_559_cast_fp16 = matmul(transpose_x = var_559_transpose_x_1, transpose_y = var_559_transpose_y_1, x = var_557_cast_fp16, y = v_new_13_cast_fp16)[name = string("op_559_cast_fp16")]; + tensor state_15_cast_fp16 = add(x = var_555_cast_fp16, y = var_559_cast_fp16)[name = string("state_15_cast_fp16")]; + tensor var_562_begin_0 = const()[name = string("op_562_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_562_end_0 = const()[name = string("op_562_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_562_end_mask_0 = const()[name = string("op_562_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_562_squeeze_mask_0 = const()[name = string("op_562_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_562_cast_fp16 = slice_by_index(begin = var_562_begin_0, end = var_562_end_0, end_mask = var_562_end_mask_0, squeeze_mask = var_562_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_562_cast_fp16")]; + tensor var_564_begin_0 = const()[name = string("op_564_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_564_end_0 = const()[name = string("op_564_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_564_end_mask_0 = const()[name = string("op_564_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_564_squeeze_mask_0 = const()[name = string("op_564_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_564_cast_fp16 = slice_by_index(begin = var_564_begin_0, end = var_564_end_0, end_mask = var_564_end_mask_0, squeeze_mask = var_564_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_564_cast_fp16")]; + bool var_565_transpose_x_0 = const()[name = string("op_565_transpose_x_0"), val = bool(false)]; + bool var_565_transpose_y_0 = const()[name = string("op_565_transpose_y_0"), val = bool(false)]; + tensor var_565_cast_fp16 = matmul(transpose_x = var_565_transpose_x_0, transpose_y = var_565_transpose_y_0, x = var_564_cast_fp16, y = state_15_cast_fp16)[name = string("op_565_cast_fp16")]; + tensor v_new_15_cast_fp16 = sub(x = var_562_cast_fp16, y = var_565_cast_fp16)[name = string("v_new_15_cast_fp16")]; + tensor var_568_begin_0 = const()[name = string("op_568_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_568_end_0 = const()[name = string("op_568_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_568_end_mask_0 = const()[name = string("op_568_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_568_squeeze_mask_0 = const()[name = string("op_568_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_568_cast_fp16 = slice_by_index(begin = var_568_begin_0, end = var_568_end_0, end_mask = var_568_end_mask_0, squeeze_mask = var_568_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_568_cast_fp16")]; + bool var_569_transpose_x_0 = const()[name = string("op_569_transpose_x_0"), val = bool(false)]; + bool var_569_transpose_y_0 = const()[name = string("op_569_transpose_y_0"), val = bool(false)]; + tensor var_569_cast_fp16 = matmul(transpose_x = var_569_transpose_x_0, transpose_y = var_569_transpose_y_0, x = var_568_cast_fp16, y = state_15_cast_fp16)[name = string("op_569_cast_fp16")]; + tensor var_571_begin_0 = const()[name = string("op_571_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_571_end_0 = const()[name = string("op_571_end_0"), val = tensor([32, 8, 64, 64])]; + tensor var_571_end_mask_0 = const()[name = string("op_571_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_571_squeeze_mask_0 = const()[name = string("op_571_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_571_cast_fp16 = slice_by_index(begin = var_571_begin_0, end = var_571_end_0, end_mask = var_571_end_mask_0, squeeze_mask = var_571_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_571_cast_fp16")]; + bool var_572_transpose_x_0 = const()[name = string("op_572_transpose_x_0"), val = bool(false)]; + bool var_572_transpose_y_0 = const()[name = string("op_572_transpose_y_0"), val = bool(false)]; + tensor var_572_cast_fp16 = matmul(transpose_x = var_572_transpose_x_0, transpose_y = var_572_transpose_y_0, x = var_571_cast_fp16, y = v_new_15_cast_fp16)[name = string("op_572_cast_fp16")]; + tensor var_573_cast_fp16 = add(x = var_569_cast_fp16, y = var_572_cast_fp16)[name = string("op_573_cast_fp16")]; + tensor var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_575_end_0 = const()[name = string("op_575_end_0"), val = tensor([32, 8, 1, 1])]; + tensor var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_575_squeeze_mask_0 = const()[name = string("op_575_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = var_575_end_0, end_mask = var_575_end_mask_0, squeeze_mask = var_575_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_575_cast_fp16")]; + tensor var_576_cast_fp16 = mul(x = state_15_cast_fp16, y = var_575_cast_fp16)[name = string("op_576_cast_fp16")]; + tensor var_578_begin_0 = const()[name = string("op_578_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_578_end_0 = const()[name = string("op_578_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_578_end_mask_0 = const()[name = string("op_578_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_578_squeeze_mask_0 = const()[name = string("op_578_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_578_cast_fp16 = slice_by_index(begin = var_578_begin_0, end = var_578_end_0, end_mask = var_578_end_mask_0, squeeze_mask = var_578_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_578_cast_fp16")]; + bool var_580_transpose_x_1 = const()[name = string("op_580_transpose_x_1"), val = bool(true)]; + bool var_580_transpose_y_1 = const()[name = string("op_580_transpose_y_1"), val = bool(false)]; + tensor var_580_cast_fp16 = matmul(transpose_x = var_580_transpose_x_1, transpose_y = var_580_transpose_y_1, x = var_578_cast_fp16, y = v_new_15_cast_fp16)[name = string("op_580_cast_fp16")]; + tensor state_17_cast_fp16 = add(x = var_576_cast_fp16, y = var_580_cast_fp16)[name = string("state_17_cast_fp16")]; + tensor var_583_begin_0 = const()[name = string("op_583_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_583_end_0 = const()[name = string("op_583_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_583_end_mask_0 = const()[name = string("op_583_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_583_squeeze_mask_0 = const()[name = string("op_583_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_583_cast_fp16 = slice_by_index(begin = var_583_begin_0, end = var_583_end_0, end_mask = var_583_end_mask_0, squeeze_mask = var_583_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_583_cast_fp16")]; + tensor var_585_begin_0 = const()[name = string("op_585_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_585_end_0 = const()[name = string("op_585_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_585_end_mask_0 = const()[name = string("op_585_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_585_squeeze_mask_0 = const()[name = string("op_585_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_585_cast_fp16 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, squeeze_mask = var_585_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_585_cast_fp16")]; + bool var_586_transpose_x_0 = const()[name = string("op_586_transpose_x_0"), val = bool(false)]; + bool var_586_transpose_y_0 = const()[name = string("op_586_transpose_y_0"), val = bool(false)]; + tensor var_586_cast_fp16 = matmul(transpose_x = var_586_transpose_x_0, transpose_y = var_586_transpose_y_0, x = var_585_cast_fp16, y = state_17_cast_fp16)[name = string("op_586_cast_fp16")]; + tensor v_new_17_cast_fp16 = sub(x = var_583_cast_fp16, y = var_586_cast_fp16)[name = string("v_new_17_cast_fp16")]; + tensor var_589_begin_0 = const()[name = string("op_589_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_589_end_0 = const()[name = string("op_589_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_589_end_mask_0 = const()[name = string("op_589_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_589_squeeze_mask_0 = const()[name = string("op_589_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_589_cast_fp16 = slice_by_index(begin = var_589_begin_0, end = var_589_end_0, end_mask = var_589_end_mask_0, squeeze_mask = var_589_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_589_cast_fp16")]; + bool var_590_transpose_x_0 = const()[name = string("op_590_transpose_x_0"), val = bool(false)]; + bool var_590_transpose_y_0 = const()[name = string("op_590_transpose_y_0"), val = bool(false)]; + tensor var_590_cast_fp16 = matmul(transpose_x = var_590_transpose_x_0, transpose_y = var_590_transpose_y_0, x = var_589_cast_fp16, y = state_17_cast_fp16)[name = string("op_590_cast_fp16")]; + tensor var_592_begin_0 = const()[name = string("op_592_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_592_end_0 = const()[name = string("op_592_end_0"), val = tensor([32, 9, 64, 64])]; + tensor var_592_end_mask_0 = const()[name = string("op_592_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_592_squeeze_mask_0 = const()[name = string("op_592_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_592_cast_fp16 = slice_by_index(begin = var_592_begin_0, end = var_592_end_0, end_mask = var_592_end_mask_0, squeeze_mask = var_592_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_592_cast_fp16")]; + bool var_593_transpose_x_0 = const()[name = string("op_593_transpose_x_0"), val = bool(false)]; + bool var_593_transpose_y_0 = const()[name = string("op_593_transpose_y_0"), val = bool(false)]; + tensor var_593_cast_fp16 = matmul(transpose_x = var_593_transpose_x_0, transpose_y = var_593_transpose_y_0, x = var_592_cast_fp16, y = v_new_17_cast_fp16)[name = string("op_593_cast_fp16")]; + tensor var_594_cast_fp16 = add(x = var_590_cast_fp16, y = var_593_cast_fp16)[name = string("op_594_cast_fp16")]; + tensor var_596_begin_0 = const()[name = string("op_596_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_596_end_0 = const()[name = string("op_596_end_0"), val = tensor([32, 9, 1, 1])]; + tensor var_596_end_mask_0 = const()[name = string("op_596_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_596_squeeze_mask_0 = const()[name = string("op_596_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_596_cast_fp16 = slice_by_index(begin = var_596_begin_0, end = var_596_end_0, end_mask = var_596_end_mask_0, squeeze_mask = var_596_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_596_cast_fp16")]; + tensor var_597_cast_fp16 = mul(x = state_17_cast_fp16, y = var_596_cast_fp16)[name = string("op_597_cast_fp16")]; + tensor var_599_begin_0 = const()[name = string("op_599_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_599_end_0 = const()[name = string("op_599_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_599_end_mask_0 = const()[name = string("op_599_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_599_squeeze_mask_0 = const()[name = string("op_599_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_599_cast_fp16 = slice_by_index(begin = var_599_begin_0, end = var_599_end_0, end_mask = var_599_end_mask_0, squeeze_mask = var_599_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_599_cast_fp16")]; + bool var_601_transpose_x_1 = const()[name = string("op_601_transpose_x_1"), val = bool(true)]; + bool var_601_transpose_y_1 = const()[name = string("op_601_transpose_y_1"), val = bool(false)]; + tensor var_601_cast_fp16 = matmul(transpose_x = var_601_transpose_x_1, transpose_y = var_601_transpose_y_1, x = var_599_cast_fp16, y = v_new_17_cast_fp16)[name = string("op_601_cast_fp16")]; + tensor state_19_cast_fp16 = add(x = var_597_cast_fp16, y = var_601_cast_fp16)[name = string("state_19_cast_fp16")]; + tensor var_604_begin_0 = const()[name = string("op_604_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_604_end_0 = const()[name = string("op_604_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_604_end_mask_0 = const()[name = string("op_604_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_604_squeeze_mask_0 = const()[name = string("op_604_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_604_cast_fp16 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, squeeze_mask = var_604_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_604_cast_fp16")]; + tensor var_606_begin_0 = const()[name = string("op_606_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_606_end_0 = const()[name = string("op_606_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_606_end_mask_0 = const()[name = string("op_606_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_606_squeeze_mask_0 = const()[name = string("op_606_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_606_cast_fp16 = slice_by_index(begin = var_606_begin_0, end = var_606_end_0, end_mask = var_606_end_mask_0, squeeze_mask = var_606_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_606_cast_fp16")]; + bool var_607_transpose_x_0 = const()[name = string("op_607_transpose_x_0"), val = bool(false)]; + bool var_607_transpose_y_0 = const()[name = string("op_607_transpose_y_0"), val = bool(false)]; + tensor var_607_cast_fp16 = matmul(transpose_x = var_607_transpose_x_0, transpose_y = var_607_transpose_y_0, x = var_606_cast_fp16, y = state_19_cast_fp16)[name = string("op_607_cast_fp16")]; + tensor v_new_19_cast_fp16 = sub(x = var_604_cast_fp16, y = var_607_cast_fp16)[name = string("v_new_19_cast_fp16")]; + tensor var_610_begin_0 = const()[name = string("op_610_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_610_end_0 = const()[name = string("op_610_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_610_end_mask_0 = const()[name = string("op_610_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_610_squeeze_mask_0 = const()[name = string("op_610_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_610_cast_fp16 = slice_by_index(begin = var_610_begin_0, end = var_610_end_0, end_mask = var_610_end_mask_0, squeeze_mask = var_610_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_610_cast_fp16")]; + bool var_611_transpose_x_0 = const()[name = string("op_611_transpose_x_0"), val = bool(false)]; + bool var_611_transpose_y_0 = const()[name = string("op_611_transpose_y_0"), val = bool(false)]; + tensor var_611_cast_fp16 = matmul(transpose_x = var_611_transpose_x_0, transpose_y = var_611_transpose_y_0, x = var_610_cast_fp16, y = state_19_cast_fp16)[name = string("op_611_cast_fp16")]; + tensor var_613_begin_0 = const()[name = string("op_613_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_613_end_0 = const()[name = string("op_613_end_0"), val = tensor([32, 10, 64, 64])]; + tensor var_613_end_mask_0 = const()[name = string("op_613_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_613_squeeze_mask_0 = const()[name = string("op_613_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_613_cast_fp16 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, squeeze_mask = var_613_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_613_cast_fp16")]; + bool var_614_transpose_x_0 = const()[name = string("op_614_transpose_x_0"), val = bool(false)]; + bool var_614_transpose_y_0 = const()[name = string("op_614_transpose_y_0"), val = bool(false)]; + tensor var_614_cast_fp16 = matmul(transpose_x = var_614_transpose_x_0, transpose_y = var_614_transpose_y_0, x = var_613_cast_fp16, y = v_new_19_cast_fp16)[name = string("op_614_cast_fp16")]; + tensor var_615_cast_fp16 = add(x = var_611_cast_fp16, y = var_614_cast_fp16)[name = string("op_615_cast_fp16")]; + tensor var_617_begin_0 = const()[name = string("op_617_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_617_end_0 = const()[name = string("op_617_end_0"), val = tensor([32, 10, 1, 1])]; + tensor var_617_end_mask_0 = const()[name = string("op_617_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_617_squeeze_mask_0 = const()[name = string("op_617_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_617_cast_fp16 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, squeeze_mask = var_617_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_617_cast_fp16")]; + tensor var_618_cast_fp16 = mul(x = state_19_cast_fp16, y = var_617_cast_fp16)[name = string("op_618_cast_fp16")]; + tensor var_620_begin_0 = const()[name = string("op_620_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_620_end_0 = const()[name = string("op_620_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_620_end_mask_0 = const()[name = string("op_620_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_620_squeeze_mask_0 = const()[name = string("op_620_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_620_cast_fp16 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, squeeze_mask = var_620_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_620_cast_fp16")]; + bool var_622_transpose_x_1 = const()[name = string("op_622_transpose_x_1"), val = bool(true)]; + bool var_622_transpose_y_1 = const()[name = string("op_622_transpose_y_1"), val = bool(false)]; + tensor var_622_cast_fp16 = matmul(transpose_x = var_622_transpose_x_1, transpose_y = var_622_transpose_y_1, x = var_620_cast_fp16, y = v_new_19_cast_fp16)[name = string("op_622_cast_fp16")]; + tensor state_21_cast_fp16 = add(x = var_618_cast_fp16, y = var_622_cast_fp16)[name = string("state_21_cast_fp16")]; + tensor var_625_begin_0 = const()[name = string("op_625_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_625_end_0 = const()[name = string("op_625_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_625_end_mask_0 = const()[name = string("op_625_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_625_squeeze_mask_0 = const()[name = string("op_625_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_625_cast_fp16 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, squeeze_mask = var_625_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_625_cast_fp16")]; + tensor var_627_begin_0 = const()[name = string("op_627_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_627_end_0 = const()[name = string("op_627_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_627_end_mask_0 = const()[name = string("op_627_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_627_squeeze_mask_0 = const()[name = string("op_627_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_627_cast_fp16 = slice_by_index(begin = var_627_begin_0, end = var_627_end_0, end_mask = var_627_end_mask_0, squeeze_mask = var_627_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_627_cast_fp16")]; + bool var_628_transpose_x_0 = const()[name = string("op_628_transpose_x_0"), val = bool(false)]; + bool var_628_transpose_y_0 = const()[name = string("op_628_transpose_y_0"), val = bool(false)]; + tensor var_628_cast_fp16 = matmul(transpose_x = var_628_transpose_x_0, transpose_y = var_628_transpose_y_0, x = var_627_cast_fp16, y = state_21_cast_fp16)[name = string("op_628_cast_fp16")]; + tensor v_new_21_cast_fp16 = sub(x = var_625_cast_fp16, y = var_628_cast_fp16)[name = string("v_new_21_cast_fp16")]; + tensor var_631_begin_0 = const()[name = string("op_631_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_631_end_0 = const()[name = string("op_631_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_631_end_mask_0 = const()[name = string("op_631_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_631_squeeze_mask_0 = const()[name = string("op_631_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_631_cast_fp16 = slice_by_index(begin = var_631_begin_0, end = var_631_end_0, end_mask = var_631_end_mask_0, squeeze_mask = var_631_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_631_cast_fp16")]; + bool var_632_transpose_x_0 = const()[name = string("op_632_transpose_x_0"), val = bool(false)]; + bool var_632_transpose_y_0 = const()[name = string("op_632_transpose_y_0"), val = bool(false)]; + tensor var_632_cast_fp16 = matmul(transpose_x = var_632_transpose_x_0, transpose_y = var_632_transpose_y_0, x = var_631_cast_fp16, y = state_21_cast_fp16)[name = string("op_632_cast_fp16")]; + tensor var_634_begin_0 = const()[name = string("op_634_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_634_end_0 = const()[name = string("op_634_end_0"), val = tensor([32, 11, 64, 64])]; + tensor var_634_end_mask_0 = const()[name = string("op_634_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_634_squeeze_mask_0 = const()[name = string("op_634_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_634_cast_fp16 = slice_by_index(begin = var_634_begin_0, end = var_634_end_0, end_mask = var_634_end_mask_0, squeeze_mask = var_634_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_634_cast_fp16")]; + bool var_635_transpose_x_0 = const()[name = string("op_635_transpose_x_0"), val = bool(false)]; + bool var_635_transpose_y_0 = const()[name = string("op_635_transpose_y_0"), val = bool(false)]; + tensor var_635_cast_fp16 = matmul(transpose_x = var_635_transpose_x_0, transpose_y = var_635_transpose_y_0, x = var_634_cast_fp16, y = v_new_21_cast_fp16)[name = string("op_635_cast_fp16")]; + tensor var_636_cast_fp16 = add(x = var_632_cast_fp16, y = var_635_cast_fp16)[name = string("op_636_cast_fp16")]; + tensor var_638_begin_0 = const()[name = string("op_638_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_638_end_0 = const()[name = string("op_638_end_0"), val = tensor([32, 11, 1, 1])]; + tensor var_638_end_mask_0 = const()[name = string("op_638_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_638_squeeze_mask_0 = const()[name = string("op_638_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_638_cast_fp16 = slice_by_index(begin = var_638_begin_0, end = var_638_end_0, end_mask = var_638_end_mask_0, squeeze_mask = var_638_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_638_cast_fp16")]; + tensor var_639_cast_fp16 = mul(x = state_21_cast_fp16, y = var_638_cast_fp16)[name = string("op_639_cast_fp16")]; + tensor var_641_begin_0 = const()[name = string("op_641_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_641_end_0 = const()[name = string("op_641_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_641_end_mask_0 = const()[name = string("op_641_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_641_squeeze_mask_0 = const()[name = string("op_641_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_641_cast_fp16 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, squeeze_mask = var_641_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_641_cast_fp16")]; + bool var_643_transpose_x_1 = const()[name = string("op_643_transpose_x_1"), val = bool(true)]; + bool var_643_transpose_y_1 = const()[name = string("op_643_transpose_y_1"), val = bool(false)]; + tensor var_643_cast_fp16 = matmul(transpose_x = var_643_transpose_x_1, transpose_y = var_643_transpose_y_1, x = var_641_cast_fp16, y = v_new_21_cast_fp16)[name = string("op_643_cast_fp16")]; + tensor state_23_cast_fp16 = add(x = var_639_cast_fp16, y = var_643_cast_fp16)[name = string("state_23_cast_fp16")]; + tensor var_646_begin_0 = const()[name = string("op_646_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_646_end_0 = const()[name = string("op_646_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_646_end_mask_0 = const()[name = string("op_646_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_646_squeeze_mask_0 = const()[name = string("op_646_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_646_cast_fp16 = slice_by_index(begin = var_646_begin_0, end = var_646_end_0, end_mask = var_646_end_mask_0, squeeze_mask = var_646_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_646_cast_fp16")]; + tensor var_648_begin_0 = const()[name = string("op_648_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_648_end_0 = const()[name = string("op_648_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_648_end_mask_0 = const()[name = string("op_648_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_648_squeeze_mask_0 = const()[name = string("op_648_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_648_cast_fp16 = slice_by_index(begin = var_648_begin_0, end = var_648_end_0, end_mask = var_648_end_mask_0, squeeze_mask = var_648_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_648_cast_fp16")]; + bool var_649_transpose_x_0 = const()[name = string("op_649_transpose_x_0"), val = bool(false)]; + bool var_649_transpose_y_0 = const()[name = string("op_649_transpose_y_0"), val = bool(false)]; + tensor var_649_cast_fp16 = matmul(transpose_x = var_649_transpose_x_0, transpose_y = var_649_transpose_y_0, x = var_648_cast_fp16, y = state_23_cast_fp16)[name = string("op_649_cast_fp16")]; + tensor v_new_23_cast_fp16 = sub(x = var_646_cast_fp16, y = var_649_cast_fp16)[name = string("v_new_23_cast_fp16")]; + tensor var_652_begin_0 = const()[name = string("op_652_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_652_end_0 = const()[name = string("op_652_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_652_end_mask_0 = const()[name = string("op_652_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_652_squeeze_mask_0 = const()[name = string("op_652_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_652_cast_fp16 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, squeeze_mask = var_652_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_652_cast_fp16")]; + bool var_653_transpose_x_0 = const()[name = string("op_653_transpose_x_0"), val = bool(false)]; + bool var_653_transpose_y_0 = const()[name = string("op_653_transpose_y_0"), val = bool(false)]; + tensor var_653_cast_fp16 = matmul(transpose_x = var_653_transpose_x_0, transpose_y = var_653_transpose_y_0, x = var_652_cast_fp16, y = state_23_cast_fp16)[name = string("op_653_cast_fp16")]; + tensor var_655_begin_0 = const()[name = string("op_655_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_655_end_0 = const()[name = string("op_655_end_0"), val = tensor([32, 12, 64, 64])]; + tensor var_655_end_mask_0 = const()[name = string("op_655_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_655_squeeze_mask_0 = const()[name = string("op_655_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_655_cast_fp16 = slice_by_index(begin = var_655_begin_0, end = var_655_end_0, end_mask = var_655_end_mask_0, squeeze_mask = var_655_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_655_cast_fp16")]; + bool var_656_transpose_x_0 = const()[name = string("op_656_transpose_x_0"), val = bool(false)]; + bool var_656_transpose_y_0 = const()[name = string("op_656_transpose_y_0"), val = bool(false)]; + tensor var_656_cast_fp16 = matmul(transpose_x = var_656_transpose_x_0, transpose_y = var_656_transpose_y_0, x = var_655_cast_fp16, y = v_new_23_cast_fp16)[name = string("op_656_cast_fp16")]; + tensor var_657_cast_fp16 = add(x = var_653_cast_fp16, y = var_656_cast_fp16)[name = string("op_657_cast_fp16")]; + tensor var_659_begin_0 = const()[name = string("op_659_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_659_end_0 = const()[name = string("op_659_end_0"), val = tensor([32, 12, 1, 1])]; + tensor var_659_end_mask_0 = const()[name = string("op_659_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_659_squeeze_mask_0 = const()[name = string("op_659_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_659_cast_fp16 = slice_by_index(begin = var_659_begin_0, end = var_659_end_0, end_mask = var_659_end_mask_0, squeeze_mask = var_659_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_659_cast_fp16")]; + tensor var_660_cast_fp16 = mul(x = state_23_cast_fp16, y = var_659_cast_fp16)[name = string("op_660_cast_fp16")]; + tensor var_662_begin_0 = const()[name = string("op_662_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_662_end_0 = const()[name = string("op_662_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_662_end_mask_0 = const()[name = string("op_662_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_662_squeeze_mask_0 = const()[name = string("op_662_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_662_cast_fp16 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, squeeze_mask = var_662_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_662_cast_fp16")]; + bool var_664_transpose_x_1 = const()[name = string("op_664_transpose_x_1"), val = bool(true)]; + bool var_664_transpose_y_1 = const()[name = string("op_664_transpose_y_1"), val = bool(false)]; + tensor var_664_cast_fp16 = matmul(transpose_x = var_664_transpose_x_1, transpose_y = var_664_transpose_y_1, x = var_662_cast_fp16, y = v_new_23_cast_fp16)[name = string("op_664_cast_fp16")]; + tensor state_25_cast_fp16 = add(x = var_660_cast_fp16, y = var_664_cast_fp16)[name = string("state_25_cast_fp16")]; + tensor var_667_begin_0 = const()[name = string("op_667_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_667_end_0 = const()[name = string("op_667_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_667_end_mask_0 = const()[name = string("op_667_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_667_squeeze_mask_0 = const()[name = string("op_667_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_667_cast_fp16 = slice_by_index(begin = var_667_begin_0, end = var_667_end_0, end_mask = var_667_end_mask_0, squeeze_mask = var_667_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_667_cast_fp16")]; + tensor var_669_begin_0 = const()[name = string("op_669_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_669_end_0 = const()[name = string("op_669_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_669_end_mask_0 = const()[name = string("op_669_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_669_squeeze_mask_0 = const()[name = string("op_669_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_669_cast_fp16 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, squeeze_mask = var_669_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_669_cast_fp16")]; + bool var_670_transpose_x_0 = const()[name = string("op_670_transpose_x_0"), val = bool(false)]; + bool var_670_transpose_y_0 = const()[name = string("op_670_transpose_y_0"), val = bool(false)]; + tensor var_670_cast_fp16 = matmul(transpose_x = var_670_transpose_x_0, transpose_y = var_670_transpose_y_0, x = var_669_cast_fp16, y = state_25_cast_fp16)[name = string("op_670_cast_fp16")]; + tensor v_new_25_cast_fp16 = sub(x = var_667_cast_fp16, y = var_670_cast_fp16)[name = string("v_new_25_cast_fp16")]; + tensor var_673_begin_0 = const()[name = string("op_673_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_673_end_0 = const()[name = string("op_673_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_673_end_mask_0 = const()[name = string("op_673_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_673_squeeze_mask_0 = const()[name = string("op_673_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_673_cast_fp16 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, squeeze_mask = var_673_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_673_cast_fp16")]; + bool var_674_transpose_x_0 = const()[name = string("op_674_transpose_x_0"), val = bool(false)]; + bool var_674_transpose_y_0 = const()[name = string("op_674_transpose_y_0"), val = bool(false)]; + tensor var_674_cast_fp16 = matmul(transpose_x = var_674_transpose_x_0, transpose_y = var_674_transpose_y_0, x = var_673_cast_fp16, y = state_25_cast_fp16)[name = string("op_674_cast_fp16")]; + tensor var_676_begin_0 = const()[name = string("op_676_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_676_end_0 = const()[name = string("op_676_end_0"), val = tensor([32, 13, 64, 64])]; + tensor var_676_end_mask_0 = const()[name = string("op_676_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_676_squeeze_mask_0 = const()[name = string("op_676_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_676_cast_fp16 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, squeeze_mask = var_676_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_676_cast_fp16")]; + bool var_677_transpose_x_0 = const()[name = string("op_677_transpose_x_0"), val = bool(false)]; + bool var_677_transpose_y_0 = const()[name = string("op_677_transpose_y_0"), val = bool(false)]; + tensor var_677_cast_fp16 = matmul(transpose_x = var_677_transpose_x_0, transpose_y = var_677_transpose_y_0, x = var_676_cast_fp16, y = v_new_25_cast_fp16)[name = string("op_677_cast_fp16")]; + tensor var_678_cast_fp16 = add(x = var_674_cast_fp16, y = var_677_cast_fp16)[name = string("op_678_cast_fp16")]; + tensor var_680_begin_0 = const()[name = string("op_680_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_680_end_0 = const()[name = string("op_680_end_0"), val = tensor([32, 13, 1, 1])]; + tensor var_680_end_mask_0 = const()[name = string("op_680_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_680_squeeze_mask_0 = const()[name = string("op_680_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_680_cast_fp16 = slice_by_index(begin = var_680_begin_0, end = var_680_end_0, end_mask = var_680_end_mask_0, squeeze_mask = var_680_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_680_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = state_25_cast_fp16, y = var_680_cast_fp16)[name = string("op_681_cast_fp16")]; + tensor var_683_begin_0 = const()[name = string("op_683_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_683_end_0 = const()[name = string("op_683_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_683_end_mask_0 = const()[name = string("op_683_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_683_squeeze_mask_0 = const()[name = string("op_683_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_683_cast_fp16 = slice_by_index(begin = var_683_begin_0, end = var_683_end_0, end_mask = var_683_end_mask_0, squeeze_mask = var_683_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_683_cast_fp16")]; + bool var_685_transpose_x_1 = const()[name = string("op_685_transpose_x_1"), val = bool(true)]; + bool var_685_transpose_y_1 = const()[name = string("op_685_transpose_y_1"), val = bool(false)]; + tensor var_685_cast_fp16 = matmul(transpose_x = var_685_transpose_x_1, transpose_y = var_685_transpose_y_1, x = var_683_cast_fp16, y = v_new_25_cast_fp16)[name = string("op_685_cast_fp16")]; + tensor state_27_cast_fp16 = add(x = var_681_cast_fp16, y = var_685_cast_fp16)[name = string("state_27_cast_fp16")]; + tensor var_688_begin_0 = const()[name = string("op_688_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_688_end_0 = const()[name = string("op_688_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_688_end_mask_0 = const()[name = string("op_688_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_688_squeeze_mask_0 = const()[name = string("op_688_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_688_cast_fp16 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, squeeze_mask = var_688_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_688_cast_fp16")]; + tensor var_690_begin_0 = const()[name = string("op_690_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_690_end_0 = const()[name = string("op_690_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_690_end_mask_0 = const()[name = string("op_690_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_690_squeeze_mask_0 = const()[name = string("op_690_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_690_cast_fp16 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, squeeze_mask = var_690_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_690_cast_fp16")]; + bool var_691_transpose_x_0 = const()[name = string("op_691_transpose_x_0"), val = bool(false)]; + bool var_691_transpose_y_0 = const()[name = string("op_691_transpose_y_0"), val = bool(false)]; + tensor var_691_cast_fp16 = matmul(transpose_x = var_691_transpose_x_0, transpose_y = var_691_transpose_y_0, x = var_690_cast_fp16, y = state_27_cast_fp16)[name = string("op_691_cast_fp16")]; + tensor v_new_27_cast_fp16 = sub(x = var_688_cast_fp16, y = var_691_cast_fp16)[name = string("v_new_27_cast_fp16")]; + tensor var_694_begin_0 = const()[name = string("op_694_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_694_end_0 = const()[name = string("op_694_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_694_end_mask_0 = const()[name = string("op_694_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_694_squeeze_mask_0 = const()[name = string("op_694_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_694_cast_fp16 = slice_by_index(begin = var_694_begin_0, end = var_694_end_0, end_mask = var_694_end_mask_0, squeeze_mask = var_694_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_694_cast_fp16")]; + bool var_695_transpose_x_0 = const()[name = string("op_695_transpose_x_0"), val = bool(false)]; + bool var_695_transpose_y_0 = const()[name = string("op_695_transpose_y_0"), val = bool(false)]; + tensor var_695_cast_fp16 = matmul(transpose_x = var_695_transpose_x_0, transpose_y = var_695_transpose_y_0, x = var_694_cast_fp16, y = state_27_cast_fp16)[name = string("op_695_cast_fp16")]; + tensor var_697_begin_0 = const()[name = string("op_697_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_697_end_0 = const()[name = string("op_697_end_0"), val = tensor([32, 14, 64, 64])]; + tensor var_697_end_mask_0 = const()[name = string("op_697_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_697_squeeze_mask_0 = const()[name = string("op_697_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_697_cast_fp16 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, squeeze_mask = var_697_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_697_cast_fp16")]; + bool var_698_transpose_x_0 = const()[name = string("op_698_transpose_x_0"), val = bool(false)]; + bool var_698_transpose_y_0 = const()[name = string("op_698_transpose_y_0"), val = bool(false)]; + tensor var_698_cast_fp16 = matmul(transpose_x = var_698_transpose_x_0, transpose_y = var_698_transpose_y_0, x = var_697_cast_fp16, y = v_new_27_cast_fp16)[name = string("op_698_cast_fp16")]; + tensor var_699_cast_fp16 = add(x = var_695_cast_fp16, y = var_698_cast_fp16)[name = string("op_699_cast_fp16")]; + tensor var_701_begin_0 = const()[name = string("op_701_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_701_end_0 = const()[name = string("op_701_end_0"), val = tensor([32, 14, 1, 1])]; + tensor var_701_end_mask_0 = const()[name = string("op_701_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_701_squeeze_mask_0 = const()[name = string("op_701_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_701_cast_fp16 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, squeeze_mask = var_701_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_701_cast_fp16")]; + tensor var_702_cast_fp16 = mul(x = state_27_cast_fp16, y = var_701_cast_fp16)[name = string("op_702_cast_fp16")]; + tensor var_704_begin_0 = const()[name = string("op_704_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_704_end_0 = const()[name = string("op_704_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_704_end_mask_0 = const()[name = string("op_704_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_704_squeeze_mask_0 = const()[name = string("op_704_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_704_cast_fp16 = slice_by_index(begin = var_704_begin_0, end = var_704_end_0, end_mask = var_704_end_mask_0, squeeze_mask = var_704_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_704_cast_fp16")]; + bool var_706_transpose_x_1 = const()[name = string("op_706_transpose_x_1"), val = bool(true)]; + bool var_706_transpose_y_1 = const()[name = string("op_706_transpose_y_1"), val = bool(false)]; + tensor var_706_cast_fp16 = matmul(transpose_x = var_706_transpose_x_1, transpose_y = var_706_transpose_y_1, x = var_704_cast_fp16, y = v_new_27_cast_fp16)[name = string("op_706_cast_fp16")]; + tensor state_29_cast_fp16 = add(x = var_702_cast_fp16, y = var_706_cast_fp16)[name = string("state_29_cast_fp16")]; + tensor var_709_begin_0 = const()[name = string("op_709_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_709_end_0 = const()[name = string("op_709_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_709_end_mask_0 = const()[name = string("op_709_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_709_squeeze_mask_0 = const()[name = string("op_709_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_709_cast_fp16 = slice_by_index(begin = var_709_begin_0, end = var_709_end_0, end_mask = var_709_end_mask_0, squeeze_mask = var_709_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_709_cast_fp16")]; + tensor var_711_begin_0 = const()[name = string("op_711_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_711_end_0 = const()[name = string("op_711_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_711_end_mask_0 = const()[name = string("op_711_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_711_squeeze_mask_0 = const()[name = string("op_711_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_711_cast_fp16 = slice_by_index(begin = var_711_begin_0, end = var_711_end_0, end_mask = var_711_end_mask_0, squeeze_mask = var_711_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_711_cast_fp16")]; + bool var_712_transpose_x_0 = const()[name = string("op_712_transpose_x_0"), val = bool(false)]; + bool var_712_transpose_y_0 = const()[name = string("op_712_transpose_y_0"), val = bool(false)]; + tensor var_712_cast_fp16 = matmul(transpose_x = var_712_transpose_x_0, transpose_y = var_712_transpose_y_0, x = var_711_cast_fp16, y = state_29_cast_fp16)[name = string("op_712_cast_fp16")]; + tensor v_new_29_cast_fp16 = sub(x = var_709_cast_fp16, y = var_712_cast_fp16)[name = string("v_new_29_cast_fp16")]; + tensor var_715_begin_0 = const()[name = string("op_715_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_715_end_0 = const()[name = string("op_715_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_715_end_mask_0 = const()[name = string("op_715_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_715_squeeze_mask_0 = const()[name = string("op_715_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_715_cast_fp16 = slice_by_index(begin = var_715_begin_0, end = var_715_end_0, end_mask = var_715_end_mask_0, squeeze_mask = var_715_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_715_cast_fp16")]; + bool var_716_transpose_x_0 = const()[name = string("op_716_transpose_x_0"), val = bool(false)]; + bool var_716_transpose_y_0 = const()[name = string("op_716_transpose_y_0"), val = bool(false)]; + tensor var_716_cast_fp16 = matmul(transpose_x = var_716_transpose_x_0, transpose_y = var_716_transpose_y_0, x = var_715_cast_fp16, y = state_29_cast_fp16)[name = string("op_716_cast_fp16")]; + tensor var_718_begin_0 = const()[name = string("op_718_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_718_end_0 = const()[name = string("op_718_end_0"), val = tensor([32, 15, 64, 64])]; + tensor var_718_end_mask_0 = const()[name = string("op_718_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_718_squeeze_mask_0 = const()[name = string("op_718_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_718_cast_fp16 = slice_by_index(begin = var_718_begin_0, end = var_718_end_0, end_mask = var_718_end_mask_0, squeeze_mask = var_718_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_718_cast_fp16")]; + bool var_719_transpose_x_0 = const()[name = string("op_719_transpose_x_0"), val = bool(false)]; + bool var_719_transpose_y_0 = const()[name = string("op_719_transpose_y_0"), val = bool(false)]; + tensor var_719_cast_fp16 = matmul(transpose_x = var_719_transpose_x_0, transpose_y = var_719_transpose_y_0, x = var_718_cast_fp16, y = v_new_29_cast_fp16)[name = string("op_719_cast_fp16")]; + tensor var_720_cast_fp16 = add(x = var_716_cast_fp16, y = var_719_cast_fp16)[name = string("op_720_cast_fp16")]; + tensor var_722_begin_0 = const()[name = string("op_722_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_722_end_0 = const()[name = string("op_722_end_0"), val = tensor([32, 15, 1, 1])]; + tensor var_722_end_mask_0 = const()[name = string("op_722_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_722_squeeze_mask_0 = const()[name = string("op_722_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_722_cast_fp16 = slice_by_index(begin = var_722_begin_0, end = var_722_end_0, end_mask = var_722_end_mask_0, squeeze_mask = var_722_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_722_cast_fp16")]; + tensor var_723_cast_fp16 = mul(x = state_29_cast_fp16, y = var_722_cast_fp16)[name = string("op_723_cast_fp16")]; + tensor var_725_begin_0 = const()[name = string("op_725_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_725_end_0 = const()[name = string("op_725_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_725_end_mask_0 = const()[name = string("op_725_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_725_squeeze_mask_0 = const()[name = string("op_725_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, squeeze_mask = var_725_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_725_cast_fp16")]; + bool var_727_transpose_x_1 = const()[name = string("op_727_transpose_x_1"), val = bool(true)]; + bool var_727_transpose_y_1 = const()[name = string("op_727_transpose_y_1"), val = bool(false)]; + tensor var_727_cast_fp16 = matmul(transpose_x = var_727_transpose_x_1, transpose_y = var_727_transpose_y_1, x = var_725_cast_fp16, y = v_new_29_cast_fp16)[name = string("op_727_cast_fp16")]; + tensor state_31_cast_fp16 = add(x = var_723_cast_fp16, y = var_727_cast_fp16)[name = string("state_31_cast_fp16")]; + tensor var_730_begin_0 = const()[name = string("op_730_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_730_end_0 = const()[name = string("op_730_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_730_end_mask_0 = const()[name = string("op_730_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_730_squeeze_mask_0 = const()[name = string("op_730_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_730_cast_fp16 = slice_by_index(begin = var_730_begin_0, end = var_730_end_0, end_mask = var_730_end_mask_0, squeeze_mask = var_730_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_730_cast_fp16")]; + tensor var_732_begin_0 = const()[name = string("op_732_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_732_end_0 = const()[name = string("op_732_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_732_end_mask_0 = const()[name = string("op_732_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_732_squeeze_mask_0 = const()[name = string("op_732_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_732_cast_fp16 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, squeeze_mask = var_732_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_732_cast_fp16")]; + bool var_733_transpose_x_0 = const()[name = string("op_733_transpose_x_0"), val = bool(false)]; + bool var_733_transpose_y_0 = const()[name = string("op_733_transpose_y_0"), val = bool(false)]; + tensor var_733_cast_fp16 = matmul(transpose_x = var_733_transpose_x_0, transpose_y = var_733_transpose_y_0, x = var_732_cast_fp16, y = state_31_cast_fp16)[name = string("op_733_cast_fp16")]; + tensor v_new_31_cast_fp16 = sub(x = var_730_cast_fp16, y = var_733_cast_fp16)[name = string("v_new_31_cast_fp16")]; + tensor var_736_begin_0 = const()[name = string("op_736_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_736_end_0 = const()[name = string("op_736_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_736_end_mask_0 = const()[name = string("op_736_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_736_squeeze_mask_0 = const()[name = string("op_736_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_736_cast_fp16 = slice_by_index(begin = var_736_begin_0, end = var_736_end_0, end_mask = var_736_end_mask_0, squeeze_mask = var_736_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_736_cast_fp16")]; + bool var_737_transpose_x_0 = const()[name = string("op_737_transpose_x_0"), val = bool(false)]; + bool var_737_transpose_y_0 = const()[name = string("op_737_transpose_y_0"), val = bool(false)]; + tensor var_737_cast_fp16 = matmul(transpose_x = var_737_transpose_x_0, transpose_y = var_737_transpose_y_0, x = var_736_cast_fp16, y = state_31_cast_fp16)[name = string("op_737_cast_fp16")]; + tensor var_739_begin_0 = const()[name = string("op_739_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_739_end_0 = const()[name = string("op_739_end_0"), val = tensor([32, 16, 64, 64])]; + tensor var_739_end_mask_0 = const()[name = string("op_739_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_739_squeeze_mask_0 = const()[name = string("op_739_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_739_cast_fp16 = slice_by_index(begin = var_739_begin_0, end = var_739_end_0, end_mask = var_739_end_mask_0, squeeze_mask = var_739_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_739_cast_fp16")]; + bool var_740_transpose_x_0 = const()[name = string("op_740_transpose_x_0"), val = bool(false)]; + bool var_740_transpose_y_0 = const()[name = string("op_740_transpose_y_0"), val = bool(false)]; + tensor var_740_cast_fp16 = matmul(transpose_x = var_740_transpose_x_0, transpose_y = var_740_transpose_y_0, x = var_739_cast_fp16, y = v_new_31_cast_fp16)[name = string("op_740_cast_fp16")]; + tensor var_741_cast_fp16 = add(x = var_737_cast_fp16, y = var_740_cast_fp16)[name = string("op_741_cast_fp16")]; + tensor var_743_begin_0 = const()[name = string("op_743_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_743_end_0 = const()[name = string("op_743_end_0"), val = tensor([32, 16, 1, 1])]; + tensor var_743_end_mask_0 = const()[name = string("op_743_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_743_squeeze_mask_0 = const()[name = string("op_743_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_743_cast_fp16 = slice_by_index(begin = var_743_begin_0, end = var_743_end_0, end_mask = var_743_end_mask_0, squeeze_mask = var_743_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_743_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = state_31_cast_fp16, y = var_743_cast_fp16)[name = string("op_744_cast_fp16")]; + tensor var_746_begin_0 = const()[name = string("op_746_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_746_end_0 = const()[name = string("op_746_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_746_end_mask_0 = const()[name = string("op_746_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_746_squeeze_mask_0 = const()[name = string("op_746_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_746_cast_fp16 = slice_by_index(begin = var_746_begin_0, end = var_746_end_0, end_mask = var_746_end_mask_0, squeeze_mask = var_746_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_746_cast_fp16")]; + bool var_748_transpose_x_1 = const()[name = string("op_748_transpose_x_1"), val = bool(true)]; + bool var_748_transpose_y_1 = const()[name = string("op_748_transpose_y_1"), val = bool(false)]; + tensor var_748_cast_fp16 = matmul(transpose_x = var_748_transpose_x_1, transpose_y = var_748_transpose_y_1, x = var_746_cast_fp16, y = v_new_31_cast_fp16)[name = string("op_748_cast_fp16")]; + tensor state_33_cast_fp16 = add(x = var_744_cast_fp16, y = var_748_cast_fp16)[name = string("state_33_cast_fp16")]; + tensor var_751_begin_0 = const()[name = string("op_751_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_751_end_0 = const()[name = string("op_751_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_751_end_mask_0 = const()[name = string("op_751_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_751_squeeze_mask_0 = const()[name = string("op_751_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_751_cast_fp16 = slice_by_index(begin = var_751_begin_0, end = var_751_end_0, end_mask = var_751_end_mask_0, squeeze_mask = var_751_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_751_cast_fp16")]; + tensor var_753_begin_0 = const()[name = string("op_753_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_753_end_0 = const()[name = string("op_753_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_753_end_mask_0 = const()[name = string("op_753_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_753_squeeze_mask_0 = const()[name = string("op_753_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_753_cast_fp16 = slice_by_index(begin = var_753_begin_0, end = var_753_end_0, end_mask = var_753_end_mask_0, squeeze_mask = var_753_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_753_cast_fp16")]; + bool var_754_transpose_x_0 = const()[name = string("op_754_transpose_x_0"), val = bool(false)]; + bool var_754_transpose_y_0 = const()[name = string("op_754_transpose_y_0"), val = bool(false)]; + tensor var_754_cast_fp16 = matmul(transpose_x = var_754_transpose_x_0, transpose_y = var_754_transpose_y_0, x = var_753_cast_fp16, y = state_33_cast_fp16)[name = string("op_754_cast_fp16")]; + tensor v_new_33_cast_fp16 = sub(x = var_751_cast_fp16, y = var_754_cast_fp16)[name = string("v_new_33_cast_fp16")]; + tensor var_757_begin_0 = const()[name = string("op_757_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_757_end_0 = const()[name = string("op_757_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_757_end_mask_0 = const()[name = string("op_757_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_757_squeeze_mask_0 = const()[name = string("op_757_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_757_cast_fp16 = slice_by_index(begin = var_757_begin_0, end = var_757_end_0, end_mask = var_757_end_mask_0, squeeze_mask = var_757_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_757_cast_fp16")]; + bool var_758_transpose_x_0 = const()[name = string("op_758_transpose_x_0"), val = bool(false)]; + bool var_758_transpose_y_0 = const()[name = string("op_758_transpose_y_0"), val = bool(false)]; + tensor var_758_cast_fp16 = matmul(transpose_x = var_758_transpose_x_0, transpose_y = var_758_transpose_y_0, x = var_757_cast_fp16, y = state_33_cast_fp16)[name = string("op_758_cast_fp16")]; + tensor var_760_begin_0 = const()[name = string("op_760_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_760_end_0 = const()[name = string("op_760_end_0"), val = tensor([32, 17, 64, 64])]; + tensor var_760_end_mask_0 = const()[name = string("op_760_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_760_squeeze_mask_0 = const()[name = string("op_760_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_760_cast_fp16 = slice_by_index(begin = var_760_begin_0, end = var_760_end_0, end_mask = var_760_end_mask_0, squeeze_mask = var_760_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_760_cast_fp16")]; + bool var_761_transpose_x_0 = const()[name = string("op_761_transpose_x_0"), val = bool(false)]; + bool var_761_transpose_y_0 = const()[name = string("op_761_transpose_y_0"), val = bool(false)]; + tensor var_761_cast_fp16 = matmul(transpose_x = var_761_transpose_x_0, transpose_y = var_761_transpose_y_0, x = var_760_cast_fp16, y = v_new_33_cast_fp16)[name = string("op_761_cast_fp16")]; + tensor var_762_cast_fp16 = add(x = var_758_cast_fp16, y = var_761_cast_fp16)[name = string("op_762_cast_fp16")]; + tensor var_764_begin_0 = const()[name = string("op_764_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_764_end_0 = const()[name = string("op_764_end_0"), val = tensor([32, 17, 1, 1])]; + tensor var_764_end_mask_0 = const()[name = string("op_764_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_764_squeeze_mask_0 = const()[name = string("op_764_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_764_cast_fp16 = slice_by_index(begin = var_764_begin_0, end = var_764_end_0, end_mask = var_764_end_mask_0, squeeze_mask = var_764_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_764_cast_fp16")]; + tensor var_765_cast_fp16 = mul(x = state_33_cast_fp16, y = var_764_cast_fp16)[name = string("op_765_cast_fp16")]; + tensor var_767_begin_0 = const()[name = string("op_767_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_767_end_0 = const()[name = string("op_767_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_767_end_mask_0 = const()[name = string("op_767_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_767_squeeze_mask_0 = const()[name = string("op_767_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_767_cast_fp16 = slice_by_index(begin = var_767_begin_0, end = var_767_end_0, end_mask = var_767_end_mask_0, squeeze_mask = var_767_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_767_cast_fp16")]; + bool var_769_transpose_x_1 = const()[name = string("op_769_transpose_x_1"), val = bool(true)]; + bool var_769_transpose_y_1 = const()[name = string("op_769_transpose_y_1"), val = bool(false)]; + tensor var_769_cast_fp16 = matmul(transpose_x = var_769_transpose_x_1, transpose_y = var_769_transpose_y_1, x = var_767_cast_fp16, y = v_new_33_cast_fp16)[name = string("op_769_cast_fp16")]; + tensor state_35_cast_fp16 = add(x = var_765_cast_fp16, y = var_769_cast_fp16)[name = string("state_35_cast_fp16")]; + tensor var_772_begin_0 = const()[name = string("op_772_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_772_end_0 = const()[name = string("op_772_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_772_end_mask_0 = const()[name = string("op_772_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_772_squeeze_mask_0 = const()[name = string("op_772_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_772_cast_fp16 = slice_by_index(begin = var_772_begin_0, end = var_772_end_0, end_mask = var_772_end_mask_0, squeeze_mask = var_772_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_772_cast_fp16")]; + tensor var_774_begin_0 = const()[name = string("op_774_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_774_end_0 = const()[name = string("op_774_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_774_end_mask_0 = const()[name = string("op_774_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_774_squeeze_mask_0 = const()[name = string("op_774_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_774_cast_fp16 = slice_by_index(begin = var_774_begin_0, end = var_774_end_0, end_mask = var_774_end_mask_0, squeeze_mask = var_774_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_774_cast_fp16")]; + bool var_775_transpose_x_0 = const()[name = string("op_775_transpose_x_0"), val = bool(false)]; + bool var_775_transpose_y_0 = const()[name = string("op_775_transpose_y_0"), val = bool(false)]; + tensor var_775_cast_fp16 = matmul(transpose_x = var_775_transpose_x_0, transpose_y = var_775_transpose_y_0, x = var_774_cast_fp16, y = state_35_cast_fp16)[name = string("op_775_cast_fp16")]; + tensor v_new_35_cast_fp16 = sub(x = var_772_cast_fp16, y = var_775_cast_fp16)[name = string("v_new_35_cast_fp16")]; + tensor var_778_begin_0 = const()[name = string("op_778_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_778_end_0 = const()[name = string("op_778_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_778_end_mask_0 = const()[name = string("op_778_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_778_squeeze_mask_0 = const()[name = string("op_778_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_778_cast_fp16 = slice_by_index(begin = var_778_begin_0, end = var_778_end_0, end_mask = var_778_end_mask_0, squeeze_mask = var_778_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_778_cast_fp16")]; + bool var_779_transpose_x_0 = const()[name = string("op_779_transpose_x_0"), val = bool(false)]; + bool var_779_transpose_y_0 = const()[name = string("op_779_transpose_y_0"), val = bool(false)]; + tensor var_779_cast_fp16 = matmul(transpose_x = var_779_transpose_x_0, transpose_y = var_779_transpose_y_0, x = var_778_cast_fp16, y = state_35_cast_fp16)[name = string("op_779_cast_fp16")]; + tensor var_781_begin_0 = const()[name = string("op_781_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_781_end_0 = const()[name = string("op_781_end_0"), val = tensor([32, 18, 64, 64])]; + tensor var_781_end_mask_0 = const()[name = string("op_781_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_781_squeeze_mask_0 = const()[name = string("op_781_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_781_cast_fp16 = slice_by_index(begin = var_781_begin_0, end = var_781_end_0, end_mask = var_781_end_mask_0, squeeze_mask = var_781_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_781_cast_fp16")]; + bool var_782_transpose_x_0 = const()[name = string("op_782_transpose_x_0"), val = bool(false)]; + bool var_782_transpose_y_0 = const()[name = string("op_782_transpose_y_0"), val = bool(false)]; + tensor var_782_cast_fp16 = matmul(transpose_x = var_782_transpose_x_0, transpose_y = var_782_transpose_y_0, x = var_781_cast_fp16, y = v_new_35_cast_fp16)[name = string("op_782_cast_fp16")]; + tensor var_783_cast_fp16 = add(x = var_779_cast_fp16, y = var_782_cast_fp16)[name = string("op_783_cast_fp16")]; + tensor var_785_begin_0 = const()[name = string("op_785_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_785_end_0 = const()[name = string("op_785_end_0"), val = tensor([32, 18, 1, 1])]; + tensor var_785_end_mask_0 = const()[name = string("op_785_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_785_squeeze_mask_0 = const()[name = string("op_785_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_785_cast_fp16 = slice_by_index(begin = var_785_begin_0, end = var_785_end_0, end_mask = var_785_end_mask_0, squeeze_mask = var_785_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_785_cast_fp16")]; + tensor var_786_cast_fp16 = mul(x = state_35_cast_fp16, y = var_785_cast_fp16)[name = string("op_786_cast_fp16")]; + tensor var_788_begin_0 = const()[name = string("op_788_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_788_end_0 = const()[name = string("op_788_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_788_end_mask_0 = const()[name = string("op_788_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_788_squeeze_mask_0 = const()[name = string("op_788_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_788_cast_fp16 = slice_by_index(begin = var_788_begin_0, end = var_788_end_0, end_mask = var_788_end_mask_0, squeeze_mask = var_788_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_788_cast_fp16")]; + bool var_790_transpose_x_1 = const()[name = string("op_790_transpose_x_1"), val = bool(true)]; + bool var_790_transpose_y_1 = const()[name = string("op_790_transpose_y_1"), val = bool(false)]; + tensor var_790_cast_fp16 = matmul(transpose_x = var_790_transpose_x_1, transpose_y = var_790_transpose_y_1, x = var_788_cast_fp16, y = v_new_35_cast_fp16)[name = string("op_790_cast_fp16")]; + tensor state_37_cast_fp16 = add(x = var_786_cast_fp16, y = var_790_cast_fp16)[name = string("state_37_cast_fp16")]; + tensor var_793_begin_0 = const()[name = string("op_793_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_793_end_0 = const()[name = string("op_793_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_793_end_mask_0 = const()[name = string("op_793_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_793_squeeze_mask_0 = const()[name = string("op_793_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_793_cast_fp16 = slice_by_index(begin = var_793_begin_0, end = var_793_end_0, end_mask = var_793_end_mask_0, squeeze_mask = var_793_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_793_cast_fp16")]; + tensor var_795_begin_0 = const()[name = string("op_795_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_795_end_0 = const()[name = string("op_795_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_795_end_mask_0 = const()[name = string("op_795_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_795_squeeze_mask_0 = const()[name = string("op_795_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_795_cast_fp16 = slice_by_index(begin = var_795_begin_0, end = var_795_end_0, end_mask = var_795_end_mask_0, squeeze_mask = var_795_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_795_cast_fp16")]; + bool var_796_transpose_x_0 = const()[name = string("op_796_transpose_x_0"), val = bool(false)]; + bool var_796_transpose_y_0 = const()[name = string("op_796_transpose_y_0"), val = bool(false)]; + tensor var_796_cast_fp16 = matmul(transpose_x = var_796_transpose_x_0, transpose_y = var_796_transpose_y_0, x = var_795_cast_fp16, y = state_37_cast_fp16)[name = string("op_796_cast_fp16")]; + tensor v_new_37_cast_fp16 = sub(x = var_793_cast_fp16, y = var_796_cast_fp16)[name = string("v_new_37_cast_fp16")]; + tensor var_799_begin_0 = const()[name = string("op_799_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_799_end_0 = const()[name = string("op_799_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_799_end_mask_0 = const()[name = string("op_799_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_799_squeeze_mask_0 = const()[name = string("op_799_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_799_cast_fp16 = slice_by_index(begin = var_799_begin_0, end = var_799_end_0, end_mask = var_799_end_mask_0, squeeze_mask = var_799_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_799_cast_fp16")]; + bool var_800_transpose_x_0 = const()[name = string("op_800_transpose_x_0"), val = bool(false)]; + bool var_800_transpose_y_0 = const()[name = string("op_800_transpose_y_0"), val = bool(false)]; + tensor var_800_cast_fp16 = matmul(transpose_x = var_800_transpose_x_0, transpose_y = var_800_transpose_y_0, x = var_799_cast_fp16, y = state_37_cast_fp16)[name = string("op_800_cast_fp16")]; + tensor var_802_begin_0 = const()[name = string("op_802_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_802_end_0 = const()[name = string("op_802_end_0"), val = tensor([32, 19, 64, 64])]; + tensor var_802_end_mask_0 = const()[name = string("op_802_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_802_squeeze_mask_0 = const()[name = string("op_802_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_802_cast_fp16 = slice_by_index(begin = var_802_begin_0, end = var_802_end_0, end_mask = var_802_end_mask_0, squeeze_mask = var_802_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_802_cast_fp16")]; + bool var_803_transpose_x_0 = const()[name = string("op_803_transpose_x_0"), val = bool(false)]; + bool var_803_transpose_y_0 = const()[name = string("op_803_transpose_y_0"), val = bool(false)]; + tensor var_803_cast_fp16 = matmul(transpose_x = var_803_transpose_x_0, transpose_y = var_803_transpose_y_0, x = var_802_cast_fp16, y = v_new_37_cast_fp16)[name = string("op_803_cast_fp16")]; + tensor var_804_cast_fp16 = add(x = var_800_cast_fp16, y = var_803_cast_fp16)[name = string("op_804_cast_fp16")]; + tensor var_806_begin_0 = const()[name = string("op_806_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_806_end_0 = const()[name = string("op_806_end_0"), val = tensor([32, 19, 1, 1])]; + tensor var_806_end_mask_0 = const()[name = string("op_806_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_806_squeeze_mask_0 = const()[name = string("op_806_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_806_cast_fp16 = slice_by_index(begin = var_806_begin_0, end = var_806_end_0, end_mask = var_806_end_mask_0, squeeze_mask = var_806_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_806_cast_fp16")]; + tensor var_807_cast_fp16 = mul(x = state_37_cast_fp16, y = var_806_cast_fp16)[name = string("op_807_cast_fp16")]; + tensor var_809_begin_0 = const()[name = string("op_809_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_809_end_0 = const()[name = string("op_809_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_809_end_mask_0 = const()[name = string("op_809_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_809_squeeze_mask_0 = const()[name = string("op_809_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_809_cast_fp16 = slice_by_index(begin = var_809_begin_0, end = var_809_end_0, end_mask = var_809_end_mask_0, squeeze_mask = var_809_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_809_cast_fp16")]; + bool var_811_transpose_x_1 = const()[name = string("op_811_transpose_x_1"), val = bool(true)]; + bool var_811_transpose_y_1 = const()[name = string("op_811_transpose_y_1"), val = bool(false)]; + tensor var_811_cast_fp16 = matmul(transpose_x = var_811_transpose_x_1, transpose_y = var_811_transpose_y_1, x = var_809_cast_fp16, y = v_new_37_cast_fp16)[name = string("op_811_cast_fp16")]; + tensor state_39_cast_fp16 = add(x = var_807_cast_fp16, y = var_811_cast_fp16)[name = string("state_39_cast_fp16")]; + tensor var_814_begin_0 = const()[name = string("op_814_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_814_end_0 = const()[name = string("op_814_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_814_end_mask_0 = const()[name = string("op_814_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_814_squeeze_mask_0 = const()[name = string("op_814_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_814_cast_fp16 = slice_by_index(begin = var_814_begin_0, end = var_814_end_0, end_mask = var_814_end_mask_0, squeeze_mask = var_814_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_814_cast_fp16")]; + tensor var_816_begin_0 = const()[name = string("op_816_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_816_end_0 = const()[name = string("op_816_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_816_end_mask_0 = const()[name = string("op_816_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_816_squeeze_mask_0 = const()[name = string("op_816_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_816_cast_fp16 = slice_by_index(begin = var_816_begin_0, end = var_816_end_0, end_mask = var_816_end_mask_0, squeeze_mask = var_816_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_816_cast_fp16")]; + bool var_817_transpose_x_0 = const()[name = string("op_817_transpose_x_0"), val = bool(false)]; + bool var_817_transpose_y_0 = const()[name = string("op_817_transpose_y_0"), val = bool(false)]; + tensor var_817_cast_fp16 = matmul(transpose_x = var_817_transpose_x_0, transpose_y = var_817_transpose_y_0, x = var_816_cast_fp16, y = state_39_cast_fp16)[name = string("op_817_cast_fp16")]; + tensor v_new_39_cast_fp16 = sub(x = var_814_cast_fp16, y = var_817_cast_fp16)[name = string("v_new_39_cast_fp16")]; + tensor var_820_begin_0 = const()[name = string("op_820_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_820_end_0 = const()[name = string("op_820_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_820_end_mask_0 = const()[name = string("op_820_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_820_squeeze_mask_0 = const()[name = string("op_820_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_820_cast_fp16 = slice_by_index(begin = var_820_begin_0, end = var_820_end_0, end_mask = var_820_end_mask_0, squeeze_mask = var_820_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_820_cast_fp16")]; + bool var_821_transpose_x_0 = const()[name = string("op_821_transpose_x_0"), val = bool(false)]; + bool var_821_transpose_y_0 = const()[name = string("op_821_transpose_y_0"), val = bool(false)]; + tensor var_821_cast_fp16 = matmul(transpose_x = var_821_transpose_x_0, transpose_y = var_821_transpose_y_0, x = var_820_cast_fp16, y = state_39_cast_fp16)[name = string("op_821_cast_fp16")]; + tensor var_823_begin_0 = const()[name = string("op_823_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_823_end_0 = const()[name = string("op_823_end_0"), val = tensor([32, 20, 64, 64])]; + tensor var_823_end_mask_0 = const()[name = string("op_823_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_823_squeeze_mask_0 = const()[name = string("op_823_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_823_cast_fp16 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, squeeze_mask = var_823_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_823_cast_fp16")]; + bool var_824_transpose_x_0 = const()[name = string("op_824_transpose_x_0"), val = bool(false)]; + bool var_824_transpose_y_0 = const()[name = string("op_824_transpose_y_0"), val = bool(false)]; + tensor var_824_cast_fp16 = matmul(transpose_x = var_824_transpose_x_0, transpose_y = var_824_transpose_y_0, x = var_823_cast_fp16, y = v_new_39_cast_fp16)[name = string("op_824_cast_fp16")]; + tensor var_825_cast_fp16 = add(x = var_821_cast_fp16, y = var_824_cast_fp16)[name = string("op_825_cast_fp16")]; + tensor var_827_begin_0 = const()[name = string("op_827_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_827_end_0 = const()[name = string("op_827_end_0"), val = tensor([32, 20, 1, 1])]; + tensor var_827_end_mask_0 = const()[name = string("op_827_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_827_squeeze_mask_0 = const()[name = string("op_827_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_827_cast_fp16 = slice_by_index(begin = var_827_begin_0, end = var_827_end_0, end_mask = var_827_end_mask_0, squeeze_mask = var_827_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_827_cast_fp16")]; + tensor var_828_cast_fp16 = mul(x = state_39_cast_fp16, y = var_827_cast_fp16)[name = string("op_828_cast_fp16")]; + tensor var_830_begin_0 = const()[name = string("op_830_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_830_end_0 = const()[name = string("op_830_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_830_end_mask_0 = const()[name = string("op_830_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_830_squeeze_mask_0 = const()[name = string("op_830_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_830_cast_fp16 = slice_by_index(begin = var_830_begin_0, end = var_830_end_0, end_mask = var_830_end_mask_0, squeeze_mask = var_830_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_830_cast_fp16")]; + bool var_832_transpose_x_1 = const()[name = string("op_832_transpose_x_1"), val = bool(true)]; + bool var_832_transpose_y_1 = const()[name = string("op_832_transpose_y_1"), val = bool(false)]; + tensor var_832_cast_fp16 = matmul(transpose_x = var_832_transpose_x_1, transpose_y = var_832_transpose_y_1, x = var_830_cast_fp16, y = v_new_39_cast_fp16)[name = string("op_832_cast_fp16")]; + tensor state_41_cast_fp16 = add(x = var_828_cast_fp16, y = var_832_cast_fp16)[name = string("state_41_cast_fp16")]; + tensor var_835_begin_0 = const()[name = string("op_835_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_835_end_0 = const()[name = string("op_835_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_835_end_mask_0 = const()[name = string("op_835_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_835_squeeze_mask_0 = const()[name = string("op_835_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_835_cast_fp16 = slice_by_index(begin = var_835_begin_0, end = var_835_end_0, end_mask = var_835_end_mask_0, squeeze_mask = var_835_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_835_cast_fp16")]; + tensor var_837_begin_0 = const()[name = string("op_837_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_837_end_0 = const()[name = string("op_837_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_837_end_mask_0 = const()[name = string("op_837_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_837_squeeze_mask_0 = const()[name = string("op_837_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_837_cast_fp16 = slice_by_index(begin = var_837_begin_0, end = var_837_end_0, end_mask = var_837_end_mask_0, squeeze_mask = var_837_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_837_cast_fp16")]; + bool var_838_transpose_x_0 = const()[name = string("op_838_transpose_x_0"), val = bool(false)]; + bool var_838_transpose_y_0 = const()[name = string("op_838_transpose_y_0"), val = bool(false)]; + tensor var_838_cast_fp16 = matmul(transpose_x = var_838_transpose_x_0, transpose_y = var_838_transpose_y_0, x = var_837_cast_fp16, y = state_41_cast_fp16)[name = string("op_838_cast_fp16")]; + tensor v_new_41_cast_fp16 = sub(x = var_835_cast_fp16, y = var_838_cast_fp16)[name = string("v_new_41_cast_fp16")]; + tensor var_841_begin_0 = const()[name = string("op_841_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_841_end_0 = const()[name = string("op_841_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_841_end_mask_0 = const()[name = string("op_841_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_841_squeeze_mask_0 = const()[name = string("op_841_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_841_cast_fp16 = slice_by_index(begin = var_841_begin_0, end = var_841_end_0, end_mask = var_841_end_mask_0, squeeze_mask = var_841_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_841_cast_fp16")]; + bool var_842_transpose_x_0 = const()[name = string("op_842_transpose_x_0"), val = bool(false)]; + bool var_842_transpose_y_0 = const()[name = string("op_842_transpose_y_0"), val = bool(false)]; + tensor var_842_cast_fp16 = matmul(transpose_x = var_842_transpose_x_0, transpose_y = var_842_transpose_y_0, x = var_841_cast_fp16, y = state_41_cast_fp16)[name = string("op_842_cast_fp16")]; + tensor var_844_begin_0 = const()[name = string("op_844_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_844_end_0 = const()[name = string("op_844_end_0"), val = tensor([32, 21, 64, 64])]; + tensor var_844_end_mask_0 = const()[name = string("op_844_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_844_squeeze_mask_0 = const()[name = string("op_844_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_844_cast_fp16 = slice_by_index(begin = var_844_begin_0, end = var_844_end_0, end_mask = var_844_end_mask_0, squeeze_mask = var_844_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_844_cast_fp16")]; + bool var_845_transpose_x_0 = const()[name = string("op_845_transpose_x_0"), val = bool(false)]; + bool var_845_transpose_y_0 = const()[name = string("op_845_transpose_y_0"), val = bool(false)]; + tensor var_845_cast_fp16 = matmul(transpose_x = var_845_transpose_x_0, transpose_y = var_845_transpose_y_0, x = var_844_cast_fp16, y = v_new_41_cast_fp16)[name = string("op_845_cast_fp16")]; + tensor var_846_cast_fp16 = add(x = var_842_cast_fp16, y = var_845_cast_fp16)[name = string("op_846_cast_fp16")]; + tensor var_848_begin_0 = const()[name = string("op_848_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_848_end_0 = const()[name = string("op_848_end_0"), val = tensor([32, 21, 1, 1])]; + tensor var_848_end_mask_0 = const()[name = string("op_848_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_848_squeeze_mask_0 = const()[name = string("op_848_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_848_cast_fp16 = slice_by_index(begin = var_848_begin_0, end = var_848_end_0, end_mask = var_848_end_mask_0, squeeze_mask = var_848_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_848_cast_fp16")]; + tensor var_849_cast_fp16 = mul(x = state_41_cast_fp16, y = var_848_cast_fp16)[name = string("op_849_cast_fp16")]; + tensor var_851_begin_0 = const()[name = string("op_851_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_851_end_0 = const()[name = string("op_851_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_851_end_mask_0 = const()[name = string("op_851_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_851_squeeze_mask_0 = const()[name = string("op_851_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, squeeze_mask = var_851_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_851_cast_fp16")]; + bool var_853_transpose_x_1 = const()[name = string("op_853_transpose_x_1"), val = bool(true)]; + bool var_853_transpose_y_1 = const()[name = string("op_853_transpose_y_1"), val = bool(false)]; + tensor var_853_cast_fp16 = matmul(transpose_x = var_853_transpose_x_1, transpose_y = var_853_transpose_y_1, x = var_851_cast_fp16, y = v_new_41_cast_fp16)[name = string("op_853_cast_fp16")]; + tensor state_43_cast_fp16 = add(x = var_849_cast_fp16, y = var_853_cast_fp16)[name = string("state_43_cast_fp16")]; + tensor var_856_begin_0 = const()[name = string("op_856_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_856_end_0 = const()[name = string("op_856_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_856_end_mask_0 = const()[name = string("op_856_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_856_squeeze_mask_0 = const()[name = string("op_856_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_856_cast_fp16 = slice_by_index(begin = var_856_begin_0, end = var_856_end_0, end_mask = var_856_end_mask_0, squeeze_mask = var_856_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_856_cast_fp16")]; + tensor var_858_begin_0 = const()[name = string("op_858_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_858_end_0 = const()[name = string("op_858_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_858_end_mask_0 = const()[name = string("op_858_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_858_squeeze_mask_0 = const()[name = string("op_858_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, squeeze_mask = var_858_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_858_cast_fp16")]; + bool var_859_transpose_x_0 = const()[name = string("op_859_transpose_x_0"), val = bool(false)]; + bool var_859_transpose_y_0 = const()[name = string("op_859_transpose_y_0"), val = bool(false)]; + tensor var_859_cast_fp16 = matmul(transpose_x = var_859_transpose_x_0, transpose_y = var_859_transpose_y_0, x = var_858_cast_fp16, y = state_43_cast_fp16)[name = string("op_859_cast_fp16")]; + tensor v_new_43_cast_fp16 = sub(x = var_856_cast_fp16, y = var_859_cast_fp16)[name = string("v_new_43_cast_fp16")]; + tensor var_862_begin_0 = const()[name = string("op_862_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_862_end_0 = const()[name = string("op_862_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_862_end_mask_0 = const()[name = string("op_862_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_862_squeeze_mask_0 = const()[name = string("op_862_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_862_cast_fp16 = slice_by_index(begin = var_862_begin_0, end = var_862_end_0, end_mask = var_862_end_mask_0, squeeze_mask = var_862_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_862_cast_fp16")]; + bool var_863_transpose_x_0 = const()[name = string("op_863_transpose_x_0"), val = bool(false)]; + bool var_863_transpose_y_0 = const()[name = string("op_863_transpose_y_0"), val = bool(false)]; + tensor var_863_cast_fp16 = matmul(transpose_x = var_863_transpose_x_0, transpose_y = var_863_transpose_y_0, x = var_862_cast_fp16, y = state_43_cast_fp16)[name = string("op_863_cast_fp16")]; + tensor var_865_begin_0 = const()[name = string("op_865_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_865_end_0 = const()[name = string("op_865_end_0"), val = tensor([32, 22, 64, 64])]; + tensor var_865_end_mask_0 = const()[name = string("op_865_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_865_squeeze_mask_0 = const()[name = string("op_865_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_865_cast_fp16 = slice_by_index(begin = var_865_begin_0, end = var_865_end_0, end_mask = var_865_end_mask_0, squeeze_mask = var_865_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_865_cast_fp16")]; + bool var_866_transpose_x_0 = const()[name = string("op_866_transpose_x_0"), val = bool(false)]; + bool var_866_transpose_y_0 = const()[name = string("op_866_transpose_y_0"), val = bool(false)]; + tensor var_866_cast_fp16 = matmul(transpose_x = var_866_transpose_x_0, transpose_y = var_866_transpose_y_0, x = var_865_cast_fp16, y = v_new_43_cast_fp16)[name = string("op_866_cast_fp16")]; + tensor var_867_cast_fp16 = add(x = var_863_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")]; + tensor var_869_begin_0 = const()[name = string("op_869_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_869_end_0 = const()[name = string("op_869_end_0"), val = tensor([32, 22, 1, 1])]; + tensor var_869_end_mask_0 = const()[name = string("op_869_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_869_squeeze_mask_0 = const()[name = string("op_869_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_869_cast_fp16 = slice_by_index(begin = var_869_begin_0, end = var_869_end_0, end_mask = var_869_end_mask_0, squeeze_mask = var_869_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_869_cast_fp16")]; + tensor var_870_cast_fp16 = mul(x = state_43_cast_fp16, y = var_869_cast_fp16)[name = string("op_870_cast_fp16")]; + tensor var_872_begin_0 = const()[name = string("op_872_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_872_end_0 = const()[name = string("op_872_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_872_end_mask_0 = const()[name = string("op_872_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_872_squeeze_mask_0 = const()[name = string("op_872_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_872_cast_fp16 = slice_by_index(begin = var_872_begin_0, end = var_872_end_0, end_mask = var_872_end_mask_0, squeeze_mask = var_872_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_872_cast_fp16")]; + bool var_874_transpose_x_1 = const()[name = string("op_874_transpose_x_1"), val = bool(true)]; + bool var_874_transpose_y_1 = const()[name = string("op_874_transpose_y_1"), val = bool(false)]; + tensor var_874_cast_fp16 = matmul(transpose_x = var_874_transpose_x_1, transpose_y = var_874_transpose_y_1, x = var_872_cast_fp16, y = v_new_43_cast_fp16)[name = string("op_874_cast_fp16")]; + tensor state_45_cast_fp16 = add(x = var_870_cast_fp16, y = var_874_cast_fp16)[name = string("state_45_cast_fp16")]; + tensor var_877_begin_0 = const()[name = string("op_877_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_877_end_0 = const()[name = string("op_877_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_877_end_mask_0 = const()[name = string("op_877_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_877_squeeze_mask_0 = const()[name = string("op_877_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_877_cast_fp16 = slice_by_index(begin = var_877_begin_0, end = var_877_end_0, end_mask = var_877_end_mask_0, squeeze_mask = var_877_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_877_cast_fp16")]; + tensor var_879_begin_0 = const()[name = string("op_879_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_879_end_0 = const()[name = string("op_879_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_879_end_mask_0 = const()[name = string("op_879_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_879_squeeze_mask_0 = const()[name = string("op_879_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_879_cast_fp16 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, squeeze_mask = var_879_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_879_cast_fp16")]; + bool var_880_transpose_x_0 = const()[name = string("op_880_transpose_x_0"), val = bool(false)]; + bool var_880_transpose_y_0 = const()[name = string("op_880_transpose_y_0"), val = bool(false)]; + tensor var_880_cast_fp16 = matmul(transpose_x = var_880_transpose_x_0, transpose_y = var_880_transpose_y_0, x = var_879_cast_fp16, y = state_45_cast_fp16)[name = string("op_880_cast_fp16")]; + tensor v_new_45_cast_fp16 = sub(x = var_877_cast_fp16, y = var_880_cast_fp16)[name = string("v_new_45_cast_fp16")]; + tensor var_883_begin_0 = const()[name = string("op_883_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_883_end_0 = const()[name = string("op_883_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_883_end_mask_0 = const()[name = string("op_883_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_883_squeeze_mask_0 = const()[name = string("op_883_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_883_cast_fp16 = slice_by_index(begin = var_883_begin_0, end = var_883_end_0, end_mask = var_883_end_mask_0, squeeze_mask = var_883_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_883_cast_fp16")]; + bool var_884_transpose_x_0 = const()[name = string("op_884_transpose_x_0"), val = bool(false)]; + bool var_884_transpose_y_0 = const()[name = string("op_884_transpose_y_0"), val = bool(false)]; + tensor var_884_cast_fp16 = matmul(transpose_x = var_884_transpose_x_0, transpose_y = var_884_transpose_y_0, x = var_883_cast_fp16, y = state_45_cast_fp16)[name = string("op_884_cast_fp16")]; + tensor var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor([32, 23, 64, 64])]; + tensor var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_886_squeeze_mask_0 = const()[name = string("op_886_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, squeeze_mask = var_886_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_886_cast_fp16")]; + bool var_887_transpose_x_0 = const()[name = string("op_887_transpose_x_0"), val = bool(false)]; + bool var_887_transpose_y_0 = const()[name = string("op_887_transpose_y_0"), val = bool(false)]; + tensor var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_0, transpose_y = var_887_transpose_y_0, x = var_886_cast_fp16, y = v_new_45_cast_fp16)[name = string("op_887_cast_fp16")]; + tensor var_888_cast_fp16 = add(x = var_884_cast_fp16, y = var_887_cast_fp16)[name = string("op_888_cast_fp16")]; + tensor var_890_begin_0 = const()[name = string("op_890_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_890_end_0 = const()[name = string("op_890_end_0"), val = tensor([32, 23, 1, 1])]; + tensor var_890_end_mask_0 = const()[name = string("op_890_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_890_squeeze_mask_0 = const()[name = string("op_890_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_890_cast_fp16 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, squeeze_mask = var_890_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_890_cast_fp16")]; + tensor var_891_cast_fp16 = mul(x = state_45_cast_fp16, y = var_890_cast_fp16)[name = string("op_891_cast_fp16")]; + tensor var_893_begin_0 = const()[name = string("op_893_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_893_end_0 = const()[name = string("op_893_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_893_end_mask_0 = const()[name = string("op_893_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_893_squeeze_mask_0 = const()[name = string("op_893_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_893_cast_fp16 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, squeeze_mask = var_893_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_893_cast_fp16")]; + bool var_895_transpose_x_1 = const()[name = string("op_895_transpose_x_1"), val = bool(true)]; + bool var_895_transpose_y_1 = const()[name = string("op_895_transpose_y_1"), val = bool(false)]; + tensor var_895_cast_fp16 = matmul(transpose_x = var_895_transpose_x_1, transpose_y = var_895_transpose_y_1, x = var_893_cast_fp16, y = v_new_45_cast_fp16)[name = string("op_895_cast_fp16")]; + tensor state_47_cast_fp16 = add(x = var_891_cast_fp16, y = var_895_cast_fp16)[name = string("state_47_cast_fp16")]; + tensor var_898_begin_0 = const()[name = string("op_898_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_898_end_0 = const()[name = string("op_898_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_898_end_mask_0 = const()[name = string("op_898_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_898_squeeze_mask_0 = const()[name = string("op_898_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_898_cast_fp16 = slice_by_index(begin = var_898_begin_0, end = var_898_end_0, end_mask = var_898_end_mask_0, squeeze_mask = var_898_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_898_cast_fp16")]; + tensor var_900_begin_0 = const()[name = string("op_900_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_900_end_0 = const()[name = string("op_900_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_900_end_mask_0 = const()[name = string("op_900_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_900_squeeze_mask_0 = const()[name = string("op_900_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_900_cast_fp16 = slice_by_index(begin = var_900_begin_0, end = var_900_end_0, end_mask = var_900_end_mask_0, squeeze_mask = var_900_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_900_cast_fp16")]; + bool var_901_transpose_x_0 = const()[name = string("op_901_transpose_x_0"), val = bool(false)]; + bool var_901_transpose_y_0 = const()[name = string("op_901_transpose_y_0"), val = bool(false)]; + tensor var_901_cast_fp16 = matmul(transpose_x = var_901_transpose_x_0, transpose_y = var_901_transpose_y_0, x = var_900_cast_fp16, y = state_47_cast_fp16)[name = string("op_901_cast_fp16")]; + tensor v_new_47_cast_fp16 = sub(x = var_898_cast_fp16, y = var_901_cast_fp16)[name = string("v_new_47_cast_fp16")]; + tensor var_904_begin_0 = const()[name = string("op_904_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_904_end_0 = const()[name = string("op_904_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_904_end_mask_0 = const()[name = string("op_904_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_904_squeeze_mask_0 = const()[name = string("op_904_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_904_cast_fp16 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, squeeze_mask = var_904_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_904_cast_fp16")]; + bool var_905_transpose_x_0 = const()[name = string("op_905_transpose_x_0"), val = bool(false)]; + bool var_905_transpose_y_0 = const()[name = string("op_905_transpose_y_0"), val = bool(false)]; + tensor var_905_cast_fp16 = matmul(transpose_x = var_905_transpose_x_0, transpose_y = var_905_transpose_y_0, x = var_904_cast_fp16, y = state_47_cast_fp16)[name = string("op_905_cast_fp16")]; + tensor var_907_begin_0 = const()[name = string("op_907_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_907_end_0 = const()[name = string("op_907_end_0"), val = tensor([32, 24, 64, 64])]; + tensor var_907_end_mask_0 = const()[name = string("op_907_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_907_squeeze_mask_0 = const()[name = string("op_907_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_907_cast_fp16 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, squeeze_mask = var_907_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_907_cast_fp16")]; + bool var_908_transpose_x_0 = const()[name = string("op_908_transpose_x_0"), val = bool(false)]; + bool var_908_transpose_y_0 = const()[name = string("op_908_transpose_y_0"), val = bool(false)]; + tensor var_908_cast_fp16 = matmul(transpose_x = var_908_transpose_x_0, transpose_y = var_908_transpose_y_0, x = var_907_cast_fp16, y = v_new_47_cast_fp16)[name = string("op_908_cast_fp16")]; + tensor var_909_cast_fp16 = add(x = var_905_cast_fp16, y = var_908_cast_fp16)[name = string("op_909_cast_fp16")]; + tensor var_911_begin_0 = const()[name = string("op_911_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_911_end_0 = const()[name = string("op_911_end_0"), val = tensor([32, 24, 1, 1])]; + tensor var_911_end_mask_0 = const()[name = string("op_911_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_911_squeeze_mask_0 = const()[name = string("op_911_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_911_cast_fp16 = slice_by_index(begin = var_911_begin_0, end = var_911_end_0, end_mask = var_911_end_mask_0, squeeze_mask = var_911_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_911_cast_fp16")]; + tensor var_912_cast_fp16 = mul(x = state_47_cast_fp16, y = var_911_cast_fp16)[name = string("op_912_cast_fp16")]; + tensor var_914_begin_0 = const()[name = string("op_914_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_914_end_0 = const()[name = string("op_914_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_914_end_mask_0 = const()[name = string("op_914_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_914_squeeze_mask_0 = const()[name = string("op_914_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_914_cast_fp16 = slice_by_index(begin = var_914_begin_0, end = var_914_end_0, end_mask = var_914_end_mask_0, squeeze_mask = var_914_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_914_cast_fp16")]; + bool var_916_transpose_x_1 = const()[name = string("op_916_transpose_x_1"), val = bool(true)]; + bool var_916_transpose_y_1 = const()[name = string("op_916_transpose_y_1"), val = bool(false)]; + tensor var_916_cast_fp16 = matmul(transpose_x = var_916_transpose_x_1, transpose_y = var_916_transpose_y_1, x = var_914_cast_fp16, y = v_new_47_cast_fp16)[name = string("op_916_cast_fp16")]; + tensor state_49_cast_fp16 = add(x = var_912_cast_fp16, y = var_916_cast_fp16)[name = string("state_49_cast_fp16")]; + tensor var_919_begin_0 = const()[name = string("op_919_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_919_end_0 = const()[name = string("op_919_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_919_end_mask_0 = const()[name = string("op_919_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_919_squeeze_mask_0 = const()[name = string("op_919_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_919_cast_fp16 = slice_by_index(begin = var_919_begin_0, end = var_919_end_0, end_mask = var_919_end_mask_0, squeeze_mask = var_919_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_919_cast_fp16")]; + tensor var_921_begin_0 = const()[name = string("op_921_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_921_end_0 = const()[name = string("op_921_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_921_end_mask_0 = const()[name = string("op_921_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_921_squeeze_mask_0 = const()[name = string("op_921_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_921_cast_fp16 = slice_by_index(begin = var_921_begin_0, end = var_921_end_0, end_mask = var_921_end_mask_0, squeeze_mask = var_921_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_921_cast_fp16")]; + bool var_922_transpose_x_0 = const()[name = string("op_922_transpose_x_0"), val = bool(false)]; + bool var_922_transpose_y_0 = const()[name = string("op_922_transpose_y_0"), val = bool(false)]; + tensor var_922_cast_fp16 = matmul(transpose_x = var_922_transpose_x_0, transpose_y = var_922_transpose_y_0, x = var_921_cast_fp16, y = state_49_cast_fp16)[name = string("op_922_cast_fp16")]; + tensor v_new_49_cast_fp16 = sub(x = var_919_cast_fp16, y = var_922_cast_fp16)[name = string("v_new_49_cast_fp16")]; + tensor var_925_begin_0 = const()[name = string("op_925_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_925_end_0 = const()[name = string("op_925_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_925_end_mask_0 = const()[name = string("op_925_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_925_squeeze_mask_0 = const()[name = string("op_925_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_925_cast_fp16 = slice_by_index(begin = var_925_begin_0, end = var_925_end_0, end_mask = var_925_end_mask_0, squeeze_mask = var_925_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_925_cast_fp16")]; + bool var_926_transpose_x_0 = const()[name = string("op_926_transpose_x_0"), val = bool(false)]; + bool var_926_transpose_y_0 = const()[name = string("op_926_transpose_y_0"), val = bool(false)]; + tensor var_926_cast_fp16 = matmul(transpose_x = var_926_transpose_x_0, transpose_y = var_926_transpose_y_0, x = var_925_cast_fp16, y = state_49_cast_fp16)[name = string("op_926_cast_fp16")]; + tensor var_928_begin_0 = const()[name = string("op_928_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_928_end_0 = const()[name = string("op_928_end_0"), val = tensor([32, 25, 64, 64])]; + tensor var_928_end_mask_0 = const()[name = string("op_928_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_928_squeeze_mask_0 = const()[name = string("op_928_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_928_cast_fp16 = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, squeeze_mask = var_928_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_928_cast_fp16")]; + bool var_929_transpose_x_0 = const()[name = string("op_929_transpose_x_0"), val = bool(false)]; + bool var_929_transpose_y_0 = const()[name = string("op_929_transpose_y_0"), val = bool(false)]; + tensor var_929_cast_fp16 = matmul(transpose_x = var_929_transpose_x_0, transpose_y = var_929_transpose_y_0, x = var_928_cast_fp16, y = v_new_49_cast_fp16)[name = string("op_929_cast_fp16")]; + tensor var_930_cast_fp16 = add(x = var_926_cast_fp16, y = var_929_cast_fp16)[name = string("op_930_cast_fp16")]; + tensor var_932_begin_0 = const()[name = string("op_932_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_932_end_0 = const()[name = string("op_932_end_0"), val = tensor([32, 25, 1, 1])]; + tensor var_932_end_mask_0 = const()[name = string("op_932_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_932_squeeze_mask_0 = const()[name = string("op_932_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_932_cast_fp16 = slice_by_index(begin = var_932_begin_0, end = var_932_end_0, end_mask = var_932_end_mask_0, squeeze_mask = var_932_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_932_cast_fp16")]; + tensor var_933_cast_fp16 = mul(x = state_49_cast_fp16, y = var_932_cast_fp16)[name = string("op_933_cast_fp16")]; + tensor var_935_begin_0 = const()[name = string("op_935_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_935_end_0 = const()[name = string("op_935_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_935_end_mask_0 = const()[name = string("op_935_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_935_squeeze_mask_0 = const()[name = string("op_935_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_935_cast_fp16 = slice_by_index(begin = var_935_begin_0, end = var_935_end_0, end_mask = var_935_end_mask_0, squeeze_mask = var_935_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_935_cast_fp16")]; + bool var_937_transpose_x_1 = const()[name = string("op_937_transpose_x_1"), val = bool(true)]; + bool var_937_transpose_y_1 = const()[name = string("op_937_transpose_y_1"), val = bool(false)]; + tensor var_937_cast_fp16 = matmul(transpose_x = var_937_transpose_x_1, transpose_y = var_937_transpose_y_1, x = var_935_cast_fp16, y = v_new_49_cast_fp16)[name = string("op_937_cast_fp16")]; + tensor state_51_cast_fp16 = add(x = var_933_cast_fp16, y = var_937_cast_fp16)[name = string("state_51_cast_fp16")]; + tensor var_940_begin_0 = const()[name = string("op_940_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_940_end_0 = const()[name = string("op_940_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_940_end_mask_0 = const()[name = string("op_940_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_940_squeeze_mask_0 = const()[name = string("op_940_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_940_cast_fp16 = slice_by_index(begin = var_940_begin_0, end = var_940_end_0, end_mask = var_940_end_mask_0, squeeze_mask = var_940_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_940_cast_fp16")]; + tensor var_942_begin_0 = const()[name = string("op_942_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_942_end_0 = const()[name = string("op_942_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_942_end_mask_0 = const()[name = string("op_942_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_942_squeeze_mask_0 = const()[name = string("op_942_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_942_cast_fp16 = slice_by_index(begin = var_942_begin_0, end = var_942_end_0, end_mask = var_942_end_mask_0, squeeze_mask = var_942_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_942_cast_fp16")]; + bool var_943_transpose_x_0 = const()[name = string("op_943_transpose_x_0"), val = bool(false)]; + bool var_943_transpose_y_0 = const()[name = string("op_943_transpose_y_0"), val = bool(false)]; + tensor var_943_cast_fp16 = matmul(transpose_x = var_943_transpose_x_0, transpose_y = var_943_transpose_y_0, x = var_942_cast_fp16, y = state_51_cast_fp16)[name = string("op_943_cast_fp16")]; + tensor v_new_51_cast_fp16 = sub(x = var_940_cast_fp16, y = var_943_cast_fp16)[name = string("v_new_51_cast_fp16")]; + tensor var_946_begin_0 = const()[name = string("op_946_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_946_end_0 = const()[name = string("op_946_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_946_end_mask_0 = const()[name = string("op_946_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_946_squeeze_mask_0 = const()[name = string("op_946_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_946_cast_fp16 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, squeeze_mask = var_946_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_946_cast_fp16")]; + bool var_947_transpose_x_0 = const()[name = string("op_947_transpose_x_0"), val = bool(false)]; + bool var_947_transpose_y_0 = const()[name = string("op_947_transpose_y_0"), val = bool(false)]; + tensor var_947_cast_fp16 = matmul(transpose_x = var_947_transpose_x_0, transpose_y = var_947_transpose_y_0, x = var_946_cast_fp16, y = state_51_cast_fp16)[name = string("op_947_cast_fp16")]; + tensor var_949_begin_0 = const()[name = string("op_949_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_949_end_0 = const()[name = string("op_949_end_0"), val = tensor([32, 26, 64, 64])]; + tensor var_949_end_mask_0 = const()[name = string("op_949_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_949_squeeze_mask_0 = const()[name = string("op_949_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_949_cast_fp16 = slice_by_index(begin = var_949_begin_0, end = var_949_end_0, end_mask = var_949_end_mask_0, squeeze_mask = var_949_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_949_cast_fp16")]; + bool var_950_transpose_x_0 = const()[name = string("op_950_transpose_x_0"), val = bool(false)]; + bool var_950_transpose_y_0 = const()[name = string("op_950_transpose_y_0"), val = bool(false)]; + tensor var_950_cast_fp16 = matmul(transpose_x = var_950_transpose_x_0, transpose_y = var_950_transpose_y_0, x = var_949_cast_fp16, y = v_new_51_cast_fp16)[name = string("op_950_cast_fp16")]; + tensor var_951_cast_fp16 = add(x = var_947_cast_fp16, y = var_950_cast_fp16)[name = string("op_951_cast_fp16")]; + tensor var_953_begin_0 = const()[name = string("op_953_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_953_end_0 = const()[name = string("op_953_end_0"), val = tensor([32, 26, 1, 1])]; + tensor var_953_end_mask_0 = const()[name = string("op_953_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_953_squeeze_mask_0 = const()[name = string("op_953_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_953_cast_fp16 = slice_by_index(begin = var_953_begin_0, end = var_953_end_0, end_mask = var_953_end_mask_0, squeeze_mask = var_953_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_953_cast_fp16")]; + tensor var_954_cast_fp16 = mul(x = state_51_cast_fp16, y = var_953_cast_fp16)[name = string("op_954_cast_fp16")]; + tensor var_956_begin_0 = const()[name = string("op_956_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_956_end_0 = const()[name = string("op_956_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_956_end_mask_0 = const()[name = string("op_956_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_956_squeeze_mask_0 = const()[name = string("op_956_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_956_cast_fp16 = slice_by_index(begin = var_956_begin_0, end = var_956_end_0, end_mask = var_956_end_mask_0, squeeze_mask = var_956_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_956_cast_fp16")]; + bool var_958_transpose_x_1 = const()[name = string("op_958_transpose_x_1"), val = bool(true)]; + bool var_958_transpose_y_1 = const()[name = string("op_958_transpose_y_1"), val = bool(false)]; + tensor var_958_cast_fp16 = matmul(transpose_x = var_958_transpose_x_1, transpose_y = var_958_transpose_y_1, x = var_956_cast_fp16, y = v_new_51_cast_fp16)[name = string("op_958_cast_fp16")]; + tensor state_53_cast_fp16 = add(x = var_954_cast_fp16, y = var_958_cast_fp16)[name = string("state_53_cast_fp16")]; + tensor var_961_begin_0 = const()[name = string("op_961_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_961_end_0 = const()[name = string("op_961_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_961_end_mask_0 = const()[name = string("op_961_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_961_squeeze_mask_0 = const()[name = string("op_961_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_961_cast_fp16 = slice_by_index(begin = var_961_begin_0, end = var_961_end_0, end_mask = var_961_end_mask_0, squeeze_mask = var_961_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_961_cast_fp16")]; + tensor var_963_begin_0 = const()[name = string("op_963_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_963_end_0 = const()[name = string("op_963_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_963_end_mask_0 = const()[name = string("op_963_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_963_squeeze_mask_0 = const()[name = string("op_963_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_963_cast_fp16 = slice_by_index(begin = var_963_begin_0, end = var_963_end_0, end_mask = var_963_end_mask_0, squeeze_mask = var_963_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_963_cast_fp16")]; + bool var_964_transpose_x_0 = const()[name = string("op_964_transpose_x_0"), val = bool(false)]; + bool var_964_transpose_y_0 = const()[name = string("op_964_transpose_y_0"), val = bool(false)]; + tensor var_964_cast_fp16 = matmul(transpose_x = var_964_transpose_x_0, transpose_y = var_964_transpose_y_0, x = var_963_cast_fp16, y = state_53_cast_fp16)[name = string("op_964_cast_fp16")]; + tensor v_new_53_cast_fp16 = sub(x = var_961_cast_fp16, y = var_964_cast_fp16)[name = string("v_new_53_cast_fp16")]; + tensor var_967_begin_0 = const()[name = string("op_967_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_967_end_0 = const()[name = string("op_967_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_967_end_mask_0 = const()[name = string("op_967_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_967_squeeze_mask_0 = const()[name = string("op_967_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_967_cast_fp16 = slice_by_index(begin = var_967_begin_0, end = var_967_end_0, end_mask = var_967_end_mask_0, squeeze_mask = var_967_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_967_cast_fp16")]; + bool var_968_transpose_x_0 = const()[name = string("op_968_transpose_x_0"), val = bool(false)]; + bool var_968_transpose_y_0 = const()[name = string("op_968_transpose_y_0"), val = bool(false)]; + tensor var_968_cast_fp16 = matmul(transpose_x = var_968_transpose_x_0, transpose_y = var_968_transpose_y_0, x = var_967_cast_fp16, y = state_53_cast_fp16)[name = string("op_968_cast_fp16")]; + tensor var_970_begin_0 = const()[name = string("op_970_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_970_end_0 = const()[name = string("op_970_end_0"), val = tensor([32, 27, 64, 64])]; + tensor var_970_end_mask_0 = const()[name = string("op_970_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_970_squeeze_mask_0 = const()[name = string("op_970_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_970_cast_fp16 = slice_by_index(begin = var_970_begin_0, end = var_970_end_0, end_mask = var_970_end_mask_0, squeeze_mask = var_970_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_970_cast_fp16")]; + bool var_971_transpose_x_0 = const()[name = string("op_971_transpose_x_0"), val = bool(false)]; + bool var_971_transpose_y_0 = const()[name = string("op_971_transpose_y_0"), val = bool(false)]; + tensor var_971_cast_fp16 = matmul(transpose_x = var_971_transpose_x_0, transpose_y = var_971_transpose_y_0, x = var_970_cast_fp16, y = v_new_53_cast_fp16)[name = string("op_971_cast_fp16")]; + tensor var_972_cast_fp16 = add(x = var_968_cast_fp16, y = var_971_cast_fp16)[name = string("op_972_cast_fp16")]; + tensor var_974_begin_0 = const()[name = string("op_974_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_974_end_0 = const()[name = string("op_974_end_0"), val = tensor([32, 27, 1, 1])]; + tensor var_974_end_mask_0 = const()[name = string("op_974_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_974_squeeze_mask_0 = const()[name = string("op_974_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_974_cast_fp16 = slice_by_index(begin = var_974_begin_0, end = var_974_end_0, end_mask = var_974_end_mask_0, squeeze_mask = var_974_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_974_cast_fp16")]; + tensor var_975_cast_fp16 = mul(x = state_53_cast_fp16, y = var_974_cast_fp16)[name = string("op_975_cast_fp16")]; + tensor var_977_begin_0 = const()[name = string("op_977_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_977_end_0 = const()[name = string("op_977_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_977_end_mask_0 = const()[name = string("op_977_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_977_squeeze_mask_0 = const()[name = string("op_977_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_977_cast_fp16 = slice_by_index(begin = var_977_begin_0, end = var_977_end_0, end_mask = var_977_end_mask_0, squeeze_mask = var_977_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_977_cast_fp16")]; + bool var_979_transpose_x_1 = const()[name = string("op_979_transpose_x_1"), val = bool(true)]; + bool var_979_transpose_y_1 = const()[name = string("op_979_transpose_y_1"), val = bool(false)]; + tensor var_979_cast_fp16 = matmul(transpose_x = var_979_transpose_x_1, transpose_y = var_979_transpose_y_1, x = var_977_cast_fp16, y = v_new_53_cast_fp16)[name = string("op_979_cast_fp16")]; + tensor state_55_cast_fp16 = add(x = var_975_cast_fp16, y = var_979_cast_fp16)[name = string("state_55_cast_fp16")]; + tensor var_982_begin_0 = const()[name = string("op_982_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_982_end_0 = const()[name = string("op_982_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_982_end_mask_0 = const()[name = string("op_982_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_982_squeeze_mask_0 = const()[name = string("op_982_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_982_cast_fp16 = slice_by_index(begin = var_982_begin_0, end = var_982_end_0, end_mask = var_982_end_mask_0, squeeze_mask = var_982_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_982_cast_fp16")]; + tensor var_984_begin_0 = const()[name = string("op_984_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_984_end_0 = const()[name = string("op_984_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_984_end_mask_0 = const()[name = string("op_984_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_984_squeeze_mask_0 = const()[name = string("op_984_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_984_cast_fp16 = slice_by_index(begin = var_984_begin_0, end = var_984_end_0, end_mask = var_984_end_mask_0, squeeze_mask = var_984_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_984_cast_fp16")]; + bool var_985_transpose_x_0 = const()[name = string("op_985_transpose_x_0"), val = bool(false)]; + bool var_985_transpose_y_0 = const()[name = string("op_985_transpose_y_0"), val = bool(false)]; + tensor var_985_cast_fp16 = matmul(transpose_x = var_985_transpose_x_0, transpose_y = var_985_transpose_y_0, x = var_984_cast_fp16, y = state_55_cast_fp16)[name = string("op_985_cast_fp16")]; + tensor v_new_55_cast_fp16 = sub(x = var_982_cast_fp16, y = var_985_cast_fp16)[name = string("v_new_55_cast_fp16")]; + tensor var_988_begin_0 = const()[name = string("op_988_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_988_end_0 = const()[name = string("op_988_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_988_end_mask_0 = const()[name = string("op_988_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_988_squeeze_mask_0 = const()[name = string("op_988_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_988_cast_fp16 = slice_by_index(begin = var_988_begin_0, end = var_988_end_0, end_mask = var_988_end_mask_0, squeeze_mask = var_988_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_988_cast_fp16")]; + bool var_989_transpose_x_0 = const()[name = string("op_989_transpose_x_0"), val = bool(false)]; + bool var_989_transpose_y_0 = const()[name = string("op_989_transpose_y_0"), val = bool(false)]; + tensor var_989_cast_fp16 = matmul(transpose_x = var_989_transpose_x_0, transpose_y = var_989_transpose_y_0, x = var_988_cast_fp16, y = state_55_cast_fp16)[name = string("op_989_cast_fp16")]; + tensor var_991_begin_0 = const()[name = string("op_991_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_991_end_0 = const()[name = string("op_991_end_0"), val = tensor([32, 28, 64, 64])]; + tensor var_991_end_mask_0 = const()[name = string("op_991_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_991_squeeze_mask_0 = const()[name = string("op_991_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_991_cast_fp16 = slice_by_index(begin = var_991_begin_0, end = var_991_end_0, end_mask = var_991_end_mask_0, squeeze_mask = var_991_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_991_cast_fp16")]; + bool var_992_transpose_x_0 = const()[name = string("op_992_transpose_x_0"), val = bool(false)]; + bool var_992_transpose_y_0 = const()[name = string("op_992_transpose_y_0"), val = bool(false)]; + tensor var_992_cast_fp16 = matmul(transpose_x = var_992_transpose_x_0, transpose_y = var_992_transpose_y_0, x = var_991_cast_fp16, y = v_new_55_cast_fp16)[name = string("op_992_cast_fp16")]; + tensor var_993_cast_fp16 = add(x = var_989_cast_fp16, y = var_992_cast_fp16)[name = string("op_993_cast_fp16")]; + tensor var_995_begin_0 = const()[name = string("op_995_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_995_end_0 = const()[name = string("op_995_end_0"), val = tensor([32, 28, 1, 1])]; + tensor var_995_end_mask_0 = const()[name = string("op_995_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_995_squeeze_mask_0 = const()[name = string("op_995_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_995_cast_fp16 = slice_by_index(begin = var_995_begin_0, end = var_995_end_0, end_mask = var_995_end_mask_0, squeeze_mask = var_995_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_995_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = state_55_cast_fp16, y = var_995_cast_fp16)[name = string("op_996_cast_fp16")]; + tensor var_998_begin_0 = const()[name = string("op_998_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_998_end_0 = const()[name = string("op_998_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_998_end_mask_0 = const()[name = string("op_998_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_998_squeeze_mask_0 = const()[name = string("op_998_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_998_cast_fp16 = slice_by_index(begin = var_998_begin_0, end = var_998_end_0, end_mask = var_998_end_mask_0, squeeze_mask = var_998_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_998_cast_fp16")]; + bool var_1000_transpose_x_1 = const()[name = string("op_1000_transpose_x_1"), val = bool(true)]; + bool var_1000_transpose_y_1 = const()[name = string("op_1000_transpose_y_1"), val = bool(false)]; + tensor var_1000_cast_fp16 = matmul(transpose_x = var_1000_transpose_x_1, transpose_y = var_1000_transpose_y_1, x = var_998_cast_fp16, y = v_new_55_cast_fp16)[name = string("op_1000_cast_fp16")]; + tensor state_57_cast_fp16 = add(x = var_996_cast_fp16, y = var_1000_cast_fp16)[name = string("state_57_cast_fp16")]; + tensor var_1003_begin_0 = const()[name = string("op_1003_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_1003_end_0 = const()[name = string("op_1003_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_1003_end_mask_0 = const()[name = string("op_1003_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1003_squeeze_mask_0 = const()[name = string("op_1003_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, squeeze_mask = var_1003_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_1003_cast_fp16")]; + tensor var_1005_begin_0 = const()[name = string("op_1005_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_1005_end_0 = const()[name = string("op_1005_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_1005_end_mask_0 = const()[name = string("op_1005_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1005_squeeze_mask_0 = const()[name = string("op_1005_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1005_cast_fp16 = slice_by_index(begin = var_1005_begin_0, end = var_1005_end_0, end_mask = var_1005_end_mask_0, squeeze_mask = var_1005_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_1005_cast_fp16")]; + bool var_1006_transpose_x_0 = const()[name = string("op_1006_transpose_x_0"), val = bool(false)]; + bool var_1006_transpose_y_0 = const()[name = string("op_1006_transpose_y_0"), val = bool(false)]; + tensor var_1006_cast_fp16 = matmul(transpose_x = var_1006_transpose_x_0, transpose_y = var_1006_transpose_y_0, x = var_1005_cast_fp16, y = state_57_cast_fp16)[name = string("op_1006_cast_fp16")]; + tensor v_new_57_cast_fp16 = sub(x = var_1003_cast_fp16, y = var_1006_cast_fp16)[name = string("v_new_57_cast_fp16")]; + tensor var_1009_begin_0 = const()[name = string("op_1009_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_1009_end_0 = const()[name = string("op_1009_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_1009_end_mask_0 = const()[name = string("op_1009_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1009_squeeze_mask_0 = const()[name = string("op_1009_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1009_cast_fp16 = slice_by_index(begin = var_1009_begin_0, end = var_1009_end_0, end_mask = var_1009_end_mask_0, squeeze_mask = var_1009_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_1009_cast_fp16")]; + bool var_1010_transpose_x_0 = const()[name = string("op_1010_transpose_x_0"), val = bool(false)]; + bool var_1010_transpose_y_0 = const()[name = string("op_1010_transpose_y_0"), val = bool(false)]; + tensor var_1010_cast_fp16 = matmul(transpose_x = var_1010_transpose_x_0, transpose_y = var_1010_transpose_y_0, x = var_1009_cast_fp16, y = state_57_cast_fp16)[name = string("op_1010_cast_fp16")]; + tensor var_1012_begin_0 = const()[name = string("op_1012_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_1012_end_0 = const()[name = string("op_1012_end_0"), val = tensor([32, 29, 64, 64])]; + tensor var_1012_end_mask_0 = const()[name = string("op_1012_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1012_squeeze_mask_0 = const()[name = string("op_1012_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1012_cast_fp16 = slice_by_index(begin = var_1012_begin_0, end = var_1012_end_0, end_mask = var_1012_end_mask_0, squeeze_mask = var_1012_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_1012_cast_fp16")]; + bool var_1013_transpose_x_0 = const()[name = string("op_1013_transpose_x_0"), val = bool(false)]; + bool var_1013_transpose_y_0 = const()[name = string("op_1013_transpose_y_0"), val = bool(false)]; + tensor var_1013_cast_fp16 = matmul(transpose_x = var_1013_transpose_x_0, transpose_y = var_1013_transpose_y_0, x = var_1012_cast_fp16, y = v_new_57_cast_fp16)[name = string("op_1013_cast_fp16")]; + tensor var_1014_cast_fp16 = add(x = var_1010_cast_fp16, y = var_1013_cast_fp16)[name = string("op_1014_cast_fp16")]; + tensor var_1016_begin_0 = const()[name = string("op_1016_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_1016_end_0 = const()[name = string("op_1016_end_0"), val = tensor([32, 29, 1, 1])]; + tensor var_1016_end_mask_0 = const()[name = string("op_1016_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1016_squeeze_mask_0 = const()[name = string("op_1016_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1016_cast_fp16 = slice_by_index(begin = var_1016_begin_0, end = var_1016_end_0, end_mask = var_1016_end_mask_0, squeeze_mask = var_1016_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_1016_cast_fp16")]; + tensor var_1017_cast_fp16 = mul(x = state_57_cast_fp16, y = var_1016_cast_fp16)[name = string("op_1017_cast_fp16")]; + tensor var_1019_begin_0 = const()[name = string("op_1019_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_1019_end_0 = const()[name = string("op_1019_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_1019_end_mask_0 = const()[name = string("op_1019_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1019_squeeze_mask_0 = const()[name = string("op_1019_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1019_cast_fp16 = slice_by_index(begin = var_1019_begin_0, end = var_1019_end_0, end_mask = var_1019_end_mask_0, squeeze_mask = var_1019_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_1019_cast_fp16")]; + bool var_1021_transpose_x_1 = const()[name = string("op_1021_transpose_x_1"), val = bool(true)]; + bool var_1021_transpose_y_1 = const()[name = string("op_1021_transpose_y_1"), val = bool(false)]; + tensor var_1021_cast_fp16 = matmul(transpose_x = var_1021_transpose_x_1, transpose_y = var_1021_transpose_y_1, x = var_1019_cast_fp16, y = v_new_57_cast_fp16)[name = string("op_1021_cast_fp16")]; + tensor state_59_cast_fp16 = add(x = var_1017_cast_fp16, y = var_1021_cast_fp16)[name = string("state_59_cast_fp16")]; + tensor var_1024_begin_0 = const()[name = string("op_1024_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_1024_end_0 = const()[name = string("op_1024_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_1024_end_mask_0 = const()[name = string("op_1024_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1024_squeeze_mask_0 = const()[name = string("op_1024_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1024_cast_fp16 = slice_by_index(begin = var_1024_begin_0, end = var_1024_end_0, end_mask = var_1024_end_mask_0, squeeze_mask = var_1024_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_1024_cast_fp16")]; + tensor var_1026_begin_0 = const()[name = string("op_1026_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_1026_end_0 = const()[name = string("op_1026_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_1026_end_mask_0 = const()[name = string("op_1026_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1026_squeeze_mask_0 = const()[name = string("op_1026_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1026_cast_fp16 = slice_by_index(begin = var_1026_begin_0, end = var_1026_end_0, end_mask = var_1026_end_mask_0, squeeze_mask = var_1026_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_1026_cast_fp16")]; + bool var_1027_transpose_x_0 = const()[name = string("op_1027_transpose_x_0"), val = bool(false)]; + bool var_1027_transpose_y_0 = const()[name = string("op_1027_transpose_y_0"), val = bool(false)]; + tensor var_1027_cast_fp16 = matmul(transpose_x = var_1027_transpose_x_0, transpose_y = var_1027_transpose_y_0, x = var_1026_cast_fp16, y = state_59_cast_fp16)[name = string("op_1027_cast_fp16")]; + tensor v_new_59_cast_fp16 = sub(x = var_1024_cast_fp16, y = var_1027_cast_fp16)[name = string("v_new_59_cast_fp16")]; + tensor var_1030_begin_0 = const()[name = string("op_1030_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_1030_end_0 = const()[name = string("op_1030_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_1030_end_mask_0 = const()[name = string("op_1030_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1030_squeeze_mask_0 = const()[name = string("op_1030_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1030_cast_fp16 = slice_by_index(begin = var_1030_begin_0, end = var_1030_end_0, end_mask = var_1030_end_mask_0, squeeze_mask = var_1030_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_1030_cast_fp16")]; + bool var_1031_transpose_x_0 = const()[name = string("op_1031_transpose_x_0"), val = bool(false)]; + bool var_1031_transpose_y_0 = const()[name = string("op_1031_transpose_y_0"), val = bool(false)]; + tensor var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_0, transpose_y = var_1031_transpose_y_0, x = var_1030_cast_fp16, y = state_59_cast_fp16)[name = string("op_1031_cast_fp16")]; + tensor var_1033_begin_0 = const()[name = string("op_1033_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_1033_end_0 = const()[name = string("op_1033_end_0"), val = tensor([32, 30, 64, 64])]; + tensor var_1033_end_mask_0 = const()[name = string("op_1033_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1033_squeeze_mask_0 = const()[name = string("op_1033_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1033_cast_fp16 = slice_by_index(begin = var_1033_begin_0, end = var_1033_end_0, end_mask = var_1033_end_mask_0, squeeze_mask = var_1033_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_1033_cast_fp16")]; + bool var_1034_transpose_x_0 = const()[name = string("op_1034_transpose_x_0"), val = bool(false)]; + bool var_1034_transpose_y_0 = const()[name = string("op_1034_transpose_y_0"), val = bool(false)]; + tensor var_1034_cast_fp16 = matmul(transpose_x = var_1034_transpose_x_0, transpose_y = var_1034_transpose_y_0, x = var_1033_cast_fp16, y = v_new_59_cast_fp16)[name = string("op_1034_cast_fp16")]; + tensor var_1035_cast_fp16 = add(x = var_1031_cast_fp16, y = var_1034_cast_fp16)[name = string("op_1035_cast_fp16")]; + tensor var_1037_begin_0 = const()[name = string("op_1037_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_1037_end_0 = const()[name = string("op_1037_end_0"), val = tensor([32, 30, 1, 1])]; + tensor var_1037_end_mask_0 = const()[name = string("op_1037_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1037_squeeze_mask_0 = const()[name = string("op_1037_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1037_cast_fp16 = slice_by_index(begin = var_1037_begin_0, end = var_1037_end_0, end_mask = var_1037_end_mask_0, squeeze_mask = var_1037_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_1037_cast_fp16")]; + tensor var_1038_cast_fp16 = mul(x = state_59_cast_fp16, y = var_1037_cast_fp16)[name = string("op_1038_cast_fp16")]; + tensor var_1040_begin_0 = const()[name = string("op_1040_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_1040_end_0 = const()[name = string("op_1040_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_1040_end_mask_0 = const()[name = string("op_1040_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1040_squeeze_mask_0 = const()[name = string("op_1040_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1040_cast_fp16 = slice_by_index(begin = var_1040_begin_0, end = var_1040_end_0, end_mask = var_1040_end_mask_0, squeeze_mask = var_1040_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_1040_cast_fp16")]; + bool var_1042_transpose_x_1 = const()[name = string("op_1042_transpose_x_1"), val = bool(true)]; + bool var_1042_transpose_y_1 = const()[name = string("op_1042_transpose_y_1"), val = bool(false)]; + tensor var_1042_cast_fp16 = matmul(transpose_x = var_1042_transpose_x_1, transpose_y = var_1042_transpose_y_1, x = var_1040_cast_fp16, y = v_new_59_cast_fp16)[name = string("op_1042_cast_fp16")]; + tensor state_61_cast_fp16 = add(x = var_1038_cast_fp16, y = var_1042_cast_fp16)[name = string("state_61_cast_fp16")]; + tensor var_1045_begin_0 = const()[name = string("op_1045_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_1045_end_0 = const()[name = string("op_1045_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_1045_end_mask_0 = const()[name = string("op_1045_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1045_squeeze_mask_0 = const()[name = string("op_1045_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1045_cast_fp16 = slice_by_index(begin = var_1045_begin_0, end = var_1045_end_0, end_mask = var_1045_end_mask_0, squeeze_mask = var_1045_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_1045_cast_fp16")]; + tensor var_1047_begin_0 = const()[name = string("op_1047_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_1047_end_0 = const()[name = string("op_1047_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_1047_end_mask_0 = const()[name = string("op_1047_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1047_squeeze_mask_0 = const()[name = string("op_1047_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1047_cast_fp16 = slice_by_index(begin = var_1047_begin_0, end = var_1047_end_0, end_mask = var_1047_end_mask_0, squeeze_mask = var_1047_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_1047_cast_fp16")]; + bool var_1048_transpose_x_0 = const()[name = string("op_1048_transpose_x_0"), val = bool(false)]; + bool var_1048_transpose_y_0 = const()[name = string("op_1048_transpose_y_0"), val = bool(false)]; + tensor var_1048_cast_fp16 = matmul(transpose_x = var_1048_transpose_x_0, transpose_y = var_1048_transpose_y_0, x = var_1047_cast_fp16, y = state_61_cast_fp16)[name = string("op_1048_cast_fp16")]; + tensor v_new_61_cast_fp16 = sub(x = var_1045_cast_fp16, y = var_1048_cast_fp16)[name = string("v_new_61_cast_fp16")]; + tensor var_1051_begin_0 = const()[name = string("op_1051_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_1051_end_0 = const()[name = string("op_1051_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_1051_end_mask_0 = const()[name = string("op_1051_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1051_squeeze_mask_0 = const()[name = string("op_1051_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1051_cast_fp16 = slice_by_index(begin = var_1051_begin_0, end = var_1051_end_0, end_mask = var_1051_end_mask_0, squeeze_mask = var_1051_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_1051_cast_fp16")]; + bool var_1052_transpose_x_0 = const()[name = string("op_1052_transpose_x_0"), val = bool(false)]; + bool var_1052_transpose_y_0 = const()[name = string("op_1052_transpose_y_0"), val = bool(false)]; + tensor var_1052_cast_fp16 = matmul(transpose_x = var_1052_transpose_x_0, transpose_y = var_1052_transpose_y_0, x = var_1051_cast_fp16, y = state_61_cast_fp16)[name = string("op_1052_cast_fp16")]; + tensor var_1054_begin_0 = const()[name = string("op_1054_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_1054_end_0 = const()[name = string("op_1054_end_0"), val = tensor([32, 31, 64, 64])]; + tensor var_1054_end_mask_0 = const()[name = string("op_1054_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1054_squeeze_mask_0 = const()[name = string("op_1054_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1054_cast_fp16 = slice_by_index(begin = var_1054_begin_0, end = var_1054_end_0, end_mask = var_1054_end_mask_0, squeeze_mask = var_1054_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_1054_cast_fp16")]; + bool var_1055_transpose_x_0 = const()[name = string("op_1055_transpose_x_0"), val = bool(false)]; + bool var_1055_transpose_y_0 = const()[name = string("op_1055_transpose_y_0"), val = bool(false)]; + tensor var_1055_cast_fp16 = matmul(transpose_x = var_1055_transpose_x_0, transpose_y = var_1055_transpose_y_0, x = var_1054_cast_fp16, y = v_new_61_cast_fp16)[name = string("op_1055_cast_fp16")]; + tensor var_1056_cast_fp16 = add(x = var_1052_cast_fp16, y = var_1055_cast_fp16)[name = string("op_1056_cast_fp16")]; + tensor var_1058_begin_0 = const()[name = string("op_1058_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_1058_end_0 = const()[name = string("op_1058_end_0"), val = tensor([32, 31, 1, 1])]; + tensor var_1058_end_mask_0 = const()[name = string("op_1058_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1058_squeeze_mask_0 = const()[name = string("op_1058_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1058_cast_fp16 = slice_by_index(begin = var_1058_begin_0, end = var_1058_end_0, end_mask = var_1058_end_mask_0, squeeze_mask = var_1058_squeeze_mask_0, x = chunk_decay_1_cast_fp16)[name = string("op_1058_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = state_61_cast_fp16, y = var_1058_cast_fp16)[name = string("op_1059_cast_fp16")]; + tensor var_1061_begin_0 = const()[name = string("op_1061_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_1061_end_0 = const()[name = string("op_1061_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_1061_end_mask_0 = const()[name = string("op_1061_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1061_squeeze_mask_0 = const()[name = string("op_1061_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, squeeze_mask = var_1061_squeeze_mask_0, x = k_dec_1_cast_fp16)[name = string("op_1061_cast_fp16")]; + bool var_1063_transpose_x_1 = const()[name = string("op_1063_transpose_x_1"), val = bool(true)]; + bool var_1063_transpose_y_1 = const()[name = string("op_1063_transpose_y_1"), val = bool(false)]; + tensor var_1063_cast_fp16 = matmul(transpose_x = var_1063_transpose_x_1, transpose_y = var_1063_transpose_y_1, x = var_1061_cast_fp16, y = v_new_61_cast_fp16)[name = string("op_1063_cast_fp16")]; + tensor state_63_cast_fp16 = add(x = var_1059_cast_fp16, y = var_1063_cast_fp16)[name = string("state_63_cast_fp16")]; + tensor var_1066_begin_0 = const()[name = string("op_1066_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_1066_end_0 = const()[name = string("op_1066_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_1066_end_mask_0 = const()[name = string("op_1066_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1066_squeeze_mask_0 = const()[name = string("op_1066_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1066_cast_fp16 = slice_by_index(begin = var_1066_begin_0, end = var_1066_end_0, end_mask = var_1066_end_mask_0, squeeze_mask = var_1066_squeeze_mask_0, x = new_v_1_cast_fp16)[name = string("op_1066_cast_fp16")]; + tensor var_1068_begin_0 = const()[name = string("op_1068_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_1068_end_0 = const()[name = string("op_1068_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_1068_end_mask_0 = const()[name = string("op_1068_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1068_squeeze_mask_0 = const()[name = string("op_1068_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1068_cast_fp16 = slice_by_index(begin = var_1068_begin_0, end = var_1068_end_0, end_mask = var_1068_end_mask_0, squeeze_mask = var_1068_squeeze_mask_0, x = k_cumdecay_1_cast_fp16)[name = string("op_1068_cast_fp16")]; + bool var_1069_transpose_x_0 = const()[name = string("op_1069_transpose_x_0"), val = bool(false)]; + bool var_1069_transpose_y_0 = const()[name = string("op_1069_transpose_y_0"), val = bool(false)]; + tensor var_1069_cast_fp16 = matmul(transpose_x = var_1069_transpose_x_0, transpose_y = var_1069_transpose_y_0, x = var_1068_cast_fp16, y = state_63_cast_fp16)[name = string("op_1069_cast_fp16")]; + tensor v_new_63_cast_fp16 = sub(x = var_1066_cast_fp16, y = var_1069_cast_fp16)[name = string("v_new_63_cast_fp16")]; + tensor var_1072_begin_0 = const()[name = string("op_1072_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_1072_end_0 = const()[name = string("op_1072_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_1072_end_mask_0 = const()[name = string("op_1072_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1072_squeeze_mask_0 = const()[name = string("op_1072_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1072_cast_fp16 = slice_by_index(begin = var_1072_begin_0, end = var_1072_end_0, end_mask = var_1072_end_mask_0, squeeze_mask = var_1072_squeeze_mask_0, x = q_dec_1_cast_fp16)[name = string("op_1072_cast_fp16")]; + bool var_1073_transpose_x_0 = const()[name = string("op_1073_transpose_x_0"), val = bool(false)]; + bool var_1073_transpose_y_0 = const()[name = string("op_1073_transpose_y_0"), val = bool(false)]; + tensor var_1073_cast_fp16 = matmul(transpose_x = var_1073_transpose_x_0, transpose_y = var_1073_transpose_y_0, x = var_1072_cast_fp16, y = state_63_cast_fp16)[name = string("op_1073_cast_fp16")]; + tensor var_1075_begin_0 = const()[name = string("op_1075_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_1075_end_0 = const()[name = string("op_1075_end_0"), val = tensor([32, 32, 64, 64])]; + tensor var_1075_end_mask_0 = const()[name = string("op_1075_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1075_squeeze_mask_0 = const()[name = string("op_1075_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1075_cast_fp16 = slice_by_index(begin = var_1075_begin_0, end = var_1075_end_0, end_mask = var_1075_end_mask_0, squeeze_mask = var_1075_squeeze_mask_0, x = attn_intra_1_cast_fp16)[name = string("op_1075_cast_fp16")]; + bool var_1076_transpose_x_0 = const()[name = string("op_1076_transpose_x_0"), val = bool(false)]; + bool var_1076_transpose_y_0 = const()[name = string("op_1076_transpose_y_0"), val = bool(false)]; + tensor var_1076_cast_fp16 = matmul(transpose_x = var_1076_transpose_x_0, transpose_y = var_1076_transpose_y_0, x = var_1075_cast_fp16, y = v_new_63_cast_fp16)[name = string("op_1076_cast_fp16")]; + tensor var_1077_cast_fp16 = add(x = var_1073_cast_fp16, y = var_1076_cast_fp16)[name = string("op_1077_cast_fp16")]; + int32 var_1079_axis_0 = const()[name = string("op_1079_axis_0"), val = int32(1)]; + tensor var_1079_cast_fp16 = stack(axis = var_1079_axis_0, values = (var_426_cast_fp16, var_447_cast_fp16, var_468_cast_fp16, var_489_cast_fp16, var_510_cast_fp16, var_531_cast_fp16, var_552_cast_fp16, var_573_cast_fp16, var_594_cast_fp16, var_615_cast_fp16, var_636_cast_fp16, var_657_cast_fp16, var_678_cast_fp16, var_699_cast_fp16, var_720_cast_fp16, var_741_cast_fp16, var_762_cast_fp16, var_783_cast_fp16, var_804_cast_fp16, var_825_cast_fp16, var_846_cast_fp16, var_867_cast_fp16, var_888_cast_fp16, var_909_cast_fp16, var_930_cast_fp16, var_951_cast_fp16, var_972_cast_fp16, var_993_cast_fp16, var_1014_cast_fp16, var_1035_cast_fp16, var_1056_cast_fp16, var_1077_cast_fp16))[name = string("op_1079_cast_fp16")]; + tensor var_1080 = const()[name = string("op_1080"), val = tensor([32, 2048, 128])]; + tensor var_1081_cast_fp16 = reshape(shape = var_1080, x = var_1079_cast_fp16)[name = string("op_1081_cast_fp16")]; + tensor var_1082 = const()[name = string("op_1082"), val = tensor([1, 0, 2])]; + tensor input_7_begin_0 = const()[name = string("input_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor input_7_end_0 = const()[name = string("input_7_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor input_7_end_mask_0 = const()[name = string("input_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor input_7_squeeze_mask_0 = const()[name = string("input_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor input_7_cast_fp16 = slice_by_index(begin = input_7_begin_0, end = input_7_end_0, end_mask = input_7_end_mask_0, squeeze_mask = input_7_squeeze_mask_0, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")]; + fp16 var_62_promoted_1_to_fp16 = const()[name = string("op_62_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor x_17_cast_fp16 = transpose(perm = var_1082, x = var_1081_cast_fp16)[name = string("transpose_48")]; + tensor var_1086_cast_fp16 = pow(x = x_17_cast_fp16, y = var_62_promoted_1_to_fp16)[name = string("op_1086_cast_fp16")]; + tensor var_1088_axes_0 = const()[name = string("op_1088_axes_0"), val = tensor([-1])]; + bool var_1088_keep_dims_0 = const()[name = string("op_1088_keep_dims_0"), val = bool(true)]; + tensor var_1088_cast_fp16 = reduce_mean(axes = var_1088_axes_0, keep_dims = var_1088_keep_dims_0, x = var_1086_cast_fp16)[name = string("op_1088_cast_fp16")]; + fp16 var_1089_to_fp16 = const()[name = string("op_1089_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_1090_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_to_fp16)[name = string("op_1090_cast_fp16")]; + fp32 var_1091_epsilon_0 = const()[name = string("op_1091_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1091_cast_fp16 = rsqrt(epsilon = var_1091_epsilon_0, x = var_1090_cast_fp16)[name = string("op_1091_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_17_cast_fp16, y = var_1091_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor layers_0_linear_attn_norm_weight_to_fp16 = const()[name = string("layers_0_linear_attn_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37693376)))]; + tensor var_1093_cast_fp16 = mul(x = layers_0_linear_attn_norm_weight_to_fp16, y = x_19_cast_fp16)[name = string("op_1093_cast_fp16")]; + tensor var_1094_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1094_cast_fp16")]; + tensor core_1_cast_fp16 = mul(x = var_1093_cast_fp16, y = var_1094_cast_fp16)[name = string("core_1_cast_fp16")]; + tensor var_1096 = const()[name = string("op_1096"), val = tensor([1, 2048, 4096])]; + tensor input_9_cast_fp16 = reshape(shape = var_1096, x = core_1_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor layers_0_linear_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37693696))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48179520))))[name = string("layers_0_linear_attn_out_proj_weight_to_fp16_quantized")]; + tensor linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48184704)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_linear_attn_out_proj_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor x_21_cast_fp16 = add(x = hidden, y = linear_4_cast_fp16)[name = string("x_21_cast_fp16")]; + fp16 var_62_promoted_2_to_fp16 = const()[name = string("op_62_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor var_1102_cast_fp16 = pow(x = x_21_cast_fp16, y = var_62_promoted_2_to_fp16)[name = string("op_1102_cast_fp16")]; + tensor var_1104_axes_0 = const()[name = string("op_1104_axes_0"), val = tensor([-1])]; + bool var_1104_keep_dims_0 = const()[name = string("op_1104_keep_dims_0"), val = bool(true)]; + tensor var_1104_cast_fp16 = reduce_mean(axes = var_1104_axes_0, keep_dims = var_1104_keep_dims_0, x = var_1102_cast_fp16)[name = string("op_1104_cast_fp16")]; + fp16 var_1105_to_fp16 = const()[name = string("op_1105_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_1106_cast_fp16 = add(x = var_1104_cast_fp16, y = var_1105_to_fp16)[name = string("op_1106_cast_fp16")]; + fp32 var_1107_epsilon_0 = const()[name = string("op_1107_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1107_cast_fp16 = rsqrt(epsilon = var_1107_epsilon_0, x = var_1106_cast_fp16)[name = string("op_1107_cast_fp16")]; + tensor x_23_cast_fp16 = mul(x = x_21_cast_fp16, y = var_1107_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor var_1110_to_fp16 = const()[name = string("op_1110_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48189888)))]; + tensor input_11_cast_fp16 = mul(x = x_23_cast_fp16, y = var_1110_to_fp16)[name = string("input_11_cast_fp16")]; + tensor layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48195072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71788096))))[name = string("layers_0_mlp_gate_proj_weight_to_fp16_quantized")]; + tensor linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71806592)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_1117_cast_fp16 = silu(x = linear_5_cast_fp16)[name = string("op_1117_cast_fp16")]; + tensor layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71825088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95418112))))[name = string("layers_0_mlp_up_proj_weight_to_fp16_quantized")]; + tensor linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_6_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = var_1117_cast_fp16, y = linear_6_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95436608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119029632))))[name = string("layers_0_mlp_down_proj_weight_to_fp16_quantized")]; + tensor linear_7_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_7_cast_fp16)[name = string("x_25_cast_fp16")]; + int32 var_1137 = const()[name = string("op_1137"), val = int32(-2)]; + int32 var_1166 = const()[name = string("op_1166"), val = int32(-1)]; + fp16 var_1165_promoted_to_fp16 = const()[name = string("op_1165_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_1175_cast_fp16 = pow(x = x_25_cast_fp16, y = var_1165_promoted_to_fp16)[name = string("op_1175_cast_fp16")]; + tensor var_1177_axes_0 = const()[name = string("op_1177_axes_0"), val = tensor([-1])]; + bool var_1177_keep_dims_0 = const()[name = string("op_1177_keep_dims_0"), val = bool(true)]; + tensor var_1177_cast_fp16 = reduce_mean(axes = var_1177_axes_0, keep_dims = var_1177_keep_dims_0, x = var_1175_cast_fp16)[name = string("op_1177_cast_fp16")]; + fp16 var_1178_to_fp16 = const()[name = string("op_1178_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_1179_cast_fp16 = add(x = var_1177_cast_fp16, y = var_1178_to_fp16)[name = string("op_1179_cast_fp16")]; + fp32 var_1180_epsilon_0 = const()[name = string("op_1180_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1180_cast_fp16 = rsqrt(epsilon = var_1180_epsilon_0, x = var_1179_cast_fp16)[name = string("op_1180_cast_fp16")]; + tensor x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_1180_cast_fp16)[name = string("x_27_cast_fp16")]; + tensor var_1183_to_fp16 = const()[name = string("op_1183_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119034816)))]; + tensor x_29_cast_fp16 = mul(x = x_27_cast_fp16, y = var_1183_to_fp16)[name = string("x_29_cast_fp16")]; + tensor layers_1_linear_attn_in_proj_qkv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119040000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140011584))))[name = string("layers_1_linear_attn_in_proj_qkv_weight_to_fp16_quantized")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_linear_attn_in_proj_qkv_weight_to_fp16_quantized, x = x_29_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_17_perm_0 = const()[name = string("input_17_perm_0"), val = tensor([0, 2, 1])]; + tensor input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor([0, 0, 0, 0, 3, 0])]; + string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("constant")]; + fp16 const_13_to_fp16 = const()[name = string("const_13_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = transpose(perm = input_17_perm_0, x = linear_8_cast_fp16)[name = string("transpose_47")]; + tensor input_19_cast_fp16 = pad(constant_val = const_13_to_fp16, mode = input_19_mode_0, pad = input_19_pad_0, x = input_17_cast_fp16)[name = string("input_19_cast_fp16")]; + string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")]; + int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(8192)]; + tensor input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor([1])]; + tensor input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor([0, 0])]; + tensor input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor([1])]; + tensor layers_1_linear_attn_conv1d_weight_to_fp16 = const()[name = string("layers_1_linear_attn_conv1d_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140028032)))]; + tensor input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_1_linear_attn_conv1d_weight_to_fp16, x = input_19_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor var_1218_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_1218_cast_fp16")]; + tensor qkv_3_perm_0 = const()[name = string("qkv_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1220_begin_0 = const()[name = string("op_1220_begin_0"), val = tensor([0, 0, 0])]; + tensor var_1220_end_0 = const()[name = string("op_1220_end_0"), val = tensor([1, 2048, 2048])]; + tensor var_1220_end_mask_0 = const()[name = string("op_1220_end_mask_0"), val = tensor([true, true, false])]; + tensor qkv_3_cast_fp16 = transpose(perm = qkv_3_perm_0, x = var_1218_cast_fp16)[name = string("transpose_46")]; + tensor var_1220_cast_fp16 = slice_by_index(begin = var_1220_begin_0, end = var_1220_end_0, end_mask = var_1220_end_mask_0, x = qkv_3_cast_fp16)[name = string("op_1220_cast_fp16")]; + tensor var_1221 = const()[name = string("op_1221"), val = tensor([1, 2048, 16, 128])]; + tensor q_11_cast_fp16 = reshape(shape = var_1221, x = var_1220_cast_fp16)[name = string("q_11_cast_fp16")]; + tensor var_1223_begin_0 = const()[name = string("op_1223_begin_0"), val = tensor([0, 0, 2048])]; + tensor var_1223_end_0 = const()[name = string("op_1223_end_0"), val = tensor([1, 2048, 4096])]; + tensor var_1223_end_mask_0 = const()[name = string("op_1223_end_mask_0"), val = tensor([true, true, false])]; + tensor var_1223_cast_fp16 = slice_by_index(begin = var_1223_begin_0, end = var_1223_end_0, end_mask = var_1223_end_mask_0, x = qkv_3_cast_fp16)[name = string("op_1223_cast_fp16")]; + tensor var_1224 = const()[name = string("op_1224"), val = tensor([1, 2048, 16, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_1224, x = var_1223_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor var_1226_begin_0 = const()[name = string("op_1226_begin_0"), val = tensor([0, 0, 4096])]; + tensor var_1226_end_0 = const()[name = string("op_1226_end_0"), val = tensor([1, 2048, 8192])]; + tensor var_1226_end_mask_0 = const()[name = string("op_1226_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1226_cast_fp16 = slice_by_index(begin = var_1226_begin_0, end = var_1226_end_0, end_mask = var_1226_end_mask_0, x = qkv_3_cast_fp16)[name = string("op_1226_cast_fp16")]; + tensor var_1227 = const()[name = string("op_1227"), val = tensor([1, 2048, 32, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_1227, x = var_1226_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor layers_1_linear_attn_in_proj_z_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140093632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150579456))))[name = string("layers_1_linear_attn_in_proj_z_weight_to_fp16_quantized")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_linear_attn_in_proj_z_weight_to_fp16_quantized, x = x_29_cast_fp16)[name = string("linear_9_cast_fp16")]; + tensor var_1231 = const()[name = string("op_1231"), val = tensor([1, 2048, 32, 128])]; + tensor z_3_cast_fp16 = reshape(shape = var_1231, x = linear_9_cast_fp16)[name = string("z_3_cast_fp16")]; + tensor layers_1_linear_attn_in_proj_b_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150587712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150669696))))[name = string("layers_1_linear_attn_in_proj_b_weight_to_fp16_quantized")]; + tensor linear_10_bias_0_to_fp16 = const()[name = string("linear_10_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150669824)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = layers_1_linear_attn_in_proj_b_weight_to_fp16_quantized, x = x_29_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor beta_5_cast_fp16 = sigmoid(x = linear_10_cast_fp16)[name = string("beta_5_cast_fp16")]; + tensor var_1240_weight_0_to_fp16 = const()[name = string("op_1240_weight_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150669952)))]; + tensor var_1240_bias_0_to_fp16 = const()[name = string("op_1240_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150833856)))]; + tensor var_1240_cast_fp16 = linear(bias = var_1240_bias_0_to_fp16, weight = var_1240_weight_0_to_fp16, x = x_29_cast_fp16)[name = string("op_1240_cast_fp16")]; + tensor var_1241_cast_fp16 = softplus(x = var_1240_cast_fp16)[name = string("op_1241_cast_fp16")]; + tensor var_1237_to_fp16 = const()[name = string("op_1237_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150833984)))]; + tensor g_5_cast_fp16 = mul(x = var_1237_to_fp16, y = var_1241_cast_fp16)[name = string("g_5_cast_fp16")]; + tensor var_1243_cast_fp16 = mul(x = q_11_cast_fp16, y = q_11_cast_fp16)[name = string("op_1243_cast_fp16")]; + tensor var_1245_axes_0 = const()[name = string("op_1245_axes_0"), val = tensor([-1])]; + bool var_1245_keep_dims_0 = const()[name = string("op_1245_keep_dims_0"), val = bool(true)]; + tensor var_1245_cast_fp16 = reduce_sum(axes = var_1245_axes_0, keep_dims = var_1245_keep_dims_0, x = var_1243_cast_fp16)[name = string("op_1245_cast_fp16")]; + fp16 var_1246_to_fp16 = const()[name = string("op_1246_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_1247_cast_fp16 = add(x = var_1245_cast_fp16, y = var_1246_to_fp16)[name = string("op_1247_cast_fp16")]; + fp32 var_1248_epsilon_0 = const()[name = string("op_1248_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1248_cast_fp16 = rsqrt(epsilon = var_1248_epsilon_0, x = var_1247_cast_fp16)[name = string("op_1248_cast_fp16")]; + tensor q_13_cast_fp16 = mul(x = q_11_cast_fp16, y = var_1248_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor var_1250_cast_fp16 = mul(x = k_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_1250_cast_fp16")]; + tensor var_1252_axes_0 = const()[name = string("op_1252_axes_0"), val = tensor([-1])]; + bool var_1252_keep_dims_0 = const()[name = string("op_1252_keep_dims_0"), val = bool(true)]; + tensor var_1252_cast_fp16 = reduce_sum(axes = var_1252_axes_0, keep_dims = var_1252_keep_dims_0, x = var_1250_cast_fp16)[name = string("op_1252_cast_fp16")]; + fp16 var_1253_to_fp16 = const()[name = string("op_1253_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_1254_cast_fp16 = add(x = var_1252_cast_fp16, y = var_1253_to_fp16)[name = string("op_1254_cast_fp16")]; + fp32 var_1255_epsilon_0 = const()[name = string("op_1255_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_1254_cast_fp16)[name = string("op_1255_cast_fp16")]; + tensor k_11_cast_fp16 = mul(x = k_9_cast_fp16, y = var_1255_cast_fp16)[name = string("k_11_cast_fp16")]; + tensor var_1260_axes_0 = const()[name = string("op_1260_axes_0"), val = tensor([3])]; + tensor var_1260_cast_fp16 = expand_dims(axes = var_1260_axes_0, x = q_13_cast_fp16)[name = string("op_1260_cast_fp16")]; + tensor var_1262_reps_0 = const()[name = string("op_1262_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_1262_cast_fp16 = tile(reps = var_1262_reps_0, x = var_1260_cast_fp16)[name = string("op_1262_cast_fp16")]; + tensor var_1263 = const()[name = string("op_1263"), val = tensor([1, 2048, 32, 128])]; + tensor q_15_cast_fp16 = reshape(shape = var_1263, x = var_1262_cast_fp16)[name = string("q_15_cast_fp16")]; + tensor var_1268_axes_0 = const()[name = string("op_1268_axes_0"), val = tensor([3])]; + tensor var_1268_cast_fp16 = expand_dims(axes = var_1268_axes_0, x = k_11_cast_fp16)[name = string("op_1268_cast_fp16")]; + tensor var_1270_reps_0 = const()[name = string("op_1270_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_1270_cast_fp16 = tile(reps = var_1270_reps_0, x = var_1268_cast_fp16)[name = string("op_1270_cast_fp16")]; + tensor var_1271 = const()[name = string("op_1271"), val = tensor([1, 2048, 32, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1271, x = var_1270_cast_fp16)[name = string("k_13_cast_fp16")]; + fp16 var_1273_to_fp16 = const()[name = string("op_1273_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor q_17_cast_fp16 = mul(x = q_15_cast_fp16, y = var_1273_to_fp16)[name = string("q_17_cast_fp16")]; + tensor var_1275_begin_0 = const()[name = string("op_1275_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1275_end_0 = const()[name = string("op_1275_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_1275_end_mask_0 = const()[name = string("op_1275_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_1275_squeeze_mask_0 = const()[name = string("op_1275_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_1275_cast_fp16 = slice_by_index(begin = var_1275_begin_0, end = var_1275_end_0, end_mask = var_1275_end_mask_0, squeeze_mask = var_1275_squeeze_mask_0, x = q_17_cast_fp16)[name = string("op_1275_cast_fp16")]; + tensor var_1276 = const()[name = string("op_1276"), val = tensor([1, 0, 2])]; + tensor var_1278 = const()[name = string("op_1278"), val = tensor([32, 32, 64, 128])]; + tensor var_1277_cast_fp16 = transpose(perm = var_1276, x = var_1275_cast_fp16)[name = string("transpose_45")]; + tensor q_19_cast_fp16 = reshape(shape = var_1278, x = var_1277_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor var_1280_begin_0 = const()[name = string("op_1280_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1280_end_0 = const()[name = string("op_1280_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_1280_end_mask_0 = const()[name = string("op_1280_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_1280_squeeze_mask_0 = const()[name = string("op_1280_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_1280_cast_fp16 = slice_by_index(begin = var_1280_begin_0, end = var_1280_end_0, end_mask = var_1280_end_mask_0, squeeze_mask = var_1280_squeeze_mask_0, x = k_13_cast_fp16)[name = string("op_1280_cast_fp16")]; + tensor var_1281 = const()[name = string("op_1281"), val = tensor([1, 0, 2])]; + tensor var_1283 = const()[name = string("op_1283"), val = tensor([32, 32, 64, 128])]; + tensor var_1282_cast_fp16 = transpose(perm = var_1281, x = var_1280_cast_fp16)[name = string("transpose_44")]; + tensor k_15_cast_fp16 = reshape(shape = var_1283, x = var_1282_cast_fp16)[name = string("k_15_cast_fp16")]; + tensor var_1285_begin_0 = const()[name = string("op_1285_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1285_end_0 = const()[name = string("op_1285_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_1285_end_mask_0 = const()[name = string("op_1285_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_1285_squeeze_mask_0 = const()[name = string("op_1285_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_1285_cast_fp16 = slice_by_index(begin = var_1285_begin_0, end = var_1285_end_0, end_mask = var_1285_end_mask_0, squeeze_mask = var_1285_squeeze_mask_0, x = v_5_cast_fp16)[name = string("op_1285_cast_fp16")]; + tensor var_1286 = const()[name = string("op_1286"), val = tensor([1, 0, 2])]; + tensor var_1288 = const()[name = string("op_1288"), val = tensor([32, 32, 64, 128])]; + tensor var_1287_cast_fp16 = transpose(perm = var_1286, x = var_1285_cast_fp16)[name = string("transpose_43")]; + tensor v_7_cast_fp16 = reshape(shape = var_1288, x = var_1287_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor var_1290_begin_0 = const()[name = string("op_1290_begin_0"), val = tensor([0, 0, 0])]; + tensor var_1290_end_0 = const()[name = string("op_1290_end_0"), val = tensor([1, 2048, 32])]; + tensor var_1290_end_mask_0 = const()[name = string("op_1290_end_mask_0"), val = tensor([false, true, true])]; + tensor var_1290_squeeze_mask_0 = const()[name = string("op_1290_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_1290_cast_fp16 = slice_by_index(begin = var_1290_begin_0, end = var_1290_end_0, end_mask = var_1290_end_mask_0, squeeze_mask = var_1290_squeeze_mask_0, x = beta_5_cast_fp16)[name = string("op_1290_cast_fp16")]; + tensor var_1291_perm_0 = const()[name = string("op_1291_perm_0"), val = tensor([1, 0])]; + tensor var_1292 = const()[name = string("op_1292"), val = tensor([32, 32, 64, 1])]; + tensor var_1291_cast_fp16 = transpose(perm = var_1291_perm_0, x = var_1290_cast_fp16)[name = string("transpose_42")]; + tensor beta_7_cast_fp16 = reshape(shape = var_1292, x = var_1291_cast_fp16)[name = string("beta_7_cast_fp16")]; + tensor var_1294_begin_0 = const()[name = string("op_1294_begin_0"), val = tensor([0, 0, 0])]; + tensor var_1294_end_0 = const()[name = string("op_1294_end_0"), val = tensor([1, 2048, 32])]; + tensor var_1294_end_mask_0 = const()[name = string("op_1294_end_mask_0"), val = tensor([false, true, true])]; + tensor var_1294_squeeze_mask_0 = const()[name = string("op_1294_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_1294_cast_fp16 = slice_by_index(begin = var_1294_begin_0, end = var_1294_end_0, end_mask = var_1294_end_mask_0, squeeze_mask = var_1294_squeeze_mask_0, x = g_5_cast_fp16)[name = string("op_1294_cast_fp16")]; + tensor var_1295_perm_0 = const()[name = string("op_1295_perm_0"), val = tensor([1, 0])]; + tensor var_1296 = const()[name = string("op_1296"), val = tensor([32, 32, 64])]; + tensor var_1295_cast_fp16 = transpose(perm = var_1295_perm_0, x = var_1294_cast_fp16)[name = string("transpose_41")]; + tensor g_7_cast_fp16 = reshape(shape = var_1296, x = var_1295_cast_fp16)[name = string("g_7_cast_fp16")]; + tensor g_col_3_axes_0 = const()[name = string("g_col_3_axes_0"), val = tensor([-1])]; + tensor g_col_3_cast_fp16 = expand_dims(axes = g_col_3_axes_0, x = g_7_cast_fp16)[name = string("g_col_3_cast_fp16")]; + bool var_1299_transpose_x_0 = const()[name = string("op_1299_transpose_x_0"), val = bool(false)]; + bool var_1299_transpose_y_0 = const()[name = string("op_1299_transpose_y_0"), val = bool(false)]; + tensor var_1299_cast_fp16 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = layers_0_linear_attn_lower_incl_to_fp16, y = g_col_3_cast_fp16)[name = string("op_1299_cast_fp16")]; + tensor cum_3_axes_0 = const()[name = string("cum_3_axes_0"), val = tensor([-1])]; + tensor cum_3_cast_fp16 = squeeze(axes = cum_3_axes_0, x = var_1299_cast_fp16)[name = string("cum_3_cast_fp16")]; + bool var_1301_transpose_x_0 = const()[name = string("op_1301_transpose_x_0"), val = bool(false)]; + bool var_1301_transpose_y_0 = const()[name = string("op_1301_transpose_y_0"), val = bool(false)]; + tensor var_1301_cast_fp16 = matmul(transpose_x = var_1301_transpose_x_0, transpose_y = var_1301_transpose_y_0, x = layers_0_linear_attn_suffix_sum_t_to_fp16, y = g_col_3_cast_fp16)[name = string("op_1301_cast_fp16")]; + tensor to_end_3_axes_0 = const()[name = string("to_end_3_axes_0"), val = tensor([-1])]; + tensor to_end_3_cast_fp16 = squeeze(axes = to_end_3_axes_0, x = var_1301_cast_fp16)[name = string("to_end_3_cast_fp16")]; + bool var_1303_transpose_x_0 = const()[name = string("op_1303_transpose_x_0"), val = bool(false)]; + bool var_1303_transpose_y_0 = const()[name = string("op_1303_transpose_y_0"), val = bool(false)]; + tensor var_1303_cast_fp16 = matmul(transpose_x = var_1303_transpose_x_0, transpose_y = var_1303_transpose_y_0, x = layers_0_linear_attn_pair_sum_t_to_fp16, y = g_col_3_cast_fp16)[name = string("op_1303_cast_fp16")]; + tensor var_1304 = const()[name = string("op_1304"), val = tensor([32, 32, 64, 64])]; + tensor pair_15_cast_fp16 = reshape(shape = var_1304, x = var_1303_cast_fp16)[name = string("pair_15_cast_fp16")]; + tensor var_1306_cast_fp16 = exp(x = pair_15_cast_fp16)[name = string("op_1306_cast_fp16")]; + tensor pair_decay_3_cast_fp16 = mul(x = var_1306_cast_fp16, y = layers_0_linear_attn_lower_incl_to_fp16)[name = string("pair_decay_3_cast_fp16")]; + tensor k_beta_3_cast_fp16 = mul(x = k_15_cast_fp16, y = beta_7_cast_fp16)[name = string("k_beta_3_cast_fp16")]; + tensor v_beta_3_cast_fp16 = mul(x = v_7_cast_fp16, y = beta_7_cast_fp16)[name = string("v_beta_3_cast_fp16")]; + bool var_1311_transpose_x_1 = const()[name = string("op_1311_transpose_x_1"), val = bool(false)]; + bool var_1311_transpose_y_1 = const()[name = string("op_1311_transpose_y_1"), val = bool(true)]; + tensor var_1311_cast_fp16 = matmul(transpose_x = var_1311_transpose_x_1, transpose_y = var_1311_transpose_y_1, x = k_beta_3_cast_fp16, y = k_15_cast_fp16)[name = string("op_1311_cast_fp16")]; + tensor kkt_3_cast_fp16 = mul(x = var_1311_cast_fp16, y = pair_decay_3_cast_fp16)[name = string("kkt_3_cast_fp16")]; + bool var_1314_transpose_x_1 = const()[name = string("op_1314_transpose_x_1"), val = bool(false)]; + bool var_1314_transpose_y_1 = const()[name = string("op_1314_transpose_y_1"), val = bool(true)]; + tensor var_1314_cast_fp16 = matmul(transpose_x = var_1314_transpose_x_1, transpose_y = var_1314_transpose_y_1, x = q_19_cast_fp16, y = k_15_cast_fp16)[name = string("op_1314_cast_fp16")]; + tensor attn_intra_3_cast_fp16 = mul(x = var_1314_cast_fp16, y = pair_decay_3_cast_fp16)[name = string("attn_intra_3_cast_fp16")]; + tensor var_1316_cast_fp16 = mul(x = kkt_3_cast_fp16, y = layers_0_linear_attn_strict_lower_to_fp16)[name = string("op_1316_cast_fp16")]; + tensor t_5_cast_fp16 = add(x = layers_0_linear_attn_eye_to_fp16, y = var_1316_cast_fp16)[name = string("t_5_cast_fp16")]; + tensor var_1320 = const()[name = string("op_1320"), val = tensor([-1, 64, 64])]; + tensor t_7_cast_fp16 = reshape(shape = var_1320, x = t_5_cast_fp16)[name = string("t_7_cast_fp16")]; + tensor var_1325 = const()[name = string("op_1325"), val = tensor([1024, 32, 2, 32, 2])]; + tensor var_1326_cast_fp16 = reshape(shape = var_1325, x = t_7_cast_fp16)[name = string("op_1326_cast_fp16")]; + tensor var_1331_cast_fp16 = mul(x = var_1326_cast_fp16, y = var_227_to_fp16)[name = string("op_1331_cast_fp16")]; + tensor blocks_13_axes_0 = const()[name = string("blocks_13_axes_0"), val = tensor([-2])]; + bool blocks_13_keep_dims_0 = const()[name = string("blocks_13_keep_dims_0"), val = bool(false)]; + tensor blocks_13_cast_fp16 = reduce_sum(axes = blocks_13_axes_0, keep_dims = blocks_13_keep_dims_0, x = var_1331_cast_fp16)[name = string("blocks_13_cast_fp16")]; + tensor var_1334_begin_0 = const()[name = string("op_1334_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1334_end_0 = const()[name = string("op_1334_end_0"), val = tensor([1024, 32, 2, 2])]; + tensor var_1334_end_mask_0 = const()[name = string("op_1334_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1334_cast_fp16 = slice_by_index(begin = var_1334_begin_0, end = var_1334_end_0, end_mask = var_1334_end_mask_0, x = blocks_13_cast_fp16)[name = string("op_1334_cast_fp16")]; + tensor x_31_begin_0 = const()[name = string("x_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_31_end_0 = const()[name = string("x_31_end_0"), val = tensor([1024, 32, 1, 1])]; + tensor x_31_end_mask_0 = const()[name = string("x_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, x = var_1334_cast_fp16)[name = string("x_31_cast_fp16")]; + bool var_1344_transpose_x_0 = const()[name = string("op_1344_transpose_x_0"), val = bool(false)]; + bool var_1344_transpose_y_0 = const()[name = string("op_1344_transpose_y_0"), val = bool(false)]; + tensor var_1344_cast_fp16 = matmul(transpose_x = var_1344_transpose_x_0, transpose_y = var_1344_transpose_y_0, x = x_31_cast_fp16, y = var_235_to_fp16)[name = string("op_1344_cast_fp16")]; + bool var_1345_transpose_x_0 = const()[name = string("op_1345_transpose_x_0"), val = bool(false)]; + bool var_1345_transpose_y_0 = const()[name = string("op_1345_transpose_y_0"), val = bool(false)]; + tensor var_1345_cast_fp16 = matmul(transpose_x = var_1345_transpose_x_0, transpose_y = var_1345_transpose_y_0, x = var_235_to_fp16, y = var_1344_cast_fp16)[name = string("op_1345_cast_fp16")]; + fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_13_cast_fp16 = mul(x = var_1345_cast_fp16, y = const_18_promoted_to_fp16)[name = string("lower_13_cast_fp16")]; + bool bottom_13_interleave_0 = const()[name = string("bottom_13_interleave_0"), val = bool(false)]; + tensor bottom_13_cast_fp16 = concat(axis = var_1166, interleave = bottom_13_interleave_0, values = (lower_13_cast_fp16, var_235_to_fp16))[name = string("bottom_13_cast_fp16")]; + bool inv_19_interleave_0 = const()[name = string("inv_19_interleave_0"), val = bool(false)]; + tensor inv_19_cast_fp16 = concat(axis = var_1137, interleave = inv_19_interleave_0, values = (top_1_to_fp16, bottom_13_cast_fp16))[name = string("inv_19_cast_fp16")]; + tensor var_1354 = const()[name = string("op_1354"), val = tensor([1024, 16, 4, 16, 4])]; + tensor var_1355_cast_fp16 = reshape(shape = var_1354, x = t_7_cast_fp16)[name = string("op_1355_cast_fp16")]; + tensor var_1360_cast_fp16 = mul(x = var_1355_cast_fp16, y = var_256_to_fp16)[name = string("op_1360_cast_fp16")]; + tensor blocks_15_axes_0 = const()[name = string("blocks_15_axes_0"), val = tensor([-2])]; + bool blocks_15_keep_dims_0 = const()[name = string("blocks_15_keep_dims_0"), val = bool(false)]; + tensor blocks_15_cast_fp16 = reduce_sum(axes = blocks_15_axes_0, keep_dims = blocks_15_keep_dims_0, x = var_1360_cast_fp16)[name = string("blocks_15_cast_fp16")]; + tensor var_1363_begin_0 = const()[name = string("op_1363_begin_0"), val = tensor([0, 0, 2, 0])]; + tensor var_1363_end_0 = const()[name = string("op_1363_end_0"), val = tensor([1024, 16, 4, 4])]; + tensor var_1363_end_mask_0 = const()[name = string("op_1363_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1363_cast_fp16 = slice_by_index(begin = var_1363_begin_0, end = var_1363_end_0, end_mask = var_1363_end_mask_0, x = blocks_15_cast_fp16)[name = string("op_1363_cast_fp16")]; + tensor x_33_begin_0 = const()[name = string("x_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = string("x_33_end_0"), val = tensor([1024, 16, 2, 2])]; + tensor x_33_end_mask_0 = const()[name = string("x_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, x = var_1363_cast_fp16)[name = string("x_33_cast_fp16")]; + tensor var_1365 = const()[name = string("op_1365"), val = tensor([1024, 16, 2, 2, 2])]; + tensor pair_19_cast_fp16 = reshape(shape = var_1365, x = inv_19_cast_fp16)[name = string("pair_19_cast_fp16")]; + tensor var_1367_begin_0 = const()[name = string("op_1367_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1367_end_0 = const()[name = string("op_1367_end_0"), val = tensor([1024, 16, 1, 2, 2])]; + tensor var_1367_end_mask_0 = const()[name = string("op_1367_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1367_squeeze_mask_0 = const()[name = string("op_1367_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1367_cast_fp16 = slice_by_index(begin = var_1367_begin_0, end = var_1367_end_0, end_mask = var_1367_end_mask_0, squeeze_mask = var_1367_squeeze_mask_0, x = pair_19_cast_fp16)[name = string("op_1367_cast_fp16")]; + tensor var_1370_begin_0 = const()[name = string("op_1370_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_1370_end_0 = const()[name = string("op_1370_end_0"), val = tensor([1024, 16, 2, 2, 2])]; + tensor var_1370_end_mask_0 = const()[name = string("op_1370_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1370_squeeze_mask_0 = const()[name = string("op_1370_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1370_cast_fp16 = slice_by_index(begin = var_1370_begin_0, end = var_1370_end_0, end_mask = var_1370_end_mask_0, squeeze_mask = var_1370_squeeze_mask_0, x = pair_19_cast_fp16)[name = string("op_1370_cast_fp16")]; + bool var_1373_transpose_x_0 = const()[name = string("op_1373_transpose_x_0"), val = bool(false)]; + bool var_1373_transpose_y_0 = const()[name = string("op_1373_transpose_y_0"), val = bool(false)]; + tensor var_1373_cast_fp16 = matmul(transpose_x = var_1373_transpose_x_0, transpose_y = var_1373_transpose_y_0, x = x_33_cast_fp16, y = var_1367_cast_fp16)[name = string("op_1373_cast_fp16")]; + bool var_1374_transpose_x_0 = const()[name = string("op_1374_transpose_x_0"), val = bool(false)]; + bool var_1374_transpose_y_0 = const()[name = string("op_1374_transpose_y_0"), val = bool(false)]; + tensor var_1374_cast_fp16 = matmul(transpose_x = var_1374_transpose_x_0, transpose_y = var_1374_transpose_y_0, x = var_1370_cast_fp16, y = var_1373_cast_fp16)[name = string("op_1374_cast_fp16")]; + fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_15_cast_fp16 = mul(x = var_1374_cast_fp16, y = const_19_promoted_to_fp16)[name = string("lower_15_cast_fp16")]; + bool top_15_interleave_0 = const()[name = string("top_15_interleave_0"), val = bool(false)]; + tensor top_15_cast_fp16 = concat(axis = var_1166, interleave = top_15_interleave_0, values = (var_1367_cast_fp16, var_273_to_fp16))[name = string("top_15_cast_fp16")]; + bool bottom_15_interleave_0 = const()[name = string("bottom_15_interleave_0"), val = bool(false)]; + tensor bottom_15_cast_fp16 = concat(axis = var_1166, interleave = bottom_15_interleave_0, values = (lower_15_cast_fp16, var_1370_cast_fp16))[name = string("bottom_15_cast_fp16")]; + bool inv_21_interleave_0 = const()[name = string("inv_21_interleave_0"), val = bool(false)]; + tensor inv_21_cast_fp16 = concat(axis = var_1137, interleave = inv_21_interleave_0, values = (top_15_cast_fp16, bottom_15_cast_fp16))[name = string("inv_21_cast_fp16")]; + tensor var_1383 = const()[name = string("op_1383"), val = tensor([1024, 8, 8, 8, 8])]; + tensor var_1384_cast_fp16 = reshape(shape = var_1383, x = t_7_cast_fp16)[name = string("op_1384_cast_fp16")]; + tensor var_1388_to_fp16 = const()[name = string("op_1388_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150834112)))]; + tensor var_1389_cast_fp16 = mul(x = var_1384_cast_fp16, y = var_1388_to_fp16)[name = string("op_1389_cast_fp16")]; + tensor blocks_17_axes_0 = const()[name = string("blocks_17_axes_0"), val = tensor([-2])]; + bool blocks_17_keep_dims_0 = const()[name = string("blocks_17_keep_dims_0"), val = bool(false)]; + tensor blocks_17_cast_fp16 = reduce_sum(axes = blocks_17_axes_0, keep_dims = blocks_17_keep_dims_0, x = var_1389_cast_fp16)[name = string("blocks_17_cast_fp16")]; + tensor var_1392_begin_0 = const()[name = string("op_1392_begin_0"), val = tensor([0, 0, 4, 0])]; + tensor var_1392_end_0 = const()[name = string("op_1392_end_0"), val = tensor([1024, 8, 8, 8])]; + tensor var_1392_end_mask_0 = const()[name = string("op_1392_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1392_cast_fp16 = slice_by_index(begin = var_1392_begin_0, end = var_1392_end_0, end_mask = var_1392_end_mask_0, x = blocks_17_cast_fp16)[name = string("op_1392_cast_fp16")]; + tensor x_35_begin_0 = const()[name = string("x_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_35_end_0 = const()[name = string("x_35_end_0"), val = tensor([1024, 8, 4, 4])]; + tensor x_35_end_mask_0 = const()[name = string("x_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, x = var_1392_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_1394 = const()[name = string("op_1394"), val = tensor([1024, 8, 2, 4, 4])]; + tensor pair_21_cast_fp16 = reshape(shape = var_1394, x = inv_21_cast_fp16)[name = string("pair_21_cast_fp16")]; + tensor var_1396_begin_0 = const()[name = string("op_1396_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1396_end_0 = const()[name = string("op_1396_end_0"), val = tensor([1024, 8, 1, 4, 4])]; + tensor var_1396_end_mask_0 = const()[name = string("op_1396_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1396_squeeze_mask_0 = const()[name = string("op_1396_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1396_cast_fp16 = slice_by_index(begin = var_1396_begin_0, end = var_1396_end_0, end_mask = var_1396_end_mask_0, squeeze_mask = var_1396_squeeze_mask_0, x = pair_21_cast_fp16)[name = string("op_1396_cast_fp16")]; + tensor var_1399_begin_0 = const()[name = string("op_1399_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_1399_end_0 = const()[name = string("op_1399_end_0"), val = tensor([1024, 8, 2, 4, 4])]; + tensor var_1399_end_mask_0 = const()[name = string("op_1399_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1399_squeeze_mask_0 = const()[name = string("op_1399_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1399_cast_fp16 = slice_by_index(begin = var_1399_begin_0, end = var_1399_end_0, end_mask = var_1399_end_mask_0, squeeze_mask = var_1399_squeeze_mask_0, x = pair_21_cast_fp16)[name = string("op_1399_cast_fp16")]; + bool var_1402_transpose_x_0 = const()[name = string("op_1402_transpose_x_0"), val = bool(false)]; + bool var_1402_transpose_y_0 = const()[name = string("op_1402_transpose_y_0"), val = bool(false)]; + tensor var_1402_cast_fp16 = matmul(transpose_x = var_1402_transpose_x_0, transpose_y = var_1402_transpose_y_0, x = x_35_cast_fp16, y = var_1396_cast_fp16)[name = string("op_1402_cast_fp16")]; + bool var_1403_transpose_x_0 = const()[name = string("op_1403_transpose_x_0"), val = bool(false)]; + bool var_1403_transpose_y_0 = const()[name = string("op_1403_transpose_y_0"), val = bool(false)]; + tensor var_1403_cast_fp16 = matmul(transpose_x = var_1403_transpose_x_0, transpose_y = var_1403_transpose_y_0, x = var_1399_cast_fp16, y = var_1402_cast_fp16)[name = string("op_1403_cast_fp16")]; + fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_17_cast_fp16 = mul(x = var_1403_cast_fp16, y = const_20_promoted_to_fp16)[name = string("lower_17_cast_fp16")]; + bool top_17_interleave_0 = const()[name = string("top_17_interleave_0"), val = bool(false)]; + tensor top_17_cast_fp16 = concat(axis = var_1166, interleave = top_17_interleave_0, values = (var_1396_cast_fp16, var_302_to_fp16))[name = string("top_17_cast_fp16")]; + bool bottom_17_interleave_0 = const()[name = string("bottom_17_interleave_0"), val = bool(false)]; + tensor bottom_17_cast_fp16 = concat(axis = var_1166, interleave = bottom_17_interleave_0, values = (lower_17_cast_fp16, var_1399_cast_fp16))[name = string("bottom_17_cast_fp16")]; + bool inv_23_interleave_0 = const()[name = string("inv_23_interleave_0"), val = bool(false)]; + tensor inv_23_cast_fp16 = concat(axis = var_1137, interleave = inv_23_interleave_0, values = (top_17_cast_fp16, bottom_17_cast_fp16))[name = string("inv_23_cast_fp16")]; + tensor var_1412 = const()[name = string("op_1412"), val = tensor([1024, 4, 16, 4, 16])]; + tensor var_1413_cast_fp16 = reshape(shape = var_1412, x = t_7_cast_fp16)[name = string("op_1413_cast_fp16")]; + tensor var_1417_to_fp16 = const()[name = string("op_1417_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150834304)))]; + tensor var_1418_cast_fp16 = mul(x = var_1413_cast_fp16, y = var_1417_to_fp16)[name = string("op_1418_cast_fp16")]; + tensor blocks_19_axes_0 = const()[name = string("blocks_19_axes_0"), val = tensor([-2])]; + bool blocks_19_keep_dims_0 = const()[name = string("blocks_19_keep_dims_0"), val = bool(false)]; + tensor blocks_19_cast_fp16 = reduce_sum(axes = blocks_19_axes_0, keep_dims = blocks_19_keep_dims_0, x = var_1418_cast_fp16)[name = string("blocks_19_cast_fp16")]; + tensor var_1421_begin_0 = const()[name = string("op_1421_begin_0"), val = tensor([0, 0, 8, 0])]; + tensor var_1421_end_0 = const()[name = string("op_1421_end_0"), val = tensor([1024, 4, 16, 16])]; + tensor var_1421_end_mask_0 = const()[name = string("op_1421_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1421_cast_fp16 = slice_by_index(begin = var_1421_begin_0, end = var_1421_end_0, end_mask = var_1421_end_mask_0, x = blocks_19_cast_fp16)[name = string("op_1421_cast_fp16")]; + tensor x_37_begin_0 = const()[name = string("x_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = string("x_37_end_0"), val = tensor([1024, 4, 8, 8])]; + tensor x_37_end_mask_0 = const()[name = string("x_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, x = var_1421_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_1423 = const()[name = string("op_1423"), val = tensor([1024, 4, 2, 8, 8])]; + tensor pair_23_cast_fp16 = reshape(shape = var_1423, x = inv_23_cast_fp16)[name = string("pair_23_cast_fp16")]; + tensor var_1425_begin_0 = const()[name = string("op_1425_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1425_end_0 = const()[name = string("op_1425_end_0"), val = tensor([1024, 4, 1, 8, 8])]; + tensor var_1425_end_mask_0 = const()[name = string("op_1425_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1425_squeeze_mask_0 = const()[name = string("op_1425_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1425_cast_fp16 = slice_by_index(begin = var_1425_begin_0, end = var_1425_end_0, end_mask = var_1425_end_mask_0, squeeze_mask = var_1425_squeeze_mask_0, x = pair_23_cast_fp16)[name = string("op_1425_cast_fp16")]; + tensor var_1428_begin_0 = const()[name = string("op_1428_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_1428_end_0 = const()[name = string("op_1428_end_0"), val = tensor([1024, 4, 2, 8, 8])]; + tensor var_1428_end_mask_0 = const()[name = string("op_1428_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1428_squeeze_mask_0 = const()[name = string("op_1428_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1428_cast_fp16 = slice_by_index(begin = var_1428_begin_0, end = var_1428_end_0, end_mask = var_1428_end_mask_0, squeeze_mask = var_1428_squeeze_mask_0, x = pair_23_cast_fp16)[name = string("op_1428_cast_fp16")]; + bool var_1431_transpose_x_0 = const()[name = string("op_1431_transpose_x_0"), val = bool(false)]; + bool var_1431_transpose_y_0 = const()[name = string("op_1431_transpose_y_0"), val = bool(false)]; + tensor var_1431_cast_fp16 = matmul(transpose_x = var_1431_transpose_x_0, transpose_y = var_1431_transpose_y_0, x = x_37_cast_fp16, y = var_1425_cast_fp16)[name = string("op_1431_cast_fp16")]; + bool var_1432_transpose_x_0 = const()[name = string("op_1432_transpose_x_0"), val = bool(false)]; + bool var_1432_transpose_y_0 = const()[name = string("op_1432_transpose_y_0"), val = bool(false)]; + tensor var_1432_cast_fp16 = matmul(transpose_x = var_1432_transpose_x_0, transpose_y = var_1432_transpose_y_0, x = var_1428_cast_fp16, y = var_1431_cast_fp16)[name = string("op_1432_cast_fp16")]; + fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_19_cast_fp16 = mul(x = var_1432_cast_fp16, y = const_21_promoted_to_fp16)[name = string("lower_19_cast_fp16")]; + bool top_19_interleave_0 = const()[name = string("top_19_interleave_0"), val = bool(false)]; + tensor top_19_cast_fp16 = concat(axis = var_1166, interleave = top_19_interleave_0, values = (var_1425_cast_fp16, var_331_to_fp16))[name = string("top_19_cast_fp16")]; + bool bottom_19_interleave_0 = const()[name = string("bottom_19_interleave_0"), val = bool(false)]; + tensor bottom_19_cast_fp16 = concat(axis = var_1166, interleave = bottom_19_interleave_0, values = (lower_19_cast_fp16, var_1428_cast_fp16))[name = string("bottom_19_cast_fp16")]; + bool inv_25_interleave_0 = const()[name = string("inv_25_interleave_0"), val = bool(false)]; + tensor inv_25_cast_fp16 = concat(axis = var_1137, interleave = inv_25_interleave_0, values = (top_19_cast_fp16, bottom_19_cast_fp16))[name = string("inv_25_cast_fp16")]; + tensor var_1441 = const()[name = string("op_1441"), val = tensor([1024, 2, 32, 2, 32])]; + tensor var_1442_cast_fp16 = reshape(shape = var_1441, x = t_7_cast_fp16)[name = string("op_1442_cast_fp16")]; + tensor var_1446_to_fp16 = const()[name = string("op_1446_to_fp16"), val = tensor([[[[0x1p+0], [0x0p+0]]], [[[0x0p+0], [0x1p+0]]]])]; + tensor var_1447_cast_fp16 = mul(x = var_1442_cast_fp16, y = var_1446_to_fp16)[name = string("op_1447_cast_fp16")]; + tensor blocks_21_axes_0 = const()[name = string("blocks_21_axes_0"), val = tensor([-2])]; + bool blocks_21_keep_dims_0 = const()[name = string("blocks_21_keep_dims_0"), val = bool(false)]; + tensor blocks_21_cast_fp16 = reduce_sum(axes = blocks_21_axes_0, keep_dims = blocks_21_keep_dims_0, x = var_1447_cast_fp16)[name = string("blocks_21_cast_fp16")]; + tensor var_1450_begin_0 = const()[name = string("op_1450_begin_0"), val = tensor([0, 0, 16, 0])]; + tensor var_1450_end_0 = const()[name = string("op_1450_end_0"), val = tensor([1024, 2, 32, 32])]; + tensor var_1450_end_mask_0 = const()[name = string("op_1450_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1450_cast_fp16 = slice_by_index(begin = var_1450_begin_0, end = var_1450_end_0, end_mask = var_1450_end_mask_0, x = blocks_21_cast_fp16)[name = string("op_1450_cast_fp16")]; + tensor x_39_begin_0 = const()[name = string("x_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_39_end_0 = const()[name = string("x_39_end_0"), val = tensor([1024, 2, 16, 16])]; + tensor x_39_end_mask_0 = const()[name = string("x_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, x = var_1450_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_1452 = const()[name = string("op_1452"), val = tensor([1024, 2, 2, 16, 16])]; + tensor pair_25_cast_fp16 = reshape(shape = var_1452, x = inv_25_cast_fp16)[name = string("pair_25_cast_fp16")]; + tensor var_1454_begin_0 = const()[name = string("op_1454_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1454_end_0 = const()[name = string("op_1454_end_0"), val = tensor([1024, 2, 1, 16, 16])]; + tensor var_1454_end_mask_0 = const()[name = string("op_1454_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1454_squeeze_mask_0 = const()[name = string("op_1454_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1454_cast_fp16 = slice_by_index(begin = var_1454_begin_0, end = var_1454_end_0, end_mask = var_1454_end_mask_0, squeeze_mask = var_1454_squeeze_mask_0, x = pair_25_cast_fp16)[name = string("op_1454_cast_fp16")]; + tensor var_1457_begin_0 = const()[name = string("op_1457_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_1457_end_0 = const()[name = string("op_1457_end_0"), val = tensor([1024, 2, 2, 16, 16])]; + tensor var_1457_end_mask_0 = const()[name = string("op_1457_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1457_squeeze_mask_0 = const()[name = string("op_1457_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1457_cast_fp16 = slice_by_index(begin = var_1457_begin_0, end = var_1457_end_0, end_mask = var_1457_end_mask_0, squeeze_mask = var_1457_squeeze_mask_0, x = pair_25_cast_fp16)[name = string("op_1457_cast_fp16")]; + bool var_1460_transpose_x_0 = const()[name = string("op_1460_transpose_x_0"), val = bool(false)]; + bool var_1460_transpose_y_0 = const()[name = string("op_1460_transpose_y_0"), val = bool(false)]; + tensor var_1460_cast_fp16 = matmul(transpose_x = var_1460_transpose_x_0, transpose_y = var_1460_transpose_y_0, x = x_39_cast_fp16, y = var_1454_cast_fp16)[name = string("op_1460_cast_fp16")]; + bool var_1461_transpose_x_0 = const()[name = string("op_1461_transpose_x_0"), val = bool(false)]; + bool var_1461_transpose_y_0 = const()[name = string("op_1461_transpose_y_0"), val = bool(false)]; + tensor var_1461_cast_fp16 = matmul(transpose_x = var_1461_transpose_x_0, transpose_y = var_1461_transpose_y_0, x = var_1457_cast_fp16, y = var_1460_cast_fp16)[name = string("op_1461_cast_fp16")]; + fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_21_cast_fp16 = mul(x = var_1461_cast_fp16, y = const_22_promoted_to_fp16)[name = string("lower_21_cast_fp16")]; + bool top_21_interleave_0 = const()[name = string("top_21_interleave_0"), val = bool(false)]; + tensor top_21_cast_fp16 = concat(axis = var_1166, interleave = top_21_interleave_0, values = (var_1454_cast_fp16, var_360_to_fp16))[name = string("top_21_cast_fp16")]; + bool bottom_21_interleave_0 = const()[name = string("bottom_21_interleave_0"), val = bool(false)]; + tensor bottom_21_cast_fp16 = concat(axis = var_1166, interleave = bottom_21_interleave_0, values = (lower_21_cast_fp16, var_1457_cast_fp16))[name = string("bottom_21_cast_fp16")]; + bool inv_27_interleave_0 = const()[name = string("inv_27_interleave_0"), val = bool(false)]; + tensor inv_27_cast_fp16 = concat(axis = var_1137, interleave = inv_27_interleave_0, values = (top_21_cast_fp16, bottom_21_cast_fp16))[name = string("inv_27_cast_fp16")]; + tensor var_1470 = const()[name = string("op_1470"), val = tensor([1024, 1, 64, 1, 64])]; + tensor var_1471_cast_fp16 = reshape(shape = var_1470, x = t_7_cast_fp16)[name = string("op_1471_cast_fp16")]; + tensor blocks_23_axes_0 = const()[name = string("blocks_23_axes_0"), val = tensor([-2])]; + bool blocks_23_keep_dims_0 = const()[name = string("blocks_23_keep_dims_0"), val = bool(false)]; + tensor blocks_23_cast_fp16 = reduce_sum(axes = blocks_23_axes_0, keep_dims = blocks_23_keep_dims_0, x = var_1471_cast_fp16)[name = string("blocks_23_cast_fp16")]; + tensor var_1479_begin_0 = const()[name = string("op_1479_begin_0"), val = tensor([0, 0, 32, 0])]; + tensor var_1479_end_0 = const()[name = string("op_1479_end_0"), val = tensor([1024, 1, 64, 64])]; + tensor var_1479_end_mask_0 = const()[name = string("op_1479_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1479_cast_fp16 = slice_by_index(begin = var_1479_begin_0, end = var_1479_end_0, end_mask = var_1479_end_mask_0, x = blocks_23_cast_fp16)[name = string("op_1479_cast_fp16")]; + tensor x_41_begin_0 = const()[name = string("x_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = string("x_41_end_0"), val = tensor([1024, 1, 32, 32])]; + tensor x_41_end_mask_0 = const()[name = string("x_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, x = var_1479_cast_fp16)[name = string("x_41_cast_fp16")]; + tensor var_1481 = const()[name = string("op_1481"), val = tensor([1024, 1, 2, 32, 32])]; + tensor pair_27_cast_fp16 = reshape(shape = var_1481, x = inv_27_cast_fp16)[name = string("pair_27_cast_fp16")]; + tensor var_1483_begin_0 = const()[name = string("op_1483_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1483_end_0 = const()[name = string("op_1483_end_0"), val = tensor([1024, 1, 1, 32, 32])]; + tensor var_1483_end_mask_0 = const()[name = string("op_1483_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1483_squeeze_mask_0 = const()[name = string("op_1483_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1483_cast_fp16 = slice_by_index(begin = var_1483_begin_0, end = var_1483_end_0, end_mask = var_1483_end_mask_0, squeeze_mask = var_1483_squeeze_mask_0, x = pair_27_cast_fp16)[name = string("op_1483_cast_fp16")]; + tensor var_1486_begin_0 = const()[name = string("op_1486_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_1486_end_0 = const()[name = string("op_1486_end_0"), val = tensor([1024, 1, 2, 32, 32])]; + tensor var_1486_end_mask_0 = const()[name = string("op_1486_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_1486_squeeze_mask_0 = const()[name = string("op_1486_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_1486_cast_fp16 = slice_by_index(begin = var_1486_begin_0, end = var_1486_end_0, end_mask = var_1486_end_mask_0, squeeze_mask = var_1486_squeeze_mask_0, x = pair_27_cast_fp16)[name = string("op_1486_cast_fp16")]; + bool var_1489_transpose_x_0 = const()[name = string("op_1489_transpose_x_0"), val = bool(false)]; + bool var_1489_transpose_y_0 = const()[name = string("op_1489_transpose_y_0"), val = bool(false)]; + tensor var_1489_cast_fp16 = matmul(transpose_x = var_1489_transpose_x_0, transpose_y = var_1489_transpose_y_0, x = x_41_cast_fp16, y = var_1483_cast_fp16)[name = string("op_1489_cast_fp16")]; + bool var_1490_transpose_x_0 = const()[name = string("op_1490_transpose_x_0"), val = bool(false)]; + bool var_1490_transpose_y_0 = const()[name = string("op_1490_transpose_y_0"), val = bool(false)]; + tensor var_1490_cast_fp16 = matmul(transpose_x = var_1490_transpose_x_0, transpose_y = var_1490_transpose_y_0, x = var_1486_cast_fp16, y = var_1489_cast_fp16)[name = string("op_1490_cast_fp16")]; + fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_23_cast_fp16 = mul(x = var_1490_cast_fp16, y = const_23_promoted_to_fp16)[name = string("lower_23_cast_fp16")]; + bool top_23_interleave_0 = const()[name = string("top_23_interleave_0"), val = bool(false)]; + tensor top_23_cast_fp16 = concat(axis = var_1166, interleave = top_23_interleave_0, values = (var_1483_cast_fp16, var_389_to_fp16))[name = string("top_23_cast_fp16")]; + bool bottom_23_interleave_0 = const()[name = string("bottom_23_interleave_0"), val = bool(false)]; + tensor bottom_23_cast_fp16 = concat(axis = var_1166, interleave = bottom_23_interleave_0, values = (lower_23_cast_fp16, var_1486_cast_fp16))[name = string("bottom_23_cast_fp16")]; + bool inv_29_interleave_0 = const()[name = string("inv_29_interleave_0"), val = bool(false)]; + tensor inv_29_cast_fp16 = concat(axis = var_1137, interleave = inv_29_interleave_0, values = (top_23_cast_fp16, bottom_23_cast_fp16))[name = string("inv_29_cast_fp16")]; + tensor var_1499 = const()[name = string("op_1499"), val = tensor([32, 32, 64, 64])]; + tensor inv_31_cast_fp16 = reshape(shape = var_1499, x = inv_29_cast_fp16)[name = string("inv_31_cast_fp16")]; + bool new_v_3_transpose_x_0 = const()[name = string("new_v_3_transpose_x_0"), val = bool(false)]; + bool new_v_3_transpose_y_0 = const()[name = string("new_v_3_transpose_y_0"), val = bool(false)]; + tensor new_v_3_cast_fp16 = matmul(transpose_x = new_v_3_transpose_x_0, transpose_y = new_v_3_transpose_y_0, x = inv_31_cast_fp16, y = v_beta_3_cast_fp16)[name = string("new_v_3_cast_fp16")]; + tensor var_1502_cast_fp16 = exp(x = cum_3_cast_fp16)[name = string("op_1502_cast_fp16")]; + tensor var_1503_axes_0 = const()[name = string("op_1503_axes_0"), val = tensor([3])]; + tensor var_1503_cast_fp16 = expand_dims(axes = var_1503_axes_0, x = var_1502_cast_fp16)[name = string("op_1503_cast_fp16")]; + tensor var_1504_cast_fp16 = mul(x = k_beta_3_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")]; + bool k_cumdecay_3_transpose_x_0 = const()[name = string("k_cumdecay_3_transpose_x_0"), val = bool(false)]; + bool k_cumdecay_3_transpose_y_0 = const()[name = string("k_cumdecay_3_transpose_y_0"), val = bool(false)]; + tensor k_cumdecay_3_cast_fp16 = matmul(transpose_x = k_cumdecay_3_transpose_x_0, transpose_y = k_cumdecay_3_transpose_y_0, x = inv_31_cast_fp16, y = var_1504_cast_fp16)[name = string("k_cumdecay_3_cast_fp16")]; + tensor q_dec_3_cast_fp16 = mul(x = q_19_cast_fp16, y = var_1503_cast_fp16)[name = string("q_dec_3_cast_fp16")]; + tensor var_1509_cast_fp16 = exp(x = to_end_3_cast_fp16)[name = string("op_1509_cast_fp16")]; + tensor var_1510_axes_0 = const()[name = string("op_1510_axes_0"), val = tensor([3])]; + tensor var_1510_cast_fp16 = expand_dims(axes = var_1510_axes_0, x = var_1509_cast_fp16)[name = string("op_1510_cast_fp16")]; + tensor k_dec_3_cast_fp16 = mul(x = k_15_cast_fp16, y = var_1510_cast_fp16)[name = string("k_dec_3_cast_fp16")]; + tensor var_1512_begin_0 = const()[name = string("op_1512_begin_0"), val = tensor([0, 0, -1])]; + tensor var_1512_end_0 = const()[name = string("op_1512_end_0"), val = tensor([32, 32, 64])]; + tensor var_1512_end_mask_0 = const()[name = string("op_1512_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1512_cast_fp16 = slice_by_index(begin = var_1512_begin_0, end = var_1512_end_0, end_mask = var_1512_end_mask_0, x = cum_3_cast_fp16)[name = string("op_1512_cast_fp16")]; + tensor var_1513_cast_fp16 = exp(x = var_1512_cast_fp16)[name = string("op_1513_cast_fp16")]; + tensor chunk_decay_3_axes_0 = const()[name = string("chunk_decay_3_axes_0"), val = tensor([3])]; + tensor chunk_decay_3_cast_fp16 = expand_dims(axes = chunk_decay_3_axes_0, x = var_1513_cast_fp16)[name = string("chunk_decay_3_cast_fp16")]; + tensor var_1518_begin_0 = const()[name = string("op_1518_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1518_end_0 = const()[name = string("op_1518_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_1518_end_mask_0 = const()[name = string("op_1518_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1518_squeeze_mask_0 = const()[name = string("op_1518_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1518_cast_fp16 = slice_by_index(begin = var_1518_begin_0, end = var_1518_end_0, end_mask = var_1518_end_mask_0, squeeze_mask = var_1518_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1518_cast_fp16")]; + tensor var_1520_begin_0 = const()[name = string("op_1520_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1520_end_0 = const()[name = string("op_1520_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_1520_end_mask_0 = const()[name = string("op_1520_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1520_squeeze_mask_0 = const()[name = string("op_1520_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1520_cast_fp16 = slice_by_index(begin = var_1520_begin_0, end = var_1520_end_0, end_mask = var_1520_end_mask_0, squeeze_mask = var_1520_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1520_cast_fp16")]; + bool var_1521_transpose_x_0 = const()[name = string("op_1521_transpose_x_0"), val = bool(false)]; + bool var_1521_transpose_y_0 = const()[name = string("op_1521_transpose_y_0"), val = bool(false)]; + tensor var_1521_cast_fp16 = matmul(transpose_x = var_1521_transpose_x_0, transpose_y = var_1521_transpose_y_0, x = var_1520_cast_fp16, y = state_1_to_fp16)[name = string("op_1521_cast_fp16")]; + tensor v_new_65_cast_fp16 = sub(x = var_1518_cast_fp16, y = var_1521_cast_fp16)[name = string("v_new_65_cast_fp16")]; + tensor var_1524_begin_0 = const()[name = string("op_1524_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1524_end_0 = const()[name = string("op_1524_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_1524_end_mask_0 = const()[name = string("op_1524_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1524_squeeze_mask_0 = const()[name = string("op_1524_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1524_cast_fp16 = slice_by_index(begin = var_1524_begin_0, end = var_1524_end_0, end_mask = var_1524_end_mask_0, squeeze_mask = var_1524_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1524_cast_fp16")]; + bool var_1525_transpose_x_0 = const()[name = string("op_1525_transpose_x_0"), val = bool(false)]; + bool var_1525_transpose_y_0 = const()[name = string("op_1525_transpose_y_0"), val = bool(false)]; + tensor var_1525_cast_fp16 = matmul(transpose_x = var_1525_transpose_x_0, transpose_y = var_1525_transpose_y_0, x = var_1524_cast_fp16, y = state_1_to_fp16)[name = string("op_1525_cast_fp16")]; + tensor var_1527_begin_0 = const()[name = string("op_1527_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1527_end_0 = const()[name = string("op_1527_end_0"), val = tensor([32, 1, 64, 64])]; + tensor var_1527_end_mask_0 = const()[name = string("op_1527_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1527_squeeze_mask_0 = const()[name = string("op_1527_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1527_cast_fp16 = slice_by_index(begin = var_1527_begin_0, end = var_1527_end_0, end_mask = var_1527_end_mask_0, squeeze_mask = var_1527_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1527_cast_fp16")]; + bool var_1528_transpose_x_0 = const()[name = string("op_1528_transpose_x_0"), val = bool(false)]; + bool var_1528_transpose_y_0 = const()[name = string("op_1528_transpose_y_0"), val = bool(false)]; + tensor var_1528_cast_fp16 = matmul(transpose_x = var_1528_transpose_x_0, transpose_y = var_1528_transpose_y_0, x = var_1527_cast_fp16, y = v_new_65_cast_fp16)[name = string("op_1528_cast_fp16")]; + tensor var_1529_cast_fp16 = add(x = var_1525_cast_fp16, y = var_1528_cast_fp16)[name = string("op_1529_cast_fp16")]; + tensor var_1531_begin_0 = const()[name = string("op_1531_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1531_end_0 = const()[name = string("op_1531_end_0"), val = tensor([32, 1, 1, 1])]; + tensor var_1531_end_mask_0 = const()[name = string("op_1531_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1531_squeeze_mask_0 = const()[name = string("op_1531_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1531_cast_fp16 = slice_by_index(begin = var_1531_begin_0, end = var_1531_end_0, end_mask = var_1531_end_mask_0, squeeze_mask = var_1531_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1531_cast_fp16")]; + tensor var_1532_cast_fp16 = mul(x = state_1_to_fp16, y = var_1531_cast_fp16)[name = string("op_1532_cast_fp16")]; + tensor var_1534_begin_0 = const()[name = string("op_1534_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1534_end_0 = const()[name = string("op_1534_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_1534_end_mask_0 = const()[name = string("op_1534_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1534_squeeze_mask_0 = const()[name = string("op_1534_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1534_cast_fp16 = slice_by_index(begin = var_1534_begin_0, end = var_1534_end_0, end_mask = var_1534_end_mask_0, squeeze_mask = var_1534_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1534_cast_fp16")]; + bool var_1536_transpose_x_1 = const()[name = string("op_1536_transpose_x_1"), val = bool(true)]; + bool var_1536_transpose_y_1 = const()[name = string("op_1536_transpose_y_1"), val = bool(false)]; + tensor var_1536_cast_fp16 = matmul(transpose_x = var_1536_transpose_x_1, transpose_y = var_1536_transpose_y_1, x = var_1534_cast_fp16, y = v_new_65_cast_fp16)[name = string("op_1536_cast_fp16")]; + tensor state_67_cast_fp16 = add(x = var_1532_cast_fp16, y = var_1536_cast_fp16)[name = string("state_67_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = string("op_1539_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1539_end_0 = const()[name = string("op_1539_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_1539_end_mask_0 = const()[name = string("op_1539_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1539_squeeze_mask_0 = const()[name = string("op_1539_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1539_cast_fp16")]; + tensor var_1541_begin_0 = const()[name = string("op_1541_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1541_end_0 = const()[name = string("op_1541_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_1541_end_mask_0 = const()[name = string("op_1541_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1541_squeeze_mask_0 = const()[name = string("op_1541_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1541_cast_fp16 = slice_by_index(begin = var_1541_begin_0, end = var_1541_end_0, end_mask = var_1541_end_mask_0, squeeze_mask = var_1541_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1541_cast_fp16")]; + bool var_1542_transpose_x_0 = const()[name = string("op_1542_transpose_x_0"), val = bool(false)]; + bool var_1542_transpose_y_0 = const()[name = string("op_1542_transpose_y_0"), val = bool(false)]; + tensor var_1542_cast_fp16 = matmul(transpose_x = var_1542_transpose_x_0, transpose_y = var_1542_transpose_y_0, x = var_1541_cast_fp16, y = state_67_cast_fp16)[name = string("op_1542_cast_fp16")]; + tensor v_new_67_cast_fp16 = sub(x = var_1539_cast_fp16, y = var_1542_cast_fp16)[name = string("v_new_67_cast_fp16")]; + tensor var_1545_begin_0 = const()[name = string("op_1545_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1545_end_0 = const()[name = string("op_1545_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_1545_end_mask_0 = const()[name = string("op_1545_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1545_squeeze_mask_0 = const()[name = string("op_1545_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1545_cast_fp16 = slice_by_index(begin = var_1545_begin_0, end = var_1545_end_0, end_mask = var_1545_end_mask_0, squeeze_mask = var_1545_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1545_cast_fp16")]; + bool var_1546_transpose_x_0 = const()[name = string("op_1546_transpose_x_0"), val = bool(false)]; + bool var_1546_transpose_y_0 = const()[name = string("op_1546_transpose_y_0"), val = bool(false)]; + tensor var_1546_cast_fp16 = matmul(transpose_x = var_1546_transpose_x_0, transpose_y = var_1546_transpose_y_0, x = var_1545_cast_fp16, y = state_67_cast_fp16)[name = string("op_1546_cast_fp16")]; + tensor var_1548_begin_0 = const()[name = string("op_1548_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1548_end_0 = const()[name = string("op_1548_end_0"), val = tensor([32, 2, 64, 64])]; + tensor var_1548_end_mask_0 = const()[name = string("op_1548_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1548_squeeze_mask_0 = const()[name = string("op_1548_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1548_cast_fp16 = slice_by_index(begin = var_1548_begin_0, end = var_1548_end_0, end_mask = var_1548_end_mask_0, squeeze_mask = var_1548_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1548_cast_fp16")]; + bool var_1549_transpose_x_0 = const()[name = string("op_1549_transpose_x_0"), val = bool(false)]; + bool var_1549_transpose_y_0 = const()[name = string("op_1549_transpose_y_0"), val = bool(false)]; + tensor var_1549_cast_fp16 = matmul(transpose_x = var_1549_transpose_x_0, transpose_y = var_1549_transpose_y_0, x = var_1548_cast_fp16, y = v_new_67_cast_fp16)[name = string("op_1549_cast_fp16")]; + tensor var_1550_cast_fp16 = add(x = var_1546_cast_fp16, y = var_1549_cast_fp16)[name = string("op_1550_cast_fp16")]; + tensor var_1552_begin_0 = const()[name = string("op_1552_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1552_end_0 = const()[name = string("op_1552_end_0"), val = tensor([32, 2, 1, 1])]; + tensor var_1552_end_mask_0 = const()[name = string("op_1552_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1552_squeeze_mask_0 = const()[name = string("op_1552_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1552_cast_fp16 = slice_by_index(begin = var_1552_begin_0, end = var_1552_end_0, end_mask = var_1552_end_mask_0, squeeze_mask = var_1552_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1552_cast_fp16")]; + tensor var_1553_cast_fp16 = mul(x = state_67_cast_fp16, y = var_1552_cast_fp16)[name = string("op_1553_cast_fp16")]; + tensor var_1555_begin_0 = const()[name = string("op_1555_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1555_end_0 = const()[name = string("op_1555_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_1555_end_mask_0 = const()[name = string("op_1555_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1555_squeeze_mask_0 = const()[name = string("op_1555_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1555_cast_fp16 = slice_by_index(begin = var_1555_begin_0, end = var_1555_end_0, end_mask = var_1555_end_mask_0, squeeze_mask = var_1555_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1555_cast_fp16")]; + bool var_1557_transpose_x_1 = const()[name = string("op_1557_transpose_x_1"), val = bool(true)]; + bool var_1557_transpose_y_1 = const()[name = string("op_1557_transpose_y_1"), val = bool(false)]; + tensor var_1557_cast_fp16 = matmul(transpose_x = var_1557_transpose_x_1, transpose_y = var_1557_transpose_y_1, x = var_1555_cast_fp16, y = v_new_67_cast_fp16)[name = string("op_1557_cast_fp16")]; + tensor state_69_cast_fp16 = add(x = var_1553_cast_fp16, y = var_1557_cast_fp16)[name = string("state_69_cast_fp16")]; + tensor var_1560_begin_0 = const()[name = string("op_1560_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1560_end_0 = const()[name = string("op_1560_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_1560_end_mask_0 = const()[name = string("op_1560_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1560_squeeze_mask_0 = const()[name = string("op_1560_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1560_cast_fp16 = slice_by_index(begin = var_1560_begin_0, end = var_1560_end_0, end_mask = var_1560_end_mask_0, squeeze_mask = var_1560_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1560_cast_fp16")]; + tensor var_1562_begin_0 = const()[name = string("op_1562_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1562_end_0 = const()[name = string("op_1562_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_1562_end_mask_0 = const()[name = string("op_1562_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1562_squeeze_mask_0 = const()[name = string("op_1562_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1562_cast_fp16 = slice_by_index(begin = var_1562_begin_0, end = var_1562_end_0, end_mask = var_1562_end_mask_0, squeeze_mask = var_1562_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1562_cast_fp16")]; + bool var_1563_transpose_x_0 = const()[name = string("op_1563_transpose_x_0"), val = bool(false)]; + bool var_1563_transpose_y_0 = const()[name = string("op_1563_transpose_y_0"), val = bool(false)]; + tensor var_1563_cast_fp16 = matmul(transpose_x = var_1563_transpose_x_0, transpose_y = var_1563_transpose_y_0, x = var_1562_cast_fp16, y = state_69_cast_fp16)[name = string("op_1563_cast_fp16")]; + tensor v_new_69_cast_fp16 = sub(x = var_1560_cast_fp16, y = var_1563_cast_fp16)[name = string("v_new_69_cast_fp16")]; + tensor var_1566_begin_0 = const()[name = string("op_1566_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1566_end_0 = const()[name = string("op_1566_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_1566_end_mask_0 = const()[name = string("op_1566_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1566_squeeze_mask_0 = const()[name = string("op_1566_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1566_cast_fp16 = slice_by_index(begin = var_1566_begin_0, end = var_1566_end_0, end_mask = var_1566_end_mask_0, squeeze_mask = var_1566_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1566_cast_fp16")]; + bool var_1567_transpose_x_0 = const()[name = string("op_1567_transpose_x_0"), val = bool(false)]; + bool var_1567_transpose_y_0 = const()[name = string("op_1567_transpose_y_0"), val = bool(false)]; + tensor var_1567_cast_fp16 = matmul(transpose_x = var_1567_transpose_x_0, transpose_y = var_1567_transpose_y_0, x = var_1566_cast_fp16, y = state_69_cast_fp16)[name = string("op_1567_cast_fp16")]; + tensor var_1569_begin_0 = const()[name = string("op_1569_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1569_end_0 = const()[name = string("op_1569_end_0"), val = tensor([32, 3, 64, 64])]; + tensor var_1569_end_mask_0 = const()[name = string("op_1569_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1569_squeeze_mask_0 = const()[name = string("op_1569_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1569_cast_fp16 = slice_by_index(begin = var_1569_begin_0, end = var_1569_end_0, end_mask = var_1569_end_mask_0, squeeze_mask = var_1569_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1569_cast_fp16")]; + bool var_1570_transpose_x_0 = const()[name = string("op_1570_transpose_x_0"), val = bool(false)]; + bool var_1570_transpose_y_0 = const()[name = string("op_1570_transpose_y_0"), val = bool(false)]; + tensor var_1570_cast_fp16 = matmul(transpose_x = var_1570_transpose_x_0, transpose_y = var_1570_transpose_y_0, x = var_1569_cast_fp16, y = v_new_69_cast_fp16)[name = string("op_1570_cast_fp16")]; + tensor var_1571_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1570_cast_fp16)[name = string("op_1571_cast_fp16")]; + tensor var_1573_begin_0 = const()[name = string("op_1573_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1573_end_0 = const()[name = string("op_1573_end_0"), val = tensor([32, 3, 1, 1])]; + tensor var_1573_end_mask_0 = const()[name = string("op_1573_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1573_squeeze_mask_0 = const()[name = string("op_1573_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1573_cast_fp16 = slice_by_index(begin = var_1573_begin_0, end = var_1573_end_0, end_mask = var_1573_end_mask_0, squeeze_mask = var_1573_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1573_cast_fp16")]; + tensor var_1574_cast_fp16 = mul(x = state_69_cast_fp16, y = var_1573_cast_fp16)[name = string("op_1574_cast_fp16")]; + tensor var_1576_begin_0 = const()[name = string("op_1576_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1576_end_0 = const()[name = string("op_1576_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_1576_end_mask_0 = const()[name = string("op_1576_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1576_squeeze_mask_0 = const()[name = string("op_1576_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1576_cast_fp16 = slice_by_index(begin = var_1576_begin_0, end = var_1576_end_0, end_mask = var_1576_end_mask_0, squeeze_mask = var_1576_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1576_cast_fp16")]; + bool var_1578_transpose_x_1 = const()[name = string("op_1578_transpose_x_1"), val = bool(true)]; + bool var_1578_transpose_y_1 = const()[name = string("op_1578_transpose_y_1"), val = bool(false)]; + tensor var_1578_cast_fp16 = matmul(transpose_x = var_1578_transpose_x_1, transpose_y = var_1578_transpose_y_1, x = var_1576_cast_fp16, y = v_new_69_cast_fp16)[name = string("op_1578_cast_fp16")]; + tensor state_71_cast_fp16 = add(x = var_1574_cast_fp16, y = var_1578_cast_fp16)[name = string("state_71_cast_fp16")]; + tensor var_1581_begin_0 = const()[name = string("op_1581_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1581_end_0 = const()[name = string("op_1581_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_1581_end_mask_0 = const()[name = string("op_1581_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1581_squeeze_mask_0 = const()[name = string("op_1581_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1581_cast_fp16 = slice_by_index(begin = var_1581_begin_0, end = var_1581_end_0, end_mask = var_1581_end_mask_0, squeeze_mask = var_1581_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1581_cast_fp16")]; + tensor var_1583_begin_0 = const()[name = string("op_1583_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1583_end_0 = const()[name = string("op_1583_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_1583_end_mask_0 = const()[name = string("op_1583_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1583_squeeze_mask_0 = const()[name = string("op_1583_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1583_cast_fp16 = slice_by_index(begin = var_1583_begin_0, end = var_1583_end_0, end_mask = var_1583_end_mask_0, squeeze_mask = var_1583_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1583_cast_fp16")]; + bool var_1584_transpose_x_0 = const()[name = string("op_1584_transpose_x_0"), val = bool(false)]; + bool var_1584_transpose_y_0 = const()[name = string("op_1584_transpose_y_0"), val = bool(false)]; + tensor var_1584_cast_fp16 = matmul(transpose_x = var_1584_transpose_x_0, transpose_y = var_1584_transpose_y_0, x = var_1583_cast_fp16, y = state_71_cast_fp16)[name = string("op_1584_cast_fp16")]; + tensor v_new_71_cast_fp16 = sub(x = var_1581_cast_fp16, y = var_1584_cast_fp16)[name = string("v_new_71_cast_fp16")]; + tensor var_1587_begin_0 = const()[name = string("op_1587_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1587_end_0 = const()[name = string("op_1587_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_1587_end_mask_0 = const()[name = string("op_1587_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1587_squeeze_mask_0 = const()[name = string("op_1587_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1587_cast_fp16 = slice_by_index(begin = var_1587_begin_0, end = var_1587_end_0, end_mask = var_1587_end_mask_0, squeeze_mask = var_1587_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1587_cast_fp16")]; + bool var_1588_transpose_x_0 = const()[name = string("op_1588_transpose_x_0"), val = bool(false)]; + bool var_1588_transpose_y_0 = const()[name = string("op_1588_transpose_y_0"), val = bool(false)]; + tensor var_1588_cast_fp16 = matmul(transpose_x = var_1588_transpose_x_0, transpose_y = var_1588_transpose_y_0, x = var_1587_cast_fp16, y = state_71_cast_fp16)[name = string("op_1588_cast_fp16")]; + tensor var_1590_begin_0 = const()[name = string("op_1590_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1590_end_0 = const()[name = string("op_1590_end_0"), val = tensor([32, 4, 64, 64])]; + tensor var_1590_end_mask_0 = const()[name = string("op_1590_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1590_squeeze_mask_0 = const()[name = string("op_1590_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1590_cast_fp16 = slice_by_index(begin = var_1590_begin_0, end = var_1590_end_0, end_mask = var_1590_end_mask_0, squeeze_mask = var_1590_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1590_cast_fp16")]; + bool var_1591_transpose_x_0 = const()[name = string("op_1591_transpose_x_0"), val = bool(false)]; + bool var_1591_transpose_y_0 = const()[name = string("op_1591_transpose_y_0"), val = bool(false)]; + tensor var_1591_cast_fp16 = matmul(transpose_x = var_1591_transpose_x_0, transpose_y = var_1591_transpose_y_0, x = var_1590_cast_fp16, y = v_new_71_cast_fp16)[name = string("op_1591_cast_fp16")]; + tensor var_1592_cast_fp16 = add(x = var_1588_cast_fp16, y = var_1591_cast_fp16)[name = string("op_1592_cast_fp16")]; + tensor var_1594_begin_0 = const()[name = string("op_1594_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1594_end_0 = const()[name = string("op_1594_end_0"), val = tensor([32, 4, 1, 1])]; + tensor var_1594_end_mask_0 = const()[name = string("op_1594_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1594_squeeze_mask_0 = const()[name = string("op_1594_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1594_cast_fp16 = slice_by_index(begin = var_1594_begin_0, end = var_1594_end_0, end_mask = var_1594_end_mask_0, squeeze_mask = var_1594_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1594_cast_fp16")]; + tensor var_1595_cast_fp16 = mul(x = state_71_cast_fp16, y = var_1594_cast_fp16)[name = string("op_1595_cast_fp16")]; + tensor var_1597_begin_0 = const()[name = string("op_1597_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1597_end_0 = const()[name = string("op_1597_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_1597_end_mask_0 = const()[name = string("op_1597_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1597_squeeze_mask_0 = const()[name = string("op_1597_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1597_cast_fp16 = slice_by_index(begin = var_1597_begin_0, end = var_1597_end_0, end_mask = var_1597_end_mask_0, squeeze_mask = var_1597_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1597_cast_fp16")]; + bool var_1599_transpose_x_1 = const()[name = string("op_1599_transpose_x_1"), val = bool(true)]; + bool var_1599_transpose_y_1 = const()[name = string("op_1599_transpose_y_1"), val = bool(false)]; + tensor var_1599_cast_fp16 = matmul(transpose_x = var_1599_transpose_x_1, transpose_y = var_1599_transpose_y_1, x = var_1597_cast_fp16, y = v_new_71_cast_fp16)[name = string("op_1599_cast_fp16")]; + tensor state_73_cast_fp16 = add(x = var_1595_cast_fp16, y = var_1599_cast_fp16)[name = string("state_73_cast_fp16")]; + tensor var_1602_begin_0 = const()[name = string("op_1602_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1602_end_0 = const()[name = string("op_1602_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_1602_end_mask_0 = const()[name = string("op_1602_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1602_squeeze_mask_0 = const()[name = string("op_1602_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1602_cast_fp16 = slice_by_index(begin = var_1602_begin_0, end = var_1602_end_0, end_mask = var_1602_end_mask_0, squeeze_mask = var_1602_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1602_cast_fp16")]; + tensor var_1604_begin_0 = const()[name = string("op_1604_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1604_end_0 = const()[name = string("op_1604_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_1604_end_mask_0 = const()[name = string("op_1604_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1604_squeeze_mask_0 = const()[name = string("op_1604_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1604_cast_fp16 = slice_by_index(begin = var_1604_begin_0, end = var_1604_end_0, end_mask = var_1604_end_mask_0, squeeze_mask = var_1604_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1604_cast_fp16")]; + bool var_1605_transpose_x_0 = const()[name = string("op_1605_transpose_x_0"), val = bool(false)]; + bool var_1605_transpose_y_0 = const()[name = string("op_1605_transpose_y_0"), val = bool(false)]; + tensor var_1605_cast_fp16 = matmul(transpose_x = var_1605_transpose_x_0, transpose_y = var_1605_transpose_y_0, x = var_1604_cast_fp16, y = state_73_cast_fp16)[name = string("op_1605_cast_fp16")]; + tensor v_new_73_cast_fp16 = sub(x = var_1602_cast_fp16, y = var_1605_cast_fp16)[name = string("v_new_73_cast_fp16")]; + tensor var_1608_begin_0 = const()[name = string("op_1608_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1608_end_0 = const()[name = string("op_1608_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_1608_end_mask_0 = const()[name = string("op_1608_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1608_squeeze_mask_0 = const()[name = string("op_1608_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1608_cast_fp16 = slice_by_index(begin = var_1608_begin_0, end = var_1608_end_0, end_mask = var_1608_end_mask_0, squeeze_mask = var_1608_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1608_cast_fp16")]; + bool var_1609_transpose_x_0 = const()[name = string("op_1609_transpose_x_0"), val = bool(false)]; + bool var_1609_transpose_y_0 = const()[name = string("op_1609_transpose_y_0"), val = bool(false)]; + tensor var_1609_cast_fp16 = matmul(transpose_x = var_1609_transpose_x_0, transpose_y = var_1609_transpose_y_0, x = var_1608_cast_fp16, y = state_73_cast_fp16)[name = string("op_1609_cast_fp16")]; + tensor var_1611_begin_0 = const()[name = string("op_1611_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1611_end_0 = const()[name = string("op_1611_end_0"), val = tensor([32, 5, 64, 64])]; + tensor var_1611_end_mask_0 = const()[name = string("op_1611_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1611_squeeze_mask_0 = const()[name = string("op_1611_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1611_cast_fp16 = slice_by_index(begin = var_1611_begin_0, end = var_1611_end_0, end_mask = var_1611_end_mask_0, squeeze_mask = var_1611_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1611_cast_fp16")]; + bool var_1612_transpose_x_0 = const()[name = string("op_1612_transpose_x_0"), val = bool(false)]; + bool var_1612_transpose_y_0 = const()[name = string("op_1612_transpose_y_0"), val = bool(false)]; + tensor var_1612_cast_fp16 = matmul(transpose_x = var_1612_transpose_x_0, transpose_y = var_1612_transpose_y_0, x = var_1611_cast_fp16, y = v_new_73_cast_fp16)[name = string("op_1612_cast_fp16")]; + tensor var_1613_cast_fp16 = add(x = var_1609_cast_fp16, y = var_1612_cast_fp16)[name = string("op_1613_cast_fp16")]; + tensor var_1615_begin_0 = const()[name = string("op_1615_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1615_end_0 = const()[name = string("op_1615_end_0"), val = tensor([32, 5, 1, 1])]; + tensor var_1615_end_mask_0 = const()[name = string("op_1615_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1615_squeeze_mask_0 = const()[name = string("op_1615_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1615_cast_fp16 = slice_by_index(begin = var_1615_begin_0, end = var_1615_end_0, end_mask = var_1615_end_mask_0, squeeze_mask = var_1615_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1615_cast_fp16")]; + tensor var_1616_cast_fp16 = mul(x = state_73_cast_fp16, y = var_1615_cast_fp16)[name = string("op_1616_cast_fp16")]; + tensor var_1618_begin_0 = const()[name = string("op_1618_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1618_end_0 = const()[name = string("op_1618_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_1618_end_mask_0 = const()[name = string("op_1618_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1618_squeeze_mask_0 = const()[name = string("op_1618_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1618_cast_fp16 = slice_by_index(begin = var_1618_begin_0, end = var_1618_end_0, end_mask = var_1618_end_mask_0, squeeze_mask = var_1618_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1618_cast_fp16")]; + bool var_1620_transpose_x_1 = const()[name = string("op_1620_transpose_x_1"), val = bool(true)]; + bool var_1620_transpose_y_1 = const()[name = string("op_1620_transpose_y_1"), val = bool(false)]; + tensor var_1620_cast_fp16 = matmul(transpose_x = var_1620_transpose_x_1, transpose_y = var_1620_transpose_y_1, x = var_1618_cast_fp16, y = v_new_73_cast_fp16)[name = string("op_1620_cast_fp16")]; + tensor state_75_cast_fp16 = add(x = var_1616_cast_fp16, y = var_1620_cast_fp16)[name = string("state_75_cast_fp16")]; + tensor var_1623_begin_0 = const()[name = string("op_1623_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_1623_end_0 = const()[name = string("op_1623_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_1623_end_mask_0 = const()[name = string("op_1623_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1623_squeeze_mask_0 = const()[name = string("op_1623_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1623_cast_fp16 = slice_by_index(begin = var_1623_begin_0, end = var_1623_end_0, end_mask = var_1623_end_mask_0, squeeze_mask = var_1623_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1623_cast_fp16")]; + tensor var_1625_begin_0 = const()[name = string("op_1625_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_1625_end_0 = const()[name = string("op_1625_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_1625_end_mask_0 = const()[name = string("op_1625_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1625_squeeze_mask_0 = const()[name = string("op_1625_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1625_cast_fp16 = slice_by_index(begin = var_1625_begin_0, end = var_1625_end_0, end_mask = var_1625_end_mask_0, squeeze_mask = var_1625_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1625_cast_fp16")]; + bool var_1626_transpose_x_0 = const()[name = string("op_1626_transpose_x_0"), val = bool(false)]; + bool var_1626_transpose_y_0 = const()[name = string("op_1626_transpose_y_0"), val = bool(false)]; + tensor var_1626_cast_fp16 = matmul(transpose_x = var_1626_transpose_x_0, transpose_y = var_1626_transpose_y_0, x = var_1625_cast_fp16, y = state_75_cast_fp16)[name = string("op_1626_cast_fp16")]; + tensor v_new_75_cast_fp16 = sub(x = var_1623_cast_fp16, y = var_1626_cast_fp16)[name = string("v_new_75_cast_fp16")]; + tensor var_1629_begin_0 = const()[name = string("op_1629_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_1629_end_0 = const()[name = string("op_1629_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_1629_end_mask_0 = const()[name = string("op_1629_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1629_squeeze_mask_0 = const()[name = string("op_1629_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1629_cast_fp16 = slice_by_index(begin = var_1629_begin_0, end = var_1629_end_0, end_mask = var_1629_end_mask_0, squeeze_mask = var_1629_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1629_cast_fp16")]; + bool var_1630_transpose_x_0 = const()[name = string("op_1630_transpose_x_0"), val = bool(false)]; + bool var_1630_transpose_y_0 = const()[name = string("op_1630_transpose_y_0"), val = bool(false)]; + tensor var_1630_cast_fp16 = matmul(transpose_x = var_1630_transpose_x_0, transpose_y = var_1630_transpose_y_0, x = var_1629_cast_fp16, y = state_75_cast_fp16)[name = string("op_1630_cast_fp16")]; + tensor var_1632_begin_0 = const()[name = string("op_1632_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_1632_end_0 = const()[name = string("op_1632_end_0"), val = tensor([32, 6, 64, 64])]; + tensor var_1632_end_mask_0 = const()[name = string("op_1632_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1632_squeeze_mask_0 = const()[name = string("op_1632_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1632_cast_fp16 = slice_by_index(begin = var_1632_begin_0, end = var_1632_end_0, end_mask = var_1632_end_mask_0, squeeze_mask = var_1632_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1632_cast_fp16")]; + bool var_1633_transpose_x_0 = const()[name = string("op_1633_transpose_x_0"), val = bool(false)]; + bool var_1633_transpose_y_0 = const()[name = string("op_1633_transpose_y_0"), val = bool(false)]; + tensor var_1633_cast_fp16 = matmul(transpose_x = var_1633_transpose_x_0, transpose_y = var_1633_transpose_y_0, x = var_1632_cast_fp16, y = v_new_75_cast_fp16)[name = string("op_1633_cast_fp16")]; + tensor var_1634_cast_fp16 = add(x = var_1630_cast_fp16, y = var_1633_cast_fp16)[name = string("op_1634_cast_fp16")]; + tensor var_1636_begin_0 = const()[name = string("op_1636_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_1636_end_0 = const()[name = string("op_1636_end_0"), val = tensor([32, 6, 1, 1])]; + tensor var_1636_end_mask_0 = const()[name = string("op_1636_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1636_squeeze_mask_0 = const()[name = string("op_1636_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1636_cast_fp16 = slice_by_index(begin = var_1636_begin_0, end = var_1636_end_0, end_mask = var_1636_end_mask_0, squeeze_mask = var_1636_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1636_cast_fp16")]; + tensor var_1637_cast_fp16 = mul(x = state_75_cast_fp16, y = var_1636_cast_fp16)[name = string("op_1637_cast_fp16")]; + tensor var_1639_begin_0 = const()[name = string("op_1639_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_1639_end_0 = const()[name = string("op_1639_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_1639_end_mask_0 = const()[name = string("op_1639_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1639_squeeze_mask_0 = const()[name = string("op_1639_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1639_cast_fp16 = slice_by_index(begin = var_1639_begin_0, end = var_1639_end_0, end_mask = var_1639_end_mask_0, squeeze_mask = var_1639_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1639_cast_fp16")]; + bool var_1641_transpose_x_1 = const()[name = string("op_1641_transpose_x_1"), val = bool(true)]; + bool var_1641_transpose_y_1 = const()[name = string("op_1641_transpose_y_1"), val = bool(false)]; + tensor var_1641_cast_fp16 = matmul(transpose_x = var_1641_transpose_x_1, transpose_y = var_1641_transpose_y_1, x = var_1639_cast_fp16, y = v_new_75_cast_fp16)[name = string("op_1641_cast_fp16")]; + tensor state_77_cast_fp16 = add(x = var_1637_cast_fp16, y = var_1641_cast_fp16)[name = string("state_77_cast_fp16")]; + tensor var_1644_begin_0 = const()[name = string("op_1644_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1644_end_0 = const()[name = string("op_1644_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_1644_end_mask_0 = const()[name = string("op_1644_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1644_squeeze_mask_0 = const()[name = string("op_1644_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1644_cast_fp16 = slice_by_index(begin = var_1644_begin_0, end = var_1644_end_0, end_mask = var_1644_end_mask_0, squeeze_mask = var_1644_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1644_cast_fp16")]; + tensor var_1646_begin_0 = const()[name = string("op_1646_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1646_end_0 = const()[name = string("op_1646_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_1646_end_mask_0 = const()[name = string("op_1646_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1646_squeeze_mask_0 = const()[name = string("op_1646_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1646_cast_fp16 = slice_by_index(begin = var_1646_begin_0, end = var_1646_end_0, end_mask = var_1646_end_mask_0, squeeze_mask = var_1646_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1646_cast_fp16")]; + bool var_1647_transpose_x_0 = const()[name = string("op_1647_transpose_x_0"), val = bool(false)]; + bool var_1647_transpose_y_0 = const()[name = string("op_1647_transpose_y_0"), val = bool(false)]; + tensor var_1647_cast_fp16 = matmul(transpose_x = var_1647_transpose_x_0, transpose_y = var_1647_transpose_y_0, x = var_1646_cast_fp16, y = state_77_cast_fp16)[name = string("op_1647_cast_fp16")]; + tensor v_new_77_cast_fp16 = sub(x = var_1644_cast_fp16, y = var_1647_cast_fp16)[name = string("v_new_77_cast_fp16")]; + tensor var_1650_begin_0 = const()[name = string("op_1650_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1650_end_0 = const()[name = string("op_1650_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_1650_end_mask_0 = const()[name = string("op_1650_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1650_squeeze_mask_0 = const()[name = string("op_1650_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1650_cast_fp16 = slice_by_index(begin = var_1650_begin_0, end = var_1650_end_0, end_mask = var_1650_end_mask_0, squeeze_mask = var_1650_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1650_cast_fp16")]; + bool var_1651_transpose_x_0 = const()[name = string("op_1651_transpose_x_0"), val = bool(false)]; + bool var_1651_transpose_y_0 = const()[name = string("op_1651_transpose_y_0"), val = bool(false)]; + tensor var_1651_cast_fp16 = matmul(transpose_x = var_1651_transpose_x_0, transpose_y = var_1651_transpose_y_0, x = var_1650_cast_fp16, y = state_77_cast_fp16)[name = string("op_1651_cast_fp16")]; + tensor var_1653_begin_0 = const()[name = string("op_1653_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1653_end_0 = const()[name = string("op_1653_end_0"), val = tensor([32, 7, 64, 64])]; + tensor var_1653_end_mask_0 = const()[name = string("op_1653_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1653_squeeze_mask_0 = const()[name = string("op_1653_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, squeeze_mask = var_1653_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1653_cast_fp16")]; + bool var_1654_transpose_x_0 = const()[name = string("op_1654_transpose_x_0"), val = bool(false)]; + bool var_1654_transpose_y_0 = const()[name = string("op_1654_transpose_y_0"), val = bool(false)]; + tensor var_1654_cast_fp16 = matmul(transpose_x = var_1654_transpose_x_0, transpose_y = var_1654_transpose_y_0, x = var_1653_cast_fp16, y = v_new_77_cast_fp16)[name = string("op_1654_cast_fp16")]; + tensor var_1655_cast_fp16 = add(x = var_1651_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")]; + tensor var_1657_begin_0 = const()[name = string("op_1657_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1657_end_0 = const()[name = string("op_1657_end_0"), val = tensor([32, 7, 1, 1])]; + tensor var_1657_end_mask_0 = const()[name = string("op_1657_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1657_squeeze_mask_0 = const()[name = string("op_1657_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1657_cast_fp16 = slice_by_index(begin = var_1657_begin_0, end = var_1657_end_0, end_mask = var_1657_end_mask_0, squeeze_mask = var_1657_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1657_cast_fp16")]; + tensor var_1658_cast_fp16 = mul(x = state_77_cast_fp16, y = var_1657_cast_fp16)[name = string("op_1658_cast_fp16")]; + tensor var_1660_begin_0 = const()[name = string("op_1660_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1660_end_0 = const()[name = string("op_1660_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_1660_end_mask_0 = const()[name = string("op_1660_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1660_squeeze_mask_0 = const()[name = string("op_1660_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1660_cast_fp16 = slice_by_index(begin = var_1660_begin_0, end = var_1660_end_0, end_mask = var_1660_end_mask_0, squeeze_mask = var_1660_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1660_cast_fp16")]; + bool var_1662_transpose_x_1 = const()[name = string("op_1662_transpose_x_1"), val = bool(true)]; + bool var_1662_transpose_y_1 = const()[name = string("op_1662_transpose_y_1"), val = bool(false)]; + tensor var_1662_cast_fp16 = matmul(transpose_x = var_1662_transpose_x_1, transpose_y = var_1662_transpose_y_1, x = var_1660_cast_fp16, y = v_new_77_cast_fp16)[name = string("op_1662_cast_fp16")]; + tensor state_79_cast_fp16 = add(x = var_1658_cast_fp16, y = var_1662_cast_fp16)[name = string("state_79_cast_fp16")]; + tensor var_1665_begin_0 = const()[name = string("op_1665_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_1665_end_0 = const()[name = string("op_1665_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_1665_end_mask_0 = const()[name = string("op_1665_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1665_squeeze_mask_0 = const()[name = string("op_1665_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1665_cast_fp16 = slice_by_index(begin = var_1665_begin_0, end = var_1665_end_0, end_mask = var_1665_end_mask_0, squeeze_mask = var_1665_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1665_cast_fp16")]; + tensor var_1667_begin_0 = const()[name = string("op_1667_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_1667_end_0 = const()[name = string("op_1667_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_1667_end_mask_0 = const()[name = string("op_1667_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1667_squeeze_mask_0 = const()[name = string("op_1667_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1667_cast_fp16 = slice_by_index(begin = var_1667_begin_0, end = var_1667_end_0, end_mask = var_1667_end_mask_0, squeeze_mask = var_1667_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1667_cast_fp16")]; + bool var_1668_transpose_x_0 = const()[name = string("op_1668_transpose_x_0"), val = bool(false)]; + bool var_1668_transpose_y_0 = const()[name = string("op_1668_transpose_y_0"), val = bool(false)]; + tensor var_1668_cast_fp16 = matmul(transpose_x = var_1668_transpose_x_0, transpose_y = var_1668_transpose_y_0, x = var_1667_cast_fp16, y = state_79_cast_fp16)[name = string("op_1668_cast_fp16")]; + tensor v_new_79_cast_fp16 = sub(x = var_1665_cast_fp16, y = var_1668_cast_fp16)[name = string("v_new_79_cast_fp16")]; + tensor var_1671_begin_0 = const()[name = string("op_1671_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_1671_end_0 = const()[name = string("op_1671_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_1671_end_mask_0 = const()[name = string("op_1671_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1671_squeeze_mask_0 = const()[name = string("op_1671_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1671_cast_fp16 = slice_by_index(begin = var_1671_begin_0, end = var_1671_end_0, end_mask = var_1671_end_mask_0, squeeze_mask = var_1671_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1671_cast_fp16")]; + bool var_1672_transpose_x_0 = const()[name = string("op_1672_transpose_x_0"), val = bool(false)]; + bool var_1672_transpose_y_0 = const()[name = string("op_1672_transpose_y_0"), val = bool(false)]; + tensor var_1672_cast_fp16 = matmul(transpose_x = var_1672_transpose_x_0, transpose_y = var_1672_transpose_y_0, x = var_1671_cast_fp16, y = state_79_cast_fp16)[name = string("op_1672_cast_fp16")]; + tensor var_1674_begin_0 = const()[name = string("op_1674_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_1674_end_0 = const()[name = string("op_1674_end_0"), val = tensor([32, 8, 64, 64])]; + tensor var_1674_end_mask_0 = const()[name = string("op_1674_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1674_squeeze_mask_0 = const()[name = string("op_1674_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1674_cast_fp16 = slice_by_index(begin = var_1674_begin_0, end = var_1674_end_0, end_mask = var_1674_end_mask_0, squeeze_mask = var_1674_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1674_cast_fp16")]; + bool var_1675_transpose_x_0 = const()[name = string("op_1675_transpose_x_0"), val = bool(false)]; + bool var_1675_transpose_y_0 = const()[name = string("op_1675_transpose_y_0"), val = bool(false)]; + tensor var_1675_cast_fp16 = matmul(transpose_x = var_1675_transpose_x_0, transpose_y = var_1675_transpose_y_0, x = var_1674_cast_fp16, y = v_new_79_cast_fp16)[name = string("op_1675_cast_fp16")]; + tensor var_1676_cast_fp16 = add(x = var_1672_cast_fp16, y = var_1675_cast_fp16)[name = string("op_1676_cast_fp16")]; + tensor var_1678_begin_0 = const()[name = string("op_1678_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_1678_end_0 = const()[name = string("op_1678_end_0"), val = tensor([32, 8, 1, 1])]; + tensor var_1678_end_mask_0 = const()[name = string("op_1678_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1678_squeeze_mask_0 = const()[name = string("op_1678_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1678_cast_fp16 = slice_by_index(begin = var_1678_begin_0, end = var_1678_end_0, end_mask = var_1678_end_mask_0, squeeze_mask = var_1678_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1678_cast_fp16")]; + tensor var_1679_cast_fp16 = mul(x = state_79_cast_fp16, y = var_1678_cast_fp16)[name = string("op_1679_cast_fp16")]; + tensor var_1681_begin_0 = const()[name = string("op_1681_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_1681_end_0 = const()[name = string("op_1681_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_1681_end_mask_0 = const()[name = string("op_1681_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1681_squeeze_mask_0 = const()[name = string("op_1681_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1681_cast_fp16 = slice_by_index(begin = var_1681_begin_0, end = var_1681_end_0, end_mask = var_1681_end_mask_0, squeeze_mask = var_1681_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1681_cast_fp16")]; + bool var_1683_transpose_x_1 = const()[name = string("op_1683_transpose_x_1"), val = bool(true)]; + bool var_1683_transpose_y_1 = const()[name = string("op_1683_transpose_y_1"), val = bool(false)]; + tensor var_1683_cast_fp16 = matmul(transpose_x = var_1683_transpose_x_1, transpose_y = var_1683_transpose_y_1, x = var_1681_cast_fp16, y = v_new_79_cast_fp16)[name = string("op_1683_cast_fp16")]; + tensor state_81_cast_fp16 = add(x = var_1679_cast_fp16, y = var_1683_cast_fp16)[name = string("state_81_cast_fp16")]; + tensor var_1686_begin_0 = const()[name = string("op_1686_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_1686_end_0 = const()[name = string("op_1686_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_1686_end_mask_0 = const()[name = string("op_1686_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1686_squeeze_mask_0 = const()[name = string("op_1686_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1686_cast_fp16 = slice_by_index(begin = var_1686_begin_0, end = var_1686_end_0, end_mask = var_1686_end_mask_0, squeeze_mask = var_1686_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1686_cast_fp16")]; + tensor var_1688_begin_0 = const()[name = string("op_1688_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_1688_end_0 = const()[name = string("op_1688_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_1688_end_mask_0 = const()[name = string("op_1688_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1688_squeeze_mask_0 = const()[name = string("op_1688_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1688_cast_fp16 = slice_by_index(begin = var_1688_begin_0, end = var_1688_end_0, end_mask = var_1688_end_mask_0, squeeze_mask = var_1688_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1688_cast_fp16")]; + bool var_1689_transpose_x_0 = const()[name = string("op_1689_transpose_x_0"), val = bool(false)]; + bool var_1689_transpose_y_0 = const()[name = string("op_1689_transpose_y_0"), val = bool(false)]; + tensor var_1689_cast_fp16 = matmul(transpose_x = var_1689_transpose_x_0, transpose_y = var_1689_transpose_y_0, x = var_1688_cast_fp16, y = state_81_cast_fp16)[name = string("op_1689_cast_fp16")]; + tensor v_new_81_cast_fp16 = sub(x = var_1686_cast_fp16, y = var_1689_cast_fp16)[name = string("v_new_81_cast_fp16")]; + tensor var_1692_begin_0 = const()[name = string("op_1692_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_1692_end_0 = const()[name = string("op_1692_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_1692_end_mask_0 = const()[name = string("op_1692_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1692_squeeze_mask_0 = const()[name = string("op_1692_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1692_cast_fp16 = slice_by_index(begin = var_1692_begin_0, end = var_1692_end_0, end_mask = var_1692_end_mask_0, squeeze_mask = var_1692_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1692_cast_fp16")]; + bool var_1693_transpose_x_0 = const()[name = string("op_1693_transpose_x_0"), val = bool(false)]; + bool var_1693_transpose_y_0 = const()[name = string("op_1693_transpose_y_0"), val = bool(false)]; + tensor var_1693_cast_fp16 = matmul(transpose_x = var_1693_transpose_x_0, transpose_y = var_1693_transpose_y_0, x = var_1692_cast_fp16, y = state_81_cast_fp16)[name = string("op_1693_cast_fp16")]; + tensor var_1695_begin_0 = const()[name = string("op_1695_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_1695_end_0 = const()[name = string("op_1695_end_0"), val = tensor([32, 9, 64, 64])]; + tensor var_1695_end_mask_0 = const()[name = string("op_1695_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1695_squeeze_mask_0 = const()[name = string("op_1695_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1695_cast_fp16 = slice_by_index(begin = var_1695_begin_0, end = var_1695_end_0, end_mask = var_1695_end_mask_0, squeeze_mask = var_1695_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1695_cast_fp16")]; + bool var_1696_transpose_x_0 = const()[name = string("op_1696_transpose_x_0"), val = bool(false)]; + bool var_1696_transpose_y_0 = const()[name = string("op_1696_transpose_y_0"), val = bool(false)]; + tensor var_1696_cast_fp16 = matmul(transpose_x = var_1696_transpose_x_0, transpose_y = var_1696_transpose_y_0, x = var_1695_cast_fp16, y = v_new_81_cast_fp16)[name = string("op_1696_cast_fp16")]; + tensor var_1697_cast_fp16 = add(x = var_1693_cast_fp16, y = var_1696_cast_fp16)[name = string("op_1697_cast_fp16")]; + tensor var_1699_begin_0 = const()[name = string("op_1699_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_1699_end_0 = const()[name = string("op_1699_end_0"), val = tensor([32, 9, 1, 1])]; + tensor var_1699_end_mask_0 = const()[name = string("op_1699_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1699_squeeze_mask_0 = const()[name = string("op_1699_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1699_cast_fp16 = slice_by_index(begin = var_1699_begin_0, end = var_1699_end_0, end_mask = var_1699_end_mask_0, squeeze_mask = var_1699_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1699_cast_fp16")]; + tensor var_1700_cast_fp16 = mul(x = state_81_cast_fp16, y = var_1699_cast_fp16)[name = string("op_1700_cast_fp16")]; + tensor var_1702_begin_0 = const()[name = string("op_1702_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_1702_end_0 = const()[name = string("op_1702_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_1702_end_mask_0 = const()[name = string("op_1702_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1702_squeeze_mask_0 = const()[name = string("op_1702_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1702_cast_fp16 = slice_by_index(begin = var_1702_begin_0, end = var_1702_end_0, end_mask = var_1702_end_mask_0, squeeze_mask = var_1702_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1702_cast_fp16")]; + bool var_1704_transpose_x_1 = const()[name = string("op_1704_transpose_x_1"), val = bool(true)]; + bool var_1704_transpose_y_1 = const()[name = string("op_1704_transpose_y_1"), val = bool(false)]; + tensor var_1704_cast_fp16 = matmul(transpose_x = var_1704_transpose_x_1, transpose_y = var_1704_transpose_y_1, x = var_1702_cast_fp16, y = v_new_81_cast_fp16)[name = string("op_1704_cast_fp16")]; + tensor state_83_cast_fp16 = add(x = var_1700_cast_fp16, y = var_1704_cast_fp16)[name = string("state_83_cast_fp16")]; + tensor var_1707_begin_0 = const()[name = string("op_1707_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_1707_end_0 = const()[name = string("op_1707_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_1707_end_mask_0 = const()[name = string("op_1707_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1707_squeeze_mask_0 = const()[name = string("op_1707_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1707_cast_fp16 = slice_by_index(begin = var_1707_begin_0, end = var_1707_end_0, end_mask = var_1707_end_mask_0, squeeze_mask = var_1707_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1707_cast_fp16")]; + tensor var_1709_begin_0 = const()[name = string("op_1709_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_1709_end_0 = const()[name = string("op_1709_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_1709_end_mask_0 = const()[name = string("op_1709_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1709_squeeze_mask_0 = const()[name = string("op_1709_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1709_cast_fp16 = slice_by_index(begin = var_1709_begin_0, end = var_1709_end_0, end_mask = var_1709_end_mask_0, squeeze_mask = var_1709_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1709_cast_fp16")]; + bool var_1710_transpose_x_0 = const()[name = string("op_1710_transpose_x_0"), val = bool(false)]; + bool var_1710_transpose_y_0 = const()[name = string("op_1710_transpose_y_0"), val = bool(false)]; + tensor var_1710_cast_fp16 = matmul(transpose_x = var_1710_transpose_x_0, transpose_y = var_1710_transpose_y_0, x = var_1709_cast_fp16, y = state_83_cast_fp16)[name = string("op_1710_cast_fp16")]; + tensor v_new_83_cast_fp16 = sub(x = var_1707_cast_fp16, y = var_1710_cast_fp16)[name = string("v_new_83_cast_fp16")]; + tensor var_1713_begin_0 = const()[name = string("op_1713_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_1713_end_0 = const()[name = string("op_1713_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_1713_end_mask_0 = const()[name = string("op_1713_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1713_squeeze_mask_0 = const()[name = string("op_1713_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1713_cast_fp16 = slice_by_index(begin = var_1713_begin_0, end = var_1713_end_0, end_mask = var_1713_end_mask_0, squeeze_mask = var_1713_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1713_cast_fp16")]; + bool var_1714_transpose_x_0 = const()[name = string("op_1714_transpose_x_0"), val = bool(false)]; + bool var_1714_transpose_y_0 = const()[name = string("op_1714_transpose_y_0"), val = bool(false)]; + tensor var_1714_cast_fp16 = matmul(transpose_x = var_1714_transpose_x_0, transpose_y = var_1714_transpose_y_0, x = var_1713_cast_fp16, y = state_83_cast_fp16)[name = string("op_1714_cast_fp16")]; + tensor var_1716_begin_0 = const()[name = string("op_1716_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_1716_end_0 = const()[name = string("op_1716_end_0"), val = tensor([32, 10, 64, 64])]; + tensor var_1716_end_mask_0 = const()[name = string("op_1716_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1716_squeeze_mask_0 = const()[name = string("op_1716_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1716_cast_fp16 = slice_by_index(begin = var_1716_begin_0, end = var_1716_end_0, end_mask = var_1716_end_mask_0, squeeze_mask = var_1716_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1716_cast_fp16")]; + bool var_1717_transpose_x_0 = const()[name = string("op_1717_transpose_x_0"), val = bool(false)]; + bool var_1717_transpose_y_0 = const()[name = string("op_1717_transpose_y_0"), val = bool(false)]; + tensor var_1717_cast_fp16 = matmul(transpose_x = var_1717_transpose_x_0, transpose_y = var_1717_transpose_y_0, x = var_1716_cast_fp16, y = v_new_83_cast_fp16)[name = string("op_1717_cast_fp16")]; + tensor var_1718_cast_fp16 = add(x = var_1714_cast_fp16, y = var_1717_cast_fp16)[name = string("op_1718_cast_fp16")]; + tensor var_1720_begin_0 = const()[name = string("op_1720_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_1720_end_0 = const()[name = string("op_1720_end_0"), val = tensor([32, 10, 1, 1])]; + tensor var_1720_end_mask_0 = const()[name = string("op_1720_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1720_squeeze_mask_0 = const()[name = string("op_1720_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1720_cast_fp16 = slice_by_index(begin = var_1720_begin_0, end = var_1720_end_0, end_mask = var_1720_end_mask_0, squeeze_mask = var_1720_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1720_cast_fp16")]; + tensor var_1721_cast_fp16 = mul(x = state_83_cast_fp16, y = var_1720_cast_fp16)[name = string("op_1721_cast_fp16")]; + tensor var_1723_begin_0 = const()[name = string("op_1723_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_1723_end_0 = const()[name = string("op_1723_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_1723_end_mask_0 = const()[name = string("op_1723_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1723_squeeze_mask_0 = const()[name = string("op_1723_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1723_cast_fp16 = slice_by_index(begin = var_1723_begin_0, end = var_1723_end_0, end_mask = var_1723_end_mask_0, squeeze_mask = var_1723_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1723_cast_fp16")]; + bool var_1725_transpose_x_1 = const()[name = string("op_1725_transpose_x_1"), val = bool(true)]; + bool var_1725_transpose_y_1 = const()[name = string("op_1725_transpose_y_1"), val = bool(false)]; + tensor var_1725_cast_fp16 = matmul(transpose_x = var_1725_transpose_x_1, transpose_y = var_1725_transpose_y_1, x = var_1723_cast_fp16, y = v_new_83_cast_fp16)[name = string("op_1725_cast_fp16")]; + tensor state_85_cast_fp16 = add(x = var_1721_cast_fp16, y = var_1725_cast_fp16)[name = string("state_85_cast_fp16")]; + tensor var_1728_begin_0 = const()[name = string("op_1728_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_1728_end_0 = const()[name = string("op_1728_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_1728_end_mask_0 = const()[name = string("op_1728_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1728_squeeze_mask_0 = const()[name = string("op_1728_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1728_cast_fp16 = slice_by_index(begin = var_1728_begin_0, end = var_1728_end_0, end_mask = var_1728_end_mask_0, squeeze_mask = var_1728_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1728_cast_fp16")]; + tensor var_1730_begin_0 = const()[name = string("op_1730_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_1730_end_0 = const()[name = string("op_1730_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_1730_end_mask_0 = const()[name = string("op_1730_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1730_squeeze_mask_0 = const()[name = string("op_1730_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1730_cast_fp16 = slice_by_index(begin = var_1730_begin_0, end = var_1730_end_0, end_mask = var_1730_end_mask_0, squeeze_mask = var_1730_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1730_cast_fp16")]; + bool var_1731_transpose_x_0 = const()[name = string("op_1731_transpose_x_0"), val = bool(false)]; + bool var_1731_transpose_y_0 = const()[name = string("op_1731_transpose_y_0"), val = bool(false)]; + tensor var_1731_cast_fp16 = matmul(transpose_x = var_1731_transpose_x_0, transpose_y = var_1731_transpose_y_0, x = var_1730_cast_fp16, y = state_85_cast_fp16)[name = string("op_1731_cast_fp16")]; + tensor v_new_85_cast_fp16 = sub(x = var_1728_cast_fp16, y = var_1731_cast_fp16)[name = string("v_new_85_cast_fp16")]; + tensor var_1734_begin_0 = const()[name = string("op_1734_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_1734_end_0 = const()[name = string("op_1734_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_1734_end_mask_0 = const()[name = string("op_1734_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1734_squeeze_mask_0 = const()[name = string("op_1734_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1734_cast_fp16 = slice_by_index(begin = var_1734_begin_0, end = var_1734_end_0, end_mask = var_1734_end_mask_0, squeeze_mask = var_1734_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1734_cast_fp16")]; + bool var_1735_transpose_x_0 = const()[name = string("op_1735_transpose_x_0"), val = bool(false)]; + bool var_1735_transpose_y_0 = const()[name = string("op_1735_transpose_y_0"), val = bool(false)]; + tensor var_1735_cast_fp16 = matmul(transpose_x = var_1735_transpose_x_0, transpose_y = var_1735_transpose_y_0, x = var_1734_cast_fp16, y = state_85_cast_fp16)[name = string("op_1735_cast_fp16")]; + tensor var_1737_begin_0 = const()[name = string("op_1737_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_1737_end_0 = const()[name = string("op_1737_end_0"), val = tensor([32, 11, 64, 64])]; + tensor var_1737_end_mask_0 = const()[name = string("op_1737_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1737_squeeze_mask_0 = const()[name = string("op_1737_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1737_cast_fp16 = slice_by_index(begin = var_1737_begin_0, end = var_1737_end_0, end_mask = var_1737_end_mask_0, squeeze_mask = var_1737_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1737_cast_fp16")]; + bool var_1738_transpose_x_0 = const()[name = string("op_1738_transpose_x_0"), val = bool(false)]; + bool var_1738_transpose_y_0 = const()[name = string("op_1738_transpose_y_0"), val = bool(false)]; + tensor var_1738_cast_fp16 = matmul(transpose_x = var_1738_transpose_x_0, transpose_y = var_1738_transpose_y_0, x = var_1737_cast_fp16, y = v_new_85_cast_fp16)[name = string("op_1738_cast_fp16")]; + tensor var_1739_cast_fp16 = add(x = var_1735_cast_fp16, y = var_1738_cast_fp16)[name = string("op_1739_cast_fp16")]; + tensor var_1741_begin_0 = const()[name = string("op_1741_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_1741_end_0 = const()[name = string("op_1741_end_0"), val = tensor([32, 11, 1, 1])]; + tensor var_1741_end_mask_0 = const()[name = string("op_1741_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1741_squeeze_mask_0 = const()[name = string("op_1741_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1741_cast_fp16 = slice_by_index(begin = var_1741_begin_0, end = var_1741_end_0, end_mask = var_1741_end_mask_0, squeeze_mask = var_1741_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1741_cast_fp16")]; + tensor var_1742_cast_fp16 = mul(x = state_85_cast_fp16, y = var_1741_cast_fp16)[name = string("op_1742_cast_fp16")]; + tensor var_1744_begin_0 = const()[name = string("op_1744_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_1744_end_0 = const()[name = string("op_1744_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_1744_end_mask_0 = const()[name = string("op_1744_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1744_squeeze_mask_0 = const()[name = string("op_1744_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1744_cast_fp16 = slice_by_index(begin = var_1744_begin_0, end = var_1744_end_0, end_mask = var_1744_end_mask_0, squeeze_mask = var_1744_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1744_cast_fp16")]; + bool var_1746_transpose_x_1 = const()[name = string("op_1746_transpose_x_1"), val = bool(true)]; + bool var_1746_transpose_y_1 = const()[name = string("op_1746_transpose_y_1"), val = bool(false)]; + tensor var_1746_cast_fp16 = matmul(transpose_x = var_1746_transpose_x_1, transpose_y = var_1746_transpose_y_1, x = var_1744_cast_fp16, y = v_new_85_cast_fp16)[name = string("op_1746_cast_fp16")]; + tensor state_87_cast_fp16 = add(x = var_1742_cast_fp16, y = var_1746_cast_fp16)[name = string("state_87_cast_fp16")]; + tensor var_1749_begin_0 = const()[name = string("op_1749_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_1749_end_0 = const()[name = string("op_1749_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_1749_end_mask_0 = const()[name = string("op_1749_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1749_squeeze_mask_0 = const()[name = string("op_1749_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1749_cast_fp16 = slice_by_index(begin = var_1749_begin_0, end = var_1749_end_0, end_mask = var_1749_end_mask_0, squeeze_mask = var_1749_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1749_cast_fp16")]; + tensor var_1751_begin_0 = const()[name = string("op_1751_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_1751_end_0 = const()[name = string("op_1751_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_1751_end_mask_0 = const()[name = string("op_1751_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1751_squeeze_mask_0 = const()[name = string("op_1751_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1751_cast_fp16 = slice_by_index(begin = var_1751_begin_0, end = var_1751_end_0, end_mask = var_1751_end_mask_0, squeeze_mask = var_1751_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1751_cast_fp16")]; + bool var_1752_transpose_x_0 = const()[name = string("op_1752_transpose_x_0"), val = bool(false)]; + bool var_1752_transpose_y_0 = const()[name = string("op_1752_transpose_y_0"), val = bool(false)]; + tensor var_1752_cast_fp16 = matmul(transpose_x = var_1752_transpose_x_0, transpose_y = var_1752_transpose_y_0, x = var_1751_cast_fp16, y = state_87_cast_fp16)[name = string("op_1752_cast_fp16")]; + tensor v_new_87_cast_fp16 = sub(x = var_1749_cast_fp16, y = var_1752_cast_fp16)[name = string("v_new_87_cast_fp16")]; + tensor var_1755_begin_0 = const()[name = string("op_1755_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_1755_end_0 = const()[name = string("op_1755_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_1755_end_mask_0 = const()[name = string("op_1755_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1755_squeeze_mask_0 = const()[name = string("op_1755_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1755_cast_fp16 = slice_by_index(begin = var_1755_begin_0, end = var_1755_end_0, end_mask = var_1755_end_mask_0, squeeze_mask = var_1755_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1755_cast_fp16")]; + bool var_1756_transpose_x_0 = const()[name = string("op_1756_transpose_x_0"), val = bool(false)]; + bool var_1756_transpose_y_0 = const()[name = string("op_1756_transpose_y_0"), val = bool(false)]; + tensor var_1756_cast_fp16 = matmul(transpose_x = var_1756_transpose_x_0, transpose_y = var_1756_transpose_y_0, x = var_1755_cast_fp16, y = state_87_cast_fp16)[name = string("op_1756_cast_fp16")]; + tensor var_1758_begin_0 = const()[name = string("op_1758_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_1758_end_0 = const()[name = string("op_1758_end_0"), val = tensor([32, 12, 64, 64])]; + tensor var_1758_end_mask_0 = const()[name = string("op_1758_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1758_squeeze_mask_0 = const()[name = string("op_1758_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1758_cast_fp16 = slice_by_index(begin = var_1758_begin_0, end = var_1758_end_0, end_mask = var_1758_end_mask_0, squeeze_mask = var_1758_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1758_cast_fp16")]; + bool var_1759_transpose_x_0 = const()[name = string("op_1759_transpose_x_0"), val = bool(false)]; + bool var_1759_transpose_y_0 = const()[name = string("op_1759_transpose_y_0"), val = bool(false)]; + tensor var_1759_cast_fp16 = matmul(transpose_x = var_1759_transpose_x_0, transpose_y = var_1759_transpose_y_0, x = var_1758_cast_fp16, y = v_new_87_cast_fp16)[name = string("op_1759_cast_fp16")]; + tensor var_1760_cast_fp16 = add(x = var_1756_cast_fp16, y = var_1759_cast_fp16)[name = string("op_1760_cast_fp16")]; + tensor var_1762_begin_0 = const()[name = string("op_1762_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_1762_end_0 = const()[name = string("op_1762_end_0"), val = tensor([32, 12, 1, 1])]; + tensor var_1762_end_mask_0 = const()[name = string("op_1762_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1762_squeeze_mask_0 = const()[name = string("op_1762_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1762_cast_fp16 = slice_by_index(begin = var_1762_begin_0, end = var_1762_end_0, end_mask = var_1762_end_mask_0, squeeze_mask = var_1762_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1762_cast_fp16")]; + tensor var_1763_cast_fp16 = mul(x = state_87_cast_fp16, y = var_1762_cast_fp16)[name = string("op_1763_cast_fp16")]; + tensor var_1765_begin_0 = const()[name = string("op_1765_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_1765_end_0 = const()[name = string("op_1765_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_1765_end_mask_0 = const()[name = string("op_1765_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1765_squeeze_mask_0 = const()[name = string("op_1765_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1765_cast_fp16 = slice_by_index(begin = var_1765_begin_0, end = var_1765_end_0, end_mask = var_1765_end_mask_0, squeeze_mask = var_1765_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1765_cast_fp16")]; + bool var_1767_transpose_x_1 = const()[name = string("op_1767_transpose_x_1"), val = bool(true)]; + bool var_1767_transpose_y_1 = const()[name = string("op_1767_transpose_y_1"), val = bool(false)]; + tensor var_1767_cast_fp16 = matmul(transpose_x = var_1767_transpose_x_1, transpose_y = var_1767_transpose_y_1, x = var_1765_cast_fp16, y = v_new_87_cast_fp16)[name = string("op_1767_cast_fp16")]; + tensor state_89_cast_fp16 = add(x = var_1763_cast_fp16, y = var_1767_cast_fp16)[name = string("state_89_cast_fp16")]; + tensor var_1770_begin_0 = const()[name = string("op_1770_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_1770_end_0 = const()[name = string("op_1770_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_1770_end_mask_0 = const()[name = string("op_1770_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1770_squeeze_mask_0 = const()[name = string("op_1770_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1770_cast_fp16 = slice_by_index(begin = var_1770_begin_0, end = var_1770_end_0, end_mask = var_1770_end_mask_0, squeeze_mask = var_1770_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1770_cast_fp16")]; + tensor var_1772_begin_0 = const()[name = string("op_1772_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_1772_end_0 = const()[name = string("op_1772_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_1772_end_mask_0 = const()[name = string("op_1772_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1772_squeeze_mask_0 = const()[name = string("op_1772_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1772_cast_fp16 = slice_by_index(begin = var_1772_begin_0, end = var_1772_end_0, end_mask = var_1772_end_mask_0, squeeze_mask = var_1772_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1772_cast_fp16")]; + bool var_1773_transpose_x_0 = const()[name = string("op_1773_transpose_x_0"), val = bool(false)]; + bool var_1773_transpose_y_0 = const()[name = string("op_1773_transpose_y_0"), val = bool(false)]; + tensor var_1773_cast_fp16 = matmul(transpose_x = var_1773_transpose_x_0, transpose_y = var_1773_transpose_y_0, x = var_1772_cast_fp16, y = state_89_cast_fp16)[name = string("op_1773_cast_fp16")]; + tensor v_new_89_cast_fp16 = sub(x = var_1770_cast_fp16, y = var_1773_cast_fp16)[name = string("v_new_89_cast_fp16")]; + tensor var_1776_begin_0 = const()[name = string("op_1776_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_1776_end_0 = const()[name = string("op_1776_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_1776_end_mask_0 = const()[name = string("op_1776_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1776_squeeze_mask_0 = const()[name = string("op_1776_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1776_cast_fp16 = slice_by_index(begin = var_1776_begin_0, end = var_1776_end_0, end_mask = var_1776_end_mask_0, squeeze_mask = var_1776_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1776_cast_fp16")]; + bool var_1777_transpose_x_0 = const()[name = string("op_1777_transpose_x_0"), val = bool(false)]; + bool var_1777_transpose_y_0 = const()[name = string("op_1777_transpose_y_0"), val = bool(false)]; + tensor var_1777_cast_fp16 = matmul(transpose_x = var_1777_transpose_x_0, transpose_y = var_1777_transpose_y_0, x = var_1776_cast_fp16, y = state_89_cast_fp16)[name = string("op_1777_cast_fp16")]; + tensor var_1779_begin_0 = const()[name = string("op_1779_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_1779_end_0 = const()[name = string("op_1779_end_0"), val = tensor([32, 13, 64, 64])]; + tensor var_1779_end_mask_0 = const()[name = string("op_1779_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1779_squeeze_mask_0 = const()[name = string("op_1779_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1779_cast_fp16 = slice_by_index(begin = var_1779_begin_0, end = var_1779_end_0, end_mask = var_1779_end_mask_0, squeeze_mask = var_1779_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1779_cast_fp16")]; + bool var_1780_transpose_x_0 = const()[name = string("op_1780_transpose_x_0"), val = bool(false)]; + bool var_1780_transpose_y_0 = const()[name = string("op_1780_transpose_y_0"), val = bool(false)]; + tensor var_1780_cast_fp16 = matmul(transpose_x = var_1780_transpose_x_0, transpose_y = var_1780_transpose_y_0, x = var_1779_cast_fp16, y = v_new_89_cast_fp16)[name = string("op_1780_cast_fp16")]; + tensor var_1781_cast_fp16 = add(x = var_1777_cast_fp16, y = var_1780_cast_fp16)[name = string("op_1781_cast_fp16")]; + tensor var_1783_begin_0 = const()[name = string("op_1783_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_1783_end_0 = const()[name = string("op_1783_end_0"), val = tensor([32, 13, 1, 1])]; + tensor var_1783_end_mask_0 = const()[name = string("op_1783_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1783_squeeze_mask_0 = const()[name = string("op_1783_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1783_cast_fp16 = slice_by_index(begin = var_1783_begin_0, end = var_1783_end_0, end_mask = var_1783_end_mask_0, squeeze_mask = var_1783_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1783_cast_fp16")]; + tensor var_1784_cast_fp16 = mul(x = state_89_cast_fp16, y = var_1783_cast_fp16)[name = string("op_1784_cast_fp16")]; + tensor var_1786_begin_0 = const()[name = string("op_1786_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_1786_end_0 = const()[name = string("op_1786_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_1786_end_mask_0 = const()[name = string("op_1786_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1786_squeeze_mask_0 = const()[name = string("op_1786_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1786_cast_fp16 = slice_by_index(begin = var_1786_begin_0, end = var_1786_end_0, end_mask = var_1786_end_mask_0, squeeze_mask = var_1786_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1786_cast_fp16")]; + bool var_1788_transpose_x_1 = const()[name = string("op_1788_transpose_x_1"), val = bool(true)]; + bool var_1788_transpose_y_1 = const()[name = string("op_1788_transpose_y_1"), val = bool(false)]; + tensor var_1788_cast_fp16 = matmul(transpose_x = var_1788_transpose_x_1, transpose_y = var_1788_transpose_y_1, x = var_1786_cast_fp16, y = v_new_89_cast_fp16)[name = string("op_1788_cast_fp16")]; + tensor state_91_cast_fp16 = add(x = var_1784_cast_fp16, y = var_1788_cast_fp16)[name = string("state_91_cast_fp16")]; + tensor var_1791_begin_0 = const()[name = string("op_1791_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_1791_end_0 = const()[name = string("op_1791_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_1791_end_mask_0 = const()[name = string("op_1791_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1791_squeeze_mask_0 = const()[name = string("op_1791_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1791_cast_fp16 = slice_by_index(begin = var_1791_begin_0, end = var_1791_end_0, end_mask = var_1791_end_mask_0, squeeze_mask = var_1791_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1791_cast_fp16")]; + tensor var_1793_begin_0 = const()[name = string("op_1793_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_1793_end_0 = const()[name = string("op_1793_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_1793_end_mask_0 = const()[name = string("op_1793_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1793_squeeze_mask_0 = const()[name = string("op_1793_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1793_cast_fp16 = slice_by_index(begin = var_1793_begin_0, end = var_1793_end_0, end_mask = var_1793_end_mask_0, squeeze_mask = var_1793_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1793_cast_fp16")]; + bool var_1794_transpose_x_0 = const()[name = string("op_1794_transpose_x_0"), val = bool(false)]; + bool var_1794_transpose_y_0 = const()[name = string("op_1794_transpose_y_0"), val = bool(false)]; + tensor var_1794_cast_fp16 = matmul(transpose_x = var_1794_transpose_x_0, transpose_y = var_1794_transpose_y_0, x = var_1793_cast_fp16, y = state_91_cast_fp16)[name = string("op_1794_cast_fp16")]; + tensor v_new_91_cast_fp16 = sub(x = var_1791_cast_fp16, y = var_1794_cast_fp16)[name = string("v_new_91_cast_fp16")]; + tensor var_1797_begin_0 = const()[name = string("op_1797_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_1797_end_0 = const()[name = string("op_1797_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_1797_end_mask_0 = const()[name = string("op_1797_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1797_squeeze_mask_0 = const()[name = string("op_1797_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1797_cast_fp16 = slice_by_index(begin = var_1797_begin_0, end = var_1797_end_0, end_mask = var_1797_end_mask_0, squeeze_mask = var_1797_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1797_cast_fp16")]; + bool var_1798_transpose_x_0 = const()[name = string("op_1798_transpose_x_0"), val = bool(false)]; + bool var_1798_transpose_y_0 = const()[name = string("op_1798_transpose_y_0"), val = bool(false)]; + tensor var_1798_cast_fp16 = matmul(transpose_x = var_1798_transpose_x_0, transpose_y = var_1798_transpose_y_0, x = var_1797_cast_fp16, y = state_91_cast_fp16)[name = string("op_1798_cast_fp16")]; + tensor var_1800_begin_0 = const()[name = string("op_1800_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_1800_end_0 = const()[name = string("op_1800_end_0"), val = tensor([32, 14, 64, 64])]; + tensor var_1800_end_mask_0 = const()[name = string("op_1800_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1800_squeeze_mask_0 = const()[name = string("op_1800_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1800_cast_fp16 = slice_by_index(begin = var_1800_begin_0, end = var_1800_end_0, end_mask = var_1800_end_mask_0, squeeze_mask = var_1800_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1800_cast_fp16")]; + bool var_1801_transpose_x_0 = const()[name = string("op_1801_transpose_x_0"), val = bool(false)]; + bool var_1801_transpose_y_0 = const()[name = string("op_1801_transpose_y_0"), val = bool(false)]; + tensor var_1801_cast_fp16 = matmul(transpose_x = var_1801_transpose_x_0, transpose_y = var_1801_transpose_y_0, x = var_1800_cast_fp16, y = v_new_91_cast_fp16)[name = string("op_1801_cast_fp16")]; + tensor var_1802_cast_fp16 = add(x = var_1798_cast_fp16, y = var_1801_cast_fp16)[name = string("op_1802_cast_fp16")]; + tensor var_1804_begin_0 = const()[name = string("op_1804_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_1804_end_0 = const()[name = string("op_1804_end_0"), val = tensor([32, 14, 1, 1])]; + tensor var_1804_end_mask_0 = const()[name = string("op_1804_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1804_squeeze_mask_0 = const()[name = string("op_1804_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1804_cast_fp16 = slice_by_index(begin = var_1804_begin_0, end = var_1804_end_0, end_mask = var_1804_end_mask_0, squeeze_mask = var_1804_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1804_cast_fp16")]; + tensor var_1805_cast_fp16 = mul(x = state_91_cast_fp16, y = var_1804_cast_fp16)[name = string("op_1805_cast_fp16")]; + tensor var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1807_squeeze_mask_0 = const()[name = string("op_1807_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, squeeze_mask = var_1807_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1807_cast_fp16")]; + bool var_1809_transpose_x_1 = const()[name = string("op_1809_transpose_x_1"), val = bool(true)]; + bool var_1809_transpose_y_1 = const()[name = string("op_1809_transpose_y_1"), val = bool(false)]; + tensor var_1809_cast_fp16 = matmul(transpose_x = var_1809_transpose_x_1, transpose_y = var_1809_transpose_y_1, x = var_1807_cast_fp16, y = v_new_91_cast_fp16)[name = string("op_1809_cast_fp16")]; + tensor state_93_cast_fp16 = add(x = var_1805_cast_fp16, y = var_1809_cast_fp16)[name = string("state_93_cast_fp16")]; + tensor var_1812_begin_0 = const()[name = string("op_1812_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_1812_end_0 = const()[name = string("op_1812_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_1812_end_mask_0 = const()[name = string("op_1812_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1812_squeeze_mask_0 = const()[name = string("op_1812_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1812_cast_fp16 = slice_by_index(begin = var_1812_begin_0, end = var_1812_end_0, end_mask = var_1812_end_mask_0, squeeze_mask = var_1812_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1812_cast_fp16")]; + tensor var_1814_begin_0 = const()[name = string("op_1814_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_1814_end_0 = const()[name = string("op_1814_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_1814_end_mask_0 = const()[name = string("op_1814_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1814_squeeze_mask_0 = const()[name = string("op_1814_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1814_cast_fp16 = slice_by_index(begin = var_1814_begin_0, end = var_1814_end_0, end_mask = var_1814_end_mask_0, squeeze_mask = var_1814_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1814_cast_fp16")]; + bool var_1815_transpose_x_0 = const()[name = string("op_1815_transpose_x_0"), val = bool(false)]; + bool var_1815_transpose_y_0 = const()[name = string("op_1815_transpose_y_0"), val = bool(false)]; + tensor var_1815_cast_fp16 = matmul(transpose_x = var_1815_transpose_x_0, transpose_y = var_1815_transpose_y_0, x = var_1814_cast_fp16, y = state_93_cast_fp16)[name = string("op_1815_cast_fp16")]; + tensor v_new_93_cast_fp16 = sub(x = var_1812_cast_fp16, y = var_1815_cast_fp16)[name = string("v_new_93_cast_fp16")]; + tensor var_1818_begin_0 = const()[name = string("op_1818_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_1818_end_0 = const()[name = string("op_1818_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_1818_end_mask_0 = const()[name = string("op_1818_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1818_squeeze_mask_0 = const()[name = string("op_1818_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1818_cast_fp16 = slice_by_index(begin = var_1818_begin_0, end = var_1818_end_0, end_mask = var_1818_end_mask_0, squeeze_mask = var_1818_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1818_cast_fp16")]; + bool var_1819_transpose_x_0 = const()[name = string("op_1819_transpose_x_0"), val = bool(false)]; + bool var_1819_transpose_y_0 = const()[name = string("op_1819_transpose_y_0"), val = bool(false)]; + tensor var_1819_cast_fp16 = matmul(transpose_x = var_1819_transpose_x_0, transpose_y = var_1819_transpose_y_0, x = var_1818_cast_fp16, y = state_93_cast_fp16)[name = string("op_1819_cast_fp16")]; + tensor var_1821_begin_0 = const()[name = string("op_1821_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_1821_end_0 = const()[name = string("op_1821_end_0"), val = tensor([32, 15, 64, 64])]; + tensor var_1821_end_mask_0 = const()[name = string("op_1821_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1821_squeeze_mask_0 = const()[name = string("op_1821_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, squeeze_mask = var_1821_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1821_cast_fp16")]; + bool var_1822_transpose_x_0 = const()[name = string("op_1822_transpose_x_0"), val = bool(false)]; + bool var_1822_transpose_y_0 = const()[name = string("op_1822_transpose_y_0"), val = bool(false)]; + tensor var_1822_cast_fp16 = matmul(transpose_x = var_1822_transpose_x_0, transpose_y = var_1822_transpose_y_0, x = var_1821_cast_fp16, y = v_new_93_cast_fp16)[name = string("op_1822_cast_fp16")]; + tensor var_1823_cast_fp16 = add(x = var_1819_cast_fp16, y = var_1822_cast_fp16)[name = string("op_1823_cast_fp16")]; + tensor var_1825_begin_0 = const()[name = string("op_1825_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_1825_end_0 = const()[name = string("op_1825_end_0"), val = tensor([32, 15, 1, 1])]; + tensor var_1825_end_mask_0 = const()[name = string("op_1825_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1825_squeeze_mask_0 = const()[name = string("op_1825_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1825_cast_fp16 = slice_by_index(begin = var_1825_begin_0, end = var_1825_end_0, end_mask = var_1825_end_mask_0, squeeze_mask = var_1825_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1825_cast_fp16")]; + tensor var_1826_cast_fp16 = mul(x = state_93_cast_fp16, y = var_1825_cast_fp16)[name = string("op_1826_cast_fp16")]; + tensor var_1828_begin_0 = const()[name = string("op_1828_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_1828_end_0 = const()[name = string("op_1828_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_1828_end_mask_0 = const()[name = string("op_1828_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1828_squeeze_mask_0 = const()[name = string("op_1828_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1828_cast_fp16 = slice_by_index(begin = var_1828_begin_0, end = var_1828_end_0, end_mask = var_1828_end_mask_0, squeeze_mask = var_1828_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1828_cast_fp16")]; + bool var_1830_transpose_x_1 = const()[name = string("op_1830_transpose_x_1"), val = bool(true)]; + bool var_1830_transpose_y_1 = const()[name = string("op_1830_transpose_y_1"), val = bool(false)]; + tensor var_1830_cast_fp16 = matmul(transpose_x = var_1830_transpose_x_1, transpose_y = var_1830_transpose_y_1, x = var_1828_cast_fp16, y = v_new_93_cast_fp16)[name = string("op_1830_cast_fp16")]; + tensor state_95_cast_fp16 = add(x = var_1826_cast_fp16, y = var_1830_cast_fp16)[name = string("state_95_cast_fp16")]; + tensor var_1833_begin_0 = const()[name = string("op_1833_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_1833_end_0 = const()[name = string("op_1833_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_1833_end_mask_0 = const()[name = string("op_1833_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1833_squeeze_mask_0 = const()[name = string("op_1833_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1833_cast_fp16 = slice_by_index(begin = var_1833_begin_0, end = var_1833_end_0, end_mask = var_1833_end_mask_0, squeeze_mask = var_1833_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1833_cast_fp16")]; + tensor var_1835_begin_0 = const()[name = string("op_1835_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_1835_end_0 = const()[name = string("op_1835_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_1835_end_mask_0 = const()[name = string("op_1835_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1835_squeeze_mask_0 = const()[name = string("op_1835_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1835_cast_fp16 = slice_by_index(begin = var_1835_begin_0, end = var_1835_end_0, end_mask = var_1835_end_mask_0, squeeze_mask = var_1835_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1835_cast_fp16")]; + bool var_1836_transpose_x_0 = const()[name = string("op_1836_transpose_x_0"), val = bool(false)]; + bool var_1836_transpose_y_0 = const()[name = string("op_1836_transpose_y_0"), val = bool(false)]; + tensor var_1836_cast_fp16 = matmul(transpose_x = var_1836_transpose_x_0, transpose_y = var_1836_transpose_y_0, x = var_1835_cast_fp16, y = state_95_cast_fp16)[name = string("op_1836_cast_fp16")]; + tensor v_new_95_cast_fp16 = sub(x = var_1833_cast_fp16, y = var_1836_cast_fp16)[name = string("v_new_95_cast_fp16")]; + tensor var_1839_begin_0 = const()[name = string("op_1839_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_1839_end_0 = const()[name = string("op_1839_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_1839_end_mask_0 = const()[name = string("op_1839_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1839_squeeze_mask_0 = const()[name = string("op_1839_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1839_cast_fp16 = slice_by_index(begin = var_1839_begin_0, end = var_1839_end_0, end_mask = var_1839_end_mask_0, squeeze_mask = var_1839_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1839_cast_fp16")]; + bool var_1840_transpose_x_0 = const()[name = string("op_1840_transpose_x_0"), val = bool(false)]; + bool var_1840_transpose_y_0 = const()[name = string("op_1840_transpose_y_0"), val = bool(false)]; + tensor var_1840_cast_fp16 = matmul(transpose_x = var_1840_transpose_x_0, transpose_y = var_1840_transpose_y_0, x = var_1839_cast_fp16, y = state_95_cast_fp16)[name = string("op_1840_cast_fp16")]; + tensor var_1842_begin_0 = const()[name = string("op_1842_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_1842_end_0 = const()[name = string("op_1842_end_0"), val = tensor([32, 16, 64, 64])]; + tensor var_1842_end_mask_0 = const()[name = string("op_1842_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1842_squeeze_mask_0 = const()[name = string("op_1842_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1842_cast_fp16 = slice_by_index(begin = var_1842_begin_0, end = var_1842_end_0, end_mask = var_1842_end_mask_0, squeeze_mask = var_1842_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1842_cast_fp16")]; + bool var_1843_transpose_x_0 = const()[name = string("op_1843_transpose_x_0"), val = bool(false)]; + bool var_1843_transpose_y_0 = const()[name = string("op_1843_transpose_y_0"), val = bool(false)]; + tensor var_1843_cast_fp16 = matmul(transpose_x = var_1843_transpose_x_0, transpose_y = var_1843_transpose_y_0, x = var_1842_cast_fp16, y = v_new_95_cast_fp16)[name = string("op_1843_cast_fp16")]; + tensor var_1844_cast_fp16 = add(x = var_1840_cast_fp16, y = var_1843_cast_fp16)[name = string("op_1844_cast_fp16")]; + tensor var_1846_begin_0 = const()[name = string("op_1846_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_1846_end_0 = const()[name = string("op_1846_end_0"), val = tensor([32, 16, 1, 1])]; + tensor var_1846_end_mask_0 = const()[name = string("op_1846_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1846_squeeze_mask_0 = const()[name = string("op_1846_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1846_cast_fp16 = slice_by_index(begin = var_1846_begin_0, end = var_1846_end_0, end_mask = var_1846_end_mask_0, squeeze_mask = var_1846_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1846_cast_fp16")]; + tensor var_1847_cast_fp16 = mul(x = state_95_cast_fp16, y = var_1846_cast_fp16)[name = string("op_1847_cast_fp16")]; + tensor var_1849_begin_0 = const()[name = string("op_1849_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_1849_end_0 = const()[name = string("op_1849_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_1849_end_mask_0 = const()[name = string("op_1849_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1849_squeeze_mask_0 = const()[name = string("op_1849_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, squeeze_mask = var_1849_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1849_cast_fp16")]; + bool var_1851_transpose_x_1 = const()[name = string("op_1851_transpose_x_1"), val = bool(true)]; + bool var_1851_transpose_y_1 = const()[name = string("op_1851_transpose_y_1"), val = bool(false)]; + tensor var_1851_cast_fp16 = matmul(transpose_x = var_1851_transpose_x_1, transpose_y = var_1851_transpose_y_1, x = var_1849_cast_fp16, y = v_new_95_cast_fp16)[name = string("op_1851_cast_fp16")]; + tensor state_97_cast_fp16 = add(x = var_1847_cast_fp16, y = var_1851_cast_fp16)[name = string("state_97_cast_fp16")]; + tensor var_1854_begin_0 = const()[name = string("op_1854_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_1854_end_0 = const()[name = string("op_1854_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_1854_end_mask_0 = const()[name = string("op_1854_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1854_squeeze_mask_0 = const()[name = string("op_1854_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1854_cast_fp16 = slice_by_index(begin = var_1854_begin_0, end = var_1854_end_0, end_mask = var_1854_end_mask_0, squeeze_mask = var_1854_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1854_cast_fp16")]; + tensor var_1856_begin_0 = const()[name = string("op_1856_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_1856_end_0 = const()[name = string("op_1856_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_1856_end_mask_0 = const()[name = string("op_1856_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1856_squeeze_mask_0 = const()[name = string("op_1856_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1856_cast_fp16 = slice_by_index(begin = var_1856_begin_0, end = var_1856_end_0, end_mask = var_1856_end_mask_0, squeeze_mask = var_1856_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1856_cast_fp16")]; + bool var_1857_transpose_x_0 = const()[name = string("op_1857_transpose_x_0"), val = bool(false)]; + bool var_1857_transpose_y_0 = const()[name = string("op_1857_transpose_y_0"), val = bool(false)]; + tensor var_1857_cast_fp16 = matmul(transpose_x = var_1857_transpose_x_0, transpose_y = var_1857_transpose_y_0, x = var_1856_cast_fp16, y = state_97_cast_fp16)[name = string("op_1857_cast_fp16")]; + tensor v_new_97_cast_fp16 = sub(x = var_1854_cast_fp16, y = var_1857_cast_fp16)[name = string("v_new_97_cast_fp16")]; + tensor var_1860_begin_0 = const()[name = string("op_1860_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_1860_end_0 = const()[name = string("op_1860_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_1860_end_mask_0 = const()[name = string("op_1860_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1860_squeeze_mask_0 = const()[name = string("op_1860_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1860_cast_fp16 = slice_by_index(begin = var_1860_begin_0, end = var_1860_end_0, end_mask = var_1860_end_mask_0, squeeze_mask = var_1860_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1860_cast_fp16")]; + bool var_1861_transpose_x_0 = const()[name = string("op_1861_transpose_x_0"), val = bool(false)]; + bool var_1861_transpose_y_0 = const()[name = string("op_1861_transpose_y_0"), val = bool(false)]; + tensor var_1861_cast_fp16 = matmul(transpose_x = var_1861_transpose_x_0, transpose_y = var_1861_transpose_y_0, x = var_1860_cast_fp16, y = state_97_cast_fp16)[name = string("op_1861_cast_fp16")]; + tensor var_1863_begin_0 = const()[name = string("op_1863_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_1863_end_0 = const()[name = string("op_1863_end_0"), val = tensor([32, 17, 64, 64])]; + tensor var_1863_end_mask_0 = const()[name = string("op_1863_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1863_squeeze_mask_0 = const()[name = string("op_1863_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1863_cast_fp16 = slice_by_index(begin = var_1863_begin_0, end = var_1863_end_0, end_mask = var_1863_end_mask_0, squeeze_mask = var_1863_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1863_cast_fp16")]; + bool var_1864_transpose_x_0 = const()[name = string("op_1864_transpose_x_0"), val = bool(false)]; + bool var_1864_transpose_y_0 = const()[name = string("op_1864_transpose_y_0"), val = bool(false)]; + tensor var_1864_cast_fp16 = matmul(transpose_x = var_1864_transpose_x_0, transpose_y = var_1864_transpose_y_0, x = var_1863_cast_fp16, y = v_new_97_cast_fp16)[name = string("op_1864_cast_fp16")]; + tensor var_1865_cast_fp16 = add(x = var_1861_cast_fp16, y = var_1864_cast_fp16)[name = string("op_1865_cast_fp16")]; + tensor var_1867_begin_0 = const()[name = string("op_1867_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_1867_end_0 = const()[name = string("op_1867_end_0"), val = tensor([32, 17, 1, 1])]; + tensor var_1867_end_mask_0 = const()[name = string("op_1867_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1867_squeeze_mask_0 = const()[name = string("op_1867_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1867_cast_fp16 = slice_by_index(begin = var_1867_begin_0, end = var_1867_end_0, end_mask = var_1867_end_mask_0, squeeze_mask = var_1867_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1867_cast_fp16")]; + tensor var_1868_cast_fp16 = mul(x = state_97_cast_fp16, y = var_1867_cast_fp16)[name = string("op_1868_cast_fp16")]; + tensor var_1870_begin_0 = const()[name = string("op_1870_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_1870_end_0 = const()[name = string("op_1870_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_1870_end_mask_0 = const()[name = string("op_1870_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1870_squeeze_mask_0 = const()[name = string("op_1870_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1870_cast_fp16 = slice_by_index(begin = var_1870_begin_0, end = var_1870_end_0, end_mask = var_1870_end_mask_0, squeeze_mask = var_1870_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1870_cast_fp16")]; + bool var_1872_transpose_x_1 = const()[name = string("op_1872_transpose_x_1"), val = bool(true)]; + bool var_1872_transpose_y_1 = const()[name = string("op_1872_transpose_y_1"), val = bool(false)]; + tensor var_1872_cast_fp16 = matmul(transpose_x = var_1872_transpose_x_1, transpose_y = var_1872_transpose_y_1, x = var_1870_cast_fp16, y = v_new_97_cast_fp16)[name = string("op_1872_cast_fp16")]; + tensor state_99_cast_fp16 = add(x = var_1868_cast_fp16, y = var_1872_cast_fp16)[name = string("state_99_cast_fp16")]; + tensor var_1875_begin_0 = const()[name = string("op_1875_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_1875_end_0 = const()[name = string("op_1875_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_1875_end_mask_0 = const()[name = string("op_1875_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1875_squeeze_mask_0 = const()[name = string("op_1875_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1875_cast_fp16 = slice_by_index(begin = var_1875_begin_0, end = var_1875_end_0, end_mask = var_1875_end_mask_0, squeeze_mask = var_1875_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1875_cast_fp16")]; + tensor var_1877_begin_0 = const()[name = string("op_1877_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_1877_end_0 = const()[name = string("op_1877_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_1877_end_mask_0 = const()[name = string("op_1877_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1877_squeeze_mask_0 = const()[name = string("op_1877_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1877_cast_fp16 = slice_by_index(begin = var_1877_begin_0, end = var_1877_end_0, end_mask = var_1877_end_mask_0, squeeze_mask = var_1877_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1877_cast_fp16")]; + bool var_1878_transpose_x_0 = const()[name = string("op_1878_transpose_x_0"), val = bool(false)]; + bool var_1878_transpose_y_0 = const()[name = string("op_1878_transpose_y_0"), val = bool(false)]; + tensor var_1878_cast_fp16 = matmul(transpose_x = var_1878_transpose_x_0, transpose_y = var_1878_transpose_y_0, x = var_1877_cast_fp16, y = state_99_cast_fp16)[name = string("op_1878_cast_fp16")]; + tensor v_new_99_cast_fp16 = sub(x = var_1875_cast_fp16, y = var_1878_cast_fp16)[name = string("v_new_99_cast_fp16")]; + tensor var_1881_begin_0 = const()[name = string("op_1881_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_1881_end_0 = const()[name = string("op_1881_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_1881_end_mask_0 = const()[name = string("op_1881_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1881_squeeze_mask_0 = const()[name = string("op_1881_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1881_cast_fp16 = slice_by_index(begin = var_1881_begin_0, end = var_1881_end_0, end_mask = var_1881_end_mask_0, squeeze_mask = var_1881_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1881_cast_fp16")]; + bool var_1882_transpose_x_0 = const()[name = string("op_1882_transpose_x_0"), val = bool(false)]; + bool var_1882_transpose_y_0 = const()[name = string("op_1882_transpose_y_0"), val = bool(false)]; + tensor var_1882_cast_fp16 = matmul(transpose_x = var_1882_transpose_x_0, transpose_y = var_1882_transpose_y_0, x = var_1881_cast_fp16, y = state_99_cast_fp16)[name = string("op_1882_cast_fp16")]; + tensor var_1884_begin_0 = const()[name = string("op_1884_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_1884_end_0 = const()[name = string("op_1884_end_0"), val = tensor([32, 18, 64, 64])]; + tensor var_1884_end_mask_0 = const()[name = string("op_1884_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1884_squeeze_mask_0 = const()[name = string("op_1884_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1884_cast_fp16 = slice_by_index(begin = var_1884_begin_0, end = var_1884_end_0, end_mask = var_1884_end_mask_0, squeeze_mask = var_1884_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1884_cast_fp16")]; + bool var_1885_transpose_x_0 = const()[name = string("op_1885_transpose_x_0"), val = bool(false)]; + bool var_1885_transpose_y_0 = const()[name = string("op_1885_transpose_y_0"), val = bool(false)]; + tensor var_1885_cast_fp16 = matmul(transpose_x = var_1885_transpose_x_0, transpose_y = var_1885_transpose_y_0, x = var_1884_cast_fp16, y = v_new_99_cast_fp16)[name = string("op_1885_cast_fp16")]; + tensor var_1886_cast_fp16 = add(x = var_1882_cast_fp16, y = var_1885_cast_fp16)[name = string("op_1886_cast_fp16")]; + tensor var_1888_begin_0 = const()[name = string("op_1888_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_1888_end_0 = const()[name = string("op_1888_end_0"), val = tensor([32, 18, 1, 1])]; + tensor var_1888_end_mask_0 = const()[name = string("op_1888_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1888_squeeze_mask_0 = const()[name = string("op_1888_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1888_cast_fp16 = slice_by_index(begin = var_1888_begin_0, end = var_1888_end_0, end_mask = var_1888_end_mask_0, squeeze_mask = var_1888_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1888_cast_fp16")]; + tensor var_1889_cast_fp16 = mul(x = state_99_cast_fp16, y = var_1888_cast_fp16)[name = string("op_1889_cast_fp16")]; + tensor var_1891_begin_0 = const()[name = string("op_1891_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_1891_end_0 = const()[name = string("op_1891_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_1891_end_mask_0 = const()[name = string("op_1891_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1891_squeeze_mask_0 = const()[name = string("op_1891_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1891_cast_fp16 = slice_by_index(begin = var_1891_begin_0, end = var_1891_end_0, end_mask = var_1891_end_mask_0, squeeze_mask = var_1891_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1891_cast_fp16")]; + bool var_1893_transpose_x_1 = const()[name = string("op_1893_transpose_x_1"), val = bool(true)]; + bool var_1893_transpose_y_1 = const()[name = string("op_1893_transpose_y_1"), val = bool(false)]; + tensor var_1893_cast_fp16 = matmul(transpose_x = var_1893_transpose_x_1, transpose_y = var_1893_transpose_y_1, x = var_1891_cast_fp16, y = v_new_99_cast_fp16)[name = string("op_1893_cast_fp16")]; + tensor state_101_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1893_cast_fp16)[name = string("state_101_cast_fp16")]; + tensor var_1896_begin_0 = const()[name = string("op_1896_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_1896_end_0 = const()[name = string("op_1896_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_1896_end_mask_0 = const()[name = string("op_1896_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1896_squeeze_mask_0 = const()[name = string("op_1896_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1896_cast_fp16 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, squeeze_mask = var_1896_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1896_cast_fp16")]; + tensor var_1898_begin_0 = const()[name = string("op_1898_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_1898_end_0 = const()[name = string("op_1898_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_1898_end_mask_0 = const()[name = string("op_1898_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1898_squeeze_mask_0 = const()[name = string("op_1898_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1898_cast_fp16 = slice_by_index(begin = var_1898_begin_0, end = var_1898_end_0, end_mask = var_1898_end_mask_0, squeeze_mask = var_1898_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1898_cast_fp16")]; + bool var_1899_transpose_x_0 = const()[name = string("op_1899_transpose_x_0"), val = bool(false)]; + bool var_1899_transpose_y_0 = const()[name = string("op_1899_transpose_y_0"), val = bool(false)]; + tensor var_1899_cast_fp16 = matmul(transpose_x = var_1899_transpose_x_0, transpose_y = var_1899_transpose_y_0, x = var_1898_cast_fp16, y = state_101_cast_fp16)[name = string("op_1899_cast_fp16")]; + tensor v_new_101_cast_fp16 = sub(x = var_1896_cast_fp16, y = var_1899_cast_fp16)[name = string("v_new_101_cast_fp16")]; + tensor var_1902_begin_0 = const()[name = string("op_1902_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_1902_end_0 = const()[name = string("op_1902_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_1902_end_mask_0 = const()[name = string("op_1902_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1902_squeeze_mask_0 = const()[name = string("op_1902_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1902_cast_fp16 = slice_by_index(begin = var_1902_begin_0, end = var_1902_end_0, end_mask = var_1902_end_mask_0, squeeze_mask = var_1902_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1902_cast_fp16")]; + bool var_1903_transpose_x_0 = const()[name = string("op_1903_transpose_x_0"), val = bool(false)]; + bool var_1903_transpose_y_0 = const()[name = string("op_1903_transpose_y_0"), val = bool(false)]; + tensor var_1903_cast_fp16 = matmul(transpose_x = var_1903_transpose_x_0, transpose_y = var_1903_transpose_y_0, x = var_1902_cast_fp16, y = state_101_cast_fp16)[name = string("op_1903_cast_fp16")]; + tensor var_1905_begin_0 = const()[name = string("op_1905_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_1905_end_0 = const()[name = string("op_1905_end_0"), val = tensor([32, 19, 64, 64])]; + tensor var_1905_end_mask_0 = const()[name = string("op_1905_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1905_squeeze_mask_0 = const()[name = string("op_1905_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1905_cast_fp16 = slice_by_index(begin = var_1905_begin_0, end = var_1905_end_0, end_mask = var_1905_end_mask_0, squeeze_mask = var_1905_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1905_cast_fp16")]; + bool var_1906_transpose_x_0 = const()[name = string("op_1906_transpose_x_0"), val = bool(false)]; + bool var_1906_transpose_y_0 = const()[name = string("op_1906_transpose_y_0"), val = bool(false)]; + tensor var_1906_cast_fp16 = matmul(transpose_x = var_1906_transpose_x_0, transpose_y = var_1906_transpose_y_0, x = var_1905_cast_fp16, y = v_new_101_cast_fp16)[name = string("op_1906_cast_fp16")]; + tensor var_1907_cast_fp16 = add(x = var_1903_cast_fp16, y = var_1906_cast_fp16)[name = string("op_1907_cast_fp16")]; + tensor var_1909_begin_0 = const()[name = string("op_1909_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_1909_end_0 = const()[name = string("op_1909_end_0"), val = tensor([32, 19, 1, 1])]; + tensor var_1909_end_mask_0 = const()[name = string("op_1909_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1909_squeeze_mask_0 = const()[name = string("op_1909_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1909_cast_fp16 = slice_by_index(begin = var_1909_begin_0, end = var_1909_end_0, end_mask = var_1909_end_mask_0, squeeze_mask = var_1909_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1909_cast_fp16")]; + tensor var_1910_cast_fp16 = mul(x = state_101_cast_fp16, y = var_1909_cast_fp16)[name = string("op_1910_cast_fp16")]; + tensor var_1912_begin_0 = const()[name = string("op_1912_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_1912_end_0 = const()[name = string("op_1912_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_1912_end_mask_0 = const()[name = string("op_1912_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1912_squeeze_mask_0 = const()[name = string("op_1912_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1912_cast_fp16 = slice_by_index(begin = var_1912_begin_0, end = var_1912_end_0, end_mask = var_1912_end_mask_0, squeeze_mask = var_1912_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1912_cast_fp16")]; + bool var_1914_transpose_x_1 = const()[name = string("op_1914_transpose_x_1"), val = bool(true)]; + bool var_1914_transpose_y_1 = const()[name = string("op_1914_transpose_y_1"), val = bool(false)]; + tensor var_1914_cast_fp16 = matmul(transpose_x = var_1914_transpose_x_1, transpose_y = var_1914_transpose_y_1, x = var_1912_cast_fp16, y = v_new_101_cast_fp16)[name = string("op_1914_cast_fp16")]; + tensor state_103_cast_fp16 = add(x = var_1910_cast_fp16, y = var_1914_cast_fp16)[name = string("state_103_cast_fp16")]; + tensor var_1917_begin_0 = const()[name = string("op_1917_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_1917_end_0 = const()[name = string("op_1917_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_1917_end_mask_0 = const()[name = string("op_1917_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1917_squeeze_mask_0 = const()[name = string("op_1917_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1917_cast_fp16 = slice_by_index(begin = var_1917_begin_0, end = var_1917_end_0, end_mask = var_1917_end_mask_0, squeeze_mask = var_1917_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1917_cast_fp16")]; + tensor var_1919_begin_0 = const()[name = string("op_1919_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_1919_end_0 = const()[name = string("op_1919_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_1919_end_mask_0 = const()[name = string("op_1919_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1919_squeeze_mask_0 = const()[name = string("op_1919_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1919_cast_fp16 = slice_by_index(begin = var_1919_begin_0, end = var_1919_end_0, end_mask = var_1919_end_mask_0, squeeze_mask = var_1919_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1919_cast_fp16")]; + bool var_1920_transpose_x_0 = const()[name = string("op_1920_transpose_x_0"), val = bool(false)]; + bool var_1920_transpose_y_0 = const()[name = string("op_1920_transpose_y_0"), val = bool(false)]; + tensor var_1920_cast_fp16 = matmul(transpose_x = var_1920_transpose_x_0, transpose_y = var_1920_transpose_y_0, x = var_1919_cast_fp16, y = state_103_cast_fp16)[name = string("op_1920_cast_fp16")]; + tensor v_new_103_cast_fp16 = sub(x = var_1917_cast_fp16, y = var_1920_cast_fp16)[name = string("v_new_103_cast_fp16")]; + tensor var_1923_begin_0 = const()[name = string("op_1923_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_1923_end_0 = const()[name = string("op_1923_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_1923_end_mask_0 = const()[name = string("op_1923_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1923_squeeze_mask_0 = const()[name = string("op_1923_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1923_cast_fp16 = slice_by_index(begin = var_1923_begin_0, end = var_1923_end_0, end_mask = var_1923_end_mask_0, squeeze_mask = var_1923_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1923_cast_fp16")]; + bool var_1924_transpose_x_0 = const()[name = string("op_1924_transpose_x_0"), val = bool(false)]; + bool var_1924_transpose_y_0 = const()[name = string("op_1924_transpose_y_0"), val = bool(false)]; + tensor var_1924_cast_fp16 = matmul(transpose_x = var_1924_transpose_x_0, transpose_y = var_1924_transpose_y_0, x = var_1923_cast_fp16, y = state_103_cast_fp16)[name = string("op_1924_cast_fp16")]; + tensor var_1926_begin_0 = const()[name = string("op_1926_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_1926_end_0 = const()[name = string("op_1926_end_0"), val = tensor([32, 20, 64, 64])]; + tensor var_1926_end_mask_0 = const()[name = string("op_1926_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1926_squeeze_mask_0 = const()[name = string("op_1926_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1926_cast_fp16 = slice_by_index(begin = var_1926_begin_0, end = var_1926_end_0, end_mask = var_1926_end_mask_0, squeeze_mask = var_1926_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1926_cast_fp16")]; + bool var_1927_transpose_x_0 = const()[name = string("op_1927_transpose_x_0"), val = bool(false)]; + bool var_1927_transpose_y_0 = const()[name = string("op_1927_transpose_y_0"), val = bool(false)]; + tensor var_1927_cast_fp16 = matmul(transpose_x = var_1927_transpose_x_0, transpose_y = var_1927_transpose_y_0, x = var_1926_cast_fp16, y = v_new_103_cast_fp16)[name = string("op_1927_cast_fp16")]; + tensor var_1928_cast_fp16 = add(x = var_1924_cast_fp16, y = var_1927_cast_fp16)[name = string("op_1928_cast_fp16")]; + tensor var_1930_begin_0 = const()[name = string("op_1930_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_1930_end_0 = const()[name = string("op_1930_end_0"), val = tensor([32, 20, 1, 1])]; + tensor var_1930_end_mask_0 = const()[name = string("op_1930_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1930_squeeze_mask_0 = const()[name = string("op_1930_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1930_cast_fp16 = slice_by_index(begin = var_1930_begin_0, end = var_1930_end_0, end_mask = var_1930_end_mask_0, squeeze_mask = var_1930_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1930_cast_fp16")]; + tensor var_1931_cast_fp16 = mul(x = state_103_cast_fp16, y = var_1930_cast_fp16)[name = string("op_1931_cast_fp16")]; + tensor var_1933_begin_0 = const()[name = string("op_1933_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_1933_end_0 = const()[name = string("op_1933_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_1933_end_mask_0 = const()[name = string("op_1933_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1933_squeeze_mask_0 = const()[name = string("op_1933_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1933_cast_fp16 = slice_by_index(begin = var_1933_begin_0, end = var_1933_end_0, end_mask = var_1933_end_mask_0, squeeze_mask = var_1933_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1933_cast_fp16")]; + bool var_1935_transpose_x_1 = const()[name = string("op_1935_transpose_x_1"), val = bool(true)]; + bool var_1935_transpose_y_1 = const()[name = string("op_1935_transpose_y_1"), val = bool(false)]; + tensor var_1935_cast_fp16 = matmul(transpose_x = var_1935_transpose_x_1, transpose_y = var_1935_transpose_y_1, x = var_1933_cast_fp16, y = v_new_103_cast_fp16)[name = string("op_1935_cast_fp16")]; + tensor state_105_cast_fp16 = add(x = var_1931_cast_fp16, y = var_1935_cast_fp16)[name = string("state_105_cast_fp16")]; + tensor var_1938_begin_0 = const()[name = string("op_1938_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_1938_end_0 = const()[name = string("op_1938_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_1938_end_mask_0 = const()[name = string("op_1938_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1938_squeeze_mask_0 = const()[name = string("op_1938_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1938_cast_fp16 = slice_by_index(begin = var_1938_begin_0, end = var_1938_end_0, end_mask = var_1938_end_mask_0, squeeze_mask = var_1938_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1938_cast_fp16")]; + tensor var_1940_begin_0 = const()[name = string("op_1940_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_1940_end_0 = const()[name = string("op_1940_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_1940_end_mask_0 = const()[name = string("op_1940_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1940_squeeze_mask_0 = const()[name = string("op_1940_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1940_cast_fp16 = slice_by_index(begin = var_1940_begin_0, end = var_1940_end_0, end_mask = var_1940_end_mask_0, squeeze_mask = var_1940_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1940_cast_fp16")]; + bool var_1941_transpose_x_0 = const()[name = string("op_1941_transpose_x_0"), val = bool(false)]; + bool var_1941_transpose_y_0 = const()[name = string("op_1941_transpose_y_0"), val = bool(false)]; + tensor var_1941_cast_fp16 = matmul(transpose_x = var_1941_transpose_x_0, transpose_y = var_1941_transpose_y_0, x = var_1940_cast_fp16, y = state_105_cast_fp16)[name = string("op_1941_cast_fp16")]; + tensor v_new_105_cast_fp16 = sub(x = var_1938_cast_fp16, y = var_1941_cast_fp16)[name = string("v_new_105_cast_fp16")]; + tensor var_1944_begin_0 = const()[name = string("op_1944_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_1944_end_0 = const()[name = string("op_1944_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_1944_end_mask_0 = const()[name = string("op_1944_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1944_squeeze_mask_0 = const()[name = string("op_1944_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1944_cast_fp16 = slice_by_index(begin = var_1944_begin_0, end = var_1944_end_0, end_mask = var_1944_end_mask_0, squeeze_mask = var_1944_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1944_cast_fp16")]; + bool var_1945_transpose_x_0 = const()[name = string("op_1945_transpose_x_0"), val = bool(false)]; + bool var_1945_transpose_y_0 = const()[name = string("op_1945_transpose_y_0"), val = bool(false)]; + tensor var_1945_cast_fp16 = matmul(transpose_x = var_1945_transpose_x_0, transpose_y = var_1945_transpose_y_0, x = var_1944_cast_fp16, y = state_105_cast_fp16)[name = string("op_1945_cast_fp16")]; + tensor var_1947_begin_0 = const()[name = string("op_1947_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_1947_end_0 = const()[name = string("op_1947_end_0"), val = tensor([32, 21, 64, 64])]; + tensor var_1947_end_mask_0 = const()[name = string("op_1947_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1947_squeeze_mask_0 = const()[name = string("op_1947_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1947_cast_fp16 = slice_by_index(begin = var_1947_begin_0, end = var_1947_end_0, end_mask = var_1947_end_mask_0, squeeze_mask = var_1947_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1947_cast_fp16")]; + bool var_1948_transpose_x_0 = const()[name = string("op_1948_transpose_x_0"), val = bool(false)]; + bool var_1948_transpose_y_0 = const()[name = string("op_1948_transpose_y_0"), val = bool(false)]; + tensor var_1948_cast_fp16 = matmul(transpose_x = var_1948_transpose_x_0, transpose_y = var_1948_transpose_y_0, x = var_1947_cast_fp16, y = v_new_105_cast_fp16)[name = string("op_1948_cast_fp16")]; + tensor var_1949_cast_fp16 = add(x = var_1945_cast_fp16, y = var_1948_cast_fp16)[name = string("op_1949_cast_fp16")]; + tensor var_1951_begin_0 = const()[name = string("op_1951_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_1951_end_0 = const()[name = string("op_1951_end_0"), val = tensor([32, 21, 1, 1])]; + tensor var_1951_end_mask_0 = const()[name = string("op_1951_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1951_squeeze_mask_0 = const()[name = string("op_1951_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1951_cast_fp16 = slice_by_index(begin = var_1951_begin_0, end = var_1951_end_0, end_mask = var_1951_end_mask_0, squeeze_mask = var_1951_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1951_cast_fp16")]; + tensor var_1952_cast_fp16 = mul(x = state_105_cast_fp16, y = var_1951_cast_fp16)[name = string("op_1952_cast_fp16")]; + tensor var_1954_begin_0 = const()[name = string("op_1954_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_1954_end_0 = const()[name = string("op_1954_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_1954_end_mask_0 = const()[name = string("op_1954_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1954_squeeze_mask_0 = const()[name = string("op_1954_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1954_cast_fp16 = slice_by_index(begin = var_1954_begin_0, end = var_1954_end_0, end_mask = var_1954_end_mask_0, squeeze_mask = var_1954_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1954_cast_fp16")]; + bool var_1956_transpose_x_1 = const()[name = string("op_1956_transpose_x_1"), val = bool(true)]; + bool var_1956_transpose_y_1 = const()[name = string("op_1956_transpose_y_1"), val = bool(false)]; + tensor var_1956_cast_fp16 = matmul(transpose_x = var_1956_transpose_x_1, transpose_y = var_1956_transpose_y_1, x = var_1954_cast_fp16, y = v_new_105_cast_fp16)[name = string("op_1956_cast_fp16")]; + tensor state_107_cast_fp16 = add(x = var_1952_cast_fp16, y = var_1956_cast_fp16)[name = string("state_107_cast_fp16")]; + tensor var_1959_begin_0 = const()[name = string("op_1959_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_1959_end_0 = const()[name = string("op_1959_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_1959_end_mask_0 = const()[name = string("op_1959_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1959_squeeze_mask_0 = const()[name = string("op_1959_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1959_cast_fp16 = slice_by_index(begin = var_1959_begin_0, end = var_1959_end_0, end_mask = var_1959_end_mask_0, squeeze_mask = var_1959_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1959_cast_fp16")]; + tensor var_1961_begin_0 = const()[name = string("op_1961_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_1961_end_0 = const()[name = string("op_1961_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_1961_end_mask_0 = const()[name = string("op_1961_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1961_squeeze_mask_0 = const()[name = string("op_1961_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1961_cast_fp16 = slice_by_index(begin = var_1961_begin_0, end = var_1961_end_0, end_mask = var_1961_end_mask_0, squeeze_mask = var_1961_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1961_cast_fp16")]; + bool var_1962_transpose_x_0 = const()[name = string("op_1962_transpose_x_0"), val = bool(false)]; + bool var_1962_transpose_y_0 = const()[name = string("op_1962_transpose_y_0"), val = bool(false)]; + tensor var_1962_cast_fp16 = matmul(transpose_x = var_1962_transpose_x_0, transpose_y = var_1962_transpose_y_0, x = var_1961_cast_fp16, y = state_107_cast_fp16)[name = string("op_1962_cast_fp16")]; + tensor v_new_107_cast_fp16 = sub(x = var_1959_cast_fp16, y = var_1962_cast_fp16)[name = string("v_new_107_cast_fp16")]; + tensor var_1965_begin_0 = const()[name = string("op_1965_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_1965_end_0 = const()[name = string("op_1965_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_1965_end_mask_0 = const()[name = string("op_1965_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1965_squeeze_mask_0 = const()[name = string("op_1965_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1965_cast_fp16 = slice_by_index(begin = var_1965_begin_0, end = var_1965_end_0, end_mask = var_1965_end_mask_0, squeeze_mask = var_1965_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1965_cast_fp16")]; + bool var_1966_transpose_x_0 = const()[name = string("op_1966_transpose_x_0"), val = bool(false)]; + bool var_1966_transpose_y_0 = const()[name = string("op_1966_transpose_y_0"), val = bool(false)]; + tensor var_1966_cast_fp16 = matmul(transpose_x = var_1966_transpose_x_0, transpose_y = var_1966_transpose_y_0, x = var_1965_cast_fp16, y = state_107_cast_fp16)[name = string("op_1966_cast_fp16")]; + tensor var_1968_begin_0 = const()[name = string("op_1968_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_1968_end_0 = const()[name = string("op_1968_end_0"), val = tensor([32, 22, 64, 64])]; + tensor var_1968_end_mask_0 = const()[name = string("op_1968_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1968_squeeze_mask_0 = const()[name = string("op_1968_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1968_cast_fp16 = slice_by_index(begin = var_1968_begin_0, end = var_1968_end_0, end_mask = var_1968_end_mask_0, squeeze_mask = var_1968_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1968_cast_fp16")]; + bool var_1969_transpose_x_0 = const()[name = string("op_1969_transpose_x_0"), val = bool(false)]; + bool var_1969_transpose_y_0 = const()[name = string("op_1969_transpose_y_0"), val = bool(false)]; + tensor var_1969_cast_fp16 = matmul(transpose_x = var_1969_transpose_x_0, transpose_y = var_1969_transpose_y_0, x = var_1968_cast_fp16, y = v_new_107_cast_fp16)[name = string("op_1969_cast_fp16")]; + tensor var_1970_cast_fp16 = add(x = var_1966_cast_fp16, y = var_1969_cast_fp16)[name = string("op_1970_cast_fp16")]; + tensor var_1972_begin_0 = const()[name = string("op_1972_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_1972_end_0 = const()[name = string("op_1972_end_0"), val = tensor([32, 22, 1, 1])]; + tensor var_1972_end_mask_0 = const()[name = string("op_1972_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1972_squeeze_mask_0 = const()[name = string("op_1972_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1972_cast_fp16 = slice_by_index(begin = var_1972_begin_0, end = var_1972_end_0, end_mask = var_1972_end_mask_0, squeeze_mask = var_1972_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1972_cast_fp16")]; + tensor var_1973_cast_fp16 = mul(x = state_107_cast_fp16, y = var_1972_cast_fp16)[name = string("op_1973_cast_fp16")]; + tensor var_1975_begin_0 = const()[name = string("op_1975_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_1975_end_0 = const()[name = string("op_1975_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_1975_end_mask_0 = const()[name = string("op_1975_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1975_squeeze_mask_0 = const()[name = string("op_1975_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1975_cast_fp16 = slice_by_index(begin = var_1975_begin_0, end = var_1975_end_0, end_mask = var_1975_end_mask_0, squeeze_mask = var_1975_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1975_cast_fp16")]; + bool var_1977_transpose_x_1 = const()[name = string("op_1977_transpose_x_1"), val = bool(true)]; + bool var_1977_transpose_y_1 = const()[name = string("op_1977_transpose_y_1"), val = bool(false)]; + tensor var_1977_cast_fp16 = matmul(transpose_x = var_1977_transpose_x_1, transpose_y = var_1977_transpose_y_1, x = var_1975_cast_fp16, y = v_new_107_cast_fp16)[name = string("op_1977_cast_fp16")]; + tensor state_109_cast_fp16 = add(x = var_1973_cast_fp16, y = var_1977_cast_fp16)[name = string("state_109_cast_fp16")]; + tensor var_1980_begin_0 = const()[name = string("op_1980_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_1980_end_0 = const()[name = string("op_1980_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_1980_end_mask_0 = const()[name = string("op_1980_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1980_squeeze_mask_0 = const()[name = string("op_1980_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1980_cast_fp16 = slice_by_index(begin = var_1980_begin_0, end = var_1980_end_0, end_mask = var_1980_end_mask_0, squeeze_mask = var_1980_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_1980_cast_fp16")]; + tensor var_1982_begin_0 = const()[name = string("op_1982_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_1982_end_0 = const()[name = string("op_1982_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_1982_end_mask_0 = const()[name = string("op_1982_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1982_squeeze_mask_0 = const()[name = string("op_1982_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1982_cast_fp16 = slice_by_index(begin = var_1982_begin_0, end = var_1982_end_0, end_mask = var_1982_end_mask_0, squeeze_mask = var_1982_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_1982_cast_fp16")]; + bool var_1983_transpose_x_0 = const()[name = string("op_1983_transpose_x_0"), val = bool(false)]; + bool var_1983_transpose_y_0 = const()[name = string("op_1983_transpose_y_0"), val = bool(false)]; + tensor var_1983_cast_fp16 = matmul(transpose_x = var_1983_transpose_x_0, transpose_y = var_1983_transpose_y_0, x = var_1982_cast_fp16, y = state_109_cast_fp16)[name = string("op_1983_cast_fp16")]; + tensor v_new_109_cast_fp16 = sub(x = var_1980_cast_fp16, y = var_1983_cast_fp16)[name = string("v_new_109_cast_fp16")]; + tensor var_1986_begin_0 = const()[name = string("op_1986_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_1986_end_0 = const()[name = string("op_1986_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_1986_end_mask_0 = const()[name = string("op_1986_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1986_squeeze_mask_0 = const()[name = string("op_1986_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1986_cast_fp16 = slice_by_index(begin = var_1986_begin_0, end = var_1986_end_0, end_mask = var_1986_end_mask_0, squeeze_mask = var_1986_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_1986_cast_fp16")]; + bool var_1987_transpose_x_0 = const()[name = string("op_1987_transpose_x_0"), val = bool(false)]; + bool var_1987_transpose_y_0 = const()[name = string("op_1987_transpose_y_0"), val = bool(false)]; + tensor var_1987_cast_fp16 = matmul(transpose_x = var_1987_transpose_x_0, transpose_y = var_1987_transpose_y_0, x = var_1986_cast_fp16, y = state_109_cast_fp16)[name = string("op_1987_cast_fp16")]; + tensor var_1989_begin_0 = const()[name = string("op_1989_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_1989_end_0 = const()[name = string("op_1989_end_0"), val = tensor([32, 23, 64, 64])]; + tensor var_1989_end_mask_0 = const()[name = string("op_1989_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1989_squeeze_mask_0 = const()[name = string("op_1989_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1989_cast_fp16 = slice_by_index(begin = var_1989_begin_0, end = var_1989_end_0, end_mask = var_1989_end_mask_0, squeeze_mask = var_1989_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_1989_cast_fp16")]; + bool var_1990_transpose_x_0 = const()[name = string("op_1990_transpose_x_0"), val = bool(false)]; + bool var_1990_transpose_y_0 = const()[name = string("op_1990_transpose_y_0"), val = bool(false)]; + tensor var_1990_cast_fp16 = matmul(transpose_x = var_1990_transpose_x_0, transpose_y = var_1990_transpose_y_0, x = var_1989_cast_fp16, y = v_new_109_cast_fp16)[name = string("op_1990_cast_fp16")]; + tensor var_1991_cast_fp16 = add(x = var_1987_cast_fp16, y = var_1990_cast_fp16)[name = string("op_1991_cast_fp16")]; + tensor var_1993_begin_0 = const()[name = string("op_1993_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_1993_end_0 = const()[name = string("op_1993_end_0"), val = tensor([32, 23, 1, 1])]; + tensor var_1993_end_mask_0 = const()[name = string("op_1993_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1993_squeeze_mask_0 = const()[name = string("op_1993_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1993_cast_fp16 = slice_by_index(begin = var_1993_begin_0, end = var_1993_end_0, end_mask = var_1993_end_mask_0, squeeze_mask = var_1993_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_1993_cast_fp16")]; + tensor var_1994_cast_fp16 = mul(x = state_109_cast_fp16, y = var_1993_cast_fp16)[name = string("op_1994_cast_fp16")]; + tensor var_1996_begin_0 = const()[name = string("op_1996_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_1996_end_0 = const()[name = string("op_1996_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_1996_end_mask_0 = const()[name = string("op_1996_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1996_squeeze_mask_0 = const()[name = string("op_1996_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_1996_cast_fp16 = slice_by_index(begin = var_1996_begin_0, end = var_1996_end_0, end_mask = var_1996_end_mask_0, squeeze_mask = var_1996_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_1996_cast_fp16")]; + bool var_1998_transpose_x_1 = const()[name = string("op_1998_transpose_x_1"), val = bool(true)]; + bool var_1998_transpose_y_1 = const()[name = string("op_1998_transpose_y_1"), val = bool(false)]; + tensor var_1998_cast_fp16 = matmul(transpose_x = var_1998_transpose_x_1, transpose_y = var_1998_transpose_y_1, x = var_1996_cast_fp16, y = v_new_109_cast_fp16)[name = string("op_1998_cast_fp16")]; + tensor state_111_cast_fp16 = add(x = var_1994_cast_fp16, y = var_1998_cast_fp16)[name = string("state_111_cast_fp16")]; + tensor var_2001_begin_0 = const()[name = string("op_2001_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_2001_end_0 = const()[name = string("op_2001_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_2001_end_mask_0 = const()[name = string("op_2001_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2001_squeeze_mask_0 = const()[name = string("op_2001_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2001_cast_fp16 = slice_by_index(begin = var_2001_begin_0, end = var_2001_end_0, end_mask = var_2001_end_mask_0, squeeze_mask = var_2001_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2001_cast_fp16")]; + tensor var_2003_begin_0 = const()[name = string("op_2003_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_2003_end_0 = const()[name = string("op_2003_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_2003_end_mask_0 = const()[name = string("op_2003_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2003_squeeze_mask_0 = const()[name = string("op_2003_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2003_cast_fp16 = slice_by_index(begin = var_2003_begin_0, end = var_2003_end_0, end_mask = var_2003_end_mask_0, squeeze_mask = var_2003_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2003_cast_fp16")]; + bool var_2004_transpose_x_0 = const()[name = string("op_2004_transpose_x_0"), val = bool(false)]; + bool var_2004_transpose_y_0 = const()[name = string("op_2004_transpose_y_0"), val = bool(false)]; + tensor var_2004_cast_fp16 = matmul(transpose_x = var_2004_transpose_x_0, transpose_y = var_2004_transpose_y_0, x = var_2003_cast_fp16, y = state_111_cast_fp16)[name = string("op_2004_cast_fp16")]; + tensor v_new_111_cast_fp16 = sub(x = var_2001_cast_fp16, y = var_2004_cast_fp16)[name = string("v_new_111_cast_fp16")]; + tensor var_2007_begin_0 = const()[name = string("op_2007_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_2007_end_0 = const()[name = string("op_2007_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_2007_end_mask_0 = const()[name = string("op_2007_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2007_squeeze_mask_0 = const()[name = string("op_2007_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2007_cast_fp16 = slice_by_index(begin = var_2007_begin_0, end = var_2007_end_0, end_mask = var_2007_end_mask_0, squeeze_mask = var_2007_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2007_cast_fp16")]; + bool var_2008_transpose_x_0 = const()[name = string("op_2008_transpose_x_0"), val = bool(false)]; + bool var_2008_transpose_y_0 = const()[name = string("op_2008_transpose_y_0"), val = bool(false)]; + tensor var_2008_cast_fp16 = matmul(transpose_x = var_2008_transpose_x_0, transpose_y = var_2008_transpose_y_0, x = var_2007_cast_fp16, y = state_111_cast_fp16)[name = string("op_2008_cast_fp16")]; + tensor var_2010_begin_0 = const()[name = string("op_2010_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_2010_end_0 = const()[name = string("op_2010_end_0"), val = tensor([32, 24, 64, 64])]; + tensor var_2010_end_mask_0 = const()[name = string("op_2010_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2010_squeeze_mask_0 = const()[name = string("op_2010_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2010_cast_fp16 = slice_by_index(begin = var_2010_begin_0, end = var_2010_end_0, end_mask = var_2010_end_mask_0, squeeze_mask = var_2010_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2010_cast_fp16")]; + bool var_2011_transpose_x_0 = const()[name = string("op_2011_transpose_x_0"), val = bool(false)]; + bool var_2011_transpose_y_0 = const()[name = string("op_2011_transpose_y_0"), val = bool(false)]; + tensor var_2011_cast_fp16 = matmul(transpose_x = var_2011_transpose_x_0, transpose_y = var_2011_transpose_y_0, x = var_2010_cast_fp16, y = v_new_111_cast_fp16)[name = string("op_2011_cast_fp16")]; + tensor var_2012_cast_fp16 = add(x = var_2008_cast_fp16, y = var_2011_cast_fp16)[name = string("op_2012_cast_fp16")]; + tensor var_2014_begin_0 = const()[name = string("op_2014_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_2014_end_0 = const()[name = string("op_2014_end_0"), val = tensor([32, 24, 1, 1])]; + tensor var_2014_end_mask_0 = const()[name = string("op_2014_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2014_squeeze_mask_0 = const()[name = string("op_2014_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2014_cast_fp16 = slice_by_index(begin = var_2014_begin_0, end = var_2014_end_0, end_mask = var_2014_end_mask_0, squeeze_mask = var_2014_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_2014_cast_fp16")]; + tensor var_2015_cast_fp16 = mul(x = state_111_cast_fp16, y = var_2014_cast_fp16)[name = string("op_2015_cast_fp16")]; + tensor var_2017_begin_0 = const()[name = string("op_2017_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_2017_end_0 = const()[name = string("op_2017_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_2017_end_mask_0 = const()[name = string("op_2017_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2017_squeeze_mask_0 = const()[name = string("op_2017_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2017_cast_fp16 = slice_by_index(begin = var_2017_begin_0, end = var_2017_end_0, end_mask = var_2017_end_mask_0, squeeze_mask = var_2017_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_2017_cast_fp16")]; + bool var_2019_transpose_x_1 = const()[name = string("op_2019_transpose_x_1"), val = bool(true)]; + bool var_2019_transpose_y_1 = const()[name = string("op_2019_transpose_y_1"), val = bool(false)]; + tensor var_2019_cast_fp16 = matmul(transpose_x = var_2019_transpose_x_1, transpose_y = var_2019_transpose_y_1, x = var_2017_cast_fp16, y = v_new_111_cast_fp16)[name = string("op_2019_cast_fp16")]; + tensor state_113_cast_fp16 = add(x = var_2015_cast_fp16, y = var_2019_cast_fp16)[name = string("state_113_cast_fp16")]; + tensor var_2022_begin_0 = const()[name = string("op_2022_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_2022_end_0 = const()[name = string("op_2022_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_2022_end_mask_0 = const()[name = string("op_2022_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2022_squeeze_mask_0 = const()[name = string("op_2022_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2022_cast_fp16 = slice_by_index(begin = var_2022_begin_0, end = var_2022_end_0, end_mask = var_2022_end_mask_0, squeeze_mask = var_2022_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2022_cast_fp16")]; + tensor var_2024_begin_0 = const()[name = string("op_2024_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_2024_end_0 = const()[name = string("op_2024_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_2024_end_mask_0 = const()[name = string("op_2024_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2024_squeeze_mask_0 = const()[name = string("op_2024_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2024_cast_fp16 = slice_by_index(begin = var_2024_begin_0, end = var_2024_end_0, end_mask = var_2024_end_mask_0, squeeze_mask = var_2024_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2024_cast_fp16")]; + bool var_2025_transpose_x_0 = const()[name = string("op_2025_transpose_x_0"), val = bool(false)]; + bool var_2025_transpose_y_0 = const()[name = string("op_2025_transpose_y_0"), val = bool(false)]; + tensor var_2025_cast_fp16 = matmul(transpose_x = var_2025_transpose_x_0, transpose_y = var_2025_transpose_y_0, x = var_2024_cast_fp16, y = state_113_cast_fp16)[name = string("op_2025_cast_fp16")]; + tensor v_new_113_cast_fp16 = sub(x = var_2022_cast_fp16, y = var_2025_cast_fp16)[name = string("v_new_113_cast_fp16")]; + tensor var_2028_begin_0 = const()[name = string("op_2028_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_2028_end_0 = const()[name = string("op_2028_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_2028_end_mask_0 = const()[name = string("op_2028_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2028_squeeze_mask_0 = const()[name = string("op_2028_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2028_cast_fp16 = slice_by_index(begin = var_2028_begin_0, end = var_2028_end_0, end_mask = var_2028_end_mask_0, squeeze_mask = var_2028_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2028_cast_fp16")]; + bool var_2029_transpose_x_0 = const()[name = string("op_2029_transpose_x_0"), val = bool(false)]; + bool var_2029_transpose_y_0 = const()[name = string("op_2029_transpose_y_0"), val = bool(false)]; + tensor var_2029_cast_fp16 = matmul(transpose_x = var_2029_transpose_x_0, transpose_y = var_2029_transpose_y_0, x = var_2028_cast_fp16, y = state_113_cast_fp16)[name = string("op_2029_cast_fp16")]; + tensor var_2031_begin_0 = const()[name = string("op_2031_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_2031_end_0 = const()[name = string("op_2031_end_0"), val = tensor([32, 25, 64, 64])]; + tensor var_2031_end_mask_0 = const()[name = string("op_2031_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2031_squeeze_mask_0 = const()[name = string("op_2031_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2031_cast_fp16 = slice_by_index(begin = var_2031_begin_0, end = var_2031_end_0, end_mask = var_2031_end_mask_0, squeeze_mask = var_2031_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2031_cast_fp16")]; + bool var_2032_transpose_x_0 = const()[name = string("op_2032_transpose_x_0"), val = bool(false)]; + bool var_2032_transpose_y_0 = const()[name = string("op_2032_transpose_y_0"), val = bool(false)]; + tensor var_2032_cast_fp16 = matmul(transpose_x = var_2032_transpose_x_0, transpose_y = var_2032_transpose_y_0, x = var_2031_cast_fp16, y = v_new_113_cast_fp16)[name = string("op_2032_cast_fp16")]; + tensor var_2033_cast_fp16 = add(x = var_2029_cast_fp16, y = var_2032_cast_fp16)[name = string("op_2033_cast_fp16")]; + tensor var_2035_begin_0 = const()[name = string("op_2035_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_2035_end_0 = const()[name = string("op_2035_end_0"), val = tensor([32, 25, 1, 1])]; + tensor var_2035_end_mask_0 = const()[name = string("op_2035_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2035_squeeze_mask_0 = const()[name = string("op_2035_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2035_cast_fp16 = slice_by_index(begin = var_2035_begin_0, end = var_2035_end_0, end_mask = var_2035_end_mask_0, squeeze_mask = var_2035_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_2035_cast_fp16")]; + tensor var_2036_cast_fp16 = mul(x = state_113_cast_fp16, y = var_2035_cast_fp16)[name = string("op_2036_cast_fp16")]; + tensor var_2038_begin_0 = const()[name = string("op_2038_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_2038_end_0 = const()[name = string("op_2038_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_2038_end_mask_0 = const()[name = string("op_2038_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2038_squeeze_mask_0 = const()[name = string("op_2038_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2038_cast_fp16 = slice_by_index(begin = var_2038_begin_0, end = var_2038_end_0, end_mask = var_2038_end_mask_0, squeeze_mask = var_2038_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_2038_cast_fp16")]; + bool var_2040_transpose_x_1 = const()[name = string("op_2040_transpose_x_1"), val = bool(true)]; + bool var_2040_transpose_y_1 = const()[name = string("op_2040_transpose_y_1"), val = bool(false)]; + tensor var_2040_cast_fp16 = matmul(transpose_x = var_2040_transpose_x_1, transpose_y = var_2040_transpose_y_1, x = var_2038_cast_fp16, y = v_new_113_cast_fp16)[name = string("op_2040_cast_fp16")]; + tensor state_115_cast_fp16 = add(x = var_2036_cast_fp16, y = var_2040_cast_fp16)[name = string("state_115_cast_fp16")]; + tensor var_2043_begin_0 = const()[name = string("op_2043_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_2043_end_0 = const()[name = string("op_2043_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_2043_end_mask_0 = const()[name = string("op_2043_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2043_squeeze_mask_0 = const()[name = string("op_2043_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2043_cast_fp16 = slice_by_index(begin = var_2043_begin_0, end = var_2043_end_0, end_mask = var_2043_end_mask_0, squeeze_mask = var_2043_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2043_cast_fp16")]; + tensor var_2045_begin_0 = const()[name = string("op_2045_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_2045_end_0 = const()[name = string("op_2045_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_2045_end_mask_0 = const()[name = string("op_2045_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2045_squeeze_mask_0 = const()[name = string("op_2045_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2045_cast_fp16 = slice_by_index(begin = var_2045_begin_0, end = var_2045_end_0, end_mask = var_2045_end_mask_0, squeeze_mask = var_2045_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2045_cast_fp16")]; + bool var_2046_transpose_x_0 = const()[name = string("op_2046_transpose_x_0"), val = bool(false)]; + bool var_2046_transpose_y_0 = const()[name = string("op_2046_transpose_y_0"), val = bool(false)]; + tensor var_2046_cast_fp16 = matmul(transpose_x = var_2046_transpose_x_0, transpose_y = var_2046_transpose_y_0, x = var_2045_cast_fp16, y = state_115_cast_fp16)[name = string("op_2046_cast_fp16")]; + tensor v_new_115_cast_fp16 = sub(x = var_2043_cast_fp16, y = var_2046_cast_fp16)[name = string("v_new_115_cast_fp16")]; + tensor var_2049_begin_0 = const()[name = string("op_2049_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_2049_end_0 = const()[name = string("op_2049_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_2049_end_mask_0 = const()[name = string("op_2049_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2049_squeeze_mask_0 = const()[name = string("op_2049_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2049_cast_fp16 = slice_by_index(begin = var_2049_begin_0, end = var_2049_end_0, end_mask = var_2049_end_mask_0, squeeze_mask = var_2049_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2049_cast_fp16")]; + bool var_2050_transpose_x_0 = const()[name = string("op_2050_transpose_x_0"), val = bool(false)]; + bool var_2050_transpose_y_0 = const()[name = string("op_2050_transpose_y_0"), val = bool(false)]; + tensor var_2050_cast_fp16 = matmul(transpose_x = var_2050_transpose_x_0, transpose_y = var_2050_transpose_y_0, x = var_2049_cast_fp16, y = state_115_cast_fp16)[name = string("op_2050_cast_fp16")]; + tensor var_2052_begin_0 = const()[name = string("op_2052_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_2052_end_0 = const()[name = string("op_2052_end_0"), val = tensor([32, 26, 64, 64])]; + tensor var_2052_end_mask_0 = const()[name = string("op_2052_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2052_squeeze_mask_0 = const()[name = string("op_2052_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2052_cast_fp16 = slice_by_index(begin = var_2052_begin_0, end = var_2052_end_0, end_mask = var_2052_end_mask_0, squeeze_mask = var_2052_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2052_cast_fp16")]; + bool var_2053_transpose_x_0 = const()[name = string("op_2053_transpose_x_0"), val = bool(false)]; + bool var_2053_transpose_y_0 = const()[name = string("op_2053_transpose_y_0"), val = bool(false)]; + tensor var_2053_cast_fp16 = matmul(transpose_x = var_2053_transpose_x_0, transpose_y = var_2053_transpose_y_0, x = var_2052_cast_fp16, y = v_new_115_cast_fp16)[name = string("op_2053_cast_fp16")]; + tensor var_2054_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2053_cast_fp16)[name = string("op_2054_cast_fp16")]; + tensor var_2056_begin_0 = const()[name = string("op_2056_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_2056_end_0 = const()[name = string("op_2056_end_0"), val = tensor([32, 26, 1, 1])]; + tensor var_2056_end_mask_0 = const()[name = string("op_2056_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2056_squeeze_mask_0 = const()[name = string("op_2056_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2056_cast_fp16 = slice_by_index(begin = var_2056_begin_0, end = var_2056_end_0, end_mask = var_2056_end_mask_0, squeeze_mask = var_2056_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_2056_cast_fp16")]; + tensor var_2057_cast_fp16 = mul(x = state_115_cast_fp16, y = var_2056_cast_fp16)[name = string("op_2057_cast_fp16")]; + tensor var_2059_begin_0 = const()[name = string("op_2059_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_2059_end_0 = const()[name = string("op_2059_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_2059_end_mask_0 = const()[name = string("op_2059_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2059_squeeze_mask_0 = const()[name = string("op_2059_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2059_cast_fp16 = slice_by_index(begin = var_2059_begin_0, end = var_2059_end_0, end_mask = var_2059_end_mask_0, squeeze_mask = var_2059_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_2059_cast_fp16")]; + bool var_2061_transpose_x_1 = const()[name = string("op_2061_transpose_x_1"), val = bool(true)]; + bool var_2061_transpose_y_1 = const()[name = string("op_2061_transpose_y_1"), val = bool(false)]; + tensor var_2061_cast_fp16 = matmul(transpose_x = var_2061_transpose_x_1, transpose_y = var_2061_transpose_y_1, x = var_2059_cast_fp16, y = v_new_115_cast_fp16)[name = string("op_2061_cast_fp16")]; + tensor state_117_cast_fp16 = add(x = var_2057_cast_fp16, y = var_2061_cast_fp16)[name = string("state_117_cast_fp16")]; + tensor var_2064_begin_0 = const()[name = string("op_2064_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_2064_end_0 = const()[name = string("op_2064_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_2064_end_mask_0 = const()[name = string("op_2064_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2064_squeeze_mask_0 = const()[name = string("op_2064_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2064_cast_fp16 = slice_by_index(begin = var_2064_begin_0, end = var_2064_end_0, end_mask = var_2064_end_mask_0, squeeze_mask = var_2064_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2064_cast_fp16")]; + tensor var_2066_begin_0 = const()[name = string("op_2066_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_2066_end_0 = const()[name = string("op_2066_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_2066_end_mask_0 = const()[name = string("op_2066_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2066_squeeze_mask_0 = const()[name = string("op_2066_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2066_cast_fp16 = slice_by_index(begin = var_2066_begin_0, end = var_2066_end_0, end_mask = var_2066_end_mask_0, squeeze_mask = var_2066_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2066_cast_fp16")]; + bool var_2067_transpose_x_0 = const()[name = string("op_2067_transpose_x_0"), val = bool(false)]; + bool var_2067_transpose_y_0 = const()[name = string("op_2067_transpose_y_0"), val = bool(false)]; + tensor var_2067_cast_fp16 = matmul(transpose_x = var_2067_transpose_x_0, transpose_y = var_2067_transpose_y_0, x = var_2066_cast_fp16, y = state_117_cast_fp16)[name = string("op_2067_cast_fp16")]; + tensor v_new_117_cast_fp16 = sub(x = var_2064_cast_fp16, y = var_2067_cast_fp16)[name = string("v_new_117_cast_fp16")]; + tensor var_2070_begin_0 = const()[name = string("op_2070_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_2070_end_0 = const()[name = string("op_2070_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_2070_end_mask_0 = const()[name = string("op_2070_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2070_squeeze_mask_0 = const()[name = string("op_2070_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2070_cast_fp16 = slice_by_index(begin = var_2070_begin_0, end = var_2070_end_0, end_mask = var_2070_end_mask_0, squeeze_mask = var_2070_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2070_cast_fp16")]; + bool var_2071_transpose_x_0 = const()[name = string("op_2071_transpose_x_0"), val = bool(false)]; + bool var_2071_transpose_y_0 = const()[name = string("op_2071_transpose_y_0"), val = bool(false)]; + tensor var_2071_cast_fp16 = matmul(transpose_x = var_2071_transpose_x_0, transpose_y = var_2071_transpose_y_0, x = var_2070_cast_fp16, y = state_117_cast_fp16)[name = string("op_2071_cast_fp16")]; + tensor var_2073_begin_0 = const()[name = string("op_2073_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_2073_end_0 = const()[name = string("op_2073_end_0"), val = tensor([32, 27, 64, 64])]; + tensor var_2073_end_mask_0 = const()[name = string("op_2073_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2073_squeeze_mask_0 = const()[name = string("op_2073_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2073_cast_fp16 = slice_by_index(begin = var_2073_begin_0, end = var_2073_end_0, end_mask = var_2073_end_mask_0, squeeze_mask = var_2073_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2073_cast_fp16")]; + bool var_2074_transpose_x_0 = const()[name = string("op_2074_transpose_x_0"), val = bool(false)]; + bool var_2074_transpose_y_0 = const()[name = string("op_2074_transpose_y_0"), val = bool(false)]; + tensor var_2074_cast_fp16 = matmul(transpose_x = var_2074_transpose_x_0, transpose_y = var_2074_transpose_y_0, x = var_2073_cast_fp16, y = v_new_117_cast_fp16)[name = string("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = add(x = var_2071_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")]; + tensor var_2077_begin_0 = const()[name = string("op_2077_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_2077_end_0 = const()[name = string("op_2077_end_0"), val = tensor([32, 27, 1, 1])]; + tensor var_2077_end_mask_0 = const()[name = string("op_2077_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2077_squeeze_mask_0 = const()[name = string("op_2077_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2077_cast_fp16 = slice_by_index(begin = var_2077_begin_0, end = var_2077_end_0, end_mask = var_2077_end_mask_0, squeeze_mask = var_2077_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_2077_cast_fp16")]; + tensor var_2078_cast_fp16 = mul(x = state_117_cast_fp16, y = var_2077_cast_fp16)[name = string("op_2078_cast_fp16")]; + tensor var_2080_begin_0 = const()[name = string("op_2080_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_2080_end_0 = const()[name = string("op_2080_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_2080_end_mask_0 = const()[name = string("op_2080_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2080_squeeze_mask_0 = const()[name = string("op_2080_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2080_cast_fp16 = slice_by_index(begin = var_2080_begin_0, end = var_2080_end_0, end_mask = var_2080_end_mask_0, squeeze_mask = var_2080_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_2080_cast_fp16")]; + bool var_2082_transpose_x_1 = const()[name = string("op_2082_transpose_x_1"), val = bool(true)]; + bool var_2082_transpose_y_1 = const()[name = string("op_2082_transpose_y_1"), val = bool(false)]; + tensor var_2082_cast_fp16 = matmul(transpose_x = var_2082_transpose_x_1, transpose_y = var_2082_transpose_y_1, x = var_2080_cast_fp16, y = v_new_117_cast_fp16)[name = string("op_2082_cast_fp16")]; + tensor state_119_cast_fp16 = add(x = var_2078_cast_fp16, y = var_2082_cast_fp16)[name = string("state_119_cast_fp16")]; + tensor var_2085_begin_0 = const()[name = string("op_2085_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_2085_end_0 = const()[name = string("op_2085_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_2085_end_mask_0 = const()[name = string("op_2085_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2085_squeeze_mask_0 = const()[name = string("op_2085_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2085_cast_fp16 = slice_by_index(begin = var_2085_begin_0, end = var_2085_end_0, end_mask = var_2085_end_mask_0, squeeze_mask = var_2085_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2085_cast_fp16")]; + tensor var_2087_begin_0 = const()[name = string("op_2087_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_2087_end_0 = const()[name = string("op_2087_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_2087_end_mask_0 = const()[name = string("op_2087_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2087_squeeze_mask_0 = const()[name = string("op_2087_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2087_cast_fp16 = slice_by_index(begin = var_2087_begin_0, end = var_2087_end_0, end_mask = var_2087_end_mask_0, squeeze_mask = var_2087_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2087_cast_fp16")]; + bool var_2088_transpose_x_0 = const()[name = string("op_2088_transpose_x_0"), val = bool(false)]; + bool var_2088_transpose_y_0 = const()[name = string("op_2088_transpose_y_0"), val = bool(false)]; + tensor var_2088_cast_fp16 = matmul(transpose_x = var_2088_transpose_x_0, transpose_y = var_2088_transpose_y_0, x = var_2087_cast_fp16, y = state_119_cast_fp16)[name = string("op_2088_cast_fp16")]; + tensor v_new_119_cast_fp16 = sub(x = var_2085_cast_fp16, y = var_2088_cast_fp16)[name = string("v_new_119_cast_fp16")]; + tensor var_2091_begin_0 = const()[name = string("op_2091_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_2091_end_0 = const()[name = string("op_2091_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_2091_end_mask_0 = const()[name = string("op_2091_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2091_squeeze_mask_0 = const()[name = string("op_2091_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2091_cast_fp16 = slice_by_index(begin = var_2091_begin_0, end = var_2091_end_0, end_mask = var_2091_end_mask_0, squeeze_mask = var_2091_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2091_cast_fp16")]; + bool var_2092_transpose_x_0 = const()[name = string("op_2092_transpose_x_0"), val = bool(false)]; + bool var_2092_transpose_y_0 = const()[name = string("op_2092_transpose_y_0"), val = bool(false)]; + tensor var_2092_cast_fp16 = matmul(transpose_x = var_2092_transpose_x_0, transpose_y = var_2092_transpose_y_0, x = var_2091_cast_fp16, y = state_119_cast_fp16)[name = string("op_2092_cast_fp16")]; + tensor var_2094_begin_0 = const()[name = string("op_2094_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_2094_end_0 = const()[name = string("op_2094_end_0"), val = tensor([32, 28, 64, 64])]; + tensor var_2094_end_mask_0 = const()[name = string("op_2094_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2094_squeeze_mask_0 = const()[name = string("op_2094_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2094_cast_fp16 = slice_by_index(begin = var_2094_begin_0, end = var_2094_end_0, end_mask = var_2094_end_mask_0, squeeze_mask = var_2094_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2094_cast_fp16")]; + bool var_2095_transpose_x_0 = const()[name = string("op_2095_transpose_x_0"), val = bool(false)]; + bool var_2095_transpose_y_0 = const()[name = string("op_2095_transpose_y_0"), val = bool(false)]; + tensor var_2095_cast_fp16 = matmul(transpose_x = var_2095_transpose_x_0, transpose_y = var_2095_transpose_y_0, x = var_2094_cast_fp16, y = v_new_119_cast_fp16)[name = string("op_2095_cast_fp16")]; + tensor var_2096_cast_fp16 = add(x = var_2092_cast_fp16, y = var_2095_cast_fp16)[name = string("op_2096_cast_fp16")]; + tensor var_2098_begin_0 = const()[name = string("op_2098_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_2098_end_0 = const()[name = string("op_2098_end_0"), val = tensor([32, 28, 1, 1])]; + tensor var_2098_end_mask_0 = const()[name = string("op_2098_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2098_squeeze_mask_0 = const()[name = string("op_2098_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2098_cast_fp16 = slice_by_index(begin = var_2098_begin_0, end = var_2098_end_0, end_mask = var_2098_end_mask_0, squeeze_mask = var_2098_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_2098_cast_fp16")]; + tensor var_2099_cast_fp16 = mul(x = state_119_cast_fp16, y = var_2098_cast_fp16)[name = string("op_2099_cast_fp16")]; + tensor var_2101_begin_0 = const()[name = string("op_2101_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_2101_end_0 = const()[name = string("op_2101_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_2101_end_mask_0 = const()[name = string("op_2101_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2101_squeeze_mask_0 = const()[name = string("op_2101_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2101_cast_fp16 = slice_by_index(begin = var_2101_begin_0, end = var_2101_end_0, end_mask = var_2101_end_mask_0, squeeze_mask = var_2101_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_2101_cast_fp16")]; + bool var_2103_transpose_x_1 = const()[name = string("op_2103_transpose_x_1"), val = bool(true)]; + bool var_2103_transpose_y_1 = const()[name = string("op_2103_transpose_y_1"), val = bool(false)]; + tensor var_2103_cast_fp16 = matmul(transpose_x = var_2103_transpose_x_1, transpose_y = var_2103_transpose_y_1, x = var_2101_cast_fp16, y = v_new_119_cast_fp16)[name = string("op_2103_cast_fp16")]; + tensor state_121_cast_fp16 = add(x = var_2099_cast_fp16, y = var_2103_cast_fp16)[name = string("state_121_cast_fp16")]; + tensor var_2106_begin_0 = const()[name = string("op_2106_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_2106_end_0 = const()[name = string("op_2106_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_2106_end_mask_0 = const()[name = string("op_2106_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2106_squeeze_mask_0 = const()[name = string("op_2106_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2106_cast_fp16 = slice_by_index(begin = var_2106_begin_0, end = var_2106_end_0, end_mask = var_2106_end_mask_0, squeeze_mask = var_2106_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2106_cast_fp16")]; + tensor var_2108_begin_0 = const()[name = string("op_2108_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_2108_end_0 = const()[name = string("op_2108_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_2108_end_mask_0 = const()[name = string("op_2108_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2108_squeeze_mask_0 = const()[name = string("op_2108_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2108_cast_fp16 = slice_by_index(begin = var_2108_begin_0, end = var_2108_end_0, end_mask = var_2108_end_mask_0, squeeze_mask = var_2108_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2108_cast_fp16")]; + bool var_2109_transpose_x_0 = const()[name = string("op_2109_transpose_x_0"), val = bool(false)]; + bool var_2109_transpose_y_0 = const()[name = string("op_2109_transpose_y_0"), val = bool(false)]; + tensor var_2109_cast_fp16 = matmul(transpose_x = var_2109_transpose_x_0, transpose_y = var_2109_transpose_y_0, x = var_2108_cast_fp16, y = state_121_cast_fp16)[name = string("op_2109_cast_fp16")]; + tensor v_new_121_cast_fp16 = sub(x = var_2106_cast_fp16, y = var_2109_cast_fp16)[name = string("v_new_121_cast_fp16")]; + tensor var_2112_begin_0 = const()[name = string("op_2112_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_2112_end_0 = const()[name = string("op_2112_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_2112_end_mask_0 = const()[name = string("op_2112_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2112_squeeze_mask_0 = const()[name = string("op_2112_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2112_cast_fp16 = slice_by_index(begin = var_2112_begin_0, end = var_2112_end_0, end_mask = var_2112_end_mask_0, squeeze_mask = var_2112_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2112_cast_fp16")]; + bool var_2113_transpose_x_0 = const()[name = string("op_2113_transpose_x_0"), val = bool(false)]; + bool var_2113_transpose_y_0 = const()[name = string("op_2113_transpose_y_0"), val = bool(false)]; + tensor var_2113_cast_fp16 = matmul(transpose_x = var_2113_transpose_x_0, transpose_y = var_2113_transpose_y_0, x = var_2112_cast_fp16, y = state_121_cast_fp16)[name = string("op_2113_cast_fp16")]; + tensor var_2115_begin_0 = const()[name = string("op_2115_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_2115_end_0 = const()[name = string("op_2115_end_0"), val = tensor([32, 29, 64, 64])]; + tensor var_2115_end_mask_0 = const()[name = string("op_2115_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2115_squeeze_mask_0 = const()[name = string("op_2115_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2115_cast_fp16 = slice_by_index(begin = var_2115_begin_0, end = var_2115_end_0, end_mask = var_2115_end_mask_0, squeeze_mask = var_2115_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2115_cast_fp16")]; + bool var_2116_transpose_x_0 = const()[name = string("op_2116_transpose_x_0"), val = bool(false)]; + bool var_2116_transpose_y_0 = const()[name = string("op_2116_transpose_y_0"), val = bool(false)]; + tensor var_2116_cast_fp16 = matmul(transpose_x = var_2116_transpose_x_0, transpose_y = var_2116_transpose_y_0, x = var_2115_cast_fp16, y = v_new_121_cast_fp16)[name = string("op_2116_cast_fp16")]; + tensor var_2117_cast_fp16 = add(x = var_2113_cast_fp16, y = var_2116_cast_fp16)[name = string("op_2117_cast_fp16")]; + tensor var_2119_begin_0 = const()[name = string("op_2119_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_2119_end_0 = const()[name = string("op_2119_end_0"), val = tensor([32, 29, 1, 1])]; + tensor var_2119_end_mask_0 = const()[name = string("op_2119_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2119_squeeze_mask_0 = const()[name = string("op_2119_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2119_cast_fp16 = slice_by_index(begin = var_2119_begin_0, end = var_2119_end_0, end_mask = var_2119_end_mask_0, squeeze_mask = var_2119_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_2119_cast_fp16")]; + tensor var_2120_cast_fp16 = mul(x = state_121_cast_fp16, y = var_2119_cast_fp16)[name = string("op_2120_cast_fp16")]; + tensor var_2122_begin_0 = const()[name = string("op_2122_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_2122_end_0 = const()[name = string("op_2122_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_2122_end_mask_0 = const()[name = string("op_2122_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2122_squeeze_mask_0 = const()[name = string("op_2122_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2122_cast_fp16 = slice_by_index(begin = var_2122_begin_0, end = var_2122_end_0, end_mask = var_2122_end_mask_0, squeeze_mask = var_2122_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_2122_cast_fp16")]; + bool var_2124_transpose_x_1 = const()[name = string("op_2124_transpose_x_1"), val = bool(true)]; + bool var_2124_transpose_y_1 = const()[name = string("op_2124_transpose_y_1"), val = bool(false)]; + tensor var_2124_cast_fp16 = matmul(transpose_x = var_2124_transpose_x_1, transpose_y = var_2124_transpose_y_1, x = var_2122_cast_fp16, y = v_new_121_cast_fp16)[name = string("op_2124_cast_fp16")]; + tensor state_123_cast_fp16 = add(x = var_2120_cast_fp16, y = var_2124_cast_fp16)[name = string("state_123_cast_fp16")]; + tensor var_2127_begin_0 = const()[name = string("op_2127_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_2127_end_0 = const()[name = string("op_2127_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_2127_end_mask_0 = const()[name = string("op_2127_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2127_squeeze_mask_0 = const()[name = string("op_2127_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2127_cast_fp16 = slice_by_index(begin = var_2127_begin_0, end = var_2127_end_0, end_mask = var_2127_end_mask_0, squeeze_mask = var_2127_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2127_cast_fp16")]; + tensor var_2129_begin_0 = const()[name = string("op_2129_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_2129_end_0 = const()[name = string("op_2129_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_2129_end_mask_0 = const()[name = string("op_2129_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2129_squeeze_mask_0 = const()[name = string("op_2129_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2129_cast_fp16 = slice_by_index(begin = var_2129_begin_0, end = var_2129_end_0, end_mask = var_2129_end_mask_0, squeeze_mask = var_2129_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2129_cast_fp16")]; + bool var_2130_transpose_x_0 = const()[name = string("op_2130_transpose_x_0"), val = bool(false)]; + bool var_2130_transpose_y_0 = const()[name = string("op_2130_transpose_y_0"), val = bool(false)]; + tensor var_2130_cast_fp16 = matmul(transpose_x = var_2130_transpose_x_0, transpose_y = var_2130_transpose_y_0, x = var_2129_cast_fp16, y = state_123_cast_fp16)[name = string("op_2130_cast_fp16")]; + tensor v_new_123_cast_fp16 = sub(x = var_2127_cast_fp16, y = var_2130_cast_fp16)[name = string("v_new_123_cast_fp16")]; + tensor var_2133_begin_0 = const()[name = string("op_2133_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_2133_end_0 = const()[name = string("op_2133_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_2133_end_mask_0 = const()[name = string("op_2133_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2133_squeeze_mask_0 = const()[name = string("op_2133_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2133_cast_fp16 = slice_by_index(begin = var_2133_begin_0, end = var_2133_end_0, end_mask = var_2133_end_mask_0, squeeze_mask = var_2133_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2133_cast_fp16")]; + bool var_2134_transpose_x_0 = const()[name = string("op_2134_transpose_x_0"), val = bool(false)]; + bool var_2134_transpose_y_0 = const()[name = string("op_2134_transpose_y_0"), val = bool(false)]; + tensor var_2134_cast_fp16 = matmul(transpose_x = var_2134_transpose_x_0, transpose_y = var_2134_transpose_y_0, x = var_2133_cast_fp16, y = state_123_cast_fp16)[name = string("op_2134_cast_fp16")]; + tensor var_2136_begin_0 = const()[name = string("op_2136_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_2136_end_0 = const()[name = string("op_2136_end_0"), val = tensor([32, 30, 64, 64])]; + tensor var_2136_end_mask_0 = const()[name = string("op_2136_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2136_squeeze_mask_0 = const()[name = string("op_2136_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2136_cast_fp16 = slice_by_index(begin = var_2136_begin_0, end = var_2136_end_0, end_mask = var_2136_end_mask_0, squeeze_mask = var_2136_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2136_cast_fp16")]; + bool var_2137_transpose_x_0 = const()[name = string("op_2137_transpose_x_0"), val = bool(false)]; + bool var_2137_transpose_y_0 = const()[name = string("op_2137_transpose_y_0"), val = bool(false)]; + tensor var_2137_cast_fp16 = matmul(transpose_x = var_2137_transpose_x_0, transpose_y = var_2137_transpose_y_0, x = var_2136_cast_fp16, y = v_new_123_cast_fp16)[name = string("op_2137_cast_fp16")]; + tensor var_2138_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2137_cast_fp16)[name = string("op_2138_cast_fp16")]; + tensor var_2140_begin_0 = const()[name = string("op_2140_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_2140_end_0 = const()[name = string("op_2140_end_0"), val = tensor([32, 30, 1, 1])]; + tensor var_2140_end_mask_0 = const()[name = string("op_2140_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2140_squeeze_mask_0 = const()[name = string("op_2140_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2140_cast_fp16 = slice_by_index(begin = var_2140_begin_0, end = var_2140_end_0, end_mask = var_2140_end_mask_0, squeeze_mask = var_2140_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_2140_cast_fp16")]; + tensor var_2141_cast_fp16 = mul(x = state_123_cast_fp16, y = var_2140_cast_fp16)[name = string("op_2141_cast_fp16")]; + tensor var_2143_begin_0 = const()[name = string("op_2143_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_2143_end_0 = const()[name = string("op_2143_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_2143_end_mask_0 = const()[name = string("op_2143_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2143_squeeze_mask_0 = const()[name = string("op_2143_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2143_cast_fp16 = slice_by_index(begin = var_2143_begin_0, end = var_2143_end_0, end_mask = var_2143_end_mask_0, squeeze_mask = var_2143_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_2143_cast_fp16")]; + bool var_2145_transpose_x_1 = const()[name = string("op_2145_transpose_x_1"), val = bool(true)]; + bool var_2145_transpose_y_1 = const()[name = string("op_2145_transpose_y_1"), val = bool(false)]; + tensor var_2145_cast_fp16 = matmul(transpose_x = var_2145_transpose_x_1, transpose_y = var_2145_transpose_y_1, x = var_2143_cast_fp16, y = v_new_123_cast_fp16)[name = string("op_2145_cast_fp16")]; + tensor state_125_cast_fp16 = add(x = var_2141_cast_fp16, y = var_2145_cast_fp16)[name = string("state_125_cast_fp16")]; + tensor var_2148_begin_0 = const()[name = string("op_2148_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_2148_end_0 = const()[name = string("op_2148_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_2148_end_mask_0 = const()[name = string("op_2148_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2148_squeeze_mask_0 = const()[name = string("op_2148_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2148_cast_fp16 = slice_by_index(begin = var_2148_begin_0, end = var_2148_end_0, end_mask = var_2148_end_mask_0, squeeze_mask = var_2148_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2148_cast_fp16")]; + tensor var_2150_begin_0 = const()[name = string("op_2150_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_2150_end_0 = const()[name = string("op_2150_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_2150_end_mask_0 = const()[name = string("op_2150_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2150_squeeze_mask_0 = const()[name = string("op_2150_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2150_cast_fp16 = slice_by_index(begin = var_2150_begin_0, end = var_2150_end_0, end_mask = var_2150_end_mask_0, squeeze_mask = var_2150_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2150_cast_fp16")]; + bool var_2151_transpose_x_0 = const()[name = string("op_2151_transpose_x_0"), val = bool(false)]; + bool var_2151_transpose_y_0 = const()[name = string("op_2151_transpose_y_0"), val = bool(false)]; + tensor var_2151_cast_fp16 = matmul(transpose_x = var_2151_transpose_x_0, transpose_y = var_2151_transpose_y_0, x = var_2150_cast_fp16, y = state_125_cast_fp16)[name = string("op_2151_cast_fp16")]; + tensor v_new_125_cast_fp16 = sub(x = var_2148_cast_fp16, y = var_2151_cast_fp16)[name = string("v_new_125_cast_fp16")]; + tensor var_2154_begin_0 = const()[name = string("op_2154_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_2154_end_0 = const()[name = string("op_2154_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_2154_end_mask_0 = const()[name = string("op_2154_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2154_squeeze_mask_0 = const()[name = string("op_2154_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2154_cast_fp16 = slice_by_index(begin = var_2154_begin_0, end = var_2154_end_0, end_mask = var_2154_end_mask_0, squeeze_mask = var_2154_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2154_cast_fp16")]; + bool var_2155_transpose_x_0 = const()[name = string("op_2155_transpose_x_0"), val = bool(false)]; + bool var_2155_transpose_y_0 = const()[name = string("op_2155_transpose_y_0"), val = bool(false)]; + tensor var_2155_cast_fp16 = matmul(transpose_x = var_2155_transpose_x_0, transpose_y = var_2155_transpose_y_0, x = var_2154_cast_fp16, y = state_125_cast_fp16)[name = string("op_2155_cast_fp16")]; + tensor var_2157_begin_0 = const()[name = string("op_2157_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_2157_end_0 = const()[name = string("op_2157_end_0"), val = tensor([32, 31, 64, 64])]; + tensor var_2157_end_mask_0 = const()[name = string("op_2157_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2157_squeeze_mask_0 = const()[name = string("op_2157_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2157_cast_fp16 = slice_by_index(begin = var_2157_begin_0, end = var_2157_end_0, end_mask = var_2157_end_mask_0, squeeze_mask = var_2157_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2157_cast_fp16")]; + bool var_2158_transpose_x_0 = const()[name = string("op_2158_transpose_x_0"), val = bool(false)]; + bool var_2158_transpose_y_0 = const()[name = string("op_2158_transpose_y_0"), val = bool(false)]; + tensor var_2158_cast_fp16 = matmul(transpose_x = var_2158_transpose_x_0, transpose_y = var_2158_transpose_y_0, x = var_2157_cast_fp16, y = v_new_125_cast_fp16)[name = string("op_2158_cast_fp16")]; + tensor var_2159_cast_fp16 = add(x = var_2155_cast_fp16, y = var_2158_cast_fp16)[name = string("op_2159_cast_fp16")]; + tensor var_2161_begin_0 = const()[name = string("op_2161_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_2161_end_0 = const()[name = string("op_2161_end_0"), val = tensor([32, 31, 1, 1])]; + tensor var_2161_end_mask_0 = const()[name = string("op_2161_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2161_squeeze_mask_0 = const()[name = string("op_2161_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2161_cast_fp16 = slice_by_index(begin = var_2161_begin_0, end = var_2161_end_0, end_mask = var_2161_end_mask_0, squeeze_mask = var_2161_squeeze_mask_0, x = chunk_decay_3_cast_fp16)[name = string("op_2161_cast_fp16")]; + tensor var_2162_cast_fp16 = mul(x = state_125_cast_fp16, y = var_2161_cast_fp16)[name = string("op_2162_cast_fp16")]; + tensor var_2164_begin_0 = const()[name = string("op_2164_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_2164_end_0 = const()[name = string("op_2164_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_2164_end_mask_0 = const()[name = string("op_2164_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2164_squeeze_mask_0 = const()[name = string("op_2164_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2164_cast_fp16 = slice_by_index(begin = var_2164_begin_0, end = var_2164_end_0, end_mask = var_2164_end_mask_0, squeeze_mask = var_2164_squeeze_mask_0, x = k_dec_3_cast_fp16)[name = string("op_2164_cast_fp16")]; + bool var_2166_transpose_x_1 = const()[name = string("op_2166_transpose_x_1"), val = bool(true)]; + bool var_2166_transpose_y_1 = const()[name = string("op_2166_transpose_y_1"), val = bool(false)]; + tensor var_2166_cast_fp16 = matmul(transpose_x = var_2166_transpose_x_1, transpose_y = var_2166_transpose_y_1, x = var_2164_cast_fp16, y = v_new_125_cast_fp16)[name = string("op_2166_cast_fp16")]; + tensor state_127_cast_fp16 = add(x = var_2162_cast_fp16, y = var_2166_cast_fp16)[name = string("state_127_cast_fp16")]; + tensor var_2169_begin_0 = const()[name = string("op_2169_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_2169_end_0 = const()[name = string("op_2169_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_2169_end_mask_0 = const()[name = string("op_2169_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2169_squeeze_mask_0 = const()[name = string("op_2169_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2169_cast_fp16 = slice_by_index(begin = var_2169_begin_0, end = var_2169_end_0, end_mask = var_2169_end_mask_0, squeeze_mask = var_2169_squeeze_mask_0, x = new_v_3_cast_fp16)[name = string("op_2169_cast_fp16")]; + tensor var_2171_begin_0 = const()[name = string("op_2171_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_2171_end_0 = const()[name = string("op_2171_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_2171_end_mask_0 = const()[name = string("op_2171_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2171_squeeze_mask_0 = const()[name = string("op_2171_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2171_cast_fp16 = slice_by_index(begin = var_2171_begin_0, end = var_2171_end_0, end_mask = var_2171_end_mask_0, squeeze_mask = var_2171_squeeze_mask_0, x = k_cumdecay_3_cast_fp16)[name = string("op_2171_cast_fp16")]; + bool var_2172_transpose_x_0 = const()[name = string("op_2172_transpose_x_0"), val = bool(false)]; + bool var_2172_transpose_y_0 = const()[name = string("op_2172_transpose_y_0"), val = bool(false)]; + tensor var_2172_cast_fp16 = matmul(transpose_x = var_2172_transpose_x_0, transpose_y = var_2172_transpose_y_0, x = var_2171_cast_fp16, y = state_127_cast_fp16)[name = string("op_2172_cast_fp16")]; + tensor v_new_127_cast_fp16 = sub(x = var_2169_cast_fp16, y = var_2172_cast_fp16)[name = string("v_new_127_cast_fp16")]; + tensor var_2175_begin_0 = const()[name = string("op_2175_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_2175_end_0 = const()[name = string("op_2175_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_2175_end_mask_0 = const()[name = string("op_2175_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2175_squeeze_mask_0 = const()[name = string("op_2175_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2175_cast_fp16 = slice_by_index(begin = var_2175_begin_0, end = var_2175_end_0, end_mask = var_2175_end_mask_0, squeeze_mask = var_2175_squeeze_mask_0, x = q_dec_3_cast_fp16)[name = string("op_2175_cast_fp16")]; + bool var_2176_transpose_x_0 = const()[name = string("op_2176_transpose_x_0"), val = bool(false)]; + bool var_2176_transpose_y_0 = const()[name = string("op_2176_transpose_y_0"), val = bool(false)]; + tensor var_2176_cast_fp16 = matmul(transpose_x = var_2176_transpose_x_0, transpose_y = var_2176_transpose_y_0, x = var_2175_cast_fp16, y = state_127_cast_fp16)[name = string("op_2176_cast_fp16")]; + tensor var_2178_begin_0 = const()[name = string("op_2178_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_2178_end_0 = const()[name = string("op_2178_end_0"), val = tensor([32, 32, 64, 64])]; + tensor var_2178_end_mask_0 = const()[name = string("op_2178_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2178_squeeze_mask_0 = const()[name = string("op_2178_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2178_cast_fp16 = slice_by_index(begin = var_2178_begin_0, end = var_2178_end_0, end_mask = var_2178_end_mask_0, squeeze_mask = var_2178_squeeze_mask_0, x = attn_intra_3_cast_fp16)[name = string("op_2178_cast_fp16")]; + bool var_2179_transpose_x_0 = const()[name = string("op_2179_transpose_x_0"), val = bool(false)]; + bool var_2179_transpose_y_0 = const()[name = string("op_2179_transpose_y_0"), val = bool(false)]; + tensor var_2179_cast_fp16 = matmul(transpose_x = var_2179_transpose_x_0, transpose_y = var_2179_transpose_y_0, x = var_2178_cast_fp16, y = v_new_127_cast_fp16)[name = string("op_2179_cast_fp16")]; + tensor var_2180_cast_fp16 = add(x = var_2176_cast_fp16, y = var_2179_cast_fp16)[name = string("op_2180_cast_fp16")]; + int32 var_2182_axis_0 = const()[name = string("op_2182_axis_0"), val = int32(1)]; + tensor var_2182_cast_fp16 = stack(axis = var_2182_axis_0, values = (var_1529_cast_fp16, var_1550_cast_fp16, var_1571_cast_fp16, var_1592_cast_fp16, var_1613_cast_fp16, var_1634_cast_fp16, var_1655_cast_fp16, var_1676_cast_fp16, var_1697_cast_fp16, var_1718_cast_fp16, var_1739_cast_fp16, var_1760_cast_fp16, var_1781_cast_fp16, var_1802_cast_fp16, var_1823_cast_fp16, var_1844_cast_fp16, var_1865_cast_fp16, var_1886_cast_fp16, var_1907_cast_fp16, var_1928_cast_fp16, var_1949_cast_fp16, var_1970_cast_fp16, var_1991_cast_fp16, var_2012_cast_fp16, var_2033_cast_fp16, var_2054_cast_fp16, var_2075_cast_fp16, var_2096_cast_fp16, var_2117_cast_fp16, var_2138_cast_fp16, var_2159_cast_fp16, var_2180_cast_fp16))[name = string("op_2182_cast_fp16")]; + tensor var_2183 = const()[name = string("op_2183"), val = tensor([32, 2048, 128])]; + tensor var_2184_cast_fp16 = reshape(shape = var_2183, x = var_2182_cast_fp16)[name = string("op_2184_cast_fp16")]; + tensor var_2185 = const()[name = string("op_2185"), val = tensor([1, 0, 2])]; + tensor input_23_begin_0 = const()[name = string("input_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor input_23_end_0 = const()[name = string("input_23_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor input_23_end_mask_0 = const()[name = string("input_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor input_23_squeeze_mask_0 = const()[name = string("input_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor input_23_cast_fp16 = slice_by_index(begin = input_23_begin_0, end = input_23_end_0, end_mask = input_23_end_mask_0, squeeze_mask = input_23_squeeze_mask_0, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")]; + fp16 var_1165_promoted_1_to_fp16 = const()[name = string("op_1165_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor x_43_cast_fp16 = transpose(perm = var_2185, x = var_2184_cast_fp16)[name = string("transpose_40")]; + tensor var_2189_cast_fp16 = pow(x = x_43_cast_fp16, y = var_1165_promoted_1_to_fp16)[name = string("op_2189_cast_fp16")]; + tensor var_2191_axes_0 = const()[name = string("op_2191_axes_0"), val = tensor([-1])]; + bool var_2191_keep_dims_0 = const()[name = string("op_2191_keep_dims_0"), val = bool(true)]; + tensor var_2191_cast_fp16 = reduce_mean(axes = var_2191_axes_0, keep_dims = var_2191_keep_dims_0, x = var_2189_cast_fp16)[name = string("op_2191_cast_fp16")]; + fp16 var_2192_to_fp16 = const()[name = string("op_2192_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_2193_cast_fp16 = add(x = var_2191_cast_fp16, y = var_2192_to_fp16)[name = string("op_2193_cast_fp16")]; + fp32 var_2194_epsilon_0 = const()[name = string("op_2194_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2194_cast_fp16 = rsqrt(epsilon = var_2194_epsilon_0, x = var_2193_cast_fp16)[name = string("op_2194_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_43_cast_fp16, y = var_2194_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor layers_1_linear_attn_norm_weight_to_fp16 = const()[name = string("layers_1_linear_attn_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150834432)))]; + tensor var_2196_cast_fp16 = mul(x = layers_1_linear_attn_norm_weight_to_fp16, y = x_45_cast_fp16)[name = string("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_2197_cast_fp16")]; + tensor core_3_cast_fp16 = mul(x = var_2196_cast_fp16, y = var_2197_cast_fp16)[name = string("core_3_cast_fp16")]; + tensor var_2199 = const()[name = string("op_2199"), val = tensor([1, 2048, 4096])]; + tensor input_25_cast_fp16 = reshape(shape = var_2199, x = core_3_cast_fp16)[name = string("input_25_cast_fp16")]; + tensor layers_1_linear_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150834752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161320576))))[name = string("layers_1_linear_attn_out_proj_weight_to_fp16_quantized")]; + tensor linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_linear_attn_out_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor x_47_cast_fp16 = add(x = x_25_cast_fp16, y = linear_12_cast_fp16)[name = string("x_47_cast_fp16")]; + fp16 var_1165_promoted_2_to_fp16 = const()[name = string("op_1165_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor var_2205_cast_fp16 = pow(x = x_47_cast_fp16, y = var_1165_promoted_2_to_fp16)[name = string("op_2205_cast_fp16")]; + tensor var_2207_axes_0 = const()[name = string("op_2207_axes_0"), val = tensor([-1])]; + bool var_2207_keep_dims_0 = const()[name = string("op_2207_keep_dims_0"), val = bool(true)]; + tensor var_2207_cast_fp16 = reduce_mean(axes = var_2207_axes_0, keep_dims = var_2207_keep_dims_0, x = var_2205_cast_fp16)[name = string("op_2207_cast_fp16")]; + fp16 var_2208_to_fp16 = const()[name = string("op_2208_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_2209_cast_fp16 = add(x = var_2207_cast_fp16, y = var_2208_to_fp16)[name = string("op_2209_cast_fp16")]; + fp32 var_2210_epsilon_0 = const()[name = string("op_2210_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2210_cast_fp16 = rsqrt(epsilon = var_2210_epsilon_0, x = var_2209_cast_fp16)[name = string("op_2210_cast_fp16")]; + tensor x_49_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2210_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor var_2213_to_fp16 = const()[name = string("op_2213_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161325760)))]; + tensor input_27_cast_fp16 = mul(x = x_49_cast_fp16, y = var_2213_to_fp16)[name = string("input_27_cast_fp16")]; + tensor layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161330944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184923968))))[name = string("layers_1_mlp_gate_proj_weight_to_fp16_quantized")]; + tensor linear_13_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_2220_cast_fp16 = silu(x = linear_13_cast_fp16)[name = string("op_2220_cast_fp16")]; + tensor layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184942464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208535488))))[name = string("layers_1_mlp_up_proj_weight_to_fp16_quantized")]; + tensor linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor input_31_cast_fp16 = mul(x = var_2220_cast_fp16, y = linear_14_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208553984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232147008))))[name = string("layers_1_mlp_down_proj_weight_to_fp16_quantized")]; + tensor linear_15_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_15_cast_fp16")]; + tensor x_51_cast_fp16 = add(x = x_47_cast_fp16, y = linear_15_cast_fp16)[name = string("x_51_cast_fp16")]; + int32 var_2240 = const()[name = string("op_2240"), val = int32(-2)]; + int32 var_2269 = const()[name = string("op_2269"), val = int32(-1)]; + fp16 var_2268_promoted_to_fp16 = const()[name = string("op_2268_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_2278_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2268_promoted_to_fp16)[name = string("op_2278_cast_fp16")]; + tensor var_2280_axes_0 = const()[name = string("op_2280_axes_0"), val = tensor([-1])]; + bool var_2280_keep_dims_0 = const()[name = string("op_2280_keep_dims_0"), val = bool(true)]; + tensor var_2280_cast_fp16 = reduce_mean(axes = var_2280_axes_0, keep_dims = var_2280_keep_dims_0, x = var_2278_cast_fp16)[name = string("op_2280_cast_fp16")]; + fp16 var_2281_to_fp16 = const()[name = string("op_2281_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_2282_cast_fp16 = add(x = var_2280_cast_fp16, y = var_2281_to_fp16)[name = string("op_2282_cast_fp16")]; + fp32 var_2283_epsilon_0 = const()[name = string("op_2283_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2283_cast_fp16 = rsqrt(epsilon = var_2283_epsilon_0, x = var_2282_cast_fp16)[name = string("op_2283_cast_fp16")]; + tensor x_53_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2283_cast_fp16)[name = string("x_53_cast_fp16")]; + tensor var_2286_to_fp16 = const()[name = string("op_2286_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232152192)))]; + tensor x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_2286_to_fp16)[name = string("x_55_cast_fp16")]; + tensor layers_2_linear_attn_in_proj_qkv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232157376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253128960))))[name = string("layers_2_linear_attn_in_proj_qkv_weight_to_fp16_quantized")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_linear_attn_in_proj_qkv_weight_to_fp16_quantized, x = x_55_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_33_perm_0 = const()[name = string("input_33_perm_0"), val = tensor([0, 2, 1])]; + tensor input_35_pad_0 = const()[name = string("input_35_pad_0"), val = tensor([0, 0, 0, 0, 3, 0])]; + string input_35_mode_0 = const()[name = string("input_35_mode_0"), val = string("constant")]; + fp16 const_25_to_fp16 = const()[name = string("const_25_to_fp16"), val = fp16(0x0p+0)]; + tensor input_33_cast_fp16 = transpose(perm = input_33_perm_0, x = linear_16_cast_fp16)[name = string("transpose_39")]; + tensor input_35_cast_fp16 = pad(constant_val = const_25_to_fp16, mode = input_35_mode_0, pad = input_35_pad_0, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")]; + string input_37_pad_type_0 = const()[name = string("input_37_pad_type_0"), val = string("valid")]; + int32 input_37_groups_0 = const()[name = string("input_37_groups_0"), val = int32(8192)]; + tensor input_37_strides_0 = const()[name = string("input_37_strides_0"), val = tensor([1])]; + tensor input_37_pad_0 = const()[name = string("input_37_pad_0"), val = tensor([0, 0])]; + tensor input_37_dilations_0 = const()[name = string("input_37_dilations_0"), val = tensor([1])]; + tensor layers_2_linear_attn_conv1d_weight_to_fp16 = const()[name = string("layers_2_linear_attn_conv1d_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253145408)))]; + tensor input_37_cast_fp16 = conv(dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_2_linear_attn_conv1d_weight_to_fp16, x = input_35_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor var_2321_cast_fp16 = silu(x = input_37_cast_fp16)[name = string("op_2321_cast_fp16")]; + tensor qkv_5_perm_0 = const()[name = string("qkv_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_2323_begin_0 = const()[name = string("op_2323_begin_0"), val = tensor([0, 0, 0])]; + tensor var_2323_end_0 = const()[name = string("op_2323_end_0"), val = tensor([1, 2048, 2048])]; + tensor var_2323_end_mask_0 = const()[name = string("op_2323_end_mask_0"), val = tensor([true, true, false])]; + tensor qkv_5_cast_fp16 = transpose(perm = qkv_5_perm_0, x = var_2321_cast_fp16)[name = string("transpose_38")]; + tensor var_2323_cast_fp16 = slice_by_index(begin = var_2323_begin_0, end = var_2323_end_0, end_mask = var_2323_end_mask_0, x = qkv_5_cast_fp16)[name = string("op_2323_cast_fp16")]; + tensor var_2324 = const()[name = string("op_2324"), val = tensor([1, 2048, 16, 128])]; + tensor q_21_cast_fp16 = reshape(shape = var_2324, x = var_2323_cast_fp16)[name = string("q_21_cast_fp16")]; + tensor var_2326_begin_0 = const()[name = string("op_2326_begin_0"), val = tensor([0, 0, 2048])]; + tensor var_2326_end_0 = const()[name = string("op_2326_end_0"), val = tensor([1, 2048, 4096])]; + tensor var_2326_end_mask_0 = const()[name = string("op_2326_end_mask_0"), val = tensor([true, true, false])]; + tensor var_2326_cast_fp16 = slice_by_index(begin = var_2326_begin_0, end = var_2326_end_0, end_mask = var_2326_end_mask_0, x = qkv_5_cast_fp16)[name = string("op_2326_cast_fp16")]; + tensor var_2327 = const()[name = string("op_2327"), val = tensor([1, 2048, 16, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_2327, x = var_2326_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor var_2329_begin_0 = const()[name = string("op_2329_begin_0"), val = tensor([0, 0, 4096])]; + tensor var_2329_end_0 = const()[name = string("op_2329_end_0"), val = tensor([1, 2048, 8192])]; + tensor var_2329_end_mask_0 = const()[name = string("op_2329_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, x = qkv_5_cast_fp16)[name = string("op_2329_cast_fp16")]; + tensor var_2330 = const()[name = string("op_2330"), val = tensor([1, 2048, 32, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_2330, x = var_2329_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor layers_2_linear_attn_in_proj_z_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253211008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263696832))))[name = string("layers_2_linear_attn_in_proj_z_weight_to_fp16_quantized")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_linear_attn_in_proj_z_weight_to_fp16_quantized, x = x_55_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor var_2334 = const()[name = string("op_2334"), val = tensor([1, 2048, 32, 128])]; + tensor z_5_cast_fp16 = reshape(shape = var_2334, x = linear_17_cast_fp16)[name = string("z_5_cast_fp16")]; + tensor layers_2_linear_attn_in_proj_b_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263705088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263787072))))[name = string("layers_2_linear_attn_in_proj_b_weight_to_fp16_quantized")]; + tensor linear_18_bias_0_to_fp16 = const()[name = string("linear_18_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263787200)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_18_bias_0_to_fp16, weight = layers_2_linear_attn_in_proj_b_weight_to_fp16_quantized, x = x_55_cast_fp16)[name = string("linear_18_cast_fp16")]; + tensor beta_9_cast_fp16 = sigmoid(x = linear_18_cast_fp16)[name = string("beta_9_cast_fp16")]; + tensor var_2343_weight_0_to_fp16 = const()[name = string("op_2343_weight_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263787328)))]; + tensor var_2343_bias_0_to_fp16 = const()[name = string("op_2343_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263951232)))]; + tensor var_2343_cast_fp16 = linear(bias = var_2343_bias_0_to_fp16, weight = var_2343_weight_0_to_fp16, x = x_55_cast_fp16)[name = string("op_2343_cast_fp16")]; + tensor var_2344_cast_fp16 = softplus(x = var_2343_cast_fp16)[name = string("op_2344_cast_fp16")]; + tensor var_2340_to_fp16 = const()[name = string("op_2340_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263951360)))]; + tensor g_9_cast_fp16 = mul(x = var_2340_to_fp16, y = var_2344_cast_fp16)[name = string("g_9_cast_fp16")]; + tensor var_2346_cast_fp16 = mul(x = q_21_cast_fp16, y = q_21_cast_fp16)[name = string("op_2346_cast_fp16")]; + tensor var_2348_axes_0 = const()[name = string("op_2348_axes_0"), val = tensor([-1])]; + bool var_2348_keep_dims_0 = const()[name = string("op_2348_keep_dims_0"), val = bool(true)]; + tensor var_2348_cast_fp16 = reduce_sum(axes = var_2348_axes_0, keep_dims = var_2348_keep_dims_0, x = var_2346_cast_fp16)[name = string("op_2348_cast_fp16")]; + fp16 var_2349_to_fp16 = const()[name = string("op_2349_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_2350_cast_fp16 = add(x = var_2348_cast_fp16, y = var_2349_to_fp16)[name = string("op_2350_cast_fp16")]; + fp32 var_2351_epsilon_0 = const()[name = string("op_2351_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2351_cast_fp16 = rsqrt(epsilon = var_2351_epsilon_0, x = var_2350_cast_fp16)[name = string("op_2351_cast_fp16")]; + tensor q_23_cast_fp16 = mul(x = q_21_cast_fp16, y = var_2351_cast_fp16)[name = string("q_23_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = k_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_2353_cast_fp16")]; + tensor var_2355_axes_0 = const()[name = string("op_2355_axes_0"), val = tensor([-1])]; + bool var_2355_keep_dims_0 = const()[name = string("op_2355_keep_dims_0"), val = bool(true)]; + tensor var_2355_cast_fp16 = reduce_sum(axes = var_2355_axes_0, keep_dims = var_2355_keep_dims_0, x = var_2353_cast_fp16)[name = string("op_2355_cast_fp16")]; + fp16 var_2356_to_fp16 = const()[name = string("op_2356_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_2357_cast_fp16 = add(x = var_2355_cast_fp16, y = var_2356_to_fp16)[name = string("op_2357_cast_fp16")]; + fp32 var_2358_epsilon_0 = const()[name = string("op_2358_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2358_cast_fp16 = rsqrt(epsilon = var_2358_epsilon_0, x = var_2357_cast_fp16)[name = string("op_2358_cast_fp16")]; + tensor k_19_cast_fp16 = mul(x = k_17_cast_fp16, y = var_2358_cast_fp16)[name = string("k_19_cast_fp16")]; + tensor var_2363_axes_0 = const()[name = string("op_2363_axes_0"), val = tensor([3])]; + tensor var_2363_cast_fp16 = expand_dims(axes = var_2363_axes_0, x = q_23_cast_fp16)[name = string("op_2363_cast_fp16")]; + tensor var_2365_reps_0 = const()[name = string("op_2365_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_2365_cast_fp16 = tile(reps = var_2365_reps_0, x = var_2363_cast_fp16)[name = string("op_2365_cast_fp16")]; + tensor var_2366 = const()[name = string("op_2366"), val = tensor([1, 2048, 32, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_2366, x = var_2365_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor var_2371_axes_0 = const()[name = string("op_2371_axes_0"), val = tensor([3])]; + tensor var_2371_cast_fp16 = expand_dims(axes = var_2371_axes_0, x = k_19_cast_fp16)[name = string("op_2371_cast_fp16")]; + tensor var_2373_reps_0 = const()[name = string("op_2373_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_2373_cast_fp16 = tile(reps = var_2373_reps_0, x = var_2371_cast_fp16)[name = string("op_2373_cast_fp16")]; + tensor var_2374 = const()[name = string("op_2374"), val = tensor([1, 2048, 32, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_2374, x = var_2373_cast_fp16)[name = string("k_21_cast_fp16")]; + fp16 var_2376_to_fp16 = const()[name = string("op_2376_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor q_27_cast_fp16 = mul(x = q_25_cast_fp16, y = var_2376_to_fp16)[name = string("q_27_cast_fp16")]; + tensor var_2378_begin_0 = const()[name = string("op_2378_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2378_end_0 = const()[name = string("op_2378_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_2378_end_mask_0 = const()[name = string("op_2378_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_2378_squeeze_mask_0 = const()[name = string("op_2378_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_2378_cast_fp16 = slice_by_index(begin = var_2378_begin_0, end = var_2378_end_0, end_mask = var_2378_end_mask_0, squeeze_mask = var_2378_squeeze_mask_0, x = q_27_cast_fp16)[name = string("op_2378_cast_fp16")]; + tensor var_2379 = const()[name = string("op_2379"), val = tensor([1, 0, 2])]; + tensor var_2381 = const()[name = string("op_2381"), val = tensor([32, 32, 64, 128])]; + tensor var_2380_cast_fp16 = transpose(perm = var_2379, x = var_2378_cast_fp16)[name = string("transpose_37")]; + tensor q_29_cast_fp16 = reshape(shape = var_2381, x = var_2380_cast_fp16)[name = string("q_29_cast_fp16")]; + tensor var_2383_begin_0 = const()[name = string("op_2383_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2383_end_0 = const()[name = string("op_2383_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_2383_end_mask_0 = const()[name = string("op_2383_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_2383_squeeze_mask_0 = const()[name = string("op_2383_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_2383_cast_fp16 = slice_by_index(begin = var_2383_begin_0, end = var_2383_end_0, end_mask = var_2383_end_mask_0, squeeze_mask = var_2383_squeeze_mask_0, x = k_21_cast_fp16)[name = string("op_2383_cast_fp16")]; + tensor var_2384 = const()[name = string("op_2384"), val = tensor([1, 0, 2])]; + tensor var_2386 = const()[name = string("op_2386"), val = tensor([32, 32, 64, 128])]; + tensor var_2385_cast_fp16 = transpose(perm = var_2384, x = var_2383_cast_fp16)[name = string("transpose_36")]; + tensor k_23_cast_fp16 = reshape(shape = var_2386, x = var_2385_cast_fp16)[name = string("k_23_cast_fp16")]; + tensor var_2388_begin_0 = const()[name = string("op_2388_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2388_end_0 = const()[name = string("op_2388_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_2388_end_mask_0 = const()[name = string("op_2388_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_2388_squeeze_mask_0 = const()[name = string("op_2388_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_2388_cast_fp16 = slice_by_index(begin = var_2388_begin_0, end = var_2388_end_0, end_mask = var_2388_end_mask_0, squeeze_mask = var_2388_squeeze_mask_0, x = v_9_cast_fp16)[name = string("op_2388_cast_fp16")]; + tensor var_2389 = const()[name = string("op_2389"), val = tensor([1, 0, 2])]; + tensor var_2391 = const()[name = string("op_2391"), val = tensor([32, 32, 64, 128])]; + tensor var_2390_cast_fp16 = transpose(perm = var_2389, x = var_2388_cast_fp16)[name = string("transpose_35")]; + tensor v_11_cast_fp16 = reshape(shape = var_2391, x = var_2390_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor var_2393_begin_0 = const()[name = string("op_2393_begin_0"), val = tensor([0, 0, 0])]; + tensor var_2393_end_0 = const()[name = string("op_2393_end_0"), val = tensor([1, 2048, 32])]; + tensor var_2393_end_mask_0 = const()[name = string("op_2393_end_mask_0"), val = tensor([false, true, true])]; + tensor var_2393_squeeze_mask_0 = const()[name = string("op_2393_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_2393_cast_fp16 = slice_by_index(begin = var_2393_begin_0, end = var_2393_end_0, end_mask = var_2393_end_mask_0, squeeze_mask = var_2393_squeeze_mask_0, x = beta_9_cast_fp16)[name = string("op_2393_cast_fp16")]; + tensor var_2394_perm_0 = const()[name = string("op_2394_perm_0"), val = tensor([1, 0])]; + tensor var_2395 = const()[name = string("op_2395"), val = tensor([32, 32, 64, 1])]; + tensor var_2394_cast_fp16 = transpose(perm = var_2394_perm_0, x = var_2393_cast_fp16)[name = string("transpose_34")]; + tensor beta_11_cast_fp16 = reshape(shape = var_2395, x = var_2394_cast_fp16)[name = string("beta_11_cast_fp16")]; + tensor var_2397_begin_0 = const()[name = string("op_2397_begin_0"), val = tensor([0, 0, 0])]; + tensor var_2397_end_0 = const()[name = string("op_2397_end_0"), val = tensor([1, 2048, 32])]; + tensor var_2397_end_mask_0 = const()[name = string("op_2397_end_mask_0"), val = tensor([false, true, true])]; + tensor var_2397_squeeze_mask_0 = const()[name = string("op_2397_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_2397_cast_fp16 = slice_by_index(begin = var_2397_begin_0, end = var_2397_end_0, end_mask = var_2397_end_mask_0, squeeze_mask = var_2397_squeeze_mask_0, x = g_9_cast_fp16)[name = string("op_2397_cast_fp16")]; + tensor var_2398_perm_0 = const()[name = string("op_2398_perm_0"), val = tensor([1, 0])]; + tensor var_2399 = const()[name = string("op_2399"), val = tensor([32, 32, 64])]; + tensor var_2398_cast_fp16 = transpose(perm = var_2398_perm_0, x = var_2397_cast_fp16)[name = string("transpose_33")]; + tensor g_11_cast_fp16 = reshape(shape = var_2399, x = var_2398_cast_fp16)[name = string("g_11_cast_fp16")]; + tensor g_col_5_axes_0 = const()[name = string("g_col_5_axes_0"), val = tensor([-1])]; + tensor g_col_5_cast_fp16 = expand_dims(axes = g_col_5_axes_0, x = g_11_cast_fp16)[name = string("g_col_5_cast_fp16")]; + bool var_2402_transpose_x_0 = const()[name = string("op_2402_transpose_x_0"), val = bool(false)]; + bool var_2402_transpose_y_0 = const()[name = string("op_2402_transpose_y_0"), val = bool(false)]; + tensor var_2402_cast_fp16 = matmul(transpose_x = var_2402_transpose_x_0, transpose_y = var_2402_transpose_y_0, x = layers_0_linear_attn_lower_incl_to_fp16, y = g_col_5_cast_fp16)[name = string("op_2402_cast_fp16")]; + tensor cum_5_axes_0 = const()[name = string("cum_5_axes_0"), val = tensor([-1])]; + tensor cum_5_cast_fp16 = squeeze(axes = cum_5_axes_0, x = var_2402_cast_fp16)[name = string("cum_5_cast_fp16")]; + bool var_2404_transpose_x_0 = const()[name = string("op_2404_transpose_x_0"), val = bool(false)]; + bool var_2404_transpose_y_0 = const()[name = string("op_2404_transpose_y_0"), val = bool(false)]; + tensor var_2404_cast_fp16 = matmul(transpose_x = var_2404_transpose_x_0, transpose_y = var_2404_transpose_y_0, x = layers_0_linear_attn_suffix_sum_t_to_fp16, y = g_col_5_cast_fp16)[name = string("op_2404_cast_fp16")]; + tensor to_end_5_axes_0 = const()[name = string("to_end_5_axes_0"), val = tensor([-1])]; + tensor to_end_5_cast_fp16 = squeeze(axes = to_end_5_axes_0, x = var_2404_cast_fp16)[name = string("to_end_5_cast_fp16")]; + bool var_2406_transpose_x_0 = const()[name = string("op_2406_transpose_x_0"), val = bool(false)]; + bool var_2406_transpose_y_0 = const()[name = string("op_2406_transpose_y_0"), val = bool(false)]; + tensor var_2406_cast_fp16 = matmul(transpose_x = var_2406_transpose_x_0, transpose_y = var_2406_transpose_y_0, x = layers_0_linear_attn_pair_sum_t_to_fp16, y = g_col_5_cast_fp16)[name = string("op_2406_cast_fp16")]; + tensor var_2407 = const()[name = string("op_2407"), val = tensor([32, 32, 64, 64])]; + tensor pair_29_cast_fp16 = reshape(shape = var_2407, x = var_2406_cast_fp16)[name = string("pair_29_cast_fp16")]; + tensor var_2409_cast_fp16 = exp(x = pair_29_cast_fp16)[name = string("op_2409_cast_fp16")]; + tensor pair_decay_5_cast_fp16 = mul(x = var_2409_cast_fp16, y = layers_0_linear_attn_lower_incl_to_fp16)[name = string("pair_decay_5_cast_fp16")]; + tensor k_beta_5_cast_fp16 = mul(x = k_23_cast_fp16, y = beta_11_cast_fp16)[name = string("k_beta_5_cast_fp16")]; + tensor v_beta_5_cast_fp16 = mul(x = v_11_cast_fp16, y = beta_11_cast_fp16)[name = string("v_beta_5_cast_fp16")]; + bool var_2414_transpose_x_1 = const()[name = string("op_2414_transpose_x_1"), val = bool(false)]; + bool var_2414_transpose_y_1 = const()[name = string("op_2414_transpose_y_1"), val = bool(true)]; + tensor var_2414_cast_fp16 = matmul(transpose_x = var_2414_transpose_x_1, transpose_y = var_2414_transpose_y_1, x = k_beta_5_cast_fp16, y = k_23_cast_fp16)[name = string("op_2414_cast_fp16")]; + tensor kkt_5_cast_fp16 = mul(x = var_2414_cast_fp16, y = pair_decay_5_cast_fp16)[name = string("kkt_5_cast_fp16")]; + bool var_2417_transpose_x_1 = const()[name = string("op_2417_transpose_x_1"), val = bool(false)]; + bool var_2417_transpose_y_1 = const()[name = string("op_2417_transpose_y_1"), val = bool(true)]; + tensor var_2417_cast_fp16 = matmul(transpose_x = var_2417_transpose_x_1, transpose_y = var_2417_transpose_y_1, x = q_29_cast_fp16, y = k_23_cast_fp16)[name = string("op_2417_cast_fp16")]; + tensor attn_intra_5_cast_fp16 = mul(x = var_2417_cast_fp16, y = pair_decay_5_cast_fp16)[name = string("attn_intra_5_cast_fp16")]; + tensor var_2419_cast_fp16 = mul(x = kkt_5_cast_fp16, y = layers_0_linear_attn_strict_lower_to_fp16)[name = string("op_2419_cast_fp16")]; + tensor t_9_cast_fp16 = add(x = layers_0_linear_attn_eye_to_fp16, y = var_2419_cast_fp16)[name = string("t_9_cast_fp16")]; + tensor var_2423 = const()[name = string("op_2423"), val = tensor([-1, 64, 64])]; + tensor t_11_cast_fp16 = reshape(shape = var_2423, x = t_9_cast_fp16)[name = string("t_11_cast_fp16")]; + tensor var_2428 = const()[name = string("op_2428"), val = tensor([1024, 32, 2, 32, 2])]; + tensor var_2429_cast_fp16 = reshape(shape = var_2428, x = t_11_cast_fp16)[name = string("op_2429_cast_fp16")]; + tensor var_2434_cast_fp16 = mul(x = var_2429_cast_fp16, y = var_227_to_fp16)[name = string("op_2434_cast_fp16")]; + tensor blocks_25_axes_0 = const()[name = string("blocks_25_axes_0"), val = tensor([-2])]; + bool blocks_25_keep_dims_0 = const()[name = string("blocks_25_keep_dims_0"), val = bool(false)]; + tensor blocks_25_cast_fp16 = reduce_sum(axes = blocks_25_axes_0, keep_dims = blocks_25_keep_dims_0, x = var_2434_cast_fp16)[name = string("blocks_25_cast_fp16")]; + tensor var_2437_begin_0 = const()[name = string("op_2437_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2437_end_0 = const()[name = string("op_2437_end_0"), val = tensor([1024, 32, 2, 2])]; + tensor var_2437_end_mask_0 = const()[name = string("op_2437_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2437_cast_fp16 = slice_by_index(begin = var_2437_begin_0, end = var_2437_end_0, end_mask = var_2437_end_mask_0, x = blocks_25_cast_fp16)[name = string("op_2437_cast_fp16")]; + tensor x_57_begin_0 = const()[name = string("x_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = string("x_57_end_0"), val = tensor([1024, 32, 1, 1])]; + tensor x_57_end_mask_0 = const()[name = string("x_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, x = var_2437_cast_fp16)[name = string("x_57_cast_fp16")]; + bool var_2447_transpose_x_0 = const()[name = string("op_2447_transpose_x_0"), val = bool(false)]; + bool var_2447_transpose_y_0 = const()[name = string("op_2447_transpose_y_0"), val = bool(false)]; + tensor var_2447_cast_fp16 = matmul(transpose_x = var_2447_transpose_x_0, transpose_y = var_2447_transpose_y_0, x = x_57_cast_fp16, y = var_235_to_fp16)[name = string("op_2447_cast_fp16")]; + bool var_2448_transpose_x_0 = const()[name = string("op_2448_transpose_x_0"), val = bool(false)]; + bool var_2448_transpose_y_0 = const()[name = string("op_2448_transpose_y_0"), val = bool(false)]; + tensor var_2448_cast_fp16 = matmul(transpose_x = var_2448_transpose_x_0, transpose_y = var_2448_transpose_y_0, x = var_235_to_fp16, y = var_2447_cast_fp16)[name = string("op_2448_cast_fp16")]; + fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_25_cast_fp16 = mul(x = var_2448_cast_fp16, y = const_30_promoted_to_fp16)[name = string("lower_25_cast_fp16")]; + bool bottom_25_interleave_0 = const()[name = string("bottom_25_interleave_0"), val = bool(false)]; + tensor bottom_25_cast_fp16 = concat(axis = var_2269, interleave = bottom_25_interleave_0, values = (lower_25_cast_fp16, var_235_to_fp16))[name = string("bottom_25_cast_fp16")]; + bool inv_35_interleave_0 = const()[name = string("inv_35_interleave_0"), val = bool(false)]; + tensor inv_35_cast_fp16 = concat(axis = var_2240, interleave = inv_35_interleave_0, values = (top_1_to_fp16, bottom_25_cast_fp16))[name = string("inv_35_cast_fp16")]; + tensor var_2457 = const()[name = string("op_2457"), val = tensor([1024, 16, 4, 16, 4])]; + tensor var_2458_cast_fp16 = reshape(shape = var_2457, x = t_11_cast_fp16)[name = string("op_2458_cast_fp16")]; + tensor var_2463_cast_fp16 = mul(x = var_2458_cast_fp16, y = var_256_to_fp16)[name = string("op_2463_cast_fp16")]; + tensor blocks_27_axes_0 = const()[name = string("blocks_27_axes_0"), val = tensor([-2])]; + bool blocks_27_keep_dims_0 = const()[name = string("blocks_27_keep_dims_0"), val = bool(false)]; + tensor blocks_27_cast_fp16 = reduce_sum(axes = blocks_27_axes_0, keep_dims = blocks_27_keep_dims_0, x = var_2463_cast_fp16)[name = string("blocks_27_cast_fp16")]; + tensor var_2466_begin_0 = const()[name = string("op_2466_begin_0"), val = tensor([0, 0, 2, 0])]; + tensor var_2466_end_0 = const()[name = string("op_2466_end_0"), val = tensor([1024, 16, 4, 4])]; + tensor var_2466_end_mask_0 = const()[name = string("op_2466_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2466_cast_fp16 = slice_by_index(begin = var_2466_begin_0, end = var_2466_end_0, end_mask = var_2466_end_mask_0, x = blocks_27_cast_fp16)[name = string("op_2466_cast_fp16")]; + tensor x_59_begin_0 = const()[name = string("x_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_59_end_0 = const()[name = string("x_59_end_0"), val = tensor([1024, 16, 2, 2])]; + tensor x_59_end_mask_0 = const()[name = string("x_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, x = var_2466_cast_fp16)[name = string("x_59_cast_fp16")]; + tensor var_2468 = const()[name = string("op_2468"), val = tensor([1024, 16, 2, 2, 2])]; + tensor pair_33_cast_fp16 = reshape(shape = var_2468, x = inv_35_cast_fp16)[name = string("pair_33_cast_fp16")]; + tensor var_2470_begin_0 = const()[name = string("op_2470_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2470_end_0 = const()[name = string("op_2470_end_0"), val = tensor([1024, 16, 1, 2, 2])]; + tensor var_2470_end_mask_0 = const()[name = string("op_2470_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2470_squeeze_mask_0 = const()[name = string("op_2470_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2470_cast_fp16 = slice_by_index(begin = var_2470_begin_0, end = var_2470_end_0, end_mask = var_2470_end_mask_0, squeeze_mask = var_2470_squeeze_mask_0, x = pair_33_cast_fp16)[name = string("op_2470_cast_fp16")]; + tensor var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor([1024, 16, 2, 2, 2])]; + tensor var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2473_squeeze_mask_0 = const()[name = string("op_2473_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, squeeze_mask = var_2473_squeeze_mask_0, x = pair_33_cast_fp16)[name = string("op_2473_cast_fp16")]; + bool var_2476_transpose_x_0 = const()[name = string("op_2476_transpose_x_0"), val = bool(false)]; + bool var_2476_transpose_y_0 = const()[name = string("op_2476_transpose_y_0"), val = bool(false)]; + tensor var_2476_cast_fp16 = matmul(transpose_x = var_2476_transpose_x_0, transpose_y = var_2476_transpose_y_0, x = x_59_cast_fp16, y = var_2470_cast_fp16)[name = string("op_2476_cast_fp16")]; + bool var_2477_transpose_x_0 = const()[name = string("op_2477_transpose_x_0"), val = bool(false)]; + bool var_2477_transpose_y_0 = const()[name = string("op_2477_transpose_y_0"), val = bool(false)]; + tensor var_2477_cast_fp16 = matmul(transpose_x = var_2477_transpose_x_0, transpose_y = var_2477_transpose_y_0, x = var_2473_cast_fp16, y = var_2476_cast_fp16)[name = string("op_2477_cast_fp16")]; + fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_27_cast_fp16 = mul(x = var_2477_cast_fp16, y = const_31_promoted_to_fp16)[name = string("lower_27_cast_fp16")]; + bool top_27_interleave_0 = const()[name = string("top_27_interleave_0"), val = bool(false)]; + tensor top_27_cast_fp16 = concat(axis = var_2269, interleave = top_27_interleave_0, values = (var_2470_cast_fp16, var_273_to_fp16))[name = string("top_27_cast_fp16")]; + bool bottom_27_interleave_0 = const()[name = string("bottom_27_interleave_0"), val = bool(false)]; + tensor bottom_27_cast_fp16 = concat(axis = var_2269, interleave = bottom_27_interleave_0, values = (lower_27_cast_fp16, var_2473_cast_fp16))[name = string("bottom_27_cast_fp16")]; + bool inv_37_interleave_0 = const()[name = string("inv_37_interleave_0"), val = bool(false)]; + tensor inv_37_cast_fp16 = concat(axis = var_2240, interleave = inv_37_interleave_0, values = (top_27_cast_fp16, bottom_27_cast_fp16))[name = string("inv_37_cast_fp16")]; + tensor var_2486 = const()[name = string("op_2486"), val = tensor([1024, 8, 8, 8, 8])]; + tensor var_2487_cast_fp16 = reshape(shape = var_2486, x = t_11_cast_fp16)[name = string("op_2487_cast_fp16")]; + tensor var_2491_to_fp16 = const()[name = string("op_2491_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263951488)))]; + tensor var_2492_cast_fp16 = mul(x = var_2487_cast_fp16, y = var_2491_to_fp16)[name = string("op_2492_cast_fp16")]; + tensor blocks_29_axes_0 = const()[name = string("blocks_29_axes_0"), val = tensor([-2])]; + bool blocks_29_keep_dims_0 = const()[name = string("blocks_29_keep_dims_0"), val = bool(false)]; + tensor blocks_29_cast_fp16 = reduce_sum(axes = blocks_29_axes_0, keep_dims = blocks_29_keep_dims_0, x = var_2492_cast_fp16)[name = string("blocks_29_cast_fp16")]; + tensor var_2495_begin_0 = const()[name = string("op_2495_begin_0"), val = tensor([0, 0, 4, 0])]; + tensor var_2495_end_0 = const()[name = string("op_2495_end_0"), val = tensor([1024, 8, 8, 8])]; + tensor var_2495_end_mask_0 = const()[name = string("op_2495_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2495_cast_fp16 = slice_by_index(begin = var_2495_begin_0, end = var_2495_end_0, end_mask = var_2495_end_mask_0, x = blocks_29_cast_fp16)[name = string("op_2495_cast_fp16")]; + tensor x_61_begin_0 = const()[name = string("x_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = string("x_61_end_0"), val = tensor([1024, 8, 4, 4])]; + tensor x_61_end_mask_0 = const()[name = string("x_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, x = var_2495_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_2497 = const()[name = string("op_2497"), val = tensor([1024, 8, 2, 4, 4])]; + tensor pair_35_cast_fp16 = reshape(shape = var_2497, x = inv_37_cast_fp16)[name = string("pair_35_cast_fp16")]; + tensor var_2499_begin_0 = const()[name = string("op_2499_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2499_end_0 = const()[name = string("op_2499_end_0"), val = tensor([1024, 8, 1, 4, 4])]; + tensor var_2499_end_mask_0 = const()[name = string("op_2499_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2499_squeeze_mask_0 = const()[name = string("op_2499_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2499_cast_fp16 = slice_by_index(begin = var_2499_begin_0, end = var_2499_end_0, end_mask = var_2499_end_mask_0, squeeze_mask = var_2499_squeeze_mask_0, x = pair_35_cast_fp16)[name = string("op_2499_cast_fp16")]; + tensor var_2502_begin_0 = const()[name = string("op_2502_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_2502_end_0 = const()[name = string("op_2502_end_0"), val = tensor([1024, 8, 2, 4, 4])]; + tensor var_2502_end_mask_0 = const()[name = string("op_2502_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2502_squeeze_mask_0 = const()[name = string("op_2502_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2502_cast_fp16 = slice_by_index(begin = var_2502_begin_0, end = var_2502_end_0, end_mask = var_2502_end_mask_0, squeeze_mask = var_2502_squeeze_mask_0, x = pair_35_cast_fp16)[name = string("op_2502_cast_fp16")]; + bool var_2505_transpose_x_0 = const()[name = string("op_2505_transpose_x_0"), val = bool(false)]; + bool var_2505_transpose_y_0 = const()[name = string("op_2505_transpose_y_0"), val = bool(false)]; + tensor var_2505_cast_fp16 = matmul(transpose_x = var_2505_transpose_x_0, transpose_y = var_2505_transpose_y_0, x = x_61_cast_fp16, y = var_2499_cast_fp16)[name = string("op_2505_cast_fp16")]; + bool var_2506_transpose_x_0 = const()[name = string("op_2506_transpose_x_0"), val = bool(false)]; + bool var_2506_transpose_y_0 = const()[name = string("op_2506_transpose_y_0"), val = bool(false)]; + tensor var_2506_cast_fp16 = matmul(transpose_x = var_2506_transpose_x_0, transpose_y = var_2506_transpose_y_0, x = var_2502_cast_fp16, y = var_2505_cast_fp16)[name = string("op_2506_cast_fp16")]; + fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_29_cast_fp16 = mul(x = var_2506_cast_fp16, y = const_32_promoted_to_fp16)[name = string("lower_29_cast_fp16")]; + bool top_29_interleave_0 = const()[name = string("top_29_interleave_0"), val = bool(false)]; + tensor top_29_cast_fp16 = concat(axis = var_2269, interleave = top_29_interleave_0, values = (var_2499_cast_fp16, var_302_to_fp16))[name = string("top_29_cast_fp16")]; + bool bottom_29_interleave_0 = const()[name = string("bottom_29_interleave_0"), val = bool(false)]; + tensor bottom_29_cast_fp16 = concat(axis = var_2269, interleave = bottom_29_interleave_0, values = (lower_29_cast_fp16, var_2502_cast_fp16))[name = string("bottom_29_cast_fp16")]; + bool inv_39_interleave_0 = const()[name = string("inv_39_interleave_0"), val = bool(false)]; + tensor inv_39_cast_fp16 = concat(axis = var_2240, interleave = inv_39_interleave_0, values = (top_29_cast_fp16, bottom_29_cast_fp16))[name = string("inv_39_cast_fp16")]; + tensor var_2515 = const()[name = string("op_2515"), val = tensor([1024, 4, 16, 4, 16])]; + tensor var_2516_cast_fp16 = reshape(shape = var_2515, x = t_11_cast_fp16)[name = string("op_2516_cast_fp16")]; + tensor var_2520_to_fp16 = const()[name = string("op_2520_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263951680)))]; + tensor var_2521_cast_fp16 = mul(x = var_2516_cast_fp16, y = var_2520_to_fp16)[name = string("op_2521_cast_fp16")]; + tensor blocks_31_axes_0 = const()[name = string("blocks_31_axes_0"), val = tensor([-2])]; + bool blocks_31_keep_dims_0 = const()[name = string("blocks_31_keep_dims_0"), val = bool(false)]; + tensor blocks_31_cast_fp16 = reduce_sum(axes = blocks_31_axes_0, keep_dims = blocks_31_keep_dims_0, x = var_2521_cast_fp16)[name = string("blocks_31_cast_fp16")]; + tensor var_2524_begin_0 = const()[name = string("op_2524_begin_0"), val = tensor([0, 0, 8, 0])]; + tensor var_2524_end_0 = const()[name = string("op_2524_end_0"), val = tensor([1024, 4, 16, 16])]; + tensor var_2524_end_mask_0 = const()[name = string("op_2524_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2524_cast_fp16 = slice_by_index(begin = var_2524_begin_0, end = var_2524_end_0, end_mask = var_2524_end_mask_0, x = blocks_31_cast_fp16)[name = string("op_2524_cast_fp16")]; + tensor x_63_begin_0 = const()[name = string("x_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_63_end_0 = const()[name = string("x_63_end_0"), val = tensor([1024, 4, 8, 8])]; + tensor x_63_end_mask_0 = const()[name = string("x_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, x = var_2524_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_2526 = const()[name = string("op_2526"), val = tensor([1024, 4, 2, 8, 8])]; + tensor pair_37_cast_fp16 = reshape(shape = var_2526, x = inv_39_cast_fp16)[name = string("pair_37_cast_fp16")]; + tensor var_2528_begin_0 = const()[name = string("op_2528_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2528_end_0 = const()[name = string("op_2528_end_0"), val = tensor([1024, 4, 1, 8, 8])]; + tensor var_2528_end_mask_0 = const()[name = string("op_2528_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2528_squeeze_mask_0 = const()[name = string("op_2528_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2528_cast_fp16 = slice_by_index(begin = var_2528_begin_0, end = var_2528_end_0, end_mask = var_2528_end_mask_0, squeeze_mask = var_2528_squeeze_mask_0, x = pair_37_cast_fp16)[name = string("op_2528_cast_fp16")]; + tensor var_2531_begin_0 = const()[name = string("op_2531_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_2531_end_0 = const()[name = string("op_2531_end_0"), val = tensor([1024, 4, 2, 8, 8])]; + tensor var_2531_end_mask_0 = const()[name = string("op_2531_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2531_squeeze_mask_0 = const()[name = string("op_2531_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2531_cast_fp16 = slice_by_index(begin = var_2531_begin_0, end = var_2531_end_0, end_mask = var_2531_end_mask_0, squeeze_mask = var_2531_squeeze_mask_0, x = pair_37_cast_fp16)[name = string("op_2531_cast_fp16")]; + bool var_2534_transpose_x_0 = const()[name = string("op_2534_transpose_x_0"), val = bool(false)]; + bool var_2534_transpose_y_0 = const()[name = string("op_2534_transpose_y_0"), val = bool(false)]; + tensor var_2534_cast_fp16 = matmul(transpose_x = var_2534_transpose_x_0, transpose_y = var_2534_transpose_y_0, x = x_63_cast_fp16, y = var_2528_cast_fp16)[name = string("op_2534_cast_fp16")]; + bool var_2535_transpose_x_0 = const()[name = string("op_2535_transpose_x_0"), val = bool(false)]; + bool var_2535_transpose_y_0 = const()[name = string("op_2535_transpose_y_0"), val = bool(false)]; + tensor var_2535_cast_fp16 = matmul(transpose_x = var_2535_transpose_x_0, transpose_y = var_2535_transpose_y_0, x = var_2531_cast_fp16, y = var_2534_cast_fp16)[name = string("op_2535_cast_fp16")]; + fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_31_cast_fp16 = mul(x = var_2535_cast_fp16, y = const_33_promoted_to_fp16)[name = string("lower_31_cast_fp16")]; + bool top_31_interleave_0 = const()[name = string("top_31_interleave_0"), val = bool(false)]; + tensor top_31_cast_fp16 = concat(axis = var_2269, interleave = top_31_interleave_0, values = (var_2528_cast_fp16, var_331_to_fp16))[name = string("top_31_cast_fp16")]; + bool bottom_31_interleave_0 = const()[name = string("bottom_31_interleave_0"), val = bool(false)]; + tensor bottom_31_cast_fp16 = concat(axis = var_2269, interleave = bottom_31_interleave_0, values = (lower_31_cast_fp16, var_2531_cast_fp16))[name = string("bottom_31_cast_fp16")]; + bool inv_41_interleave_0 = const()[name = string("inv_41_interleave_0"), val = bool(false)]; + tensor inv_41_cast_fp16 = concat(axis = var_2240, interleave = inv_41_interleave_0, values = (top_31_cast_fp16, bottom_31_cast_fp16))[name = string("inv_41_cast_fp16")]; + tensor var_2544 = const()[name = string("op_2544"), val = tensor([1024, 2, 32, 2, 32])]; + tensor var_2545_cast_fp16 = reshape(shape = var_2544, x = t_11_cast_fp16)[name = string("op_2545_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = string("op_2549_to_fp16"), val = tensor([[[[0x1p+0], [0x0p+0]]], [[[0x0p+0], [0x1p+0]]]])]; + tensor var_2550_cast_fp16 = mul(x = var_2545_cast_fp16, y = var_2549_to_fp16)[name = string("op_2550_cast_fp16")]; + tensor blocks_33_axes_0 = const()[name = string("blocks_33_axes_0"), val = tensor([-2])]; + bool blocks_33_keep_dims_0 = const()[name = string("blocks_33_keep_dims_0"), val = bool(false)]; + tensor blocks_33_cast_fp16 = reduce_sum(axes = blocks_33_axes_0, keep_dims = blocks_33_keep_dims_0, x = var_2550_cast_fp16)[name = string("blocks_33_cast_fp16")]; + tensor var_2553_begin_0 = const()[name = string("op_2553_begin_0"), val = tensor([0, 0, 16, 0])]; + tensor var_2553_end_0 = const()[name = string("op_2553_end_0"), val = tensor([1024, 2, 32, 32])]; + tensor var_2553_end_mask_0 = const()[name = string("op_2553_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2553_cast_fp16 = slice_by_index(begin = var_2553_begin_0, end = var_2553_end_0, end_mask = var_2553_end_mask_0, x = blocks_33_cast_fp16)[name = string("op_2553_cast_fp16")]; + tensor x_65_begin_0 = const()[name = string("x_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = string("x_65_end_0"), val = tensor([1024, 2, 16, 16])]; + tensor x_65_end_mask_0 = const()[name = string("x_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, x = var_2553_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_2555 = const()[name = string("op_2555"), val = tensor([1024, 2, 2, 16, 16])]; + tensor pair_39_cast_fp16 = reshape(shape = var_2555, x = inv_41_cast_fp16)[name = string("pair_39_cast_fp16")]; + tensor var_2557_begin_0 = const()[name = string("op_2557_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2557_end_0 = const()[name = string("op_2557_end_0"), val = tensor([1024, 2, 1, 16, 16])]; + tensor var_2557_end_mask_0 = const()[name = string("op_2557_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2557_squeeze_mask_0 = const()[name = string("op_2557_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2557_cast_fp16 = slice_by_index(begin = var_2557_begin_0, end = var_2557_end_0, end_mask = var_2557_end_mask_0, squeeze_mask = var_2557_squeeze_mask_0, x = pair_39_cast_fp16)[name = string("op_2557_cast_fp16")]; + tensor var_2560_begin_0 = const()[name = string("op_2560_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_2560_end_0 = const()[name = string("op_2560_end_0"), val = tensor([1024, 2, 2, 16, 16])]; + tensor var_2560_end_mask_0 = const()[name = string("op_2560_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2560_squeeze_mask_0 = const()[name = string("op_2560_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2560_cast_fp16 = slice_by_index(begin = var_2560_begin_0, end = var_2560_end_0, end_mask = var_2560_end_mask_0, squeeze_mask = var_2560_squeeze_mask_0, x = pair_39_cast_fp16)[name = string("op_2560_cast_fp16")]; + bool var_2563_transpose_x_0 = const()[name = string("op_2563_transpose_x_0"), val = bool(false)]; + bool var_2563_transpose_y_0 = const()[name = string("op_2563_transpose_y_0"), val = bool(false)]; + tensor var_2563_cast_fp16 = matmul(transpose_x = var_2563_transpose_x_0, transpose_y = var_2563_transpose_y_0, x = x_65_cast_fp16, y = var_2557_cast_fp16)[name = string("op_2563_cast_fp16")]; + bool var_2564_transpose_x_0 = const()[name = string("op_2564_transpose_x_0"), val = bool(false)]; + bool var_2564_transpose_y_0 = const()[name = string("op_2564_transpose_y_0"), val = bool(false)]; + tensor var_2564_cast_fp16 = matmul(transpose_x = var_2564_transpose_x_0, transpose_y = var_2564_transpose_y_0, x = var_2560_cast_fp16, y = var_2563_cast_fp16)[name = string("op_2564_cast_fp16")]; + fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_33_cast_fp16 = mul(x = var_2564_cast_fp16, y = const_34_promoted_to_fp16)[name = string("lower_33_cast_fp16")]; + bool top_33_interleave_0 = const()[name = string("top_33_interleave_0"), val = bool(false)]; + tensor top_33_cast_fp16 = concat(axis = var_2269, interleave = top_33_interleave_0, values = (var_2557_cast_fp16, var_360_to_fp16))[name = string("top_33_cast_fp16")]; + bool bottom_33_interleave_0 = const()[name = string("bottom_33_interleave_0"), val = bool(false)]; + tensor bottom_33_cast_fp16 = concat(axis = var_2269, interleave = bottom_33_interleave_0, values = (lower_33_cast_fp16, var_2560_cast_fp16))[name = string("bottom_33_cast_fp16")]; + bool inv_43_interleave_0 = const()[name = string("inv_43_interleave_0"), val = bool(false)]; + tensor inv_43_cast_fp16 = concat(axis = var_2240, interleave = inv_43_interleave_0, values = (top_33_cast_fp16, bottom_33_cast_fp16))[name = string("inv_43_cast_fp16")]; + tensor var_2573 = const()[name = string("op_2573"), val = tensor([1024, 1, 64, 1, 64])]; + tensor var_2574_cast_fp16 = reshape(shape = var_2573, x = t_11_cast_fp16)[name = string("op_2574_cast_fp16")]; + tensor blocks_35_axes_0 = const()[name = string("blocks_35_axes_0"), val = tensor([-2])]; + bool blocks_35_keep_dims_0 = const()[name = string("blocks_35_keep_dims_0"), val = bool(false)]; + tensor blocks_35_cast_fp16 = reduce_sum(axes = blocks_35_axes_0, keep_dims = blocks_35_keep_dims_0, x = var_2574_cast_fp16)[name = string("blocks_35_cast_fp16")]; + tensor var_2582_begin_0 = const()[name = string("op_2582_begin_0"), val = tensor([0, 0, 32, 0])]; + tensor var_2582_end_0 = const()[name = string("op_2582_end_0"), val = tensor([1024, 1, 64, 64])]; + tensor var_2582_end_mask_0 = const()[name = string("op_2582_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2582_cast_fp16 = slice_by_index(begin = var_2582_begin_0, end = var_2582_end_0, end_mask = var_2582_end_mask_0, x = blocks_35_cast_fp16)[name = string("op_2582_cast_fp16")]; + tensor x_67_begin_0 = const()[name = string("x_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_67_end_0 = const()[name = string("x_67_end_0"), val = tensor([1024, 1, 32, 32])]; + tensor x_67_end_mask_0 = const()[name = string("x_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, x = var_2582_cast_fp16)[name = string("x_67_cast_fp16")]; + tensor var_2584 = const()[name = string("op_2584"), val = tensor([1024, 1, 2, 32, 32])]; + tensor pair_41_cast_fp16 = reshape(shape = var_2584, x = inv_43_cast_fp16)[name = string("pair_41_cast_fp16")]; + tensor var_2586_begin_0 = const()[name = string("op_2586_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2586_end_0 = const()[name = string("op_2586_end_0"), val = tensor([1024, 1, 1, 32, 32])]; + tensor var_2586_end_mask_0 = const()[name = string("op_2586_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2586_squeeze_mask_0 = const()[name = string("op_2586_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2586_cast_fp16 = slice_by_index(begin = var_2586_begin_0, end = var_2586_end_0, end_mask = var_2586_end_mask_0, squeeze_mask = var_2586_squeeze_mask_0, x = pair_41_cast_fp16)[name = string("op_2586_cast_fp16")]; + tensor var_2589_begin_0 = const()[name = string("op_2589_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_2589_end_0 = const()[name = string("op_2589_end_0"), val = tensor([1024, 1, 2, 32, 32])]; + tensor var_2589_end_mask_0 = const()[name = string("op_2589_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_2589_squeeze_mask_0 = const()[name = string("op_2589_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_2589_cast_fp16 = slice_by_index(begin = var_2589_begin_0, end = var_2589_end_0, end_mask = var_2589_end_mask_0, squeeze_mask = var_2589_squeeze_mask_0, x = pair_41_cast_fp16)[name = string("op_2589_cast_fp16")]; + bool var_2592_transpose_x_0 = const()[name = string("op_2592_transpose_x_0"), val = bool(false)]; + bool var_2592_transpose_y_0 = const()[name = string("op_2592_transpose_y_0"), val = bool(false)]; + tensor var_2592_cast_fp16 = matmul(transpose_x = var_2592_transpose_x_0, transpose_y = var_2592_transpose_y_0, x = x_67_cast_fp16, y = var_2586_cast_fp16)[name = string("op_2592_cast_fp16")]; + bool var_2593_transpose_x_0 = const()[name = string("op_2593_transpose_x_0"), val = bool(false)]; + bool var_2593_transpose_y_0 = const()[name = string("op_2593_transpose_y_0"), val = bool(false)]; + tensor var_2593_cast_fp16 = matmul(transpose_x = var_2593_transpose_x_0, transpose_y = var_2593_transpose_y_0, x = var_2589_cast_fp16, y = var_2592_cast_fp16)[name = string("op_2593_cast_fp16")]; + fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_35_cast_fp16 = mul(x = var_2593_cast_fp16, y = const_35_promoted_to_fp16)[name = string("lower_35_cast_fp16")]; + bool top_35_interleave_0 = const()[name = string("top_35_interleave_0"), val = bool(false)]; + tensor top_35_cast_fp16 = concat(axis = var_2269, interleave = top_35_interleave_0, values = (var_2586_cast_fp16, var_389_to_fp16))[name = string("top_35_cast_fp16")]; + bool bottom_35_interleave_0 = const()[name = string("bottom_35_interleave_0"), val = bool(false)]; + tensor bottom_35_cast_fp16 = concat(axis = var_2269, interleave = bottom_35_interleave_0, values = (lower_35_cast_fp16, var_2589_cast_fp16))[name = string("bottom_35_cast_fp16")]; + bool inv_45_interleave_0 = const()[name = string("inv_45_interleave_0"), val = bool(false)]; + tensor inv_45_cast_fp16 = concat(axis = var_2240, interleave = inv_45_interleave_0, values = (top_35_cast_fp16, bottom_35_cast_fp16))[name = string("inv_45_cast_fp16")]; + tensor var_2602 = const()[name = string("op_2602"), val = tensor([32, 32, 64, 64])]; + tensor inv_47_cast_fp16 = reshape(shape = var_2602, x = inv_45_cast_fp16)[name = string("inv_47_cast_fp16")]; + bool new_v_5_transpose_x_0 = const()[name = string("new_v_5_transpose_x_0"), val = bool(false)]; + bool new_v_5_transpose_y_0 = const()[name = string("new_v_5_transpose_y_0"), val = bool(false)]; + tensor new_v_5_cast_fp16 = matmul(transpose_x = new_v_5_transpose_x_0, transpose_y = new_v_5_transpose_y_0, x = inv_47_cast_fp16, y = v_beta_5_cast_fp16)[name = string("new_v_5_cast_fp16")]; + tensor var_2605_cast_fp16 = exp(x = cum_5_cast_fp16)[name = string("op_2605_cast_fp16")]; + tensor var_2606_axes_0 = const()[name = string("op_2606_axes_0"), val = tensor([3])]; + tensor var_2606_cast_fp16 = expand_dims(axes = var_2606_axes_0, x = var_2605_cast_fp16)[name = string("op_2606_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = k_beta_5_cast_fp16, y = var_2606_cast_fp16)[name = string("op_2607_cast_fp16")]; + bool k_cumdecay_5_transpose_x_0 = const()[name = string("k_cumdecay_5_transpose_x_0"), val = bool(false)]; + bool k_cumdecay_5_transpose_y_0 = const()[name = string("k_cumdecay_5_transpose_y_0"), val = bool(false)]; + tensor k_cumdecay_5_cast_fp16 = matmul(transpose_x = k_cumdecay_5_transpose_x_0, transpose_y = k_cumdecay_5_transpose_y_0, x = inv_47_cast_fp16, y = var_2607_cast_fp16)[name = string("k_cumdecay_5_cast_fp16")]; + tensor q_dec_5_cast_fp16 = mul(x = q_29_cast_fp16, y = var_2606_cast_fp16)[name = string("q_dec_5_cast_fp16")]; + tensor var_2612_cast_fp16 = exp(x = to_end_5_cast_fp16)[name = string("op_2612_cast_fp16")]; + tensor var_2613_axes_0 = const()[name = string("op_2613_axes_0"), val = tensor([3])]; + tensor var_2613_cast_fp16 = expand_dims(axes = var_2613_axes_0, x = var_2612_cast_fp16)[name = string("op_2613_cast_fp16")]; + tensor k_dec_5_cast_fp16 = mul(x = k_23_cast_fp16, y = var_2613_cast_fp16)[name = string("k_dec_5_cast_fp16")]; + tensor var_2615_begin_0 = const()[name = string("op_2615_begin_0"), val = tensor([0, 0, -1])]; + tensor var_2615_end_0 = const()[name = string("op_2615_end_0"), val = tensor([32, 32, 64])]; + tensor var_2615_end_mask_0 = const()[name = string("op_2615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2615_cast_fp16 = slice_by_index(begin = var_2615_begin_0, end = var_2615_end_0, end_mask = var_2615_end_mask_0, x = cum_5_cast_fp16)[name = string("op_2615_cast_fp16")]; + tensor var_2616_cast_fp16 = exp(x = var_2615_cast_fp16)[name = string("op_2616_cast_fp16")]; + tensor chunk_decay_5_axes_0 = const()[name = string("chunk_decay_5_axes_0"), val = tensor([3])]; + tensor chunk_decay_5_cast_fp16 = expand_dims(axes = chunk_decay_5_axes_0, x = var_2616_cast_fp16)[name = string("chunk_decay_5_cast_fp16")]; + tensor var_2621_begin_0 = const()[name = string("op_2621_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2621_end_0 = const()[name = string("op_2621_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_2621_end_mask_0 = const()[name = string("op_2621_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2621_squeeze_mask_0 = const()[name = string("op_2621_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2621_cast_fp16 = slice_by_index(begin = var_2621_begin_0, end = var_2621_end_0, end_mask = var_2621_end_mask_0, squeeze_mask = var_2621_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2621_cast_fp16")]; + tensor var_2623_begin_0 = const()[name = string("op_2623_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2623_end_0 = const()[name = string("op_2623_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_2623_end_mask_0 = const()[name = string("op_2623_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2623_squeeze_mask_0 = const()[name = string("op_2623_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2623_cast_fp16 = slice_by_index(begin = var_2623_begin_0, end = var_2623_end_0, end_mask = var_2623_end_mask_0, squeeze_mask = var_2623_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2623_cast_fp16")]; + bool var_2624_transpose_x_0 = const()[name = string("op_2624_transpose_x_0"), val = bool(false)]; + bool var_2624_transpose_y_0 = const()[name = string("op_2624_transpose_y_0"), val = bool(false)]; + tensor var_2624_cast_fp16 = matmul(transpose_x = var_2624_transpose_x_0, transpose_y = var_2624_transpose_y_0, x = var_2623_cast_fp16, y = state_1_to_fp16)[name = string("op_2624_cast_fp16")]; + tensor v_new_129_cast_fp16 = sub(x = var_2621_cast_fp16, y = var_2624_cast_fp16)[name = string("v_new_129_cast_fp16")]; + tensor var_2627_begin_0 = const()[name = string("op_2627_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2627_end_0 = const()[name = string("op_2627_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_2627_end_mask_0 = const()[name = string("op_2627_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2627_squeeze_mask_0 = const()[name = string("op_2627_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2627_cast_fp16 = slice_by_index(begin = var_2627_begin_0, end = var_2627_end_0, end_mask = var_2627_end_mask_0, squeeze_mask = var_2627_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2627_cast_fp16")]; + bool var_2628_transpose_x_0 = const()[name = string("op_2628_transpose_x_0"), val = bool(false)]; + bool var_2628_transpose_y_0 = const()[name = string("op_2628_transpose_y_0"), val = bool(false)]; + tensor var_2628_cast_fp16 = matmul(transpose_x = var_2628_transpose_x_0, transpose_y = var_2628_transpose_y_0, x = var_2627_cast_fp16, y = state_1_to_fp16)[name = string("op_2628_cast_fp16")]; + tensor var_2630_begin_0 = const()[name = string("op_2630_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2630_end_0 = const()[name = string("op_2630_end_0"), val = tensor([32, 1, 64, 64])]; + tensor var_2630_end_mask_0 = const()[name = string("op_2630_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2630_squeeze_mask_0 = const()[name = string("op_2630_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2630_cast_fp16 = slice_by_index(begin = var_2630_begin_0, end = var_2630_end_0, end_mask = var_2630_end_mask_0, squeeze_mask = var_2630_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2630_cast_fp16")]; + bool var_2631_transpose_x_0 = const()[name = string("op_2631_transpose_x_0"), val = bool(false)]; + bool var_2631_transpose_y_0 = const()[name = string("op_2631_transpose_y_0"), val = bool(false)]; + tensor var_2631_cast_fp16 = matmul(transpose_x = var_2631_transpose_x_0, transpose_y = var_2631_transpose_y_0, x = var_2630_cast_fp16, y = v_new_129_cast_fp16)[name = string("op_2631_cast_fp16")]; + tensor var_2632_cast_fp16 = add(x = var_2628_cast_fp16, y = var_2631_cast_fp16)[name = string("op_2632_cast_fp16")]; + tensor var_2634_begin_0 = const()[name = string("op_2634_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2634_end_0 = const()[name = string("op_2634_end_0"), val = tensor([32, 1, 1, 1])]; + tensor var_2634_end_mask_0 = const()[name = string("op_2634_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2634_squeeze_mask_0 = const()[name = string("op_2634_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2634_cast_fp16 = slice_by_index(begin = var_2634_begin_0, end = var_2634_end_0, end_mask = var_2634_end_mask_0, squeeze_mask = var_2634_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = state_1_to_fp16, y = var_2634_cast_fp16)[name = string("op_2635_cast_fp16")]; + tensor var_2637_begin_0 = const()[name = string("op_2637_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2637_end_0 = const()[name = string("op_2637_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_2637_end_mask_0 = const()[name = string("op_2637_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2637_squeeze_mask_0 = const()[name = string("op_2637_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2637_cast_fp16 = slice_by_index(begin = var_2637_begin_0, end = var_2637_end_0, end_mask = var_2637_end_mask_0, squeeze_mask = var_2637_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2637_cast_fp16")]; + bool var_2639_transpose_x_1 = const()[name = string("op_2639_transpose_x_1"), val = bool(true)]; + bool var_2639_transpose_y_1 = const()[name = string("op_2639_transpose_y_1"), val = bool(false)]; + tensor var_2639_cast_fp16 = matmul(transpose_x = var_2639_transpose_x_1, transpose_y = var_2639_transpose_y_1, x = var_2637_cast_fp16, y = v_new_129_cast_fp16)[name = string("op_2639_cast_fp16")]; + tensor state_131_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2639_cast_fp16)[name = string("state_131_cast_fp16")]; + tensor var_2642_begin_0 = const()[name = string("op_2642_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_2642_end_0 = const()[name = string("op_2642_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_2642_end_mask_0 = const()[name = string("op_2642_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2642_squeeze_mask_0 = const()[name = string("op_2642_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2642_cast_fp16 = slice_by_index(begin = var_2642_begin_0, end = var_2642_end_0, end_mask = var_2642_end_mask_0, squeeze_mask = var_2642_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2642_cast_fp16")]; + tensor var_2644_begin_0 = const()[name = string("op_2644_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_2644_end_0 = const()[name = string("op_2644_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_2644_end_mask_0 = const()[name = string("op_2644_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2644_squeeze_mask_0 = const()[name = string("op_2644_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2644_cast_fp16 = slice_by_index(begin = var_2644_begin_0, end = var_2644_end_0, end_mask = var_2644_end_mask_0, squeeze_mask = var_2644_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2644_cast_fp16")]; + bool var_2645_transpose_x_0 = const()[name = string("op_2645_transpose_x_0"), val = bool(false)]; + bool var_2645_transpose_y_0 = const()[name = string("op_2645_transpose_y_0"), val = bool(false)]; + tensor var_2645_cast_fp16 = matmul(transpose_x = var_2645_transpose_x_0, transpose_y = var_2645_transpose_y_0, x = var_2644_cast_fp16, y = state_131_cast_fp16)[name = string("op_2645_cast_fp16")]; + tensor v_new_131_cast_fp16 = sub(x = var_2642_cast_fp16, y = var_2645_cast_fp16)[name = string("v_new_131_cast_fp16")]; + tensor var_2648_begin_0 = const()[name = string("op_2648_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_2648_end_0 = const()[name = string("op_2648_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_2648_end_mask_0 = const()[name = string("op_2648_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2648_squeeze_mask_0 = const()[name = string("op_2648_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2648_cast_fp16 = slice_by_index(begin = var_2648_begin_0, end = var_2648_end_0, end_mask = var_2648_end_mask_0, squeeze_mask = var_2648_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2648_cast_fp16")]; + bool var_2649_transpose_x_0 = const()[name = string("op_2649_transpose_x_0"), val = bool(false)]; + bool var_2649_transpose_y_0 = const()[name = string("op_2649_transpose_y_0"), val = bool(false)]; + tensor var_2649_cast_fp16 = matmul(transpose_x = var_2649_transpose_x_0, transpose_y = var_2649_transpose_y_0, x = var_2648_cast_fp16, y = state_131_cast_fp16)[name = string("op_2649_cast_fp16")]; + tensor var_2651_begin_0 = const()[name = string("op_2651_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_2651_end_0 = const()[name = string("op_2651_end_0"), val = tensor([32, 2, 64, 64])]; + tensor var_2651_end_mask_0 = const()[name = string("op_2651_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2651_squeeze_mask_0 = const()[name = string("op_2651_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2651_cast_fp16 = slice_by_index(begin = var_2651_begin_0, end = var_2651_end_0, end_mask = var_2651_end_mask_0, squeeze_mask = var_2651_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2651_cast_fp16")]; + bool var_2652_transpose_x_0 = const()[name = string("op_2652_transpose_x_0"), val = bool(false)]; + bool var_2652_transpose_y_0 = const()[name = string("op_2652_transpose_y_0"), val = bool(false)]; + tensor var_2652_cast_fp16 = matmul(transpose_x = var_2652_transpose_x_0, transpose_y = var_2652_transpose_y_0, x = var_2651_cast_fp16, y = v_new_131_cast_fp16)[name = string("op_2652_cast_fp16")]; + tensor var_2653_cast_fp16 = add(x = var_2649_cast_fp16, y = var_2652_cast_fp16)[name = string("op_2653_cast_fp16")]; + tensor var_2655_begin_0 = const()[name = string("op_2655_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_2655_end_0 = const()[name = string("op_2655_end_0"), val = tensor([32, 2, 1, 1])]; + tensor var_2655_end_mask_0 = const()[name = string("op_2655_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2655_squeeze_mask_0 = const()[name = string("op_2655_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2655_cast_fp16 = slice_by_index(begin = var_2655_begin_0, end = var_2655_end_0, end_mask = var_2655_end_mask_0, squeeze_mask = var_2655_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2655_cast_fp16")]; + tensor var_2656_cast_fp16 = mul(x = state_131_cast_fp16, y = var_2655_cast_fp16)[name = string("op_2656_cast_fp16")]; + tensor var_2658_begin_0 = const()[name = string("op_2658_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_2658_end_0 = const()[name = string("op_2658_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_2658_end_mask_0 = const()[name = string("op_2658_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2658_squeeze_mask_0 = const()[name = string("op_2658_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, squeeze_mask = var_2658_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2658_cast_fp16")]; + bool var_2660_transpose_x_1 = const()[name = string("op_2660_transpose_x_1"), val = bool(true)]; + bool var_2660_transpose_y_1 = const()[name = string("op_2660_transpose_y_1"), val = bool(false)]; + tensor var_2660_cast_fp16 = matmul(transpose_x = var_2660_transpose_x_1, transpose_y = var_2660_transpose_y_1, x = var_2658_cast_fp16, y = v_new_131_cast_fp16)[name = string("op_2660_cast_fp16")]; + tensor state_133_cast_fp16 = add(x = var_2656_cast_fp16, y = var_2660_cast_fp16)[name = string("state_133_cast_fp16")]; + tensor var_2663_begin_0 = const()[name = string("op_2663_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_2663_end_0 = const()[name = string("op_2663_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_2663_end_mask_0 = const()[name = string("op_2663_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2663_squeeze_mask_0 = const()[name = string("op_2663_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2663_cast_fp16 = slice_by_index(begin = var_2663_begin_0, end = var_2663_end_0, end_mask = var_2663_end_mask_0, squeeze_mask = var_2663_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2663_cast_fp16")]; + tensor var_2665_begin_0 = const()[name = string("op_2665_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_2665_end_0 = const()[name = string("op_2665_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_2665_end_mask_0 = const()[name = string("op_2665_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2665_squeeze_mask_0 = const()[name = string("op_2665_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2665_cast_fp16 = slice_by_index(begin = var_2665_begin_0, end = var_2665_end_0, end_mask = var_2665_end_mask_0, squeeze_mask = var_2665_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2665_cast_fp16")]; + bool var_2666_transpose_x_0 = const()[name = string("op_2666_transpose_x_0"), val = bool(false)]; + bool var_2666_transpose_y_0 = const()[name = string("op_2666_transpose_y_0"), val = bool(false)]; + tensor var_2666_cast_fp16 = matmul(transpose_x = var_2666_transpose_x_0, transpose_y = var_2666_transpose_y_0, x = var_2665_cast_fp16, y = state_133_cast_fp16)[name = string("op_2666_cast_fp16")]; + tensor v_new_133_cast_fp16 = sub(x = var_2663_cast_fp16, y = var_2666_cast_fp16)[name = string("v_new_133_cast_fp16")]; + tensor var_2669_begin_0 = const()[name = string("op_2669_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_2669_end_0 = const()[name = string("op_2669_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_2669_end_mask_0 = const()[name = string("op_2669_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2669_squeeze_mask_0 = const()[name = string("op_2669_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2669_cast_fp16 = slice_by_index(begin = var_2669_begin_0, end = var_2669_end_0, end_mask = var_2669_end_mask_0, squeeze_mask = var_2669_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2669_cast_fp16")]; + bool var_2670_transpose_x_0 = const()[name = string("op_2670_transpose_x_0"), val = bool(false)]; + bool var_2670_transpose_y_0 = const()[name = string("op_2670_transpose_y_0"), val = bool(false)]; + tensor var_2670_cast_fp16 = matmul(transpose_x = var_2670_transpose_x_0, transpose_y = var_2670_transpose_y_0, x = var_2669_cast_fp16, y = state_133_cast_fp16)[name = string("op_2670_cast_fp16")]; + tensor var_2672_begin_0 = const()[name = string("op_2672_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_2672_end_0 = const()[name = string("op_2672_end_0"), val = tensor([32, 3, 64, 64])]; + tensor var_2672_end_mask_0 = const()[name = string("op_2672_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2672_squeeze_mask_0 = const()[name = string("op_2672_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2672_cast_fp16 = slice_by_index(begin = var_2672_begin_0, end = var_2672_end_0, end_mask = var_2672_end_mask_0, squeeze_mask = var_2672_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2672_cast_fp16")]; + bool var_2673_transpose_x_0 = const()[name = string("op_2673_transpose_x_0"), val = bool(false)]; + bool var_2673_transpose_y_0 = const()[name = string("op_2673_transpose_y_0"), val = bool(false)]; + tensor var_2673_cast_fp16 = matmul(transpose_x = var_2673_transpose_x_0, transpose_y = var_2673_transpose_y_0, x = var_2672_cast_fp16, y = v_new_133_cast_fp16)[name = string("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = add(x = var_2670_cast_fp16, y = var_2673_cast_fp16)[name = string("op_2674_cast_fp16")]; + tensor var_2676_begin_0 = const()[name = string("op_2676_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_2676_end_0 = const()[name = string("op_2676_end_0"), val = tensor([32, 3, 1, 1])]; + tensor var_2676_end_mask_0 = const()[name = string("op_2676_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2676_squeeze_mask_0 = const()[name = string("op_2676_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2676_cast_fp16 = slice_by_index(begin = var_2676_begin_0, end = var_2676_end_0, end_mask = var_2676_end_mask_0, squeeze_mask = var_2676_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2676_cast_fp16")]; + tensor var_2677_cast_fp16 = mul(x = state_133_cast_fp16, y = var_2676_cast_fp16)[name = string("op_2677_cast_fp16")]; + tensor var_2679_begin_0 = const()[name = string("op_2679_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_2679_end_0 = const()[name = string("op_2679_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_2679_end_mask_0 = const()[name = string("op_2679_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2679_squeeze_mask_0 = const()[name = string("op_2679_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2679_cast_fp16 = slice_by_index(begin = var_2679_begin_0, end = var_2679_end_0, end_mask = var_2679_end_mask_0, squeeze_mask = var_2679_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2679_cast_fp16")]; + bool var_2681_transpose_x_1 = const()[name = string("op_2681_transpose_x_1"), val = bool(true)]; + bool var_2681_transpose_y_1 = const()[name = string("op_2681_transpose_y_1"), val = bool(false)]; + tensor var_2681_cast_fp16 = matmul(transpose_x = var_2681_transpose_x_1, transpose_y = var_2681_transpose_y_1, x = var_2679_cast_fp16, y = v_new_133_cast_fp16)[name = string("op_2681_cast_fp16")]; + tensor state_135_cast_fp16 = add(x = var_2677_cast_fp16, y = var_2681_cast_fp16)[name = string("state_135_cast_fp16")]; + tensor var_2684_begin_0 = const()[name = string("op_2684_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_2684_end_0 = const()[name = string("op_2684_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_2684_end_mask_0 = const()[name = string("op_2684_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2684_squeeze_mask_0 = const()[name = string("op_2684_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2684_cast_fp16 = slice_by_index(begin = var_2684_begin_0, end = var_2684_end_0, end_mask = var_2684_end_mask_0, squeeze_mask = var_2684_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2684_cast_fp16")]; + tensor var_2686_begin_0 = const()[name = string("op_2686_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_2686_end_0 = const()[name = string("op_2686_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_2686_end_mask_0 = const()[name = string("op_2686_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2686_squeeze_mask_0 = const()[name = string("op_2686_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2686_cast_fp16 = slice_by_index(begin = var_2686_begin_0, end = var_2686_end_0, end_mask = var_2686_end_mask_0, squeeze_mask = var_2686_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2686_cast_fp16")]; + bool var_2687_transpose_x_0 = const()[name = string("op_2687_transpose_x_0"), val = bool(false)]; + bool var_2687_transpose_y_0 = const()[name = string("op_2687_transpose_y_0"), val = bool(false)]; + tensor var_2687_cast_fp16 = matmul(transpose_x = var_2687_transpose_x_0, transpose_y = var_2687_transpose_y_0, x = var_2686_cast_fp16, y = state_135_cast_fp16)[name = string("op_2687_cast_fp16")]; + tensor v_new_135_cast_fp16 = sub(x = var_2684_cast_fp16, y = var_2687_cast_fp16)[name = string("v_new_135_cast_fp16")]; + tensor var_2690_begin_0 = const()[name = string("op_2690_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_2690_end_0 = const()[name = string("op_2690_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_2690_end_mask_0 = const()[name = string("op_2690_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2690_squeeze_mask_0 = const()[name = string("op_2690_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2690_cast_fp16 = slice_by_index(begin = var_2690_begin_0, end = var_2690_end_0, end_mask = var_2690_end_mask_0, squeeze_mask = var_2690_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2690_cast_fp16")]; + bool var_2691_transpose_x_0 = const()[name = string("op_2691_transpose_x_0"), val = bool(false)]; + bool var_2691_transpose_y_0 = const()[name = string("op_2691_transpose_y_0"), val = bool(false)]; + tensor var_2691_cast_fp16 = matmul(transpose_x = var_2691_transpose_x_0, transpose_y = var_2691_transpose_y_0, x = var_2690_cast_fp16, y = state_135_cast_fp16)[name = string("op_2691_cast_fp16")]; + tensor var_2693_begin_0 = const()[name = string("op_2693_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_2693_end_0 = const()[name = string("op_2693_end_0"), val = tensor([32, 4, 64, 64])]; + tensor var_2693_end_mask_0 = const()[name = string("op_2693_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2693_squeeze_mask_0 = const()[name = string("op_2693_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2693_cast_fp16 = slice_by_index(begin = var_2693_begin_0, end = var_2693_end_0, end_mask = var_2693_end_mask_0, squeeze_mask = var_2693_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2693_cast_fp16")]; + bool var_2694_transpose_x_0 = const()[name = string("op_2694_transpose_x_0"), val = bool(false)]; + bool var_2694_transpose_y_0 = const()[name = string("op_2694_transpose_y_0"), val = bool(false)]; + tensor var_2694_cast_fp16 = matmul(transpose_x = var_2694_transpose_x_0, transpose_y = var_2694_transpose_y_0, x = var_2693_cast_fp16, y = v_new_135_cast_fp16)[name = string("op_2694_cast_fp16")]; + tensor var_2695_cast_fp16 = add(x = var_2691_cast_fp16, y = var_2694_cast_fp16)[name = string("op_2695_cast_fp16")]; + tensor var_2697_begin_0 = const()[name = string("op_2697_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_2697_end_0 = const()[name = string("op_2697_end_0"), val = tensor([32, 4, 1, 1])]; + tensor var_2697_end_mask_0 = const()[name = string("op_2697_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2697_squeeze_mask_0 = const()[name = string("op_2697_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2697_cast_fp16 = slice_by_index(begin = var_2697_begin_0, end = var_2697_end_0, end_mask = var_2697_end_mask_0, squeeze_mask = var_2697_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2697_cast_fp16")]; + tensor var_2698_cast_fp16 = mul(x = state_135_cast_fp16, y = var_2697_cast_fp16)[name = string("op_2698_cast_fp16")]; + tensor var_2700_begin_0 = const()[name = string("op_2700_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_2700_end_0 = const()[name = string("op_2700_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_2700_end_mask_0 = const()[name = string("op_2700_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2700_squeeze_mask_0 = const()[name = string("op_2700_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2700_cast_fp16 = slice_by_index(begin = var_2700_begin_0, end = var_2700_end_0, end_mask = var_2700_end_mask_0, squeeze_mask = var_2700_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2700_cast_fp16")]; + bool var_2702_transpose_x_1 = const()[name = string("op_2702_transpose_x_1"), val = bool(true)]; + bool var_2702_transpose_y_1 = const()[name = string("op_2702_transpose_y_1"), val = bool(false)]; + tensor var_2702_cast_fp16 = matmul(transpose_x = var_2702_transpose_x_1, transpose_y = var_2702_transpose_y_1, x = var_2700_cast_fp16, y = v_new_135_cast_fp16)[name = string("op_2702_cast_fp16")]; + tensor state_137_cast_fp16 = add(x = var_2698_cast_fp16, y = var_2702_cast_fp16)[name = string("state_137_cast_fp16")]; + tensor var_2705_begin_0 = const()[name = string("op_2705_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_2705_end_0 = const()[name = string("op_2705_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_2705_end_mask_0 = const()[name = string("op_2705_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2705_squeeze_mask_0 = const()[name = string("op_2705_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2705_cast_fp16 = slice_by_index(begin = var_2705_begin_0, end = var_2705_end_0, end_mask = var_2705_end_mask_0, squeeze_mask = var_2705_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2705_cast_fp16")]; + tensor var_2707_begin_0 = const()[name = string("op_2707_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_2707_end_0 = const()[name = string("op_2707_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_2707_end_mask_0 = const()[name = string("op_2707_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2707_squeeze_mask_0 = const()[name = string("op_2707_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2707_cast_fp16 = slice_by_index(begin = var_2707_begin_0, end = var_2707_end_0, end_mask = var_2707_end_mask_0, squeeze_mask = var_2707_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2707_cast_fp16")]; + bool var_2708_transpose_x_0 = const()[name = string("op_2708_transpose_x_0"), val = bool(false)]; + bool var_2708_transpose_y_0 = const()[name = string("op_2708_transpose_y_0"), val = bool(false)]; + tensor var_2708_cast_fp16 = matmul(transpose_x = var_2708_transpose_x_0, transpose_y = var_2708_transpose_y_0, x = var_2707_cast_fp16, y = state_137_cast_fp16)[name = string("op_2708_cast_fp16")]; + tensor v_new_137_cast_fp16 = sub(x = var_2705_cast_fp16, y = var_2708_cast_fp16)[name = string("v_new_137_cast_fp16")]; + tensor var_2711_begin_0 = const()[name = string("op_2711_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_2711_end_0 = const()[name = string("op_2711_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_2711_end_mask_0 = const()[name = string("op_2711_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2711_squeeze_mask_0 = const()[name = string("op_2711_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2711_cast_fp16 = slice_by_index(begin = var_2711_begin_0, end = var_2711_end_0, end_mask = var_2711_end_mask_0, squeeze_mask = var_2711_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2711_cast_fp16")]; + bool var_2712_transpose_x_0 = const()[name = string("op_2712_transpose_x_0"), val = bool(false)]; + bool var_2712_transpose_y_0 = const()[name = string("op_2712_transpose_y_0"), val = bool(false)]; + tensor var_2712_cast_fp16 = matmul(transpose_x = var_2712_transpose_x_0, transpose_y = var_2712_transpose_y_0, x = var_2711_cast_fp16, y = state_137_cast_fp16)[name = string("op_2712_cast_fp16")]; + tensor var_2714_begin_0 = const()[name = string("op_2714_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_2714_end_0 = const()[name = string("op_2714_end_0"), val = tensor([32, 5, 64, 64])]; + tensor var_2714_end_mask_0 = const()[name = string("op_2714_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2714_squeeze_mask_0 = const()[name = string("op_2714_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2714_cast_fp16 = slice_by_index(begin = var_2714_begin_0, end = var_2714_end_0, end_mask = var_2714_end_mask_0, squeeze_mask = var_2714_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2714_cast_fp16")]; + bool var_2715_transpose_x_0 = const()[name = string("op_2715_transpose_x_0"), val = bool(false)]; + bool var_2715_transpose_y_0 = const()[name = string("op_2715_transpose_y_0"), val = bool(false)]; + tensor var_2715_cast_fp16 = matmul(transpose_x = var_2715_transpose_x_0, transpose_y = var_2715_transpose_y_0, x = var_2714_cast_fp16, y = v_new_137_cast_fp16)[name = string("op_2715_cast_fp16")]; + tensor var_2716_cast_fp16 = add(x = var_2712_cast_fp16, y = var_2715_cast_fp16)[name = string("op_2716_cast_fp16")]; + tensor var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor([32, 5, 1, 1])]; + tensor var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2718_squeeze_mask_0 = const()[name = string("op_2718_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, squeeze_mask = var_2718_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2718_cast_fp16")]; + tensor var_2719_cast_fp16 = mul(x = state_137_cast_fp16, y = var_2718_cast_fp16)[name = string("op_2719_cast_fp16")]; + tensor var_2721_begin_0 = const()[name = string("op_2721_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_2721_end_0 = const()[name = string("op_2721_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_2721_end_mask_0 = const()[name = string("op_2721_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2721_squeeze_mask_0 = const()[name = string("op_2721_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2721_cast_fp16 = slice_by_index(begin = var_2721_begin_0, end = var_2721_end_0, end_mask = var_2721_end_mask_0, squeeze_mask = var_2721_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2721_cast_fp16")]; + bool var_2723_transpose_x_1 = const()[name = string("op_2723_transpose_x_1"), val = bool(true)]; + bool var_2723_transpose_y_1 = const()[name = string("op_2723_transpose_y_1"), val = bool(false)]; + tensor var_2723_cast_fp16 = matmul(transpose_x = var_2723_transpose_x_1, transpose_y = var_2723_transpose_y_1, x = var_2721_cast_fp16, y = v_new_137_cast_fp16)[name = string("op_2723_cast_fp16")]; + tensor state_139_cast_fp16 = add(x = var_2719_cast_fp16, y = var_2723_cast_fp16)[name = string("state_139_cast_fp16")]; + tensor var_2726_begin_0 = const()[name = string("op_2726_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_2726_end_0 = const()[name = string("op_2726_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_2726_end_mask_0 = const()[name = string("op_2726_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2726_squeeze_mask_0 = const()[name = string("op_2726_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2726_cast_fp16 = slice_by_index(begin = var_2726_begin_0, end = var_2726_end_0, end_mask = var_2726_end_mask_0, squeeze_mask = var_2726_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2726_cast_fp16")]; + tensor var_2728_begin_0 = const()[name = string("op_2728_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_2728_end_0 = const()[name = string("op_2728_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_2728_end_mask_0 = const()[name = string("op_2728_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2728_squeeze_mask_0 = const()[name = string("op_2728_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2728_cast_fp16 = slice_by_index(begin = var_2728_begin_0, end = var_2728_end_0, end_mask = var_2728_end_mask_0, squeeze_mask = var_2728_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2728_cast_fp16")]; + bool var_2729_transpose_x_0 = const()[name = string("op_2729_transpose_x_0"), val = bool(false)]; + bool var_2729_transpose_y_0 = const()[name = string("op_2729_transpose_y_0"), val = bool(false)]; + tensor var_2729_cast_fp16 = matmul(transpose_x = var_2729_transpose_x_0, transpose_y = var_2729_transpose_y_0, x = var_2728_cast_fp16, y = state_139_cast_fp16)[name = string("op_2729_cast_fp16")]; + tensor v_new_139_cast_fp16 = sub(x = var_2726_cast_fp16, y = var_2729_cast_fp16)[name = string("v_new_139_cast_fp16")]; + tensor var_2732_begin_0 = const()[name = string("op_2732_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_2732_end_0 = const()[name = string("op_2732_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_2732_end_mask_0 = const()[name = string("op_2732_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2732_squeeze_mask_0 = const()[name = string("op_2732_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2732_cast_fp16 = slice_by_index(begin = var_2732_begin_0, end = var_2732_end_0, end_mask = var_2732_end_mask_0, squeeze_mask = var_2732_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2732_cast_fp16")]; + bool var_2733_transpose_x_0 = const()[name = string("op_2733_transpose_x_0"), val = bool(false)]; + bool var_2733_transpose_y_0 = const()[name = string("op_2733_transpose_y_0"), val = bool(false)]; + tensor var_2733_cast_fp16 = matmul(transpose_x = var_2733_transpose_x_0, transpose_y = var_2733_transpose_y_0, x = var_2732_cast_fp16, y = state_139_cast_fp16)[name = string("op_2733_cast_fp16")]; + tensor var_2735_begin_0 = const()[name = string("op_2735_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_2735_end_0 = const()[name = string("op_2735_end_0"), val = tensor([32, 6, 64, 64])]; + tensor var_2735_end_mask_0 = const()[name = string("op_2735_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2735_squeeze_mask_0 = const()[name = string("op_2735_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2735_cast_fp16 = slice_by_index(begin = var_2735_begin_0, end = var_2735_end_0, end_mask = var_2735_end_mask_0, squeeze_mask = var_2735_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2735_cast_fp16")]; + bool var_2736_transpose_x_0 = const()[name = string("op_2736_transpose_x_0"), val = bool(false)]; + bool var_2736_transpose_y_0 = const()[name = string("op_2736_transpose_y_0"), val = bool(false)]; + tensor var_2736_cast_fp16 = matmul(transpose_x = var_2736_transpose_x_0, transpose_y = var_2736_transpose_y_0, x = var_2735_cast_fp16, y = v_new_139_cast_fp16)[name = string("op_2736_cast_fp16")]; + tensor var_2737_cast_fp16 = add(x = var_2733_cast_fp16, y = var_2736_cast_fp16)[name = string("op_2737_cast_fp16")]; + tensor var_2739_begin_0 = const()[name = string("op_2739_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_2739_end_0 = const()[name = string("op_2739_end_0"), val = tensor([32, 6, 1, 1])]; + tensor var_2739_end_mask_0 = const()[name = string("op_2739_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2739_squeeze_mask_0 = const()[name = string("op_2739_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2739_cast_fp16 = slice_by_index(begin = var_2739_begin_0, end = var_2739_end_0, end_mask = var_2739_end_mask_0, squeeze_mask = var_2739_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2739_cast_fp16")]; + tensor var_2740_cast_fp16 = mul(x = state_139_cast_fp16, y = var_2739_cast_fp16)[name = string("op_2740_cast_fp16")]; + tensor var_2742_begin_0 = const()[name = string("op_2742_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_2742_end_0 = const()[name = string("op_2742_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_2742_end_mask_0 = const()[name = string("op_2742_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2742_squeeze_mask_0 = const()[name = string("op_2742_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2742_cast_fp16 = slice_by_index(begin = var_2742_begin_0, end = var_2742_end_0, end_mask = var_2742_end_mask_0, squeeze_mask = var_2742_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2742_cast_fp16")]; + bool var_2744_transpose_x_1 = const()[name = string("op_2744_transpose_x_1"), val = bool(true)]; + bool var_2744_transpose_y_1 = const()[name = string("op_2744_transpose_y_1"), val = bool(false)]; + tensor var_2744_cast_fp16 = matmul(transpose_x = var_2744_transpose_x_1, transpose_y = var_2744_transpose_y_1, x = var_2742_cast_fp16, y = v_new_139_cast_fp16)[name = string("op_2744_cast_fp16")]; + tensor state_141_cast_fp16 = add(x = var_2740_cast_fp16, y = var_2744_cast_fp16)[name = string("state_141_cast_fp16")]; + tensor var_2747_begin_0 = const()[name = string("op_2747_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_2747_end_0 = const()[name = string("op_2747_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_2747_end_mask_0 = const()[name = string("op_2747_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2747_squeeze_mask_0 = const()[name = string("op_2747_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2747_cast_fp16 = slice_by_index(begin = var_2747_begin_0, end = var_2747_end_0, end_mask = var_2747_end_mask_0, squeeze_mask = var_2747_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2747_cast_fp16")]; + tensor var_2749_begin_0 = const()[name = string("op_2749_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_2749_end_0 = const()[name = string("op_2749_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_2749_end_mask_0 = const()[name = string("op_2749_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2749_squeeze_mask_0 = const()[name = string("op_2749_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2749_cast_fp16 = slice_by_index(begin = var_2749_begin_0, end = var_2749_end_0, end_mask = var_2749_end_mask_0, squeeze_mask = var_2749_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2749_cast_fp16")]; + bool var_2750_transpose_x_0 = const()[name = string("op_2750_transpose_x_0"), val = bool(false)]; + bool var_2750_transpose_y_0 = const()[name = string("op_2750_transpose_y_0"), val = bool(false)]; + tensor var_2750_cast_fp16 = matmul(transpose_x = var_2750_transpose_x_0, transpose_y = var_2750_transpose_y_0, x = var_2749_cast_fp16, y = state_141_cast_fp16)[name = string("op_2750_cast_fp16")]; + tensor v_new_141_cast_fp16 = sub(x = var_2747_cast_fp16, y = var_2750_cast_fp16)[name = string("v_new_141_cast_fp16")]; + tensor var_2753_begin_0 = const()[name = string("op_2753_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_2753_end_0 = const()[name = string("op_2753_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_2753_end_mask_0 = const()[name = string("op_2753_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2753_squeeze_mask_0 = const()[name = string("op_2753_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2753_cast_fp16 = slice_by_index(begin = var_2753_begin_0, end = var_2753_end_0, end_mask = var_2753_end_mask_0, squeeze_mask = var_2753_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2753_cast_fp16")]; + bool var_2754_transpose_x_0 = const()[name = string("op_2754_transpose_x_0"), val = bool(false)]; + bool var_2754_transpose_y_0 = const()[name = string("op_2754_transpose_y_0"), val = bool(false)]; + tensor var_2754_cast_fp16 = matmul(transpose_x = var_2754_transpose_x_0, transpose_y = var_2754_transpose_y_0, x = var_2753_cast_fp16, y = state_141_cast_fp16)[name = string("op_2754_cast_fp16")]; + tensor var_2756_begin_0 = const()[name = string("op_2756_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_2756_end_0 = const()[name = string("op_2756_end_0"), val = tensor([32, 7, 64, 64])]; + tensor var_2756_end_mask_0 = const()[name = string("op_2756_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2756_squeeze_mask_0 = const()[name = string("op_2756_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2756_cast_fp16 = slice_by_index(begin = var_2756_begin_0, end = var_2756_end_0, end_mask = var_2756_end_mask_0, squeeze_mask = var_2756_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2756_cast_fp16")]; + bool var_2757_transpose_x_0 = const()[name = string("op_2757_transpose_x_0"), val = bool(false)]; + bool var_2757_transpose_y_0 = const()[name = string("op_2757_transpose_y_0"), val = bool(false)]; + tensor var_2757_cast_fp16 = matmul(transpose_x = var_2757_transpose_x_0, transpose_y = var_2757_transpose_y_0, x = var_2756_cast_fp16, y = v_new_141_cast_fp16)[name = string("op_2757_cast_fp16")]; + tensor var_2758_cast_fp16 = add(x = var_2754_cast_fp16, y = var_2757_cast_fp16)[name = string("op_2758_cast_fp16")]; + tensor var_2760_begin_0 = const()[name = string("op_2760_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_2760_end_0 = const()[name = string("op_2760_end_0"), val = tensor([32, 7, 1, 1])]; + tensor var_2760_end_mask_0 = const()[name = string("op_2760_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2760_squeeze_mask_0 = const()[name = string("op_2760_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2760_cast_fp16 = slice_by_index(begin = var_2760_begin_0, end = var_2760_end_0, end_mask = var_2760_end_mask_0, squeeze_mask = var_2760_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2760_cast_fp16")]; + tensor var_2761_cast_fp16 = mul(x = state_141_cast_fp16, y = var_2760_cast_fp16)[name = string("op_2761_cast_fp16")]; + tensor var_2763_begin_0 = const()[name = string("op_2763_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_2763_end_0 = const()[name = string("op_2763_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_2763_end_mask_0 = const()[name = string("op_2763_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2763_squeeze_mask_0 = const()[name = string("op_2763_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2763_cast_fp16 = slice_by_index(begin = var_2763_begin_0, end = var_2763_end_0, end_mask = var_2763_end_mask_0, squeeze_mask = var_2763_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2763_cast_fp16")]; + bool var_2765_transpose_x_1 = const()[name = string("op_2765_transpose_x_1"), val = bool(true)]; + bool var_2765_transpose_y_1 = const()[name = string("op_2765_transpose_y_1"), val = bool(false)]; + tensor var_2765_cast_fp16 = matmul(transpose_x = var_2765_transpose_x_1, transpose_y = var_2765_transpose_y_1, x = var_2763_cast_fp16, y = v_new_141_cast_fp16)[name = string("op_2765_cast_fp16")]; + tensor state_143_cast_fp16 = add(x = var_2761_cast_fp16, y = var_2765_cast_fp16)[name = string("state_143_cast_fp16")]; + tensor var_2768_begin_0 = const()[name = string("op_2768_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_2768_end_0 = const()[name = string("op_2768_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_2768_end_mask_0 = const()[name = string("op_2768_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2768_squeeze_mask_0 = const()[name = string("op_2768_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2768_cast_fp16 = slice_by_index(begin = var_2768_begin_0, end = var_2768_end_0, end_mask = var_2768_end_mask_0, squeeze_mask = var_2768_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2768_cast_fp16")]; + tensor var_2770_begin_0 = const()[name = string("op_2770_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_2770_end_0 = const()[name = string("op_2770_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_2770_end_mask_0 = const()[name = string("op_2770_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2770_squeeze_mask_0 = const()[name = string("op_2770_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2770_cast_fp16 = slice_by_index(begin = var_2770_begin_0, end = var_2770_end_0, end_mask = var_2770_end_mask_0, squeeze_mask = var_2770_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2770_cast_fp16")]; + bool var_2771_transpose_x_0 = const()[name = string("op_2771_transpose_x_0"), val = bool(false)]; + bool var_2771_transpose_y_0 = const()[name = string("op_2771_transpose_y_0"), val = bool(false)]; + tensor var_2771_cast_fp16 = matmul(transpose_x = var_2771_transpose_x_0, transpose_y = var_2771_transpose_y_0, x = var_2770_cast_fp16, y = state_143_cast_fp16)[name = string("op_2771_cast_fp16")]; + tensor v_new_143_cast_fp16 = sub(x = var_2768_cast_fp16, y = var_2771_cast_fp16)[name = string("v_new_143_cast_fp16")]; + tensor var_2774_begin_0 = const()[name = string("op_2774_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_2774_end_0 = const()[name = string("op_2774_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_2774_end_mask_0 = const()[name = string("op_2774_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2774_squeeze_mask_0 = const()[name = string("op_2774_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2774_cast_fp16 = slice_by_index(begin = var_2774_begin_0, end = var_2774_end_0, end_mask = var_2774_end_mask_0, squeeze_mask = var_2774_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2774_cast_fp16")]; + bool var_2775_transpose_x_0 = const()[name = string("op_2775_transpose_x_0"), val = bool(false)]; + bool var_2775_transpose_y_0 = const()[name = string("op_2775_transpose_y_0"), val = bool(false)]; + tensor var_2775_cast_fp16 = matmul(transpose_x = var_2775_transpose_x_0, transpose_y = var_2775_transpose_y_0, x = var_2774_cast_fp16, y = state_143_cast_fp16)[name = string("op_2775_cast_fp16")]; + tensor var_2777_begin_0 = const()[name = string("op_2777_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_2777_end_0 = const()[name = string("op_2777_end_0"), val = tensor([32, 8, 64, 64])]; + tensor var_2777_end_mask_0 = const()[name = string("op_2777_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2777_squeeze_mask_0 = const()[name = string("op_2777_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2777_cast_fp16 = slice_by_index(begin = var_2777_begin_0, end = var_2777_end_0, end_mask = var_2777_end_mask_0, squeeze_mask = var_2777_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2777_cast_fp16")]; + bool var_2778_transpose_x_0 = const()[name = string("op_2778_transpose_x_0"), val = bool(false)]; + bool var_2778_transpose_y_0 = const()[name = string("op_2778_transpose_y_0"), val = bool(false)]; + tensor var_2778_cast_fp16 = matmul(transpose_x = var_2778_transpose_x_0, transpose_y = var_2778_transpose_y_0, x = var_2777_cast_fp16, y = v_new_143_cast_fp16)[name = string("op_2778_cast_fp16")]; + tensor var_2779_cast_fp16 = add(x = var_2775_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2779_cast_fp16")]; + tensor var_2781_begin_0 = const()[name = string("op_2781_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_2781_end_0 = const()[name = string("op_2781_end_0"), val = tensor([32, 8, 1, 1])]; + tensor var_2781_end_mask_0 = const()[name = string("op_2781_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2781_squeeze_mask_0 = const()[name = string("op_2781_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2781_cast_fp16 = slice_by_index(begin = var_2781_begin_0, end = var_2781_end_0, end_mask = var_2781_end_mask_0, squeeze_mask = var_2781_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2781_cast_fp16")]; + tensor var_2782_cast_fp16 = mul(x = state_143_cast_fp16, y = var_2781_cast_fp16)[name = string("op_2782_cast_fp16")]; + tensor var_2784_begin_0 = const()[name = string("op_2784_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_2784_end_0 = const()[name = string("op_2784_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_2784_end_mask_0 = const()[name = string("op_2784_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2784_squeeze_mask_0 = const()[name = string("op_2784_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2784_cast_fp16 = slice_by_index(begin = var_2784_begin_0, end = var_2784_end_0, end_mask = var_2784_end_mask_0, squeeze_mask = var_2784_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2784_cast_fp16")]; + bool var_2786_transpose_x_1 = const()[name = string("op_2786_transpose_x_1"), val = bool(true)]; + bool var_2786_transpose_y_1 = const()[name = string("op_2786_transpose_y_1"), val = bool(false)]; + tensor var_2786_cast_fp16 = matmul(transpose_x = var_2786_transpose_x_1, transpose_y = var_2786_transpose_y_1, x = var_2784_cast_fp16, y = v_new_143_cast_fp16)[name = string("op_2786_cast_fp16")]; + tensor state_145_cast_fp16 = add(x = var_2782_cast_fp16, y = var_2786_cast_fp16)[name = string("state_145_cast_fp16")]; + tensor var_2789_begin_0 = const()[name = string("op_2789_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_2789_end_0 = const()[name = string("op_2789_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_2789_end_mask_0 = const()[name = string("op_2789_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2789_squeeze_mask_0 = const()[name = string("op_2789_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2789_cast_fp16 = slice_by_index(begin = var_2789_begin_0, end = var_2789_end_0, end_mask = var_2789_end_mask_0, squeeze_mask = var_2789_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2789_cast_fp16")]; + tensor var_2791_begin_0 = const()[name = string("op_2791_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_2791_end_0 = const()[name = string("op_2791_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_2791_end_mask_0 = const()[name = string("op_2791_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2791_squeeze_mask_0 = const()[name = string("op_2791_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2791_cast_fp16 = slice_by_index(begin = var_2791_begin_0, end = var_2791_end_0, end_mask = var_2791_end_mask_0, squeeze_mask = var_2791_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2791_cast_fp16")]; + bool var_2792_transpose_x_0 = const()[name = string("op_2792_transpose_x_0"), val = bool(false)]; + bool var_2792_transpose_y_0 = const()[name = string("op_2792_transpose_y_0"), val = bool(false)]; + tensor var_2792_cast_fp16 = matmul(transpose_x = var_2792_transpose_x_0, transpose_y = var_2792_transpose_y_0, x = var_2791_cast_fp16, y = state_145_cast_fp16)[name = string("op_2792_cast_fp16")]; + tensor v_new_145_cast_fp16 = sub(x = var_2789_cast_fp16, y = var_2792_cast_fp16)[name = string("v_new_145_cast_fp16")]; + tensor var_2795_begin_0 = const()[name = string("op_2795_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_2795_end_0 = const()[name = string("op_2795_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_2795_end_mask_0 = const()[name = string("op_2795_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2795_squeeze_mask_0 = const()[name = string("op_2795_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2795_cast_fp16 = slice_by_index(begin = var_2795_begin_0, end = var_2795_end_0, end_mask = var_2795_end_mask_0, squeeze_mask = var_2795_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2795_cast_fp16")]; + bool var_2796_transpose_x_0 = const()[name = string("op_2796_transpose_x_0"), val = bool(false)]; + bool var_2796_transpose_y_0 = const()[name = string("op_2796_transpose_y_0"), val = bool(false)]; + tensor var_2796_cast_fp16 = matmul(transpose_x = var_2796_transpose_x_0, transpose_y = var_2796_transpose_y_0, x = var_2795_cast_fp16, y = state_145_cast_fp16)[name = string("op_2796_cast_fp16")]; + tensor var_2798_begin_0 = const()[name = string("op_2798_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_2798_end_0 = const()[name = string("op_2798_end_0"), val = tensor([32, 9, 64, 64])]; + tensor var_2798_end_mask_0 = const()[name = string("op_2798_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2798_squeeze_mask_0 = const()[name = string("op_2798_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2798_cast_fp16 = slice_by_index(begin = var_2798_begin_0, end = var_2798_end_0, end_mask = var_2798_end_mask_0, squeeze_mask = var_2798_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2798_cast_fp16")]; + bool var_2799_transpose_x_0 = const()[name = string("op_2799_transpose_x_0"), val = bool(false)]; + bool var_2799_transpose_y_0 = const()[name = string("op_2799_transpose_y_0"), val = bool(false)]; + tensor var_2799_cast_fp16 = matmul(transpose_x = var_2799_transpose_x_0, transpose_y = var_2799_transpose_y_0, x = var_2798_cast_fp16, y = v_new_145_cast_fp16)[name = string("op_2799_cast_fp16")]; + tensor var_2800_cast_fp16 = add(x = var_2796_cast_fp16, y = var_2799_cast_fp16)[name = string("op_2800_cast_fp16")]; + tensor var_2802_begin_0 = const()[name = string("op_2802_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_2802_end_0 = const()[name = string("op_2802_end_0"), val = tensor([32, 9, 1, 1])]; + tensor var_2802_end_mask_0 = const()[name = string("op_2802_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2802_squeeze_mask_0 = const()[name = string("op_2802_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2802_cast_fp16 = slice_by_index(begin = var_2802_begin_0, end = var_2802_end_0, end_mask = var_2802_end_mask_0, squeeze_mask = var_2802_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2802_cast_fp16")]; + tensor var_2803_cast_fp16 = mul(x = state_145_cast_fp16, y = var_2802_cast_fp16)[name = string("op_2803_cast_fp16")]; + tensor var_2805_begin_0 = const()[name = string("op_2805_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_2805_end_0 = const()[name = string("op_2805_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_2805_end_mask_0 = const()[name = string("op_2805_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2805_squeeze_mask_0 = const()[name = string("op_2805_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2805_cast_fp16 = slice_by_index(begin = var_2805_begin_0, end = var_2805_end_0, end_mask = var_2805_end_mask_0, squeeze_mask = var_2805_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2805_cast_fp16")]; + bool var_2807_transpose_x_1 = const()[name = string("op_2807_transpose_x_1"), val = bool(true)]; + bool var_2807_transpose_y_1 = const()[name = string("op_2807_transpose_y_1"), val = bool(false)]; + tensor var_2807_cast_fp16 = matmul(transpose_x = var_2807_transpose_x_1, transpose_y = var_2807_transpose_y_1, x = var_2805_cast_fp16, y = v_new_145_cast_fp16)[name = string("op_2807_cast_fp16")]; + tensor state_147_cast_fp16 = add(x = var_2803_cast_fp16, y = var_2807_cast_fp16)[name = string("state_147_cast_fp16")]; + tensor var_2810_begin_0 = const()[name = string("op_2810_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_2810_end_0 = const()[name = string("op_2810_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_2810_end_mask_0 = const()[name = string("op_2810_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2810_squeeze_mask_0 = const()[name = string("op_2810_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2810_cast_fp16 = slice_by_index(begin = var_2810_begin_0, end = var_2810_end_0, end_mask = var_2810_end_mask_0, squeeze_mask = var_2810_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2810_cast_fp16")]; + tensor var_2812_begin_0 = const()[name = string("op_2812_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_2812_end_0 = const()[name = string("op_2812_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_2812_end_mask_0 = const()[name = string("op_2812_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2812_squeeze_mask_0 = const()[name = string("op_2812_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, squeeze_mask = var_2812_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2812_cast_fp16")]; + bool var_2813_transpose_x_0 = const()[name = string("op_2813_transpose_x_0"), val = bool(false)]; + bool var_2813_transpose_y_0 = const()[name = string("op_2813_transpose_y_0"), val = bool(false)]; + tensor var_2813_cast_fp16 = matmul(transpose_x = var_2813_transpose_x_0, transpose_y = var_2813_transpose_y_0, x = var_2812_cast_fp16, y = state_147_cast_fp16)[name = string("op_2813_cast_fp16")]; + tensor v_new_147_cast_fp16 = sub(x = var_2810_cast_fp16, y = var_2813_cast_fp16)[name = string("v_new_147_cast_fp16")]; + tensor var_2816_begin_0 = const()[name = string("op_2816_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_2816_end_0 = const()[name = string("op_2816_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_2816_end_mask_0 = const()[name = string("op_2816_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2816_squeeze_mask_0 = const()[name = string("op_2816_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2816_cast_fp16 = slice_by_index(begin = var_2816_begin_0, end = var_2816_end_0, end_mask = var_2816_end_mask_0, squeeze_mask = var_2816_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2816_cast_fp16")]; + bool var_2817_transpose_x_0 = const()[name = string("op_2817_transpose_x_0"), val = bool(false)]; + bool var_2817_transpose_y_0 = const()[name = string("op_2817_transpose_y_0"), val = bool(false)]; + tensor var_2817_cast_fp16 = matmul(transpose_x = var_2817_transpose_x_0, transpose_y = var_2817_transpose_y_0, x = var_2816_cast_fp16, y = state_147_cast_fp16)[name = string("op_2817_cast_fp16")]; + tensor var_2819_begin_0 = const()[name = string("op_2819_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_2819_end_0 = const()[name = string("op_2819_end_0"), val = tensor([32, 10, 64, 64])]; + tensor var_2819_end_mask_0 = const()[name = string("op_2819_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2819_squeeze_mask_0 = const()[name = string("op_2819_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2819_cast_fp16 = slice_by_index(begin = var_2819_begin_0, end = var_2819_end_0, end_mask = var_2819_end_mask_0, squeeze_mask = var_2819_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2819_cast_fp16")]; + bool var_2820_transpose_x_0 = const()[name = string("op_2820_transpose_x_0"), val = bool(false)]; + bool var_2820_transpose_y_0 = const()[name = string("op_2820_transpose_y_0"), val = bool(false)]; + tensor var_2820_cast_fp16 = matmul(transpose_x = var_2820_transpose_x_0, transpose_y = var_2820_transpose_y_0, x = var_2819_cast_fp16, y = v_new_147_cast_fp16)[name = string("op_2820_cast_fp16")]; + tensor var_2821_cast_fp16 = add(x = var_2817_cast_fp16, y = var_2820_cast_fp16)[name = string("op_2821_cast_fp16")]; + tensor var_2823_begin_0 = const()[name = string("op_2823_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_2823_end_0 = const()[name = string("op_2823_end_0"), val = tensor([32, 10, 1, 1])]; + tensor var_2823_end_mask_0 = const()[name = string("op_2823_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2823_squeeze_mask_0 = const()[name = string("op_2823_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2823_cast_fp16 = slice_by_index(begin = var_2823_begin_0, end = var_2823_end_0, end_mask = var_2823_end_mask_0, squeeze_mask = var_2823_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2823_cast_fp16")]; + tensor var_2824_cast_fp16 = mul(x = state_147_cast_fp16, y = var_2823_cast_fp16)[name = string("op_2824_cast_fp16")]; + tensor var_2826_begin_0 = const()[name = string("op_2826_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_2826_end_0 = const()[name = string("op_2826_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_2826_end_mask_0 = const()[name = string("op_2826_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2826_squeeze_mask_0 = const()[name = string("op_2826_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2826_cast_fp16 = slice_by_index(begin = var_2826_begin_0, end = var_2826_end_0, end_mask = var_2826_end_mask_0, squeeze_mask = var_2826_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2826_cast_fp16")]; + bool var_2828_transpose_x_1 = const()[name = string("op_2828_transpose_x_1"), val = bool(true)]; + bool var_2828_transpose_y_1 = const()[name = string("op_2828_transpose_y_1"), val = bool(false)]; + tensor var_2828_cast_fp16 = matmul(transpose_x = var_2828_transpose_x_1, transpose_y = var_2828_transpose_y_1, x = var_2826_cast_fp16, y = v_new_147_cast_fp16)[name = string("op_2828_cast_fp16")]; + tensor state_149_cast_fp16 = add(x = var_2824_cast_fp16, y = var_2828_cast_fp16)[name = string("state_149_cast_fp16")]; + tensor var_2831_begin_0 = const()[name = string("op_2831_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_2831_end_0 = const()[name = string("op_2831_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_2831_end_mask_0 = const()[name = string("op_2831_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2831_squeeze_mask_0 = const()[name = string("op_2831_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2831_cast_fp16 = slice_by_index(begin = var_2831_begin_0, end = var_2831_end_0, end_mask = var_2831_end_mask_0, squeeze_mask = var_2831_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2831_cast_fp16")]; + tensor var_2833_begin_0 = const()[name = string("op_2833_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_2833_end_0 = const()[name = string("op_2833_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_2833_end_mask_0 = const()[name = string("op_2833_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2833_squeeze_mask_0 = const()[name = string("op_2833_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2833_cast_fp16 = slice_by_index(begin = var_2833_begin_0, end = var_2833_end_0, end_mask = var_2833_end_mask_0, squeeze_mask = var_2833_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2833_cast_fp16")]; + bool var_2834_transpose_x_0 = const()[name = string("op_2834_transpose_x_0"), val = bool(false)]; + bool var_2834_transpose_y_0 = const()[name = string("op_2834_transpose_y_0"), val = bool(false)]; + tensor var_2834_cast_fp16 = matmul(transpose_x = var_2834_transpose_x_0, transpose_y = var_2834_transpose_y_0, x = var_2833_cast_fp16, y = state_149_cast_fp16)[name = string("op_2834_cast_fp16")]; + tensor v_new_149_cast_fp16 = sub(x = var_2831_cast_fp16, y = var_2834_cast_fp16)[name = string("v_new_149_cast_fp16")]; + tensor var_2837_begin_0 = const()[name = string("op_2837_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_2837_end_0 = const()[name = string("op_2837_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_2837_end_mask_0 = const()[name = string("op_2837_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2837_squeeze_mask_0 = const()[name = string("op_2837_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2837_cast_fp16 = slice_by_index(begin = var_2837_begin_0, end = var_2837_end_0, end_mask = var_2837_end_mask_0, squeeze_mask = var_2837_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2837_cast_fp16")]; + bool var_2838_transpose_x_0 = const()[name = string("op_2838_transpose_x_0"), val = bool(false)]; + bool var_2838_transpose_y_0 = const()[name = string("op_2838_transpose_y_0"), val = bool(false)]; + tensor var_2838_cast_fp16 = matmul(transpose_x = var_2838_transpose_x_0, transpose_y = var_2838_transpose_y_0, x = var_2837_cast_fp16, y = state_149_cast_fp16)[name = string("op_2838_cast_fp16")]; + tensor var_2840_begin_0 = const()[name = string("op_2840_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_2840_end_0 = const()[name = string("op_2840_end_0"), val = tensor([32, 11, 64, 64])]; + tensor var_2840_end_mask_0 = const()[name = string("op_2840_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2840_squeeze_mask_0 = const()[name = string("op_2840_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2840_cast_fp16 = slice_by_index(begin = var_2840_begin_0, end = var_2840_end_0, end_mask = var_2840_end_mask_0, squeeze_mask = var_2840_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2840_cast_fp16")]; + bool var_2841_transpose_x_0 = const()[name = string("op_2841_transpose_x_0"), val = bool(false)]; + bool var_2841_transpose_y_0 = const()[name = string("op_2841_transpose_y_0"), val = bool(false)]; + tensor var_2841_cast_fp16 = matmul(transpose_x = var_2841_transpose_x_0, transpose_y = var_2841_transpose_y_0, x = var_2840_cast_fp16, y = v_new_149_cast_fp16)[name = string("op_2841_cast_fp16")]; + tensor var_2842_cast_fp16 = add(x = var_2838_cast_fp16, y = var_2841_cast_fp16)[name = string("op_2842_cast_fp16")]; + tensor var_2844_begin_0 = const()[name = string("op_2844_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_2844_end_0 = const()[name = string("op_2844_end_0"), val = tensor([32, 11, 1, 1])]; + tensor var_2844_end_mask_0 = const()[name = string("op_2844_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2844_squeeze_mask_0 = const()[name = string("op_2844_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2844_cast_fp16 = slice_by_index(begin = var_2844_begin_0, end = var_2844_end_0, end_mask = var_2844_end_mask_0, squeeze_mask = var_2844_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2844_cast_fp16")]; + tensor var_2845_cast_fp16 = mul(x = state_149_cast_fp16, y = var_2844_cast_fp16)[name = string("op_2845_cast_fp16")]; + tensor var_2847_begin_0 = const()[name = string("op_2847_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_2847_end_0 = const()[name = string("op_2847_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_2847_end_mask_0 = const()[name = string("op_2847_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2847_squeeze_mask_0 = const()[name = string("op_2847_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2847_cast_fp16 = slice_by_index(begin = var_2847_begin_0, end = var_2847_end_0, end_mask = var_2847_end_mask_0, squeeze_mask = var_2847_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2847_cast_fp16")]; + bool var_2849_transpose_x_1 = const()[name = string("op_2849_transpose_x_1"), val = bool(true)]; + bool var_2849_transpose_y_1 = const()[name = string("op_2849_transpose_y_1"), val = bool(false)]; + tensor var_2849_cast_fp16 = matmul(transpose_x = var_2849_transpose_x_1, transpose_y = var_2849_transpose_y_1, x = var_2847_cast_fp16, y = v_new_149_cast_fp16)[name = string("op_2849_cast_fp16")]; + tensor state_151_cast_fp16 = add(x = var_2845_cast_fp16, y = var_2849_cast_fp16)[name = string("state_151_cast_fp16")]; + tensor var_2852_begin_0 = const()[name = string("op_2852_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_2852_end_0 = const()[name = string("op_2852_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_2852_end_mask_0 = const()[name = string("op_2852_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2852_squeeze_mask_0 = const()[name = string("op_2852_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2852_cast_fp16 = slice_by_index(begin = var_2852_begin_0, end = var_2852_end_0, end_mask = var_2852_end_mask_0, squeeze_mask = var_2852_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2852_cast_fp16")]; + tensor var_2854_begin_0 = const()[name = string("op_2854_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_2854_end_0 = const()[name = string("op_2854_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_2854_end_mask_0 = const()[name = string("op_2854_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2854_squeeze_mask_0 = const()[name = string("op_2854_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2854_cast_fp16 = slice_by_index(begin = var_2854_begin_0, end = var_2854_end_0, end_mask = var_2854_end_mask_0, squeeze_mask = var_2854_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2854_cast_fp16")]; + bool var_2855_transpose_x_0 = const()[name = string("op_2855_transpose_x_0"), val = bool(false)]; + bool var_2855_transpose_y_0 = const()[name = string("op_2855_transpose_y_0"), val = bool(false)]; + tensor var_2855_cast_fp16 = matmul(transpose_x = var_2855_transpose_x_0, transpose_y = var_2855_transpose_y_0, x = var_2854_cast_fp16, y = state_151_cast_fp16)[name = string("op_2855_cast_fp16")]; + tensor v_new_151_cast_fp16 = sub(x = var_2852_cast_fp16, y = var_2855_cast_fp16)[name = string("v_new_151_cast_fp16")]; + tensor var_2858_begin_0 = const()[name = string("op_2858_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_2858_end_0 = const()[name = string("op_2858_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_2858_end_mask_0 = const()[name = string("op_2858_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2858_squeeze_mask_0 = const()[name = string("op_2858_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2858_cast_fp16 = slice_by_index(begin = var_2858_begin_0, end = var_2858_end_0, end_mask = var_2858_end_mask_0, squeeze_mask = var_2858_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2858_cast_fp16")]; + bool var_2859_transpose_x_0 = const()[name = string("op_2859_transpose_x_0"), val = bool(false)]; + bool var_2859_transpose_y_0 = const()[name = string("op_2859_transpose_y_0"), val = bool(false)]; + tensor var_2859_cast_fp16 = matmul(transpose_x = var_2859_transpose_x_0, transpose_y = var_2859_transpose_y_0, x = var_2858_cast_fp16, y = state_151_cast_fp16)[name = string("op_2859_cast_fp16")]; + tensor var_2861_begin_0 = const()[name = string("op_2861_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_2861_end_0 = const()[name = string("op_2861_end_0"), val = tensor([32, 12, 64, 64])]; + tensor var_2861_end_mask_0 = const()[name = string("op_2861_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2861_squeeze_mask_0 = const()[name = string("op_2861_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2861_cast_fp16 = slice_by_index(begin = var_2861_begin_0, end = var_2861_end_0, end_mask = var_2861_end_mask_0, squeeze_mask = var_2861_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2861_cast_fp16")]; + bool var_2862_transpose_x_0 = const()[name = string("op_2862_transpose_x_0"), val = bool(false)]; + bool var_2862_transpose_y_0 = const()[name = string("op_2862_transpose_y_0"), val = bool(false)]; + tensor var_2862_cast_fp16 = matmul(transpose_x = var_2862_transpose_x_0, transpose_y = var_2862_transpose_y_0, x = var_2861_cast_fp16, y = v_new_151_cast_fp16)[name = string("op_2862_cast_fp16")]; + tensor var_2863_cast_fp16 = add(x = var_2859_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")]; + tensor var_2865_begin_0 = const()[name = string("op_2865_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_2865_end_0 = const()[name = string("op_2865_end_0"), val = tensor([32, 12, 1, 1])]; + tensor var_2865_end_mask_0 = const()[name = string("op_2865_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2865_squeeze_mask_0 = const()[name = string("op_2865_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2865_cast_fp16 = slice_by_index(begin = var_2865_begin_0, end = var_2865_end_0, end_mask = var_2865_end_mask_0, squeeze_mask = var_2865_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2865_cast_fp16")]; + tensor var_2866_cast_fp16 = mul(x = state_151_cast_fp16, y = var_2865_cast_fp16)[name = string("op_2866_cast_fp16")]; + tensor var_2868_begin_0 = const()[name = string("op_2868_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_2868_end_0 = const()[name = string("op_2868_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_2868_end_mask_0 = const()[name = string("op_2868_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2868_squeeze_mask_0 = const()[name = string("op_2868_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2868_cast_fp16 = slice_by_index(begin = var_2868_begin_0, end = var_2868_end_0, end_mask = var_2868_end_mask_0, squeeze_mask = var_2868_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2868_cast_fp16")]; + bool var_2870_transpose_x_1 = const()[name = string("op_2870_transpose_x_1"), val = bool(true)]; + bool var_2870_transpose_y_1 = const()[name = string("op_2870_transpose_y_1"), val = bool(false)]; + tensor var_2870_cast_fp16 = matmul(transpose_x = var_2870_transpose_x_1, transpose_y = var_2870_transpose_y_1, x = var_2868_cast_fp16, y = v_new_151_cast_fp16)[name = string("op_2870_cast_fp16")]; + tensor state_153_cast_fp16 = add(x = var_2866_cast_fp16, y = var_2870_cast_fp16)[name = string("state_153_cast_fp16")]; + tensor var_2873_begin_0 = const()[name = string("op_2873_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_2873_end_0 = const()[name = string("op_2873_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_2873_end_mask_0 = const()[name = string("op_2873_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2873_squeeze_mask_0 = const()[name = string("op_2873_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2873_cast_fp16 = slice_by_index(begin = var_2873_begin_0, end = var_2873_end_0, end_mask = var_2873_end_mask_0, squeeze_mask = var_2873_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2873_cast_fp16")]; + tensor var_2875_begin_0 = const()[name = string("op_2875_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_2875_end_0 = const()[name = string("op_2875_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_2875_end_mask_0 = const()[name = string("op_2875_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2875_squeeze_mask_0 = const()[name = string("op_2875_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2875_cast_fp16 = slice_by_index(begin = var_2875_begin_0, end = var_2875_end_0, end_mask = var_2875_end_mask_0, squeeze_mask = var_2875_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2875_cast_fp16")]; + bool var_2876_transpose_x_0 = const()[name = string("op_2876_transpose_x_0"), val = bool(false)]; + bool var_2876_transpose_y_0 = const()[name = string("op_2876_transpose_y_0"), val = bool(false)]; + tensor var_2876_cast_fp16 = matmul(transpose_x = var_2876_transpose_x_0, transpose_y = var_2876_transpose_y_0, x = var_2875_cast_fp16, y = state_153_cast_fp16)[name = string("op_2876_cast_fp16")]; + tensor v_new_153_cast_fp16 = sub(x = var_2873_cast_fp16, y = var_2876_cast_fp16)[name = string("v_new_153_cast_fp16")]; + tensor var_2879_begin_0 = const()[name = string("op_2879_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_2879_end_0 = const()[name = string("op_2879_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_2879_end_mask_0 = const()[name = string("op_2879_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2879_squeeze_mask_0 = const()[name = string("op_2879_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2879_cast_fp16 = slice_by_index(begin = var_2879_begin_0, end = var_2879_end_0, end_mask = var_2879_end_mask_0, squeeze_mask = var_2879_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2879_cast_fp16")]; + bool var_2880_transpose_x_0 = const()[name = string("op_2880_transpose_x_0"), val = bool(false)]; + bool var_2880_transpose_y_0 = const()[name = string("op_2880_transpose_y_0"), val = bool(false)]; + tensor var_2880_cast_fp16 = matmul(transpose_x = var_2880_transpose_x_0, transpose_y = var_2880_transpose_y_0, x = var_2879_cast_fp16, y = state_153_cast_fp16)[name = string("op_2880_cast_fp16")]; + tensor var_2882_begin_0 = const()[name = string("op_2882_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_2882_end_0 = const()[name = string("op_2882_end_0"), val = tensor([32, 13, 64, 64])]; + tensor var_2882_end_mask_0 = const()[name = string("op_2882_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2882_squeeze_mask_0 = const()[name = string("op_2882_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2882_cast_fp16 = slice_by_index(begin = var_2882_begin_0, end = var_2882_end_0, end_mask = var_2882_end_mask_0, squeeze_mask = var_2882_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2882_cast_fp16")]; + bool var_2883_transpose_x_0 = const()[name = string("op_2883_transpose_x_0"), val = bool(false)]; + bool var_2883_transpose_y_0 = const()[name = string("op_2883_transpose_y_0"), val = bool(false)]; + tensor var_2883_cast_fp16 = matmul(transpose_x = var_2883_transpose_x_0, transpose_y = var_2883_transpose_y_0, x = var_2882_cast_fp16, y = v_new_153_cast_fp16)[name = string("op_2883_cast_fp16")]; + tensor var_2884_cast_fp16 = add(x = var_2880_cast_fp16, y = var_2883_cast_fp16)[name = string("op_2884_cast_fp16")]; + tensor var_2886_begin_0 = const()[name = string("op_2886_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_2886_end_0 = const()[name = string("op_2886_end_0"), val = tensor([32, 13, 1, 1])]; + tensor var_2886_end_mask_0 = const()[name = string("op_2886_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2886_squeeze_mask_0 = const()[name = string("op_2886_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2886_cast_fp16 = slice_by_index(begin = var_2886_begin_0, end = var_2886_end_0, end_mask = var_2886_end_mask_0, squeeze_mask = var_2886_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2886_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = state_153_cast_fp16, y = var_2886_cast_fp16)[name = string("op_2887_cast_fp16")]; + tensor var_2889_begin_0 = const()[name = string("op_2889_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_2889_end_0 = const()[name = string("op_2889_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_2889_end_mask_0 = const()[name = string("op_2889_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2889_squeeze_mask_0 = const()[name = string("op_2889_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2889_cast_fp16 = slice_by_index(begin = var_2889_begin_0, end = var_2889_end_0, end_mask = var_2889_end_mask_0, squeeze_mask = var_2889_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2889_cast_fp16")]; + bool var_2891_transpose_x_1 = const()[name = string("op_2891_transpose_x_1"), val = bool(true)]; + bool var_2891_transpose_y_1 = const()[name = string("op_2891_transpose_y_1"), val = bool(false)]; + tensor var_2891_cast_fp16 = matmul(transpose_x = var_2891_transpose_x_1, transpose_y = var_2891_transpose_y_1, x = var_2889_cast_fp16, y = v_new_153_cast_fp16)[name = string("op_2891_cast_fp16")]; + tensor state_155_cast_fp16 = add(x = var_2887_cast_fp16, y = var_2891_cast_fp16)[name = string("state_155_cast_fp16")]; + tensor var_2894_begin_0 = const()[name = string("op_2894_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_2894_end_0 = const()[name = string("op_2894_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_2894_end_mask_0 = const()[name = string("op_2894_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2894_squeeze_mask_0 = const()[name = string("op_2894_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2894_cast_fp16 = slice_by_index(begin = var_2894_begin_0, end = var_2894_end_0, end_mask = var_2894_end_mask_0, squeeze_mask = var_2894_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2894_cast_fp16")]; + tensor var_2896_begin_0 = const()[name = string("op_2896_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_2896_end_0 = const()[name = string("op_2896_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_2896_end_mask_0 = const()[name = string("op_2896_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2896_squeeze_mask_0 = const()[name = string("op_2896_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2896_cast_fp16 = slice_by_index(begin = var_2896_begin_0, end = var_2896_end_0, end_mask = var_2896_end_mask_0, squeeze_mask = var_2896_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2896_cast_fp16")]; + bool var_2897_transpose_x_0 = const()[name = string("op_2897_transpose_x_0"), val = bool(false)]; + bool var_2897_transpose_y_0 = const()[name = string("op_2897_transpose_y_0"), val = bool(false)]; + tensor var_2897_cast_fp16 = matmul(transpose_x = var_2897_transpose_x_0, transpose_y = var_2897_transpose_y_0, x = var_2896_cast_fp16, y = state_155_cast_fp16)[name = string("op_2897_cast_fp16")]; + tensor v_new_155_cast_fp16 = sub(x = var_2894_cast_fp16, y = var_2897_cast_fp16)[name = string("v_new_155_cast_fp16")]; + tensor var_2900_begin_0 = const()[name = string("op_2900_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_2900_end_0 = const()[name = string("op_2900_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_2900_end_mask_0 = const()[name = string("op_2900_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2900_squeeze_mask_0 = const()[name = string("op_2900_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2900_cast_fp16 = slice_by_index(begin = var_2900_begin_0, end = var_2900_end_0, end_mask = var_2900_end_mask_0, squeeze_mask = var_2900_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2900_cast_fp16")]; + bool var_2901_transpose_x_0 = const()[name = string("op_2901_transpose_x_0"), val = bool(false)]; + bool var_2901_transpose_y_0 = const()[name = string("op_2901_transpose_y_0"), val = bool(false)]; + tensor var_2901_cast_fp16 = matmul(transpose_x = var_2901_transpose_x_0, transpose_y = var_2901_transpose_y_0, x = var_2900_cast_fp16, y = state_155_cast_fp16)[name = string("op_2901_cast_fp16")]; + tensor var_2903_begin_0 = const()[name = string("op_2903_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_2903_end_0 = const()[name = string("op_2903_end_0"), val = tensor([32, 14, 64, 64])]; + tensor var_2903_end_mask_0 = const()[name = string("op_2903_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2903_squeeze_mask_0 = const()[name = string("op_2903_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2903_cast_fp16 = slice_by_index(begin = var_2903_begin_0, end = var_2903_end_0, end_mask = var_2903_end_mask_0, squeeze_mask = var_2903_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2903_cast_fp16")]; + bool var_2904_transpose_x_0 = const()[name = string("op_2904_transpose_x_0"), val = bool(false)]; + bool var_2904_transpose_y_0 = const()[name = string("op_2904_transpose_y_0"), val = bool(false)]; + tensor var_2904_cast_fp16 = matmul(transpose_x = var_2904_transpose_x_0, transpose_y = var_2904_transpose_y_0, x = var_2903_cast_fp16, y = v_new_155_cast_fp16)[name = string("op_2904_cast_fp16")]; + tensor var_2905_cast_fp16 = add(x = var_2901_cast_fp16, y = var_2904_cast_fp16)[name = string("op_2905_cast_fp16")]; + tensor var_2907_begin_0 = const()[name = string("op_2907_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_2907_end_0 = const()[name = string("op_2907_end_0"), val = tensor([32, 14, 1, 1])]; + tensor var_2907_end_mask_0 = const()[name = string("op_2907_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2907_squeeze_mask_0 = const()[name = string("op_2907_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2907_cast_fp16 = slice_by_index(begin = var_2907_begin_0, end = var_2907_end_0, end_mask = var_2907_end_mask_0, squeeze_mask = var_2907_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2907_cast_fp16")]; + tensor var_2908_cast_fp16 = mul(x = state_155_cast_fp16, y = var_2907_cast_fp16)[name = string("op_2908_cast_fp16")]; + tensor var_2910_begin_0 = const()[name = string("op_2910_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_2910_end_0 = const()[name = string("op_2910_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_2910_end_mask_0 = const()[name = string("op_2910_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2910_squeeze_mask_0 = const()[name = string("op_2910_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2910_cast_fp16 = slice_by_index(begin = var_2910_begin_0, end = var_2910_end_0, end_mask = var_2910_end_mask_0, squeeze_mask = var_2910_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2910_cast_fp16")]; + bool var_2912_transpose_x_1 = const()[name = string("op_2912_transpose_x_1"), val = bool(true)]; + bool var_2912_transpose_y_1 = const()[name = string("op_2912_transpose_y_1"), val = bool(false)]; + tensor var_2912_cast_fp16 = matmul(transpose_x = var_2912_transpose_x_1, transpose_y = var_2912_transpose_y_1, x = var_2910_cast_fp16, y = v_new_155_cast_fp16)[name = string("op_2912_cast_fp16")]; + tensor state_157_cast_fp16 = add(x = var_2908_cast_fp16, y = var_2912_cast_fp16)[name = string("state_157_cast_fp16")]; + tensor var_2915_begin_0 = const()[name = string("op_2915_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_2915_end_0 = const()[name = string("op_2915_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_2915_end_mask_0 = const()[name = string("op_2915_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2915_squeeze_mask_0 = const()[name = string("op_2915_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2915_cast_fp16 = slice_by_index(begin = var_2915_begin_0, end = var_2915_end_0, end_mask = var_2915_end_mask_0, squeeze_mask = var_2915_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2915_cast_fp16")]; + tensor var_2917_begin_0 = const()[name = string("op_2917_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_2917_end_0 = const()[name = string("op_2917_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_2917_end_mask_0 = const()[name = string("op_2917_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2917_squeeze_mask_0 = const()[name = string("op_2917_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2917_cast_fp16 = slice_by_index(begin = var_2917_begin_0, end = var_2917_end_0, end_mask = var_2917_end_mask_0, squeeze_mask = var_2917_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2917_cast_fp16")]; + bool var_2918_transpose_x_0 = const()[name = string("op_2918_transpose_x_0"), val = bool(false)]; + bool var_2918_transpose_y_0 = const()[name = string("op_2918_transpose_y_0"), val = bool(false)]; + tensor var_2918_cast_fp16 = matmul(transpose_x = var_2918_transpose_x_0, transpose_y = var_2918_transpose_y_0, x = var_2917_cast_fp16, y = state_157_cast_fp16)[name = string("op_2918_cast_fp16")]; + tensor v_new_157_cast_fp16 = sub(x = var_2915_cast_fp16, y = var_2918_cast_fp16)[name = string("v_new_157_cast_fp16")]; + tensor var_2921_begin_0 = const()[name = string("op_2921_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_2921_end_0 = const()[name = string("op_2921_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_2921_end_mask_0 = const()[name = string("op_2921_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2921_squeeze_mask_0 = const()[name = string("op_2921_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2921_cast_fp16 = slice_by_index(begin = var_2921_begin_0, end = var_2921_end_0, end_mask = var_2921_end_mask_0, squeeze_mask = var_2921_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2921_cast_fp16")]; + bool var_2922_transpose_x_0 = const()[name = string("op_2922_transpose_x_0"), val = bool(false)]; + bool var_2922_transpose_y_0 = const()[name = string("op_2922_transpose_y_0"), val = bool(false)]; + tensor var_2922_cast_fp16 = matmul(transpose_x = var_2922_transpose_x_0, transpose_y = var_2922_transpose_y_0, x = var_2921_cast_fp16, y = state_157_cast_fp16)[name = string("op_2922_cast_fp16")]; + tensor var_2924_begin_0 = const()[name = string("op_2924_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_2924_end_0 = const()[name = string("op_2924_end_0"), val = tensor([32, 15, 64, 64])]; + tensor var_2924_end_mask_0 = const()[name = string("op_2924_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2924_squeeze_mask_0 = const()[name = string("op_2924_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2924_cast_fp16 = slice_by_index(begin = var_2924_begin_0, end = var_2924_end_0, end_mask = var_2924_end_mask_0, squeeze_mask = var_2924_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2924_cast_fp16")]; + bool var_2925_transpose_x_0 = const()[name = string("op_2925_transpose_x_0"), val = bool(false)]; + bool var_2925_transpose_y_0 = const()[name = string("op_2925_transpose_y_0"), val = bool(false)]; + tensor var_2925_cast_fp16 = matmul(transpose_x = var_2925_transpose_x_0, transpose_y = var_2925_transpose_y_0, x = var_2924_cast_fp16, y = v_new_157_cast_fp16)[name = string("op_2925_cast_fp16")]; + tensor var_2926_cast_fp16 = add(x = var_2922_cast_fp16, y = var_2925_cast_fp16)[name = string("op_2926_cast_fp16")]; + tensor var_2928_begin_0 = const()[name = string("op_2928_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_2928_end_0 = const()[name = string("op_2928_end_0"), val = tensor([32, 15, 1, 1])]; + tensor var_2928_end_mask_0 = const()[name = string("op_2928_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2928_squeeze_mask_0 = const()[name = string("op_2928_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2928_cast_fp16 = slice_by_index(begin = var_2928_begin_0, end = var_2928_end_0, end_mask = var_2928_end_mask_0, squeeze_mask = var_2928_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2928_cast_fp16")]; + tensor var_2929_cast_fp16 = mul(x = state_157_cast_fp16, y = var_2928_cast_fp16)[name = string("op_2929_cast_fp16")]; + tensor var_2931_begin_0 = const()[name = string("op_2931_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_2931_end_0 = const()[name = string("op_2931_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_2931_end_mask_0 = const()[name = string("op_2931_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2931_squeeze_mask_0 = const()[name = string("op_2931_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, squeeze_mask = var_2931_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2931_cast_fp16")]; + bool var_2933_transpose_x_1 = const()[name = string("op_2933_transpose_x_1"), val = bool(true)]; + bool var_2933_transpose_y_1 = const()[name = string("op_2933_transpose_y_1"), val = bool(false)]; + tensor var_2933_cast_fp16 = matmul(transpose_x = var_2933_transpose_x_1, transpose_y = var_2933_transpose_y_1, x = var_2931_cast_fp16, y = v_new_157_cast_fp16)[name = string("op_2933_cast_fp16")]; + tensor state_159_cast_fp16 = add(x = var_2929_cast_fp16, y = var_2933_cast_fp16)[name = string("state_159_cast_fp16")]; + tensor var_2936_begin_0 = const()[name = string("op_2936_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_2936_end_0 = const()[name = string("op_2936_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_2936_end_mask_0 = const()[name = string("op_2936_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2936_squeeze_mask_0 = const()[name = string("op_2936_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2936_cast_fp16 = slice_by_index(begin = var_2936_begin_0, end = var_2936_end_0, end_mask = var_2936_end_mask_0, squeeze_mask = var_2936_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2936_cast_fp16")]; + tensor var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2938_squeeze_mask_0 = const()[name = string("op_2938_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, squeeze_mask = var_2938_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2938_cast_fp16")]; + bool var_2939_transpose_x_0 = const()[name = string("op_2939_transpose_x_0"), val = bool(false)]; + bool var_2939_transpose_y_0 = const()[name = string("op_2939_transpose_y_0"), val = bool(false)]; + tensor var_2939_cast_fp16 = matmul(transpose_x = var_2939_transpose_x_0, transpose_y = var_2939_transpose_y_0, x = var_2938_cast_fp16, y = state_159_cast_fp16)[name = string("op_2939_cast_fp16")]; + tensor v_new_159_cast_fp16 = sub(x = var_2936_cast_fp16, y = var_2939_cast_fp16)[name = string("v_new_159_cast_fp16")]; + tensor var_2942_begin_0 = const()[name = string("op_2942_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_2942_end_0 = const()[name = string("op_2942_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_2942_end_mask_0 = const()[name = string("op_2942_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2942_squeeze_mask_0 = const()[name = string("op_2942_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2942_cast_fp16 = slice_by_index(begin = var_2942_begin_0, end = var_2942_end_0, end_mask = var_2942_end_mask_0, squeeze_mask = var_2942_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2942_cast_fp16")]; + bool var_2943_transpose_x_0 = const()[name = string("op_2943_transpose_x_0"), val = bool(false)]; + bool var_2943_transpose_y_0 = const()[name = string("op_2943_transpose_y_0"), val = bool(false)]; + tensor var_2943_cast_fp16 = matmul(transpose_x = var_2943_transpose_x_0, transpose_y = var_2943_transpose_y_0, x = var_2942_cast_fp16, y = state_159_cast_fp16)[name = string("op_2943_cast_fp16")]; + tensor var_2945_begin_0 = const()[name = string("op_2945_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_2945_end_0 = const()[name = string("op_2945_end_0"), val = tensor([32, 16, 64, 64])]; + tensor var_2945_end_mask_0 = const()[name = string("op_2945_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2945_squeeze_mask_0 = const()[name = string("op_2945_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2945_cast_fp16 = slice_by_index(begin = var_2945_begin_0, end = var_2945_end_0, end_mask = var_2945_end_mask_0, squeeze_mask = var_2945_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2945_cast_fp16")]; + bool var_2946_transpose_x_0 = const()[name = string("op_2946_transpose_x_0"), val = bool(false)]; + bool var_2946_transpose_y_0 = const()[name = string("op_2946_transpose_y_0"), val = bool(false)]; + tensor var_2946_cast_fp16 = matmul(transpose_x = var_2946_transpose_x_0, transpose_y = var_2946_transpose_y_0, x = var_2945_cast_fp16, y = v_new_159_cast_fp16)[name = string("op_2946_cast_fp16")]; + tensor var_2947_cast_fp16 = add(x = var_2943_cast_fp16, y = var_2946_cast_fp16)[name = string("op_2947_cast_fp16")]; + tensor var_2949_begin_0 = const()[name = string("op_2949_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_2949_end_0 = const()[name = string("op_2949_end_0"), val = tensor([32, 16, 1, 1])]; + tensor var_2949_end_mask_0 = const()[name = string("op_2949_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2949_squeeze_mask_0 = const()[name = string("op_2949_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2949_cast_fp16 = slice_by_index(begin = var_2949_begin_0, end = var_2949_end_0, end_mask = var_2949_end_mask_0, squeeze_mask = var_2949_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2949_cast_fp16")]; + tensor var_2950_cast_fp16 = mul(x = state_159_cast_fp16, y = var_2949_cast_fp16)[name = string("op_2950_cast_fp16")]; + tensor var_2952_begin_0 = const()[name = string("op_2952_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_2952_end_0 = const()[name = string("op_2952_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_2952_end_mask_0 = const()[name = string("op_2952_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2952_squeeze_mask_0 = const()[name = string("op_2952_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2952_cast_fp16 = slice_by_index(begin = var_2952_begin_0, end = var_2952_end_0, end_mask = var_2952_end_mask_0, squeeze_mask = var_2952_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2952_cast_fp16")]; + bool var_2954_transpose_x_1 = const()[name = string("op_2954_transpose_x_1"), val = bool(true)]; + bool var_2954_transpose_y_1 = const()[name = string("op_2954_transpose_y_1"), val = bool(false)]; + tensor var_2954_cast_fp16 = matmul(transpose_x = var_2954_transpose_x_1, transpose_y = var_2954_transpose_y_1, x = var_2952_cast_fp16, y = v_new_159_cast_fp16)[name = string("op_2954_cast_fp16")]; + tensor state_161_cast_fp16 = add(x = var_2950_cast_fp16, y = var_2954_cast_fp16)[name = string("state_161_cast_fp16")]; + tensor var_2957_begin_0 = const()[name = string("op_2957_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_2957_end_0 = const()[name = string("op_2957_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_2957_end_mask_0 = const()[name = string("op_2957_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2957_squeeze_mask_0 = const()[name = string("op_2957_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2957_cast_fp16 = slice_by_index(begin = var_2957_begin_0, end = var_2957_end_0, end_mask = var_2957_end_mask_0, squeeze_mask = var_2957_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2957_cast_fp16")]; + tensor var_2959_begin_0 = const()[name = string("op_2959_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_2959_end_0 = const()[name = string("op_2959_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_2959_end_mask_0 = const()[name = string("op_2959_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2959_squeeze_mask_0 = const()[name = string("op_2959_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2959_cast_fp16 = slice_by_index(begin = var_2959_begin_0, end = var_2959_end_0, end_mask = var_2959_end_mask_0, squeeze_mask = var_2959_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2959_cast_fp16")]; + bool var_2960_transpose_x_0 = const()[name = string("op_2960_transpose_x_0"), val = bool(false)]; + bool var_2960_transpose_y_0 = const()[name = string("op_2960_transpose_y_0"), val = bool(false)]; + tensor var_2960_cast_fp16 = matmul(transpose_x = var_2960_transpose_x_0, transpose_y = var_2960_transpose_y_0, x = var_2959_cast_fp16, y = state_161_cast_fp16)[name = string("op_2960_cast_fp16")]; + tensor v_new_161_cast_fp16 = sub(x = var_2957_cast_fp16, y = var_2960_cast_fp16)[name = string("v_new_161_cast_fp16")]; + tensor var_2963_begin_0 = const()[name = string("op_2963_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_2963_end_0 = const()[name = string("op_2963_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_2963_end_mask_0 = const()[name = string("op_2963_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2963_squeeze_mask_0 = const()[name = string("op_2963_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2963_cast_fp16 = slice_by_index(begin = var_2963_begin_0, end = var_2963_end_0, end_mask = var_2963_end_mask_0, squeeze_mask = var_2963_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2963_cast_fp16")]; + bool var_2964_transpose_x_0 = const()[name = string("op_2964_transpose_x_0"), val = bool(false)]; + bool var_2964_transpose_y_0 = const()[name = string("op_2964_transpose_y_0"), val = bool(false)]; + tensor var_2964_cast_fp16 = matmul(transpose_x = var_2964_transpose_x_0, transpose_y = var_2964_transpose_y_0, x = var_2963_cast_fp16, y = state_161_cast_fp16)[name = string("op_2964_cast_fp16")]; + tensor var_2966_begin_0 = const()[name = string("op_2966_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_2966_end_0 = const()[name = string("op_2966_end_0"), val = tensor([32, 17, 64, 64])]; + tensor var_2966_end_mask_0 = const()[name = string("op_2966_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2966_squeeze_mask_0 = const()[name = string("op_2966_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2966_cast_fp16 = slice_by_index(begin = var_2966_begin_0, end = var_2966_end_0, end_mask = var_2966_end_mask_0, squeeze_mask = var_2966_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2966_cast_fp16")]; + bool var_2967_transpose_x_0 = const()[name = string("op_2967_transpose_x_0"), val = bool(false)]; + bool var_2967_transpose_y_0 = const()[name = string("op_2967_transpose_y_0"), val = bool(false)]; + tensor var_2967_cast_fp16 = matmul(transpose_x = var_2967_transpose_x_0, transpose_y = var_2967_transpose_y_0, x = var_2966_cast_fp16, y = v_new_161_cast_fp16)[name = string("op_2967_cast_fp16")]; + tensor var_2968_cast_fp16 = add(x = var_2964_cast_fp16, y = var_2967_cast_fp16)[name = string("op_2968_cast_fp16")]; + tensor var_2970_begin_0 = const()[name = string("op_2970_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_2970_end_0 = const()[name = string("op_2970_end_0"), val = tensor([32, 17, 1, 1])]; + tensor var_2970_end_mask_0 = const()[name = string("op_2970_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2970_squeeze_mask_0 = const()[name = string("op_2970_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2970_cast_fp16 = slice_by_index(begin = var_2970_begin_0, end = var_2970_end_0, end_mask = var_2970_end_mask_0, squeeze_mask = var_2970_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = mul(x = state_161_cast_fp16, y = var_2970_cast_fp16)[name = string("op_2971_cast_fp16")]; + tensor var_2973_begin_0 = const()[name = string("op_2973_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_2973_end_0 = const()[name = string("op_2973_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_2973_end_mask_0 = const()[name = string("op_2973_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2973_squeeze_mask_0 = const()[name = string("op_2973_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2973_cast_fp16 = slice_by_index(begin = var_2973_begin_0, end = var_2973_end_0, end_mask = var_2973_end_mask_0, squeeze_mask = var_2973_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2973_cast_fp16")]; + bool var_2975_transpose_x_1 = const()[name = string("op_2975_transpose_x_1"), val = bool(true)]; + bool var_2975_transpose_y_1 = const()[name = string("op_2975_transpose_y_1"), val = bool(false)]; + tensor var_2975_cast_fp16 = matmul(transpose_x = var_2975_transpose_x_1, transpose_y = var_2975_transpose_y_1, x = var_2973_cast_fp16, y = v_new_161_cast_fp16)[name = string("op_2975_cast_fp16")]; + tensor state_163_cast_fp16 = add(x = var_2971_cast_fp16, y = var_2975_cast_fp16)[name = string("state_163_cast_fp16")]; + tensor var_2978_begin_0 = const()[name = string("op_2978_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_2978_end_0 = const()[name = string("op_2978_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_2978_end_mask_0 = const()[name = string("op_2978_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2978_squeeze_mask_0 = const()[name = string("op_2978_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2978_cast_fp16 = slice_by_index(begin = var_2978_begin_0, end = var_2978_end_0, end_mask = var_2978_end_mask_0, squeeze_mask = var_2978_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2978_cast_fp16")]; + tensor var_2980_begin_0 = const()[name = string("op_2980_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_2980_end_0 = const()[name = string("op_2980_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_2980_end_mask_0 = const()[name = string("op_2980_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2980_squeeze_mask_0 = const()[name = string("op_2980_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2980_cast_fp16 = slice_by_index(begin = var_2980_begin_0, end = var_2980_end_0, end_mask = var_2980_end_mask_0, squeeze_mask = var_2980_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_2980_cast_fp16")]; + bool var_2981_transpose_x_0 = const()[name = string("op_2981_transpose_x_0"), val = bool(false)]; + bool var_2981_transpose_y_0 = const()[name = string("op_2981_transpose_y_0"), val = bool(false)]; + tensor var_2981_cast_fp16 = matmul(transpose_x = var_2981_transpose_x_0, transpose_y = var_2981_transpose_y_0, x = var_2980_cast_fp16, y = state_163_cast_fp16)[name = string("op_2981_cast_fp16")]; + tensor v_new_163_cast_fp16 = sub(x = var_2978_cast_fp16, y = var_2981_cast_fp16)[name = string("v_new_163_cast_fp16")]; + tensor var_2984_begin_0 = const()[name = string("op_2984_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_2984_end_0 = const()[name = string("op_2984_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_2984_end_mask_0 = const()[name = string("op_2984_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2984_squeeze_mask_0 = const()[name = string("op_2984_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2984_cast_fp16 = slice_by_index(begin = var_2984_begin_0, end = var_2984_end_0, end_mask = var_2984_end_mask_0, squeeze_mask = var_2984_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_2984_cast_fp16")]; + bool var_2985_transpose_x_0 = const()[name = string("op_2985_transpose_x_0"), val = bool(false)]; + bool var_2985_transpose_y_0 = const()[name = string("op_2985_transpose_y_0"), val = bool(false)]; + tensor var_2985_cast_fp16 = matmul(transpose_x = var_2985_transpose_x_0, transpose_y = var_2985_transpose_y_0, x = var_2984_cast_fp16, y = state_163_cast_fp16)[name = string("op_2985_cast_fp16")]; + tensor var_2987_begin_0 = const()[name = string("op_2987_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_2987_end_0 = const()[name = string("op_2987_end_0"), val = tensor([32, 18, 64, 64])]; + tensor var_2987_end_mask_0 = const()[name = string("op_2987_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2987_squeeze_mask_0 = const()[name = string("op_2987_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2987_cast_fp16 = slice_by_index(begin = var_2987_begin_0, end = var_2987_end_0, end_mask = var_2987_end_mask_0, squeeze_mask = var_2987_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_2987_cast_fp16")]; + bool var_2988_transpose_x_0 = const()[name = string("op_2988_transpose_x_0"), val = bool(false)]; + bool var_2988_transpose_y_0 = const()[name = string("op_2988_transpose_y_0"), val = bool(false)]; + tensor var_2988_cast_fp16 = matmul(transpose_x = var_2988_transpose_x_0, transpose_y = var_2988_transpose_y_0, x = var_2987_cast_fp16, y = v_new_163_cast_fp16)[name = string("op_2988_cast_fp16")]; + tensor var_2989_cast_fp16 = add(x = var_2985_cast_fp16, y = var_2988_cast_fp16)[name = string("op_2989_cast_fp16")]; + tensor var_2991_begin_0 = const()[name = string("op_2991_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_2991_end_0 = const()[name = string("op_2991_end_0"), val = tensor([32, 18, 1, 1])]; + tensor var_2991_end_mask_0 = const()[name = string("op_2991_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2991_squeeze_mask_0 = const()[name = string("op_2991_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2991_cast_fp16 = slice_by_index(begin = var_2991_begin_0, end = var_2991_end_0, end_mask = var_2991_end_mask_0, squeeze_mask = var_2991_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_2991_cast_fp16")]; + tensor var_2992_cast_fp16 = mul(x = state_163_cast_fp16, y = var_2991_cast_fp16)[name = string("op_2992_cast_fp16")]; + tensor var_2994_begin_0 = const()[name = string("op_2994_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_2994_end_0 = const()[name = string("op_2994_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_2994_end_mask_0 = const()[name = string("op_2994_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2994_squeeze_mask_0 = const()[name = string("op_2994_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2994_cast_fp16 = slice_by_index(begin = var_2994_begin_0, end = var_2994_end_0, end_mask = var_2994_end_mask_0, squeeze_mask = var_2994_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_2994_cast_fp16")]; + bool var_2996_transpose_x_1 = const()[name = string("op_2996_transpose_x_1"), val = bool(true)]; + bool var_2996_transpose_y_1 = const()[name = string("op_2996_transpose_y_1"), val = bool(false)]; + tensor var_2996_cast_fp16 = matmul(transpose_x = var_2996_transpose_x_1, transpose_y = var_2996_transpose_y_1, x = var_2994_cast_fp16, y = v_new_163_cast_fp16)[name = string("op_2996_cast_fp16")]; + tensor state_165_cast_fp16 = add(x = var_2992_cast_fp16, y = var_2996_cast_fp16)[name = string("state_165_cast_fp16")]; + tensor var_2999_begin_0 = const()[name = string("op_2999_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_2999_end_0 = const()[name = string("op_2999_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_2999_end_mask_0 = const()[name = string("op_2999_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2999_squeeze_mask_0 = const()[name = string("op_2999_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_2999_cast_fp16 = slice_by_index(begin = var_2999_begin_0, end = var_2999_end_0, end_mask = var_2999_end_mask_0, squeeze_mask = var_2999_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_2999_cast_fp16")]; + tensor var_3001_begin_0 = const()[name = string("op_3001_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_3001_end_0 = const()[name = string("op_3001_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_3001_end_mask_0 = const()[name = string("op_3001_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3001_squeeze_mask_0 = const()[name = string("op_3001_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3001_cast_fp16 = slice_by_index(begin = var_3001_begin_0, end = var_3001_end_0, end_mask = var_3001_end_mask_0, squeeze_mask = var_3001_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3001_cast_fp16")]; + bool var_3002_transpose_x_0 = const()[name = string("op_3002_transpose_x_0"), val = bool(false)]; + bool var_3002_transpose_y_0 = const()[name = string("op_3002_transpose_y_0"), val = bool(false)]; + tensor var_3002_cast_fp16 = matmul(transpose_x = var_3002_transpose_x_0, transpose_y = var_3002_transpose_y_0, x = var_3001_cast_fp16, y = state_165_cast_fp16)[name = string("op_3002_cast_fp16")]; + tensor v_new_165_cast_fp16 = sub(x = var_2999_cast_fp16, y = var_3002_cast_fp16)[name = string("v_new_165_cast_fp16")]; + tensor var_3005_begin_0 = const()[name = string("op_3005_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_3005_end_0 = const()[name = string("op_3005_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_3005_end_mask_0 = const()[name = string("op_3005_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3005_squeeze_mask_0 = const()[name = string("op_3005_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3005_cast_fp16 = slice_by_index(begin = var_3005_begin_0, end = var_3005_end_0, end_mask = var_3005_end_mask_0, squeeze_mask = var_3005_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3005_cast_fp16")]; + bool var_3006_transpose_x_0 = const()[name = string("op_3006_transpose_x_0"), val = bool(false)]; + bool var_3006_transpose_y_0 = const()[name = string("op_3006_transpose_y_0"), val = bool(false)]; + tensor var_3006_cast_fp16 = matmul(transpose_x = var_3006_transpose_x_0, transpose_y = var_3006_transpose_y_0, x = var_3005_cast_fp16, y = state_165_cast_fp16)[name = string("op_3006_cast_fp16")]; + tensor var_3008_begin_0 = const()[name = string("op_3008_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_3008_end_0 = const()[name = string("op_3008_end_0"), val = tensor([32, 19, 64, 64])]; + tensor var_3008_end_mask_0 = const()[name = string("op_3008_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3008_squeeze_mask_0 = const()[name = string("op_3008_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3008_cast_fp16 = slice_by_index(begin = var_3008_begin_0, end = var_3008_end_0, end_mask = var_3008_end_mask_0, squeeze_mask = var_3008_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3008_cast_fp16")]; + bool var_3009_transpose_x_0 = const()[name = string("op_3009_transpose_x_0"), val = bool(false)]; + bool var_3009_transpose_y_0 = const()[name = string("op_3009_transpose_y_0"), val = bool(false)]; + tensor var_3009_cast_fp16 = matmul(transpose_x = var_3009_transpose_x_0, transpose_y = var_3009_transpose_y_0, x = var_3008_cast_fp16, y = v_new_165_cast_fp16)[name = string("op_3009_cast_fp16")]; + tensor var_3010_cast_fp16 = add(x = var_3006_cast_fp16, y = var_3009_cast_fp16)[name = string("op_3010_cast_fp16")]; + tensor var_3012_begin_0 = const()[name = string("op_3012_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_3012_end_0 = const()[name = string("op_3012_end_0"), val = tensor([32, 19, 1, 1])]; + tensor var_3012_end_mask_0 = const()[name = string("op_3012_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3012_squeeze_mask_0 = const()[name = string("op_3012_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3012_cast_fp16 = slice_by_index(begin = var_3012_begin_0, end = var_3012_end_0, end_mask = var_3012_end_mask_0, squeeze_mask = var_3012_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3012_cast_fp16")]; + tensor var_3013_cast_fp16 = mul(x = state_165_cast_fp16, y = var_3012_cast_fp16)[name = string("op_3013_cast_fp16")]; + tensor var_3015_begin_0 = const()[name = string("op_3015_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_3015_end_0 = const()[name = string("op_3015_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_3015_end_mask_0 = const()[name = string("op_3015_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3015_squeeze_mask_0 = const()[name = string("op_3015_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3015_cast_fp16 = slice_by_index(begin = var_3015_begin_0, end = var_3015_end_0, end_mask = var_3015_end_mask_0, squeeze_mask = var_3015_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3015_cast_fp16")]; + bool var_3017_transpose_x_1 = const()[name = string("op_3017_transpose_x_1"), val = bool(true)]; + bool var_3017_transpose_y_1 = const()[name = string("op_3017_transpose_y_1"), val = bool(false)]; + tensor var_3017_cast_fp16 = matmul(transpose_x = var_3017_transpose_x_1, transpose_y = var_3017_transpose_y_1, x = var_3015_cast_fp16, y = v_new_165_cast_fp16)[name = string("op_3017_cast_fp16")]; + tensor state_167_cast_fp16 = add(x = var_3013_cast_fp16, y = var_3017_cast_fp16)[name = string("state_167_cast_fp16")]; + tensor var_3020_begin_0 = const()[name = string("op_3020_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_3020_end_0 = const()[name = string("op_3020_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_3020_end_mask_0 = const()[name = string("op_3020_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3020_squeeze_mask_0 = const()[name = string("op_3020_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3020_cast_fp16 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, squeeze_mask = var_3020_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3020_cast_fp16")]; + tensor var_3022_begin_0 = const()[name = string("op_3022_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_3022_end_0 = const()[name = string("op_3022_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_3022_end_mask_0 = const()[name = string("op_3022_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3022_squeeze_mask_0 = const()[name = string("op_3022_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3022_cast_fp16 = slice_by_index(begin = var_3022_begin_0, end = var_3022_end_0, end_mask = var_3022_end_mask_0, squeeze_mask = var_3022_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3022_cast_fp16")]; + bool var_3023_transpose_x_0 = const()[name = string("op_3023_transpose_x_0"), val = bool(false)]; + bool var_3023_transpose_y_0 = const()[name = string("op_3023_transpose_y_0"), val = bool(false)]; + tensor var_3023_cast_fp16 = matmul(transpose_x = var_3023_transpose_x_0, transpose_y = var_3023_transpose_y_0, x = var_3022_cast_fp16, y = state_167_cast_fp16)[name = string("op_3023_cast_fp16")]; + tensor v_new_167_cast_fp16 = sub(x = var_3020_cast_fp16, y = var_3023_cast_fp16)[name = string("v_new_167_cast_fp16")]; + tensor var_3026_begin_0 = const()[name = string("op_3026_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_3026_end_0 = const()[name = string("op_3026_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_3026_end_mask_0 = const()[name = string("op_3026_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3026_squeeze_mask_0 = const()[name = string("op_3026_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3026_cast_fp16 = slice_by_index(begin = var_3026_begin_0, end = var_3026_end_0, end_mask = var_3026_end_mask_0, squeeze_mask = var_3026_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3026_cast_fp16")]; + bool var_3027_transpose_x_0 = const()[name = string("op_3027_transpose_x_0"), val = bool(false)]; + bool var_3027_transpose_y_0 = const()[name = string("op_3027_transpose_y_0"), val = bool(false)]; + tensor var_3027_cast_fp16 = matmul(transpose_x = var_3027_transpose_x_0, transpose_y = var_3027_transpose_y_0, x = var_3026_cast_fp16, y = state_167_cast_fp16)[name = string("op_3027_cast_fp16")]; + tensor var_3029_begin_0 = const()[name = string("op_3029_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_3029_end_0 = const()[name = string("op_3029_end_0"), val = tensor([32, 20, 64, 64])]; + tensor var_3029_end_mask_0 = const()[name = string("op_3029_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3029_squeeze_mask_0 = const()[name = string("op_3029_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3029_cast_fp16 = slice_by_index(begin = var_3029_begin_0, end = var_3029_end_0, end_mask = var_3029_end_mask_0, squeeze_mask = var_3029_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3029_cast_fp16")]; + bool var_3030_transpose_x_0 = const()[name = string("op_3030_transpose_x_0"), val = bool(false)]; + bool var_3030_transpose_y_0 = const()[name = string("op_3030_transpose_y_0"), val = bool(false)]; + tensor var_3030_cast_fp16 = matmul(transpose_x = var_3030_transpose_x_0, transpose_y = var_3030_transpose_y_0, x = var_3029_cast_fp16, y = v_new_167_cast_fp16)[name = string("op_3030_cast_fp16")]; + tensor var_3031_cast_fp16 = add(x = var_3027_cast_fp16, y = var_3030_cast_fp16)[name = string("op_3031_cast_fp16")]; + tensor var_3033_begin_0 = const()[name = string("op_3033_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_3033_end_0 = const()[name = string("op_3033_end_0"), val = tensor([32, 20, 1, 1])]; + tensor var_3033_end_mask_0 = const()[name = string("op_3033_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3033_squeeze_mask_0 = const()[name = string("op_3033_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3033_cast_fp16 = slice_by_index(begin = var_3033_begin_0, end = var_3033_end_0, end_mask = var_3033_end_mask_0, squeeze_mask = var_3033_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3033_cast_fp16")]; + tensor var_3034_cast_fp16 = mul(x = state_167_cast_fp16, y = var_3033_cast_fp16)[name = string("op_3034_cast_fp16")]; + tensor var_3036_begin_0 = const()[name = string("op_3036_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_3036_end_0 = const()[name = string("op_3036_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_3036_end_mask_0 = const()[name = string("op_3036_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3036_squeeze_mask_0 = const()[name = string("op_3036_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3036_cast_fp16 = slice_by_index(begin = var_3036_begin_0, end = var_3036_end_0, end_mask = var_3036_end_mask_0, squeeze_mask = var_3036_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3036_cast_fp16")]; + bool var_3038_transpose_x_1 = const()[name = string("op_3038_transpose_x_1"), val = bool(true)]; + bool var_3038_transpose_y_1 = const()[name = string("op_3038_transpose_y_1"), val = bool(false)]; + tensor var_3038_cast_fp16 = matmul(transpose_x = var_3038_transpose_x_1, transpose_y = var_3038_transpose_y_1, x = var_3036_cast_fp16, y = v_new_167_cast_fp16)[name = string("op_3038_cast_fp16")]; + tensor state_169_cast_fp16 = add(x = var_3034_cast_fp16, y = var_3038_cast_fp16)[name = string("state_169_cast_fp16")]; + tensor var_3041_begin_0 = const()[name = string("op_3041_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_3041_end_0 = const()[name = string("op_3041_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_3041_end_mask_0 = const()[name = string("op_3041_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3041_squeeze_mask_0 = const()[name = string("op_3041_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3041_cast_fp16 = slice_by_index(begin = var_3041_begin_0, end = var_3041_end_0, end_mask = var_3041_end_mask_0, squeeze_mask = var_3041_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3041_cast_fp16")]; + tensor var_3043_begin_0 = const()[name = string("op_3043_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_3043_end_0 = const()[name = string("op_3043_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_3043_end_mask_0 = const()[name = string("op_3043_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3043_squeeze_mask_0 = const()[name = string("op_3043_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3043_cast_fp16 = slice_by_index(begin = var_3043_begin_0, end = var_3043_end_0, end_mask = var_3043_end_mask_0, squeeze_mask = var_3043_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3043_cast_fp16")]; + bool var_3044_transpose_x_0 = const()[name = string("op_3044_transpose_x_0"), val = bool(false)]; + bool var_3044_transpose_y_0 = const()[name = string("op_3044_transpose_y_0"), val = bool(false)]; + tensor var_3044_cast_fp16 = matmul(transpose_x = var_3044_transpose_x_0, transpose_y = var_3044_transpose_y_0, x = var_3043_cast_fp16, y = state_169_cast_fp16)[name = string("op_3044_cast_fp16")]; + tensor v_new_169_cast_fp16 = sub(x = var_3041_cast_fp16, y = var_3044_cast_fp16)[name = string("v_new_169_cast_fp16")]; + tensor var_3047_begin_0 = const()[name = string("op_3047_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_3047_end_0 = const()[name = string("op_3047_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_3047_end_mask_0 = const()[name = string("op_3047_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3047_squeeze_mask_0 = const()[name = string("op_3047_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3047_cast_fp16 = slice_by_index(begin = var_3047_begin_0, end = var_3047_end_0, end_mask = var_3047_end_mask_0, squeeze_mask = var_3047_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3047_cast_fp16")]; + bool var_3048_transpose_x_0 = const()[name = string("op_3048_transpose_x_0"), val = bool(false)]; + bool var_3048_transpose_y_0 = const()[name = string("op_3048_transpose_y_0"), val = bool(false)]; + tensor var_3048_cast_fp16 = matmul(transpose_x = var_3048_transpose_x_0, transpose_y = var_3048_transpose_y_0, x = var_3047_cast_fp16, y = state_169_cast_fp16)[name = string("op_3048_cast_fp16")]; + tensor var_3050_begin_0 = const()[name = string("op_3050_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_3050_end_0 = const()[name = string("op_3050_end_0"), val = tensor([32, 21, 64, 64])]; + tensor var_3050_end_mask_0 = const()[name = string("op_3050_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3050_squeeze_mask_0 = const()[name = string("op_3050_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3050_cast_fp16 = slice_by_index(begin = var_3050_begin_0, end = var_3050_end_0, end_mask = var_3050_end_mask_0, squeeze_mask = var_3050_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3050_cast_fp16")]; + bool var_3051_transpose_x_0 = const()[name = string("op_3051_transpose_x_0"), val = bool(false)]; + bool var_3051_transpose_y_0 = const()[name = string("op_3051_transpose_y_0"), val = bool(false)]; + tensor var_3051_cast_fp16 = matmul(transpose_x = var_3051_transpose_x_0, transpose_y = var_3051_transpose_y_0, x = var_3050_cast_fp16, y = v_new_169_cast_fp16)[name = string("op_3051_cast_fp16")]; + tensor var_3052_cast_fp16 = add(x = var_3048_cast_fp16, y = var_3051_cast_fp16)[name = string("op_3052_cast_fp16")]; + tensor var_3054_begin_0 = const()[name = string("op_3054_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_3054_end_0 = const()[name = string("op_3054_end_0"), val = tensor([32, 21, 1, 1])]; + tensor var_3054_end_mask_0 = const()[name = string("op_3054_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3054_squeeze_mask_0 = const()[name = string("op_3054_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3054_cast_fp16 = slice_by_index(begin = var_3054_begin_0, end = var_3054_end_0, end_mask = var_3054_end_mask_0, squeeze_mask = var_3054_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3054_cast_fp16")]; + tensor var_3055_cast_fp16 = mul(x = state_169_cast_fp16, y = var_3054_cast_fp16)[name = string("op_3055_cast_fp16")]; + tensor var_3057_begin_0 = const()[name = string("op_3057_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_3057_end_0 = const()[name = string("op_3057_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_3057_end_mask_0 = const()[name = string("op_3057_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3057_squeeze_mask_0 = const()[name = string("op_3057_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3057_cast_fp16 = slice_by_index(begin = var_3057_begin_0, end = var_3057_end_0, end_mask = var_3057_end_mask_0, squeeze_mask = var_3057_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3057_cast_fp16")]; + bool var_3059_transpose_x_1 = const()[name = string("op_3059_transpose_x_1"), val = bool(true)]; + bool var_3059_transpose_y_1 = const()[name = string("op_3059_transpose_y_1"), val = bool(false)]; + tensor var_3059_cast_fp16 = matmul(transpose_x = var_3059_transpose_x_1, transpose_y = var_3059_transpose_y_1, x = var_3057_cast_fp16, y = v_new_169_cast_fp16)[name = string("op_3059_cast_fp16")]; + tensor state_171_cast_fp16 = add(x = var_3055_cast_fp16, y = var_3059_cast_fp16)[name = string("state_171_cast_fp16")]; + tensor var_3062_begin_0 = const()[name = string("op_3062_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_3062_end_0 = const()[name = string("op_3062_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_3062_end_mask_0 = const()[name = string("op_3062_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3062_squeeze_mask_0 = const()[name = string("op_3062_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3062_cast_fp16 = slice_by_index(begin = var_3062_begin_0, end = var_3062_end_0, end_mask = var_3062_end_mask_0, squeeze_mask = var_3062_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3062_cast_fp16")]; + tensor var_3064_begin_0 = const()[name = string("op_3064_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_3064_end_0 = const()[name = string("op_3064_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_3064_end_mask_0 = const()[name = string("op_3064_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3064_squeeze_mask_0 = const()[name = string("op_3064_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3064_cast_fp16 = slice_by_index(begin = var_3064_begin_0, end = var_3064_end_0, end_mask = var_3064_end_mask_0, squeeze_mask = var_3064_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3064_cast_fp16")]; + bool var_3065_transpose_x_0 = const()[name = string("op_3065_transpose_x_0"), val = bool(false)]; + bool var_3065_transpose_y_0 = const()[name = string("op_3065_transpose_y_0"), val = bool(false)]; + tensor var_3065_cast_fp16 = matmul(transpose_x = var_3065_transpose_x_0, transpose_y = var_3065_transpose_y_0, x = var_3064_cast_fp16, y = state_171_cast_fp16)[name = string("op_3065_cast_fp16")]; + tensor v_new_171_cast_fp16 = sub(x = var_3062_cast_fp16, y = var_3065_cast_fp16)[name = string("v_new_171_cast_fp16")]; + tensor var_3068_begin_0 = const()[name = string("op_3068_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_3068_end_0 = const()[name = string("op_3068_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_3068_end_mask_0 = const()[name = string("op_3068_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3068_squeeze_mask_0 = const()[name = string("op_3068_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3068_cast_fp16 = slice_by_index(begin = var_3068_begin_0, end = var_3068_end_0, end_mask = var_3068_end_mask_0, squeeze_mask = var_3068_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3068_cast_fp16")]; + bool var_3069_transpose_x_0 = const()[name = string("op_3069_transpose_x_0"), val = bool(false)]; + bool var_3069_transpose_y_0 = const()[name = string("op_3069_transpose_y_0"), val = bool(false)]; + tensor var_3069_cast_fp16 = matmul(transpose_x = var_3069_transpose_x_0, transpose_y = var_3069_transpose_y_0, x = var_3068_cast_fp16, y = state_171_cast_fp16)[name = string("op_3069_cast_fp16")]; + tensor var_3071_begin_0 = const()[name = string("op_3071_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_3071_end_0 = const()[name = string("op_3071_end_0"), val = tensor([32, 22, 64, 64])]; + tensor var_3071_end_mask_0 = const()[name = string("op_3071_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3071_squeeze_mask_0 = const()[name = string("op_3071_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3071_cast_fp16 = slice_by_index(begin = var_3071_begin_0, end = var_3071_end_0, end_mask = var_3071_end_mask_0, squeeze_mask = var_3071_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3071_cast_fp16")]; + bool var_3072_transpose_x_0 = const()[name = string("op_3072_transpose_x_0"), val = bool(false)]; + bool var_3072_transpose_y_0 = const()[name = string("op_3072_transpose_y_0"), val = bool(false)]; + tensor var_3072_cast_fp16 = matmul(transpose_x = var_3072_transpose_x_0, transpose_y = var_3072_transpose_y_0, x = var_3071_cast_fp16, y = v_new_171_cast_fp16)[name = string("op_3072_cast_fp16")]; + tensor var_3073_cast_fp16 = add(x = var_3069_cast_fp16, y = var_3072_cast_fp16)[name = string("op_3073_cast_fp16")]; + tensor var_3075_begin_0 = const()[name = string("op_3075_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_3075_end_0 = const()[name = string("op_3075_end_0"), val = tensor([32, 22, 1, 1])]; + tensor var_3075_end_mask_0 = const()[name = string("op_3075_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3075_squeeze_mask_0 = const()[name = string("op_3075_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3075_cast_fp16 = slice_by_index(begin = var_3075_begin_0, end = var_3075_end_0, end_mask = var_3075_end_mask_0, squeeze_mask = var_3075_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3075_cast_fp16")]; + tensor var_3076_cast_fp16 = mul(x = state_171_cast_fp16, y = var_3075_cast_fp16)[name = string("op_3076_cast_fp16")]; + tensor var_3078_begin_0 = const()[name = string("op_3078_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_3078_end_0 = const()[name = string("op_3078_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_3078_end_mask_0 = const()[name = string("op_3078_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3078_squeeze_mask_0 = const()[name = string("op_3078_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3078_cast_fp16 = slice_by_index(begin = var_3078_begin_0, end = var_3078_end_0, end_mask = var_3078_end_mask_0, squeeze_mask = var_3078_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3078_cast_fp16")]; + bool var_3080_transpose_x_1 = const()[name = string("op_3080_transpose_x_1"), val = bool(true)]; + bool var_3080_transpose_y_1 = const()[name = string("op_3080_transpose_y_1"), val = bool(false)]; + tensor var_3080_cast_fp16 = matmul(transpose_x = var_3080_transpose_x_1, transpose_y = var_3080_transpose_y_1, x = var_3078_cast_fp16, y = v_new_171_cast_fp16)[name = string("op_3080_cast_fp16")]; + tensor state_173_cast_fp16 = add(x = var_3076_cast_fp16, y = var_3080_cast_fp16)[name = string("state_173_cast_fp16")]; + tensor var_3083_begin_0 = const()[name = string("op_3083_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_3083_end_0 = const()[name = string("op_3083_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_3083_end_mask_0 = const()[name = string("op_3083_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3083_squeeze_mask_0 = const()[name = string("op_3083_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3083_cast_fp16 = slice_by_index(begin = var_3083_begin_0, end = var_3083_end_0, end_mask = var_3083_end_mask_0, squeeze_mask = var_3083_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3083_cast_fp16")]; + tensor var_3085_begin_0 = const()[name = string("op_3085_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_3085_end_0 = const()[name = string("op_3085_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_3085_end_mask_0 = const()[name = string("op_3085_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3085_squeeze_mask_0 = const()[name = string("op_3085_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3085_cast_fp16 = slice_by_index(begin = var_3085_begin_0, end = var_3085_end_0, end_mask = var_3085_end_mask_0, squeeze_mask = var_3085_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3085_cast_fp16")]; + bool var_3086_transpose_x_0 = const()[name = string("op_3086_transpose_x_0"), val = bool(false)]; + bool var_3086_transpose_y_0 = const()[name = string("op_3086_transpose_y_0"), val = bool(false)]; + tensor var_3086_cast_fp16 = matmul(transpose_x = var_3086_transpose_x_0, transpose_y = var_3086_transpose_y_0, x = var_3085_cast_fp16, y = state_173_cast_fp16)[name = string("op_3086_cast_fp16")]; + tensor v_new_173_cast_fp16 = sub(x = var_3083_cast_fp16, y = var_3086_cast_fp16)[name = string("v_new_173_cast_fp16")]; + tensor var_3089_begin_0 = const()[name = string("op_3089_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_3089_end_0 = const()[name = string("op_3089_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_3089_end_mask_0 = const()[name = string("op_3089_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3089_squeeze_mask_0 = const()[name = string("op_3089_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3089_cast_fp16 = slice_by_index(begin = var_3089_begin_0, end = var_3089_end_0, end_mask = var_3089_end_mask_0, squeeze_mask = var_3089_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3089_cast_fp16")]; + bool var_3090_transpose_x_0 = const()[name = string("op_3090_transpose_x_0"), val = bool(false)]; + bool var_3090_transpose_y_0 = const()[name = string("op_3090_transpose_y_0"), val = bool(false)]; + tensor var_3090_cast_fp16 = matmul(transpose_x = var_3090_transpose_x_0, transpose_y = var_3090_transpose_y_0, x = var_3089_cast_fp16, y = state_173_cast_fp16)[name = string("op_3090_cast_fp16")]; + tensor var_3092_begin_0 = const()[name = string("op_3092_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_3092_end_0 = const()[name = string("op_3092_end_0"), val = tensor([32, 23, 64, 64])]; + tensor var_3092_end_mask_0 = const()[name = string("op_3092_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3092_squeeze_mask_0 = const()[name = string("op_3092_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3092_cast_fp16 = slice_by_index(begin = var_3092_begin_0, end = var_3092_end_0, end_mask = var_3092_end_mask_0, squeeze_mask = var_3092_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3092_cast_fp16")]; + bool var_3093_transpose_x_0 = const()[name = string("op_3093_transpose_x_0"), val = bool(false)]; + bool var_3093_transpose_y_0 = const()[name = string("op_3093_transpose_y_0"), val = bool(false)]; + tensor var_3093_cast_fp16 = matmul(transpose_x = var_3093_transpose_x_0, transpose_y = var_3093_transpose_y_0, x = var_3092_cast_fp16, y = v_new_173_cast_fp16)[name = string("op_3093_cast_fp16")]; + tensor var_3094_cast_fp16 = add(x = var_3090_cast_fp16, y = var_3093_cast_fp16)[name = string("op_3094_cast_fp16")]; + tensor var_3096_begin_0 = const()[name = string("op_3096_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_3096_end_0 = const()[name = string("op_3096_end_0"), val = tensor([32, 23, 1, 1])]; + tensor var_3096_end_mask_0 = const()[name = string("op_3096_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3096_squeeze_mask_0 = const()[name = string("op_3096_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3096_cast_fp16 = slice_by_index(begin = var_3096_begin_0, end = var_3096_end_0, end_mask = var_3096_end_mask_0, squeeze_mask = var_3096_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3096_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = state_173_cast_fp16, y = var_3096_cast_fp16)[name = string("op_3097_cast_fp16")]; + tensor var_3099_begin_0 = const()[name = string("op_3099_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_3099_end_0 = const()[name = string("op_3099_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_3099_end_mask_0 = const()[name = string("op_3099_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3099_squeeze_mask_0 = const()[name = string("op_3099_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3099_cast_fp16 = slice_by_index(begin = var_3099_begin_0, end = var_3099_end_0, end_mask = var_3099_end_mask_0, squeeze_mask = var_3099_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3099_cast_fp16")]; + bool var_3101_transpose_x_1 = const()[name = string("op_3101_transpose_x_1"), val = bool(true)]; + bool var_3101_transpose_y_1 = const()[name = string("op_3101_transpose_y_1"), val = bool(false)]; + tensor var_3101_cast_fp16 = matmul(transpose_x = var_3101_transpose_x_1, transpose_y = var_3101_transpose_y_1, x = var_3099_cast_fp16, y = v_new_173_cast_fp16)[name = string("op_3101_cast_fp16")]; + tensor state_175_cast_fp16 = add(x = var_3097_cast_fp16, y = var_3101_cast_fp16)[name = string("state_175_cast_fp16")]; + tensor var_3104_begin_0 = const()[name = string("op_3104_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_3104_end_0 = const()[name = string("op_3104_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_3104_end_mask_0 = const()[name = string("op_3104_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3104_squeeze_mask_0 = const()[name = string("op_3104_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3104_cast_fp16 = slice_by_index(begin = var_3104_begin_0, end = var_3104_end_0, end_mask = var_3104_end_mask_0, squeeze_mask = var_3104_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3104_cast_fp16")]; + tensor var_3106_begin_0 = const()[name = string("op_3106_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_3106_end_0 = const()[name = string("op_3106_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_3106_end_mask_0 = const()[name = string("op_3106_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3106_squeeze_mask_0 = const()[name = string("op_3106_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3106_cast_fp16 = slice_by_index(begin = var_3106_begin_0, end = var_3106_end_0, end_mask = var_3106_end_mask_0, squeeze_mask = var_3106_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3106_cast_fp16")]; + bool var_3107_transpose_x_0 = const()[name = string("op_3107_transpose_x_0"), val = bool(false)]; + bool var_3107_transpose_y_0 = const()[name = string("op_3107_transpose_y_0"), val = bool(false)]; + tensor var_3107_cast_fp16 = matmul(transpose_x = var_3107_transpose_x_0, transpose_y = var_3107_transpose_y_0, x = var_3106_cast_fp16, y = state_175_cast_fp16)[name = string("op_3107_cast_fp16")]; + tensor v_new_175_cast_fp16 = sub(x = var_3104_cast_fp16, y = var_3107_cast_fp16)[name = string("v_new_175_cast_fp16")]; + tensor var_3110_begin_0 = const()[name = string("op_3110_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_3110_end_0 = const()[name = string("op_3110_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_3110_end_mask_0 = const()[name = string("op_3110_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3110_squeeze_mask_0 = const()[name = string("op_3110_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3110_cast_fp16 = slice_by_index(begin = var_3110_begin_0, end = var_3110_end_0, end_mask = var_3110_end_mask_0, squeeze_mask = var_3110_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3110_cast_fp16")]; + bool var_3111_transpose_x_0 = const()[name = string("op_3111_transpose_x_0"), val = bool(false)]; + bool var_3111_transpose_y_0 = const()[name = string("op_3111_transpose_y_0"), val = bool(false)]; + tensor var_3111_cast_fp16 = matmul(transpose_x = var_3111_transpose_x_0, transpose_y = var_3111_transpose_y_0, x = var_3110_cast_fp16, y = state_175_cast_fp16)[name = string("op_3111_cast_fp16")]; + tensor var_3113_begin_0 = const()[name = string("op_3113_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_3113_end_0 = const()[name = string("op_3113_end_0"), val = tensor([32, 24, 64, 64])]; + tensor var_3113_end_mask_0 = const()[name = string("op_3113_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3113_squeeze_mask_0 = const()[name = string("op_3113_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3113_cast_fp16 = slice_by_index(begin = var_3113_begin_0, end = var_3113_end_0, end_mask = var_3113_end_mask_0, squeeze_mask = var_3113_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3113_cast_fp16")]; + bool var_3114_transpose_x_0 = const()[name = string("op_3114_transpose_x_0"), val = bool(false)]; + bool var_3114_transpose_y_0 = const()[name = string("op_3114_transpose_y_0"), val = bool(false)]; + tensor var_3114_cast_fp16 = matmul(transpose_x = var_3114_transpose_x_0, transpose_y = var_3114_transpose_y_0, x = var_3113_cast_fp16, y = v_new_175_cast_fp16)[name = string("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = add(x = var_3111_cast_fp16, y = var_3114_cast_fp16)[name = string("op_3115_cast_fp16")]; + tensor var_3117_begin_0 = const()[name = string("op_3117_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_3117_end_0 = const()[name = string("op_3117_end_0"), val = tensor([32, 24, 1, 1])]; + tensor var_3117_end_mask_0 = const()[name = string("op_3117_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3117_squeeze_mask_0 = const()[name = string("op_3117_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3117_cast_fp16 = slice_by_index(begin = var_3117_begin_0, end = var_3117_end_0, end_mask = var_3117_end_mask_0, squeeze_mask = var_3117_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3117_cast_fp16")]; + tensor var_3118_cast_fp16 = mul(x = state_175_cast_fp16, y = var_3117_cast_fp16)[name = string("op_3118_cast_fp16")]; + tensor var_3120_begin_0 = const()[name = string("op_3120_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_3120_end_0 = const()[name = string("op_3120_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_3120_end_mask_0 = const()[name = string("op_3120_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3120_squeeze_mask_0 = const()[name = string("op_3120_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3120_cast_fp16 = slice_by_index(begin = var_3120_begin_0, end = var_3120_end_0, end_mask = var_3120_end_mask_0, squeeze_mask = var_3120_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3120_cast_fp16")]; + bool var_3122_transpose_x_1 = const()[name = string("op_3122_transpose_x_1"), val = bool(true)]; + bool var_3122_transpose_y_1 = const()[name = string("op_3122_transpose_y_1"), val = bool(false)]; + tensor var_3122_cast_fp16 = matmul(transpose_x = var_3122_transpose_x_1, transpose_y = var_3122_transpose_y_1, x = var_3120_cast_fp16, y = v_new_175_cast_fp16)[name = string("op_3122_cast_fp16")]; + tensor state_177_cast_fp16 = add(x = var_3118_cast_fp16, y = var_3122_cast_fp16)[name = string("state_177_cast_fp16")]; + tensor var_3125_begin_0 = const()[name = string("op_3125_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_3125_end_0 = const()[name = string("op_3125_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_3125_end_mask_0 = const()[name = string("op_3125_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3125_squeeze_mask_0 = const()[name = string("op_3125_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3125_cast_fp16 = slice_by_index(begin = var_3125_begin_0, end = var_3125_end_0, end_mask = var_3125_end_mask_0, squeeze_mask = var_3125_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3125_cast_fp16")]; + tensor var_3127_begin_0 = const()[name = string("op_3127_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_3127_end_0 = const()[name = string("op_3127_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_3127_end_mask_0 = const()[name = string("op_3127_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3127_squeeze_mask_0 = const()[name = string("op_3127_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3127_cast_fp16 = slice_by_index(begin = var_3127_begin_0, end = var_3127_end_0, end_mask = var_3127_end_mask_0, squeeze_mask = var_3127_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3127_cast_fp16")]; + bool var_3128_transpose_x_0 = const()[name = string("op_3128_transpose_x_0"), val = bool(false)]; + bool var_3128_transpose_y_0 = const()[name = string("op_3128_transpose_y_0"), val = bool(false)]; + tensor var_3128_cast_fp16 = matmul(transpose_x = var_3128_transpose_x_0, transpose_y = var_3128_transpose_y_0, x = var_3127_cast_fp16, y = state_177_cast_fp16)[name = string("op_3128_cast_fp16")]; + tensor v_new_177_cast_fp16 = sub(x = var_3125_cast_fp16, y = var_3128_cast_fp16)[name = string("v_new_177_cast_fp16")]; + tensor var_3131_begin_0 = const()[name = string("op_3131_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_3131_end_0 = const()[name = string("op_3131_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_3131_end_mask_0 = const()[name = string("op_3131_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3131_squeeze_mask_0 = const()[name = string("op_3131_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3131_cast_fp16 = slice_by_index(begin = var_3131_begin_0, end = var_3131_end_0, end_mask = var_3131_end_mask_0, squeeze_mask = var_3131_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3131_cast_fp16")]; + bool var_3132_transpose_x_0 = const()[name = string("op_3132_transpose_x_0"), val = bool(false)]; + bool var_3132_transpose_y_0 = const()[name = string("op_3132_transpose_y_0"), val = bool(false)]; + tensor var_3132_cast_fp16 = matmul(transpose_x = var_3132_transpose_x_0, transpose_y = var_3132_transpose_y_0, x = var_3131_cast_fp16, y = state_177_cast_fp16)[name = string("op_3132_cast_fp16")]; + tensor var_3134_begin_0 = const()[name = string("op_3134_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_3134_end_0 = const()[name = string("op_3134_end_0"), val = tensor([32, 25, 64, 64])]; + tensor var_3134_end_mask_0 = const()[name = string("op_3134_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3134_squeeze_mask_0 = const()[name = string("op_3134_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3134_cast_fp16 = slice_by_index(begin = var_3134_begin_0, end = var_3134_end_0, end_mask = var_3134_end_mask_0, squeeze_mask = var_3134_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3134_cast_fp16")]; + bool var_3135_transpose_x_0 = const()[name = string("op_3135_transpose_x_0"), val = bool(false)]; + bool var_3135_transpose_y_0 = const()[name = string("op_3135_transpose_y_0"), val = bool(false)]; + tensor var_3135_cast_fp16 = matmul(transpose_x = var_3135_transpose_x_0, transpose_y = var_3135_transpose_y_0, x = var_3134_cast_fp16, y = v_new_177_cast_fp16)[name = string("op_3135_cast_fp16")]; + tensor var_3136_cast_fp16 = add(x = var_3132_cast_fp16, y = var_3135_cast_fp16)[name = string("op_3136_cast_fp16")]; + tensor var_3138_begin_0 = const()[name = string("op_3138_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_3138_end_0 = const()[name = string("op_3138_end_0"), val = tensor([32, 25, 1, 1])]; + tensor var_3138_end_mask_0 = const()[name = string("op_3138_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3138_squeeze_mask_0 = const()[name = string("op_3138_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3138_cast_fp16 = slice_by_index(begin = var_3138_begin_0, end = var_3138_end_0, end_mask = var_3138_end_mask_0, squeeze_mask = var_3138_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3138_cast_fp16")]; + tensor var_3139_cast_fp16 = mul(x = state_177_cast_fp16, y = var_3138_cast_fp16)[name = string("op_3139_cast_fp16")]; + tensor var_3141_begin_0 = const()[name = string("op_3141_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_3141_end_0 = const()[name = string("op_3141_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_3141_end_mask_0 = const()[name = string("op_3141_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3141_squeeze_mask_0 = const()[name = string("op_3141_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3141_cast_fp16 = slice_by_index(begin = var_3141_begin_0, end = var_3141_end_0, end_mask = var_3141_end_mask_0, squeeze_mask = var_3141_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3141_cast_fp16")]; + bool var_3143_transpose_x_1 = const()[name = string("op_3143_transpose_x_1"), val = bool(true)]; + bool var_3143_transpose_y_1 = const()[name = string("op_3143_transpose_y_1"), val = bool(false)]; + tensor var_3143_cast_fp16 = matmul(transpose_x = var_3143_transpose_x_1, transpose_y = var_3143_transpose_y_1, x = var_3141_cast_fp16, y = v_new_177_cast_fp16)[name = string("op_3143_cast_fp16")]; + tensor state_179_cast_fp16 = add(x = var_3139_cast_fp16, y = var_3143_cast_fp16)[name = string("state_179_cast_fp16")]; + tensor var_3146_begin_0 = const()[name = string("op_3146_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_3146_end_0 = const()[name = string("op_3146_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_3146_end_mask_0 = const()[name = string("op_3146_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3146_squeeze_mask_0 = const()[name = string("op_3146_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3146_cast_fp16 = slice_by_index(begin = var_3146_begin_0, end = var_3146_end_0, end_mask = var_3146_end_mask_0, squeeze_mask = var_3146_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3146_cast_fp16")]; + tensor var_3148_begin_0 = const()[name = string("op_3148_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_3148_end_0 = const()[name = string("op_3148_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_3148_end_mask_0 = const()[name = string("op_3148_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3148_squeeze_mask_0 = const()[name = string("op_3148_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3148_cast_fp16 = slice_by_index(begin = var_3148_begin_0, end = var_3148_end_0, end_mask = var_3148_end_mask_0, squeeze_mask = var_3148_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3148_cast_fp16")]; + bool var_3149_transpose_x_0 = const()[name = string("op_3149_transpose_x_0"), val = bool(false)]; + bool var_3149_transpose_y_0 = const()[name = string("op_3149_transpose_y_0"), val = bool(false)]; + tensor var_3149_cast_fp16 = matmul(transpose_x = var_3149_transpose_x_0, transpose_y = var_3149_transpose_y_0, x = var_3148_cast_fp16, y = state_179_cast_fp16)[name = string("op_3149_cast_fp16")]; + tensor v_new_179_cast_fp16 = sub(x = var_3146_cast_fp16, y = var_3149_cast_fp16)[name = string("v_new_179_cast_fp16")]; + tensor var_3152_begin_0 = const()[name = string("op_3152_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_3152_end_0 = const()[name = string("op_3152_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_3152_end_mask_0 = const()[name = string("op_3152_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3152_squeeze_mask_0 = const()[name = string("op_3152_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3152_cast_fp16 = slice_by_index(begin = var_3152_begin_0, end = var_3152_end_0, end_mask = var_3152_end_mask_0, squeeze_mask = var_3152_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3152_cast_fp16")]; + bool var_3153_transpose_x_0 = const()[name = string("op_3153_transpose_x_0"), val = bool(false)]; + bool var_3153_transpose_y_0 = const()[name = string("op_3153_transpose_y_0"), val = bool(false)]; + tensor var_3153_cast_fp16 = matmul(transpose_x = var_3153_transpose_x_0, transpose_y = var_3153_transpose_y_0, x = var_3152_cast_fp16, y = state_179_cast_fp16)[name = string("op_3153_cast_fp16")]; + tensor var_3155_begin_0 = const()[name = string("op_3155_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_3155_end_0 = const()[name = string("op_3155_end_0"), val = tensor([32, 26, 64, 64])]; + tensor var_3155_end_mask_0 = const()[name = string("op_3155_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3155_squeeze_mask_0 = const()[name = string("op_3155_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3155_cast_fp16 = slice_by_index(begin = var_3155_begin_0, end = var_3155_end_0, end_mask = var_3155_end_mask_0, squeeze_mask = var_3155_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3155_cast_fp16")]; + bool var_3156_transpose_x_0 = const()[name = string("op_3156_transpose_x_0"), val = bool(false)]; + bool var_3156_transpose_y_0 = const()[name = string("op_3156_transpose_y_0"), val = bool(false)]; + tensor var_3156_cast_fp16 = matmul(transpose_x = var_3156_transpose_x_0, transpose_y = var_3156_transpose_y_0, x = var_3155_cast_fp16, y = v_new_179_cast_fp16)[name = string("op_3156_cast_fp16")]; + tensor var_3157_cast_fp16 = add(x = var_3153_cast_fp16, y = var_3156_cast_fp16)[name = string("op_3157_cast_fp16")]; + tensor var_3159_begin_0 = const()[name = string("op_3159_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_3159_end_0 = const()[name = string("op_3159_end_0"), val = tensor([32, 26, 1, 1])]; + tensor var_3159_end_mask_0 = const()[name = string("op_3159_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3159_squeeze_mask_0 = const()[name = string("op_3159_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3159_cast_fp16 = slice_by_index(begin = var_3159_begin_0, end = var_3159_end_0, end_mask = var_3159_end_mask_0, squeeze_mask = var_3159_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3159_cast_fp16")]; + tensor var_3160_cast_fp16 = mul(x = state_179_cast_fp16, y = var_3159_cast_fp16)[name = string("op_3160_cast_fp16")]; + tensor var_3162_begin_0 = const()[name = string("op_3162_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_3162_end_0 = const()[name = string("op_3162_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_3162_end_mask_0 = const()[name = string("op_3162_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3162_squeeze_mask_0 = const()[name = string("op_3162_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3162_cast_fp16 = slice_by_index(begin = var_3162_begin_0, end = var_3162_end_0, end_mask = var_3162_end_mask_0, squeeze_mask = var_3162_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3162_cast_fp16")]; + bool var_3164_transpose_x_1 = const()[name = string("op_3164_transpose_x_1"), val = bool(true)]; + bool var_3164_transpose_y_1 = const()[name = string("op_3164_transpose_y_1"), val = bool(false)]; + tensor var_3164_cast_fp16 = matmul(transpose_x = var_3164_transpose_x_1, transpose_y = var_3164_transpose_y_1, x = var_3162_cast_fp16, y = v_new_179_cast_fp16)[name = string("op_3164_cast_fp16")]; + tensor state_181_cast_fp16 = add(x = var_3160_cast_fp16, y = var_3164_cast_fp16)[name = string("state_181_cast_fp16")]; + tensor var_3167_begin_0 = const()[name = string("op_3167_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_3167_end_0 = const()[name = string("op_3167_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_3167_end_mask_0 = const()[name = string("op_3167_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3167_squeeze_mask_0 = const()[name = string("op_3167_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, squeeze_mask = var_3167_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3167_cast_fp16")]; + tensor var_3169_begin_0 = const()[name = string("op_3169_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_3169_end_0 = const()[name = string("op_3169_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_3169_end_mask_0 = const()[name = string("op_3169_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3169_squeeze_mask_0 = const()[name = string("op_3169_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3169_cast_fp16 = slice_by_index(begin = var_3169_begin_0, end = var_3169_end_0, end_mask = var_3169_end_mask_0, squeeze_mask = var_3169_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3169_cast_fp16")]; + bool var_3170_transpose_x_0 = const()[name = string("op_3170_transpose_x_0"), val = bool(false)]; + bool var_3170_transpose_y_0 = const()[name = string("op_3170_transpose_y_0"), val = bool(false)]; + tensor var_3170_cast_fp16 = matmul(transpose_x = var_3170_transpose_x_0, transpose_y = var_3170_transpose_y_0, x = var_3169_cast_fp16, y = state_181_cast_fp16)[name = string("op_3170_cast_fp16")]; + tensor v_new_181_cast_fp16 = sub(x = var_3167_cast_fp16, y = var_3170_cast_fp16)[name = string("v_new_181_cast_fp16")]; + tensor var_3173_begin_0 = const()[name = string("op_3173_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_3173_end_0 = const()[name = string("op_3173_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_3173_end_mask_0 = const()[name = string("op_3173_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3173_squeeze_mask_0 = const()[name = string("op_3173_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3173_cast_fp16 = slice_by_index(begin = var_3173_begin_0, end = var_3173_end_0, end_mask = var_3173_end_mask_0, squeeze_mask = var_3173_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3173_cast_fp16")]; + bool var_3174_transpose_x_0 = const()[name = string("op_3174_transpose_x_0"), val = bool(false)]; + bool var_3174_transpose_y_0 = const()[name = string("op_3174_transpose_y_0"), val = bool(false)]; + tensor var_3174_cast_fp16 = matmul(transpose_x = var_3174_transpose_x_0, transpose_y = var_3174_transpose_y_0, x = var_3173_cast_fp16, y = state_181_cast_fp16)[name = string("op_3174_cast_fp16")]; + tensor var_3176_begin_0 = const()[name = string("op_3176_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_3176_end_0 = const()[name = string("op_3176_end_0"), val = tensor([32, 27, 64, 64])]; + tensor var_3176_end_mask_0 = const()[name = string("op_3176_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3176_squeeze_mask_0 = const()[name = string("op_3176_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3176_cast_fp16 = slice_by_index(begin = var_3176_begin_0, end = var_3176_end_0, end_mask = var_3176_end_mask_0, squeeze_mask = var_3176_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3176_cast_fp16")]; + bool var_3177_transpose_x_0 = const()[name = string("op_3177_transpose_x_0"), val = bool(false)]; + bool var_3177_transpose_y_0 = const()[name = string("op_3177_transpose_y_0"), val = bool(false)]; + tensor var_3177_cast_fp16 = matmul(transpose_x = var_3177_transpose_x_0, transpose_y = var_3177_transpose_y_0, x = var_3176_cast_fp16, y = v_new_181_cast_fp16)[name = string("op_3177_cast_fp16")]; + tensor var_3178_cast_fp16 = add(x = var_3174_cast_fp16, y = var_3177_cast_fp16)[name = string("op_3178_cast_fp16")]; + tensor var_3180_begin_0 = const()[name = string("op_3180_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_3180_end_0 = const()[name = string("op_3180_end_0"), val = tensor([32, 27, 1, 1])]; + tensor var_3180_end_mask_0 = const()[name = string("op_3180_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3180_squeeze_mask_0 = const()[name = string("op_3180_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3180_cast_fp16 = slice_by_index(begin = var_3180_begin_0, end = var_3180_end_0, end_mask = var_3180_end_mask_0, squeeze_mask = var_3180_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3180_cast_fp16")]; + tensor var_3181_cast_fp16 = mul(x = state_181_cast_fp16, y = var_3180_cast_fp16)[name = string("op_3181_cast_fp16")]; + tensor var_3183_begin_0 = const()[name = string("op_3183_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_3183_end_0 = const()[name = string("op_3183_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_3183_end_mask_0 = const()[name = string("op_3183_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3183_squeeze_mask_0 = const()[name = string("op_3183_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3183_cast_fp16 = slice_by_index(begin = var_3183_begin_0, end = var_3183_end_0, end_mask = var_3183_end_mask_0, squeeze_mask = var_3183_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3183_cast_fp16")]; + bool var_3185_transpose_x_1 = const()[name = string("op_3185_transpose_x_1"), val = bool(true)]; + bool var_3185_transpose_y_1 = const()[name = string("op_3185_transpose_y_1"), val = bool(false)]; + tensor var_3185_cast_fp16 = matmul(transpose_x = var_3185_transpose_x_1, transpose_y = var_3185_transpose_y_1, x = var_3183_cast_fp16, y = v_new_181_cast_fp16)[name = string("op_3185_cast_fp16")]; + tensor state_183_cast_fp16 = add(x = var_3181_cast_fp16, y = var_3185_cast_fp16)[name = string("state_183_cast_fp16")]; + tensor var_3188_begin_0 = const()[name = string("op_3188_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_3188_end_0 = const()[name = string("op_3188_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_3188_end_mask_0 = const()[name = string("op_3188_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3188_squeeze_mask_0 = const()[name = string("op_3188_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3188_cast_fp16 = slice_by_index(begin = var_3188_begin_0, end = var_3188_end_0, end_mask = var_3188_end_mask_0, squeeze_mask = var_3188_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3188_cast_fp16")]; + tensor var_3190_begin_0 = const()[name = string("op_3190_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_3190_end_0 = const()[name = string("op_3190_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_3190_end_mask_0 = const()[name = string("op_3190_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3190_squeeze_mask_0 = const()[name = string("op_3190_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3190_cast_fp16 = slice_by_index(begin = var_3190_begin_0, end = var_3190_end_0, end_mask = var_3190_end_mask_0, squeeze_mask = var_3190_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3190_cast_fp16")]; + bool var_3191_transpose_x_0 = const()[name = string("op_3191_transpose_x_0"), val = bool(false)]; + bool var_3191_transpose_y_0 = const()[name = string("op_3191_transpose_y_0"), val = bool(false)]; + tensor var_3191_cast_fp16 = matmul(transpose_x = var_3191_transpose_x_0, transpose_y = var_3191_transpose_y_0, x = var_3190_cast_fp16, y = state_183_cast_fp16)[name = string("op_3191_cast_fp16")]; + tensor v_new_183_cast_fp16 = sub(x = var_3188_cast_fp16, y = var_3191_cast_fp16)[name = string("v_new_183_cast_fp16")]; + tensor var_3194_begin_0 = const()[name = string("op_3194_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_3194_end_0 = const()[name = string("op_3194_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_3194_end_mask_0 = const()[name = string("op_3194_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3194_squeeze_mask_0 = const()[name = string("op_3194_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3194_cast_fp16 = slice_by_index(begin = var_3194_begin_0, end = var_3194_end_0, end_mask = var_3194_end_mask_0, squeeze_mask = var_3194_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3194_cast_fp16")]; + bool var_3195_transpose_x_0 = const()[name = string("op_3195_transpose_x_0"), val = bool(false)]; + bool var_3195_transpose_y_0 = const()[name = string("op_3195_transpose_y_0"), val = bool(false)]; + tensor var_3195_cast_fp16 = matmul(transpose_x = var_3195_transpose_x_0, transpose_y = var_3195_transpose_y_0, x = var_3194_cast_fp16, y = state_183_cast_fp16)[name = string("op_3195_cast_fp16")]; + tensor var_3197_begin_0 = const()[name = string("op_3197_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_3197_end_0 = const()[name = string("op_3197_end_0"), val = tensor([32, 28, 64, 64])]; + tensor var_3197_end_mask_0 = const()[name = string("op_3197_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3197_squeeze_mask_0 = const()[name = string("op_3197_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3197_cast_fp16 = slice_by_index(begin = var_3197_begin_0, end = var_3197_end_0, end_mask = var_3197_end_mask_0, squeeze_mask = var_3197_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3197_cast_fp16")]; + bool var_3198_transpose_x_0 = const()[name = string("op_3198_transpose_x_0"), val = bool(false)]; + bool var_3198_transpose_y_0 = const()[name = string("op_3198_transpose_y_0"), val = bool(false)]; + tensor var_3198_cast_fp16 = matmul(transpose_x = var_3198_transpose_x_0, transpose_y = var_3198_transpose_y_0, x = var_3197_cast_fp16, y = v_new_183_cast_fp16)[name = string("op_3198_cast_fp16")]; + tensor var_3199_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3198_cast_fp16)[name = string("op_3199_cast_fp16")]; + tensor var_3201_begin_0 = const()[name = string("op_3201_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_3201_end_0 = const()[name = string("op_3201_end_0"), val = tensor([32, 28, 1, 1])]; + tensor var_3201_end_mask_0 = const()[name = string("op_3201_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3201_squeeze_mask_0 = const()[name = string("op_3201_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3201_cast_fp16 = slice_by_index(begin = var_3201_begin_0, end = var_3201_end_0, end_mask = var_3201_end_mask_0, squeeze_mask = var_3201_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3201_cast_fp16")]; + tensor var_3202_cast_fp16 = mul(x = state_183_cast_fp16, y = var_3201_cast_fp16)[name = string("op_3202_cast_fp16")]; + tensor var_3204_begin_0 = const()[name = string("op_3204_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_3204_end_0 = const()[name = string("op_3204_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_3204_end_mask_0 = const()[name = string("op_3204_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3204_squeeze_mask_0 = const()[name = string("op_3204_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3204_cast_fp16 = slice_by_index(begin = var_3204_begin_0, end = var_3204_end_0, end_mask = var_3204_end_mask_0, squeeze_mask = var_3204_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3204_cast_fp16")]; + bool var_3206_transpose_x_1 = const()[name = string("op_3206_transpose_x_1"), val = bool(true)]; + bool var_3206_transpose_y_1 = const()[name = string("op_3206_transpose_y_1"), val = bool(false)]; + tensor var_3206_cast_fp16 = matmul(transpose_x = var_3206_transpose_x_1, transpose_y = var_3206_transpose_y_1, x = var_3204_cast_fp16, y = v_new_183_cast_fp16)[name = string("op_3206_cast_fp16")]; + tensor state_185_cast_fp16 = add(x = var_3202_cast_fp16, y = var_3206_cast_fp16)[name = string("state_185_cast_fp16")]; + tensor var_3209_begin_0 = const()[name = string("op_3209_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_3209_end_0 = const()[name = string("op_3209_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_3209_end_mask_0 = const()[name = string("op_3209_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3209_squeeze_mask_0 = const()[name = string("op_3209_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3209_cast_fp16 = slice_by_index(begin = var_3209_begin_0, end = var_3209_end_0, end_mask = var_3209_end_mask_0, squeeze_mask = var_3209_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3209_cast_fp16")]; + tensor var_3211_begin_0 = const()[name = string("op_3211_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_3211_end_0 = const()[name = string("op_3211_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_3211_end_mask_0 = const()[name = string("op_3211_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3211_squeeze_mask_0 = const()[name = string("op_3211_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3211_cast_fp16 = slice_by_index(begin = var_3211_begin_0, end = var_3211_end_0, end_mask = var_3211_end_mask_0, squeeze_mask = var_3211_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3211_cast_fp16")]; + bool var_3212_transpose_x_0 = const()[name = string("op_3212_transpose_x_0"), val = bool(false)]; + bool var_3212_transpose_y_0 = const()[name = string("op_3212_transpose_y_0"), val = bool(false)]; + tensor var_3212_cast_fp16 = matmul(transpose_x = var_3212_transpose_x_0, transpose_y = var_3212_transpose_y_0, x = var_3211_cast_fp16, y = state_185_cast_fp16)[name = string("op_3212_cast_fp16")]; + tensor v_new_185_cast_fp16 = sub(x = var_3209_cast_fp16, y = var_3212_cast_fp16)[name = string("v_new_185_cast_fp16")]; + tensor var_3215_begin_0 = const()[name = string("op_3215_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_3215_end_0 = const()[name = string("op_3215_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_3215_end_mask_0 = const()[name = string("op_3215_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3215_squeeze_mask_0 = const()[name = string("op_3215_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3215_cast_fp16 = slice_by_index(begin = var_3215_begin_0, end = var_3215_end_0, end_mask = var_3215_end_mask_0, squeeze_mask = var_3215_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3215_cast_fp16")]; + bool var_3216_transpose_x_0 = const()[name = string("op_3216_transpose_x_0"), val = bool(false)]; + bool var_3216_transpose_y_0 = const()[name = string("op_3216_transpose_y_0"), val = bool(false)]; + tensor var_3216_cast_fp16 = matmul(transpose_x = var_3216_transpose_x_0, transpose_y = var_3216_transpose_y_0, x = var_3215_cast_fp16, y = state_185_cast_fp16)[name = string("op_3216_cast_fp16")]; + tensor var_3218_begin_0 = const()[name = string("op_3218_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_3218_end_0 = const()[name = string("op_3218_end_0"), val = tensor([32, 29, 64, 64])]; + tensor var_3218_end_mask_0 = const()[name = string("op_3218_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3218_squeeze_mask_0 = const()[name = string("op_3218_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3218_cast_fp16 = slice_by_index(begin = var_3218_begin_0, end = var_3218_end_0, end_mask = var_3218_end_mask_0, squeeze_mask = var_3218_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3218_cast_fp16")]; + bool var_3219_transpose_x_0 = const()[name = string("op_3219_transpose_x_0"), val = bool(false)]; + bool var_3219_transpose_y_0 = const()[name = string("op_3219_transpose_y_0"), val = bool(false)]; + tensor var_3219_cast_fp16 = matmul(transpose_x = var_3219_transpose_x_0, transpose_y = var_3219_transpose_y_0, x = var_3218_cast_fp16, y = v_new_185_cast_fp16)[name = string("op_3219_cast_fp16")]; + tensor var_3220_cast_fp16 = add(x = var_3216_cast_fp16, y = var_3219_cast_fp16)[name = string("op_3220_cast_fp16")]; + tensor var_3222_begin_0 = const()[name = string("op_3222_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_3222_end_0 = const()[name = string("op_3222_end_0"), val = tensor([32, 29, 1, 1])]; + tensor var_3222_end_mask_0 = const()[name = string("op_3222_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3222_squeeze_mask_0 = const()[name = string("op_3222_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3222_cast_fp16 = slice_by_index(begin = var_3222_begin_0, end = var_3222_end_0, end_mask = var_3222_end_mask_0, squeeze_mask = var_3222_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3222_cast_fp16")]; + tensor var_3223_cast_fp16 = mul(x = state_185_cast_fp16, y = var_3222_cast_fp16)[name = string("op_3223_cast_fp16")]; + tensor var_3225_begin_0 = const()[name = string("op_3225_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_3225_end_0 = const()[name = string("op_3225_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_3225_end_mask_0 = const()[name = string("op_3225_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3225_squeeze_mask_0 = const()[name = string("op_3225_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3225_cast_fp16 = slice_by_index(begin = var_3225_begin_0, end = var_3225_end_0, end_mask = var_3225_end_mask_0, squeeze_mask = var_3225_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3225_cast_fp16")]; + bool var_3227_transpose_x_1 = const()[name = string("op_3227_transpose_x_1"), val = bool(true)]; + bool var_3227_transpose_y_1 = const()[name = string("op_3227_transpose_y_1"), val = bool(false)]; + tensor var_3227_cast_fp16 = matmul(transpose_x = var_3227_transpose_x_1, transpose_y = var_3227_transpose_y_1, x = var_3225_cast_fp16, y = v_new_185_cast_fp16)[name = string("op_3227_cast_fp16")]; + tensor state_187_cast_fp16 = add(x = var_3223_cast_fp16, y = var_3227_cast_fp16)[name = string("state_187_cast_fp16")]; + tensor var_3230_begin_0 = const()[name = string("op_3230_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_3230_end_0 = const()[name = string("op_3230_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_3230_end_mask_0 = const()[name = string("op_3230_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3230_squeeze_mask_0 = const()[name = string("op_3230_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3230_cast_fp16 = slice_by_index(begin = var_3230_begin_0, end = var_3230_end_0, end_mask = var_3230_end_mask_0, squeeze_mask = var_3230_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3230_cast_fp16")]; + tensor var_3232_begin_0 = const()[name = string("op_3232_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_3232_end_0 = const()[name = string("op_3232_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_3232_end_mask_0 = const()[name = string("op_3232_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3232_squeeze_mask_0 = const()[name = string("op_3232_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3232_cast_fp16 = slice_by_index(begin = var_3232_begin_0, end = var_3232_end_0, end_mask = var_3232_end_mask_0, squeeze_mask = var_3232_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3232_cast_fp16")]; + bool var_3233_transpose_x_0 = const()[name = string("op_3233_transpose_x_0"), val = bool(false)]; + bool var_3233_transpose_y_0 = const()[name = string("op_3233_transpose_y_0"), val = bool(false)]; + tensor var_3233_cast_fp16 = matmul(transpose_x = var_3233_transpose_x_0, transpose_y = var_3233_transpose_y_0, x = var_3232_cast_fp16, y = state_187_cast_fp16)[name = string("op_3233_cast_fp16")]; + tensor v_new_187_cast_fp16 = sub(x = var_3230_cast_fp16, y = var_3233_cast_fp16)[name = string("v_new_187_cast_fp16")]; + tensor var_3236_begin_0 = const()[name = string("op_3236_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_3236_end_0 = const()[name = string("op_3236_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_3236_end_mask_0 = const()[name = string("op_3236_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3236_squeeze_mask_0 = const()[name = string("op_3236_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3236_cast_fp16 = slice_by_index(begin = var_3236_begin_0, end = var_3236_end_0, end_mask = var_3236_end_mask_0, squeeze_mask = var_3236_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3236_cast_fp16")]; + bool var_3237_transpose_x_0 = const()[name = string("op_3237_transpose_x_0"), val = bool(false)]; + bool var_3237_transpose_y_0 = const()[name = string("op_3237_transpose_y_0"), val = bool(false)]; + tensor var_3237_cast_fp16 = matmul(transpose_x = var_3237_transpose_x_0, transpose_y = var_3237_transpose_y_0, x = var_3236_cast_fp16, y = state_187_cast_fp16)[name = string("op_3237_cast_fp16")]; + tensor var_3239_begin_0 = const()[name = string("op_3239_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_3239_end_0 = const()[name = string("op_3239_end_0"), val = tensor([32, 30, 64, 64])]; + tensor var_3239_end_mask_0 = const()[name = string("op_3239_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3239_squeeze_mask_0 = const()[name = string("op_3239_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3239_cast_fp16 = slice_by_index(begin = var_3239_begin_0, end = var_3239_end_0, end_mask = var_3239_end_mask_0, squeeze_mask = var_3239_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3239_cast_fp16")]; + bool var_3240_transpose_x_0 = const()[name = string("op_3240_transpose_x_0"), val = bool(false)]; + bool var_3240_transpose_y_0 = const()[name = string("op_3240_transpose_y_0"), val = bool(false)]; + tensor var_3240_cast_fp16 = matmul(transpose_x = var_3240_transpose_x_0, transpose_y = var_3240_transpose_y_0, x = var_3239_cast_fp16, y = v_new_187_cast_fp16)[name = string("op_3240_cast_fp16")]; + tensor var_3241_cast_fp16 = add(x = var_3237_cast_fp16, y = var_3240_cast_fp16)[name = string("op_3241_cast_fp16")]; + tensor var_3243_begin_0 = const()[name = string("op_3243_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_3243_end_0 = const()[name = string("op_3243_end_0"), val = tensor([32, 30, 1, 1])]; + tensor var_3243_end_mask_0 = const()[name = string("op_3243_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3243_squeeze_mask_0 = const()[name = string("op_3243_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3243_cast_fp16 = slice_by_index(begin = var_3243_begin_0, end = var_3243_end_0, end_mask = var_3243_end_mask_0, squeeze_mask = var_3243_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3243_cast_fp16")]; + tensor var_3244_cast_fp16 = mul(x = state_187_cast_fp16, y = var_3243_cast_fp16)[name = string("op_3244_cast_fp16")]; + tensor var_3246_begin_0 = const()[name = string("op_3246_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_3246_end_0 = const()[name = string("op_3246_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_3246_end_mask_0 = const()[name = string("op_3246_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3246_squeeze_mask_0 = const()[name = string("op_3246_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3246_cast_fp16 = slice_by_index(begin = var_3246_begin_0, end = var_3246_end_0, end_mask = var_3246_end_mask_0, squeeze_mask = var_3246_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3246_cast_fp16")]; + bool var_3248_transpose_x_1 = const()[name = string("op_3248_transpose_x_1"), val = bool(true)]; + bool var_3248_transpose_y_1 = const()[name = string("op_3248_transpose_y_1"), val = bool(false)]; + tensor var_3248_cast_fp16 = matmul(transpose_x = var_3248_transpose_x_1, transpose_y = var_3248_transpose_y_1, x = var_3246_cast_fp16, y = v_new_187_cast_fp16)[name = string("op_3248_cast_fp16")]; + tensor state_189_cast_fp16 = add(x = var_3244_cast_fp16, y = var_3248_cast_fp16)[name = string("state_189_cast_fp16")]; + tensor var_3251_begin_0 = const()[name = string("op_3251_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_3251_end_0 = const()[name = string("op_3251_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_3251_end_mask_0 = const()[name = string("op_3251_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3251_squeeze_mask_0 = const()[name = string("op_3251_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3251_cast_fp16 = slice_by_index(begin = var_3251_begin_0, end = var_3251_end_0, end_mask = var_3251_end_mask_0, squeeze_mask = var_3251_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3251_cast_fp16")]; + tensor var_3253_begin_0 = const()[name = string("op_3253_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_3253_end_0 = const()[name = string("op_3253_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_3253_end_mask_0 = const()[name = string("op_3253_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3253_squeeze_mask_0 = const()[name = string("op_3253_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3253_cast_fp16 = slice_by_index(begin = var_3253_begin_0, end = var_3253_end_0, end_mask = var_3253_end_mask_0, squeeze_mask = var_3253_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3253_cast_fp16")]; + bool var_3254_transpose_x_0 = const()[name = string("op_3254_transpose_x_0"), val = bool(false)]; + bool var_3254_transpose_y_0 = const()[name = string("op_3254_transpose_y_0"), val = bool(false)]; + tensor var_3254_cast_fp16 = matmul(transpose_x = var_3254_transpose_x_0, transpose_y = var_3254_transpose_y_0, x = var_3253_cast_fp16, y = state_189_cast_fp16)[name = string("op_3254_cast_fp16")]; + tensor v_new_189_cast_fp16 = sub(x = var_3251_cast_fp16, y = var_3254_cast_fp16)[name = string("v_new_189_cast_fp16")]; + tensor var_3257_begin_0 = const()[name = string("op_3257_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_3257_end_0 = const()[name = string("op_3257_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_3257_end_mask_0 = const()[name = string("op_3257_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3257_squeeze_mask_0 = const()[name = string("op_3257_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3257_cast_fp16 = slice_by_index(begin = var_3257_begin_0, end = var_3257_end_0, end_mask = var_3257_end_mask_0, squeeze_mask = var_3257_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3257_cast_fp16")]; + bool var_3258_transpose_x_0 = const()[name = string("op_3258_transpose_x_0"), val = bool(false)]; + bool var_3258_transpose_y_0 = const()[name = string("op_3258_transpose_y_0"), val = bool(false)]; + tensor var_3258_cast_fp16 = matmul(transpose_x = var_3258_transpose_x_0, transpose_y = var_3258_transpose_y_0, x = var_3257_cast_fp16, y = state_189_cast_fp16)[name = string("op_3258_cast_fp16")]; + tensor var_3260_begin_0 = const()[name = string("op_3260_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_3260_end_0 = const()[name = string("op_3260_end_0"), val = tensor([32, 31, 64, 64])]; + tensor var_3260_end_mask_0 = const()[name = string("op_3260_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3260_squeeze_mask_0 = const()[name = string("op_3260_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3260_cast_fp16 = slice_by_index(begin = var_3260_begin_0, end = var_3260_end_0, end_mask = var_3260_end_mask_0, squeeze_mask = var_3260_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3260_cast_fp16")]; + bool var_3261_transpose_x_0 = const()[name = string("op_3261_transpose_x_0"), val = bool(false)]; + bool var_3261_transpose_y_0 = const()[name = string("op_3261_transpose_y_0"), val = bool(false)]; + tensor var_3261_cast_fp16 = matmul(transpose_x = var_3261_transpose_x_0, transpose_y = var_3261_transpose_y_0, x = var_3260_cast_fp16, y = v_new_189_cast_fp16)[name = string("op_3261_cast_fp16")]; + tensor var_3262_cast_fp16 = add(x = var_3258_cast_fp16, y = var_3261_cast_fp16)[name = string("op_3262_cast_fp16")]; + tensor var_3264_begin_0 = const()[name = string("op_3264_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_3264_end_0 = const()[name = string("op_3264_end_0"), val = tensor([32, 31, 1, 1])]; + tensor var_3264_end_mask_0 = const()[name = string("op_3264_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3264_squeeze_mask_0 = const()[name = string("op_3264_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3264_cast_fp16 = slice_by_index(begin = var_3264_begin_0, end = var_3264_end_0, end_mask = var_3264_end_mask_0, squeeze_mask = var_3264_squeeze_mask_0, x = chunk_decay_5_cast_fp16)[name = string("op_3264_cast_fp16")]; + tensor var_3265_cast_fp16 = mul(x = state_189_cast_fp16, y = var_3264_cast_fp16)[name = string("op_3265_cast_fp16")]; + tensor var_3267_begin_0 = const()[name = string("op_3267_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_3267_end_0 = const()[name = string("op_3267_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_3267_end_mask_0 = const()[name = string("op_3267_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3267_squeeze_mask_0 = const()[name = string("op_3267_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3267_cast_fp16 = slice_by_index(begin = var_3267_begin_0, end = var_3267_end_0, end_mask = var_3267_end_mask_0, squeeze_mask = var_3267_squeeze_mask_0, x = k_dec_5_cast_fp16)[name = string("op_3267_cast_fp16")]; + bool var_3269_transpose_x_1 = const()[name = string("op_3269_transpose_x_1"), val = bool(true)]; + bool var_3269_transpose_y_1 = const()[name = string("op_3269_transpose_y_1"), val = bool(false)]; + tensor var_3269_cast_fp16 = matmul(transpose_x = var_3269_transpose_x_1, transpose_y = var_3269_transpose_y_1, x = var_3267_cast_fp16, y = v_new_189_cast_fp16)[name = string("op_3269_cast_fp16")]; + tensor state_191_cast_fp16 = add(x = var_3265_cast_fp16, y = var_3269_cast_fp16)[name = string("state_191_cast_fp16")]; + tensor var_3272_begin_0 = const()[name = string("op_3272_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_3272_end_0 = const()[name = string("op_3272_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_3272_end_mask_0 = const()[name = string("op_3272_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3272_squeeze_mask_0 = const()[name = string("op_3272_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3272_cast_fp16 = slice_by_index(begin = var_3272_begin_0, end = var_3272_end_0, end_mask = var_3272_end_mask_0, squeeze_mask = var_3272_squeeze_mask_0, x = new_v_5_cast_fp16)[name = string("op_3272_cast_fp16")]; + tensor var_3274_begin_0 = const()[name = string("op_3274_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_3274_end_0 = const()[name = string("op_3274_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_3274_end_mask_0 = const()[name = string("op_3274_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3274_squeeze_mask_0 = const()[name = string("op_3274_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3274_cast_fp16 = slice_by_index(begin = var_3274_begin_0, end = var_3274_end_0, end_mask = var_3274_end_mask_0, squeeze_mask = var_3274_squeeze_mask_0, x = k_cumdecay_5_cast_fp16)[name = string("op_3274_cast_fp16")]; + bool var_3275_transpose_x_0 = const()[name = string("op_3275_transpose_x_0"), val = bool(false)]; + bool var_3275_transpose_y_0 = const()[name = string("op_3275_transpose_y_0"), val = bool(false)]; + tensor var_3275_cast_fp16 = matmul(transpose_x = var_3275_transpose_x_0, transpose_y = var_3275_transpose_y_0, x = var_3274_cast_fp16, y = state_191_cast_fp16)[name = string("op_3275_cast_fp16")]; + tensor v_new_191_cast_fp16 = sub(x = var_3272_cast_fp16, y = var_3275_cast_fp16)[name = string("v_new_191_cast_fp16")]; + tensor var_3278_begin_0 = const()[name = string("op_3278_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_3278_end_0 = const()[name = string("op_3278_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_3278_end_mask_0 = const()[name = string("op_3278_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3278_squeeze_mask_0 = const()[name = string("op_3278_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3278_cast_fp16 = slice_by_index(begin = var_3278_begin_0, end = var_3278_end_0, end_mask = var_3278_end_mask_0, squeeze_mask = var_3278_squeeze_mask_0, x = q_dec_5_cast_fp16)[name = string("op_3278_cast_fp16")]; + bool var_3279_transpose_x_0 = const()[name = string("op_3279_transpose_x_0"), val = bool(false)]; + bool var_3279_transpose_y_0 = const()[name = string("op_3279_transpose_y_0"), val = bool(false)]; + tensor var_3279_cast_fp16 = matmul(transpose_x = var_3279_transpose_x_0, transpose_y = var_3279_transpose_y_0, x = var_3278_cast_fp16, y = state_191_cast_fp16)[name = string("op_3279_cast_fp16")]; + tensor var_3281_begin_0 = const()[name = string("op_3281_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_3281_end_0 = const()[name = string("op_3281_end_0"), val = tensor([32, 32, 64, 64])]; + tensor var_3281_end_mask_0 = const()[name = string("op_3281_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3281_squeeze_mask_0 = const()[name = string("op_3281_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3281_cast_fp16 = slice_by_index(begin = var_3281_begin_0, end = var_3281_end_0, end_mask = var_3281_end_mask_0, squeeze_mask = var_3281_squeeze_mask_0, x = attn_intra_5_cast_fp16)[name = string("op_3281_cast_fp16")]; + bool var_3282_transpose_x_0 = const()[name = string("op_3282_transpose_x_0"), val = bool(false)]; + bool var_3282_transpose_y_0 = const()[name = string("op_3282_transpose_y_0"), val = bool(false)]; + tensor var_3282_cast_fp16 = matmul(transpose_x = var_3282_transpose_x_0, transpose_y = var_3282_transpose_y_0, x = var_3281_cast_fp16, y = v_new_191_cast_fp16)[name = string("op_3282_cast_fp16")]; + tensor var_3283_cast_fp16 = add(x = var_3279_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")]; + int32 var_3285_axis_0 = const()[name = string("op_3285_axis_0"), val = int32(1)]; + tensor var_3285_cast_fp16 = stack(axis = var_3285_axis_0, values = (var_2632_cast_fp16, var_2653_cast_fp16, var_2674_cast_fp16, var_2695_cast_fp16, var_2716_cast_fp16, var_2737_cast_fp16, var_2758_cast_fp16, var_2779_cast_fp16, var_2800_cast_fp16, var_2821_cast_fp16, var_2842_cast_fp16, var_2863_cast_fp16, var_2884_cast_fp16, var_2905_cast_fp16, var_2926_cast_fp16, var_2947_cast_fp16, var_2968_cast_fp16, var_2989_cast_fp16, var_3010_cast_fp16, var_3031_cast_fp16, var_3052_cast_fp16, var_3073_cast_fp16, var_3094_cast_fp16, var_3115_cast_fp16, var_3136_cast_fp16, var_3157_cast_fp16, var_3178_cast_fp16, var_3199_cast_fp16, var_3220_cast_fp16, var_3241_cast_fp16, var_3262_cast_fp16, var_3283_cast_fp16))[name = string("op_3285_cast_fp16")]; + tensor var_3286 = const()[name = string("op_3286"), val = tensor([32, 2048, 128])]; + tensor var_3287_cast_fp16 = reshape(shape = var_3286, x = var_3285_cast_fp16)[name = string("op_3287_cast_fp16")]; + tensor var_3288 = const()[name = string("op_3288"), val = tensor([1, 0, 2])]; + tensor input_39_begin_0 = const()[name = string("input_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_end_0 = const()[name = string("input_39_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor input_39_end_mask_0 = const()[name = string("input_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor input_39_squeeze_mask_0 = const()[name = string("input_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor input_39_cast_fp16 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, squeeze_mask = input_39_squeeze_mask_0, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")]; + fp16 var_2268_promoted_1_to_fp16 = const()[name = string("op_2268_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor x_69_cast_fp16 = transpose(perm = var_3288, x = var_3287_cast_fp16)[name = string("transpose_32")]; + tensor var_3292_cast_fp16 = pow(x = x_69_cast_fp16, y = var_2268_promoted_1_to_fp16)[name = string("op_3292_cast_fp16")]; + tensor var_3294_axes_0 = const()[name = string("op_3294_axes_0"), val = tensor([-1])]; + bool var_3294_keep_dims_0 = const()[name = string("op_3294_keep_dims_0"), val = bool(true)]; + tensor var_3294_cast_fp16 = reduce_mean(axes = var_3294_axes_0, keep_dims = var_3294_keep_dims_0, x = var_3292_cast_fp16)[name = string("op_3294_cast_fp16")]; + fp16 var_3295_to_fp16 = const()[name = string("op_3295_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3296_cast_fp16 = add(x = var_3294_cast_fp16, y = var_3295_to_fp16)[name = string("op_3296_cast_fp16")]; + fp32 var_3297_epsilon_0 = const()[name = string("op_3297_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3297_cast_fp16 = rsqrt(epsilon = var_3297_epsilon_0, x = var_3296_cast_fp16)[name = string("op_3297_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3297_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor layers_2_linear_attn_norm_weight_to_fp16 = const()[name = string("layers_2_linear_attn_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263951808)))]; + tensor var_3299_cast_fp16 = mul(x = layers_2_linear_attn_norm_weight_to_fp16, y = x_71_cast_fp16)[name = string("op_3299_cast_fp16")]; + tensor var_3300_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_3300_cast_fp16")]; + tensor core_5_cast_fp16 = mul(x = var_3299_cast_fp16, y = var_3300_cast_fp16)[name = string("core_5_cast_fp16")]; + tensor var_3302 = const()[name = string("op_3302"), val = tensor([1, 2048, 4096])]; + tensor input_41_cast_fp16 = reshape(shape = var_3302, x = core_5_cast_fp16)[name = string("input_41_cast_fp16")]; + tensor layers_2_linear_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263952128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274437952))))[name = string("layers_2_linear_attn_out_proj_weight_to_fp16_quantized")]; + tensor linear_20_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_linear_attn_out_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_20_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_51_cast_fp16, y = linear_20_cast_fp16)[name = string("x_73_cast_fp16")]; + fp16 var_2268_promoted_2_to_fp16 = const()[name = string("op_2268_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor var_3308_cast_fp16 = pow(x = x_73_cast_fp16, y = var_2268_promoted_2_to_fp16)[name = string("op_3308_cast_fp16")]; + tensor var_3310_axes_0 = const()[name = string("op_3310_axes_0"), val = tensor([-1])]; + bool var_3310_keep_dims_0 = const()[name = string("op_3310_keep_dims_0"), val = bool(true)]; + tensor var_3310_cast_fp16 = reduce_mean(axes = var_3310_axes_0, keep_dims = var_3310_keep_dims_0, x = var_3308_cast_fp16)[name = string("op_3310_cast_fp16")]; + fp16 var_3311_to_fp16 = const()[name = string("op_3311_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3312_cast_fp16 = add(x = var_3310_cast_fp16, y = var_3311_to_fp16)[name = string("op_3312_cast_fp16")]; + fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3313_cast_fp16 = rsqrt(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")]; + tensor x_75_cast_fp16 = mul(x = x_73_cast_fp16, y = var_3313_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor var_3316_to_fp16 = const()[name = string("op_3316_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274443136)))]; + tensor input_43_cast_fp16 = mul(x = x_75_cast_fp16, y = var_3316_to_fp16)[name = string("input_43_cast_fp16")]; + tensor layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274448320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298041344))))[name = string("layers_2_mlp_gate_proj_weight_to_fp16_quantized")]; + tensor linear_21_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_3323_cast_fp16 = silu(x = linear_21_cast_fp16)[name = string("op_3323_cast_fp16")]; + tensor layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298059840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321652864))))[name = string("layers_2_mlp_up_proj_weight_to_fp16_quantized")]; + tensor linear_22_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor input_47_cast_fp16 = mul(x = var_3323_cast_fp16, y = linear_22_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321671360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345264384))))[name = string("layers_2_mlp_down_proj_weight_to_fp16_quantized")]; + tensor linear_23_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = x_73_cast_fp16, y = linear_23_cast_fp16)[name = string("x_77_cast_fp16")]; + int32 var_3343 = const()[name = string("op_3343"), val = int32(-1)]; + fp16 var_3342_promoted_to_fp16 = const()[name = string("op_3342_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_3352_cast_fp16 = pow(x = x_77_cast_fp16, y = var_3342_promoted_to_fp16)[name = string("op_3352_cast_fp16")]; + tensor var_3354_axes_0 = const()[name = string("op_3354_axes_0"), val = tensor([-1])]; + bool var_3354_keep_dims_0 = const()[name = string("op_3354_keep_dims_0"), val = bool(true)]; + tensor var_3354_cast_fp16 = reduce_mean(axes = var_3354_axes_0, keep_dims = var_3354_keep_dims_0, x = var_3352_cast_fp16)[name = string("op_3354_cast_fp16")]; + fp16 var_3355_to_fp16 = const()[name = string("op_3355_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3356_cast_fp16 = add(x = var_3354_cast_fp16, y = var_3355_to_fp16)[name = string("op_3356_cast_fp16")]; + fp32 var_3357_epsilon_0 = const()[name = string("op_3357_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3357_cast_fp16 = rsqrt(epsilon = var_3357_epsilon_0, x = var_3356_cast_fp16)[name = string("op_3357_cast_fp16")]; + tensor x_79_cast_fp16 = mul(x = x_77_cast_fp16, y = var_3357_cast_fp16)[name = string("x_79_cast_fp16")]; + tensor var_3360_to_fp16 = const()[name = string("op_3360_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345269568)))]; + tensor x_81_cast_fp16 = mul(x = x_79_cast_fp16, y = var_3360_to_fp16)[name = string("x_81_cast_fp16")]; + tensor layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345274752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366246336))))[name = string("layers_3_self_attn_q_proj_weight_to_fp16_quantized")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = x_81_cast_fp16)[name = string("linear_24_cast_fp16")]; + tensor var_3371 = const()[name = string("op_3371"), val = tensor([1, 2048, 16, 512])]; + tensor qg_1_cast_fp16 = reshape(shape = var_3371, x = linear_24_cast_fp16)[name = string("qg_1_cast_fp16")]; + tensor x_83_begin_0 = const()[name = string("x_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_83_end_0 = const()[name = string("x_83_end_0"), val = tensor([1, 2048, 16, 256])]; + tensor x_83_end_mask_0 = const()[name = string("x_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, x = qg_1_cast_fp16)[name = string("x_83_cast_fp16")]; + tensor gate_1_begin_0 = const()[name = string("gate_1_begin_0"), val = tensor([0, 0, 0, 256])]; + tensor gate_1_end_0 = const()[name = string("gate_1_end_0"), val = tensor([1, 2048, 16, 512])]; + tensor gate_1_end_mask_0 = const()[name = string("gate_1_end_mask_0"), val = tensor([true, true, true, true])]; + tensor gate_1_cast_fp16 = slice_by_index(begin = gate_1_begin_0, end = gate_1_end_0, end_mask = gate_1_end_mask_0, x = qg_1_cast_fp16)[name = string("gate_1_cast_fp16")]; + tensor var_3375 = const()[name = string("op_3375"), val = tensor([1, 2048, 4096])]; + tensor gate_3_cast_fp16 = reshape(shape = var_3375, x = gate_1_cast_fp16)[name = string("gate_3_cast_fp16")]; + fp16 var_3342_promoted_1_to_fp16 = const()[name = string("op_3342_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor var_3378_cast_fp16 = pow(x = x_83_cast_fp16, y = var_3342_promoted_1_to_fp16)[name = string("op_3378_cast_fp16")]; + tensor var_3380_axes_0 = const()[name = string("op_3380_axes_0"), val = tensor([-1])]; + bool var_3380_keep_dims_0 = const()[name = string("op_3380_keep_dims_0"), val = bool(true)]; + tensor var_3380_cast_fp16 = reduce_mean(axes = var_3380_axes_0, keep_dims = var_3380_keep_dims_0, x = var_3378_cast_fp16)[name = string("op_3380_cast_fp16")]; + fp16 var_3381_to_fp16 = const()[name = string("op_3381_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3382_cast_fp16 = add(x = var_3380_cast_fp16, y = var_3381_to_fp16)[name = string("op_3382_cast_fp16")]; + fp32 var_3383_epsilon_0 = const()[name = string("op_3383_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3383_cast_fp16 = rsqrt(epsilon = var_3383_epsilon_0, x = var_3382_cast_fp16)[name = string("op_3383_cast_fp16")]; + tensor x_85_cast_fp16 = mul(x = x_83_cast_fp16, y = var_3383_cast_fp16)[name = string("x_85_cast_fp16")]; + tensor var_3386_to_fp16 = const()[name = string("op_3386_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366262784)))]; + tensor var_3387_cast_fp16 = mul(x = x_85_cast_fp16, y = var_3386_to_fp16)[name = string("op_3387_cast_fp16")]; + tensor q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366263360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368884864))))[name = string("layers_3_self_attn_k_proj_weight_to_fp16_quantized")]; + tensor linear_25_bias_0_to_fp16 = const()[name = string("linear_25_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368886976)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_25_bias_0_to_fp16, weight = layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = x_81_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor var_3391 = const()[name = string("op_3391"), val = tensor([1, 2048, 4, 256])]; + tensor x_87_cast_fp16 = reshape(shape = var_3391, x = linear_25_cast_fp16)[name = string("x_87_cast_fp16")]; + fp16 var_3342_promoted_2_to_fp16 = const()[name = string("op_3342_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor var_3394_cast_fp16 = pow(x = x_87_cast_fp16, y = var_3342_promoted_2_to_fp16)[name = string("op_3394_cast_fp16")]; + tensor var_3396_axes_0 = const()[name = string("op_3396_axes_0"), val = tensor([-1])]; + bool var_3396_keep_dims_0 = const()[name = string("op_3396_keep_dims_0"), val = bool(true)]; + tensor var_3396_cast_fp16 = reduce_mean(axes = var_3396_axes_0, keep_dims = var_3396_keep_dims_0, x = var_3394_cast_fp16)[name = string("op_3396_cast_fp16")]; + fp16 var_3397_to_fp16 = const()[name = string("op_3397_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3398_cast_fp16 = add(x = var_3396_cast_fp16, y = var_3397_to_fp16)[name = string("op_3398_cast_fp16")]; + fp32 var_3399_epsilon_0 = const()[name = string("op_3399_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3399_cast_fp16 = rsqrt(epsilon = var_3399_epsilon_0, x = var_3398_cast_fp16)[name = string("op_3399_cast_fp16")]; + tensor x_89_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3399_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_3402_to_fp16 = const()[name = string("op_3402_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368889088)))]; + tensor var_3403_cast_fp16 = mul(x = x_89_cast_fp16, y = var_3402_to_fp16)[name = string("op_3403_cast_fp16")]; + tensor k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368889664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371511168))))[name = string("layers_3_self_attn_v_proj_weight_to_fp16_quantized")]; + tensor linear_26_cast_fp16 = linear(bias = linear_25_bias_0_to_fp16, weight = layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = x_81_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor var_3407 = const()[name = string("op_3407"), val = tensor([1, 2048, 4, 256])]; + tensor var_3408_cast_fp16 = reshape(shape = var_3407, x = linear_26_cast_fp16)[name = string("op_3408_cast_fp16")]; + tensor v_13_perm_0 = const()[name = string("v_13_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3410_axes_0 = const()[name = string("op_3410_axes_0"), val = tensor([0])]; + tensor var_3410_cast_fp16 = expand_dims(axes = var_3410_axes_0, x = cos)[name = string("op_3410_cast_fp16")]; + tensor c_1_axes_0 = const()[name = string("c_1_axes_0"), val = tensor([1])]; + tensor c_1_cast_fp16 = expand_dims(axes = c_1_axes_0, x = var_3410_cast_fp16)[name = string("c_1_cast_fp16")]; + tensor var_3412_axes_0 = const()[name = string("op_3412_axes_0"), val = tensor([0])]; + tensor var_3412_cast_fp16 = expand_dims(axes = var_3412_axes_0, x = sin)[name = string("op_3412_cast_fp16")]; + tensor s_1_axes_0 = const()[name = string("s_1_axes_0"), val = tensor([1])]; + tensor s_1_cast_fp16 = expand_dims(axes = s_1_axes_0, x = var_3412_cast_fp16)[name = string("s_1_cast_fp16")]; + tensor var_3414_begin_0 = const()[name = string("op_3414_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3414_end_0 = const()[name = string("op_3414_end_0"), val = tensor([1, 16, 2048, 64])]; + tensor var_3414_end_mask_0 = const()[name = string("op_3414_end_mask_0"), val = tensor([true, true, true, false])]; + tensor q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_3387_cast_fp16)[name = string("transpose_31")]; + tensor var_3414_cast_fp16 = slice_by_index(begin = var_3414_begin_0, end = var_3414_end_0, end_mask = var_3414_end_mask_0, x = q_31_cast_fp16)[name = string("op_3414_cast_fp16")]; + tensor var_3415_cast_fp16 = mul(x = var_3414_cast_fp16, y = c_1_cast_fp16)[name = string("op_3415_cast_fp16")]; + tensor var_3422_begin_0 = const()[name = string("op_3422_begin_0"), val = tensor([0, 0, 0, 32])]; + tensor var_3422_end_0 = const()[name = string("op_3422_end_0"), val = tensor([1, 16, 2048, 64])]; + tensor var_3422_end_mask_0 = const()[name = string("op_3422_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3422_cast_fp16 = slice_by_index(begin = var_3422_begin_0, end = var_3422_end_0, end_mask = var_3422_end_mask_0, x = var_3414_cast_fp16)[name = string("op_3422_cast_fp16")]; + fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor var_3423_cast_fp16 = mul(x = var_3422_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_3423_cast_fp16")]; + tensor var_3424_begin_0 = const()[name = string("op_3424_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3424_end_0 = const()[name = string("op_3424_end_0"), val = tensor([1, 16, 2048, 32])]; + tensor var_3424_end_mask_0 = const()[name = string("op_3424_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3424_cast_fp16 = slice_by_index(begin = var_3424_begin_0, end = var_3424_end_0, end_mask = var_3424_end_mask_0, x = var_3414_cast_fp16)[name = string("op_3424_cast_fp16")]; + bool var_3426_interleave_0 = const()[name = string("op_3426_interleave_0"), val = bool(false)]; + tensor var_3426_cast_fp16 = concat(axis = var_3343, interleave = var_3426_interleave_0, values = (var_3423_cast_fp16, var_3424_cast_fp16))[name = string("op_3426_cast_fp16")]; + tensor var_3427_cast_fp16 = mul(x = var_3426_cast_fp16, y = s_1_cast_fp16)[name = string("op_3427_cast_fp16")]; + tensor var_3428_cast_fp16 = add(x = var_3415_cast_fp16, y = var_3427_cast_fp16)[name = string("op_3428_cast_fp16")]; + tensor var_3429_begin_0 = const()[name = string("op_3429_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor var_3429_end_0 = const()[name = string("op_3429_end_0"), val = tensor([1, 16, 2048, 256])]; + tensor var_3429_end_mask_0 = const()[name = string("op_3429_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3429_cast_fp16 = slice_by_index(begin = var_3429_begin_0, end = var_3429_end_0, end_mask = var_3429_end_mask_0, x = q_31_cast_fp16)[name = string("op_3429_cast_fp16")]; + bool q_33_interleave_0 = const()[name = string("q_33_interleave_0"), val = bool(false)]; + tensor q_33_cast_fp16 = concat(axis = var_3343, interleave = q_33_interleave_0, values = (var_3428_cast_fp16, var_3429_cast_fp16))[name = string("q_33_cast_fp16")]; + tensor var_3432_begin_0 = const()[name = string("op_3432_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3432_end_0 = const()[name = string("op_3432_end_0"), val = tensor([1, 4, 2048, 64])]; + tensor var_3432_end_mask_0 = const()[name = string("op_3432_end_mask_0"), val = tensor([true, true, true, false])]; + tensor k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_3403_cast_fp16)[name = string("transpose_30")]; + tensor var_3432_cast_fp16 = slice_by_index(begin = var_3432_begin_0, end = var_3432_end_0, end_mask = var_3432_end_mask_0, x = k_25_cast_fp16)[name = string("op_3432_cast_fp16")]; + tensor var_3433_cast_fp16 = mul(x = var_3432_cast_fp16, y = c_1_cast_fp16)[name = string("op_3433_cast_fp16")]; + tensor var_3440_begin_0 = const()[name = string("op_3440_begin_0"), val = tensor([0, 0, 0, 32])]; + tensor var_3440_end_0 = const()[name = string("op_3440_end_0"), val = tensor([1, 4, 2048, 64])]; + tensor var_3440_end_mask_0 = const()[name = string("op_3440_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3440_cast_fp16 = slice_by_index(begin = var_3440_begin_0, end = var_3440_end_0, end_mask = var_3440_end_mask_0, x = var_3432_cast_fp16)[name = string("op_3440_cast_fp16")]; + fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor var_3441_cast_fp16 = mul(x = var_3440_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_3441_cast_fp16")]; + tensor var_3442_begin_0 = const()[name = string("op_3442_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3442_end_0 = const()[name = string("op_3442_end_0"), val = tensor([1, 4, 2048, 32])]; + tensor var_3442_end_mask_0 = const()[name = string("op_3442_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3442_cast_fp16 = slice_by_index(begin = var_3442_begin_0, end = var_3442_end_0, end_mask = var_3442_end_mask_0, x = var_3432_cast_fp16)[name = string("op_3442_cast_fp16")]; + bool var_3444_interleave_0 = const()[name = string("op_3444_interleave_0"), val = bool(false)]; + tensor var_3444_cast_fp16 = concat(axis = var_3343, interleave = var_3444_interleave_0, values = (var_3441_cast_fp16, var_3442_cast_fp16))[name = string("op_3444_cast_fp16")]; + tensor var_3445_cast_fp16 = mul(x = var_3444_cast_fp16, y = s_1_cast_fp16)[name = string("op_3445_cast_fp16")]; + tensor var_3446_cast_fp16 = add(x = var_3433_cast_fp16, y = var_3445_cast_fp16)[name = string("op_3446_cast_fp16")]; + tensor var_3447_begin_0 = const()[name = string("op_3447_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor var_3447_end_0 = const()[name = string("op_3447_end_0"), val = tensor([1, 4, 2048, 256])]; + tensor var_3447_end_mask_0 = const()[name = string("op_3447_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3447_cast_fp16 = slice_by_index(begin = var_3447_begin_0, end = var_3447_end_0, end_mask = var_3447_end_mask_0, x = k_25_cast_fp16)[name = string("op_3447_cast_fp16")]; + bool k_27_interleave_0 = const()[name = string("k_27_interleave_0"), val = bool(false)]; + tensor k_27_cast_fp16 = concat(axis = var_3343, interleave = k_27_interleave_0, values = (var_3446_cast_fp16, var_3447_cast_fp16))[name = string("k_27_cast_fp16")]; + tensor var_3452_axes_0 = const()[name = string("op_3452_axes_0"), val = tensor([2])]; + tensor var_3452_cast_fp16 = expand_dims(axes = var_3452_axes_0, x = k_27_cast_fp16)[name = string("op_3452_cast_fp16")]; + tensor var_3454_reps_0 = const()[name = string("op_3454_reps_0"), val = tensor([1, 1, 4, 1, 1])]; + tensor var_3454_cast_fp16 = tile(reps = var_3454_reps_0, x = var_3452_cast_fp16)[name = string("op_3454_cast_fp16")]; + tensor var_3455 = const()[name = string("op_3455"), val = tensor([1, 16, 2048, 256])]; + tensor k_29_cast_fp16 = reshape(shape = var_3455, x = var_3454_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor var_3459_axes_0 = const()[name = string("op_3459_axes_0"), val = tensor([2])]; + tensor v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_3408_cast_fp16)[name = string("transpose_29")]; + tensor var_3459_cast_fp16 = expand_dims(axes = var_3459_axes_0, x = v_13_cast_fp16)[name = string("op_3459_cast_fp16")]; + tensor var_3461_reps_0 = const()[name = string("op_3461_reps_0"), val = tensor([1, 1, 4, 1, 1])]; + tensor var_3461_cast_fp16 = tile(reps = var_3461_reps_0, x = var_3459_cast_fp16)[name = string("op_3461_cast_fp16")]; + tensor var_3462 = const()[name = string("op_3462"), val = tensor([1, 16, 2048, 256])]; + tensor v_15_cast_fp16 = reshape(shape = var_3462, x = var_3461_cast_fp16)[name = string("v_15_cast_fp16")]; + bool var_3465_transpose_x_1 = const()[name = string("op_3465_transpose_x_1"), val = bool(false)]; + bool var_3465_transpose_y_1 = const()[name = string("op_3465_transpose_y_1"), val = bool(true)]; + tensor var_3465_cast_fp16 = matmul(transpose_x = var_3465_transpose_x_1, transpose_y = var_3465_transpose_y_1, x = q_33_cast_fp16, y = k_29_cast_fp16)[name = string("op_3465_cast_fp16")]; + fp16 var_3466_to_fp16 = const()[name = string("op_3466_to_fp16"), val = fp16(0x1p-4)]; + tensor var_3467_cast_fp16 = mul(x = var_3465_cast_fp16, y = var_3466_to_fp16)[name = string("op_3467_cast_fp16")]; + tensor mask_to_fp16 = const()[name = string("mask_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371513280)))]; + tensor scores_1_cast_fp16 = add(x = var_3467_cast_fp16, y = mask_to_fp16)[name = string("scores_1_cast_fp16")]; + tensor var_3469_cast_fp16 = softmax(axis = var_3343, x = scores_1_cast_fp16)[name = string("op_3469_cast_fp16")]; + bool out_1_transpose_x_0 = const()[name = string("out_1_transpose_x_0"), val = bool(false)]; + bool out_1_transpose_y_0 = const()[name = string("out_1_transpose_y_0"), val = bool(false)]; + tensor out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = var_3469_cast_fp16, y = v_15_cast_fp16)[name = string("out_1_cast_fp16")]; + tensor var_3471_perm_0 = const()[name = string("op_3471_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3472 = const()[name = string("op_3472"), val = tensor([1, 2048, 4096])]; + tensor var_3471_cast_fp16 = transpose(perm = var_3471_perm_0, x = out_1_cast_fp16)[name = string("transpose_28")]; + tensor out_3_cast_fp16 = reshape(shape = var_3472, x = var_3471_cast_fp16)[name = string("out_3_cast_fp16")]; + tensor var_3474_cast_fp16 = sigmoid(x = gate_3_cast_fp16)[name = string("op_3474_cast_fp16")]; + tensor input_49_cast_fp16 = mul(x = out_3_cast_fp16, y = var_3474_cast_fp16)[name = string("input_49_cast_fp16")]; + tensor layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379901952))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390387776))))[name = string("layers_3_self_attn_o_proj_weight_to_fp16_quantized")]; + tensor linear_27_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_27_cast_fp16")]; + tensor x_95_cast_fp16 = add(x = x_77_cast_fp16, y = linear_27_cast_fp16)[name = string("x_95_cast_fp16")]; + fp16 var_3342_promoted_3_to_fp16 = const()[name = string("op_3342_promoted_3_to_fp16"), val = fp16(0x1p+1)]; + tensor var_3480_cast_fp16 = pow(x = x_95_cast_fp16, y = var_3342_promoted_3_to_fp16)[name = string("op_3480_cast_fp16")]; + tensor var_3482_axes_0 = const()[name = string("op_3482_axes_0"), val = tensor([-1])]; + bool var_3482_keep_dims_0 = const()[name = string("op_3482_keep_dims_0"), val = bool(true)]; + tensor var_3482_cast_fp16 = reduce_mean(axes = var_3482_axes_0, keep_dims = var_3482_keep_dims_0, x = var_3480_cast_fp16)[name = string("op_3482_cast_fp16")]; + fp16 var_3483_to_fp16 = const()[name = string("op_3483_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3484_cast_fp16 = add(x = var_3482_cast_fp16, y = var_3483_to_fp16)[name = string("op_3484_cast_fp16")]; + fp32 var_3485_epsilon_0 = const()[name = string("op_3485_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3485_cast_fp16 = rsqrt(epsilon = var_3485_epsilon_0, x = var_3484_cast_fp16)[name = string("op_3485_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_95_cast_fp16, y = var_3485_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor var_3488_to_fp16 = const()[name = string("op_3488_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390392960)))]; + tensor input_51_cast_fp16 = mul(x = x_97_cast_fp16, y = var_3488_to_fp16)[name = string("input_51_cast_fp16")]; + tensor layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390398144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413991168))))[name = string("layers_3_mlp_gate_proj_weight_to_fp16_quantized")]; + tensor linear_28_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor var_3495_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("op_3495_cast_fp16")]; + tensor layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414009664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437602688))))[name = string("layers_3_mlp_up_proj_weight_to_fp16_quantized")]; + tensor linear_29_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_29_cast_fp16")]; + tensor input_55_cast_fp16 = mul(x = var_3495_cast_fp16, y = linear_29_cast_fp16)[name = string("input_55_cast_fp16")]; + tensor layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437621184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461214208))))[name = string("layers_3_mlp_down_proj_weight_to_fp16_quantized")]; + tensor linear_30_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = x_95_cast_fp16, y = linear_30_cast_fp16)[name = string("x_99_cast_fp16")]; + int32 var_3515 = const()[name = string("op_3515"), val = int32(-2)]; + int32 var_3544 = const()[name = string("op_3544"), val = int32(-1)]; + fp16 var_3543_promoted_to_fp16 = const()[name = string("op_3543_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_3553_cast_fp16 = pow(x = x_99_cast_fp16, y = var_3543_promoted_to_fp16)[name = string("op_3553_cast_fp16")]; + tensor var_3555_axes_0 = const()[name = string("op_3555_axes_0"), val = tensor([-1])]; + bool var_3555_keep_dims_0 = const()[name = string("op_3555_keep_dims_0"), val = bool(true)]; + tensor var_3555_cast_fp16 = reduce_mean(axes = var_3555_axes_0, keep_dims = var_3555_keep_dims_0, x = var_3553_cast_fp16)[name = string("op_3555_cast_fp16")]; + fp16 var_3556_to_fp16 = const()[name = string("op_3556_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3557_cast_fp16 = add(x = var_3555_cast_fp16, y = var_3556_to_fp16)[name = string("op_3557_cast_fp16")]; + fp32 var_3558_epsilon_0 = const()[name = string("op_3558_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3558_cast_fp16 = rsqrt(epsilon = var_3558_epsilon_0, x = var_3557_cast_fp16)[name = string("op_3558_cast_fp16")]; + tensor x_101_cast_fp16 = mul(x = x_99_cast_fp16, y = var_3558_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor var_3561_to_fp16 = const()[name = string("op_3561_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461219392)))]; + tensor x_103_cast_fp16 = mul(x = x_101_cast_fp16, y = var_3561_to_fp16)[name = string("x_103_cast_fp16")]; + tensor layers_4_linear_attn_in_proj_qkv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461224576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482196160))))[name = string("layers_4_linear_attn_in_proj_qkv_weight_to_fp16_quantized")]; + tensor linear_31_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_linear_attn_in_proj_qkv_weight_to_fp16_quantized, x = x_103_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_59_pad_0 = const()[name = string("input_59_pad_0"), val = tensor([0, 0, 0, 0, 3, 0])]; + string input_59_mode_0 = const()[name = string("input_59_mode_0"), val = string("constant")]; + fp16 const_42_to_fp16 = const()[name = string("const_42_to_fp16"), val = fp16(0x0p+0)]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = linear_31_cast_fp16)[name = string("transpose_27")]; + tensor input_59_cast_fp16 = pad(constant_val = const_42_to_fp16, mode = input_59_mode_0, pad = input_59_pad_0, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string input_61_pad_type_0 = const()[name = string("input_61_pad_type_0"), val = string("valid")]; + int32 input_61_groups_0 = const()[name = string("input_61_groups_0"), val = int32(8192)]; + tensor input_61_strides_0 = const()[name = string("input_61_strides_0"), val = tensor([1])]; + tensor input_61_pad_0 = const()[name = string("input_61_pad_0"), val = tensor([0, 0])]; + tensor input_61_dilations_0 = const()[name = string("input_61_dilations_0"), val = tensor([1])]; + tensor layers_4_linear_attn_conv1d_weight_to_fp16 = const()[name = string("layers_4_linear_attn_conv1d_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482212608)))]; + tensor input_61_cast_fp16 = conv(dilations = input_61_dilations_0, groups = input_61_groups_0, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = input_61_strides_0, weight = layers_4_linear_attn_conv1d_weight_to_fp16, x = input_59_cast_fp16)[name = string("input_61_cast_fp16")]; + tensor var_3596_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("op_3596_cast_fp16")]; + tensor qkv_7_perm_0 = const()[name = string("qkv_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_3598_begin_0 = const()[name = string("op_3598_begin_0"), val = tensor([0, 0, 0])]; + tensor var_3598_end_0 = const()[name = string("op_3598_end_0"), val = tensor([1, 2048, 2048])]; + tensor var_3598_end_mask_0 = const()[name = string("op_3598_end_mask_0"), val = tensor([true, true, false])]; + tensor qkv_7_cast_fp16 = transpose(perm = qkv_7_perm_0, x = var_3596_cast_fp16)[name = string("transpose_26")]; + tensor var_3598_cast_fp16 = slice_by_index(begin = var_3598_begin_0, end = var_3598_end_0, end_mask = var_3598_end_mask_0, x = qkv_7_cast_fp16)[name = string("op_3598_cast_fp16")]; + tensor var_3599 = const()[name = string("op_3599"), val = tensor([1, 2048, 16, 128])]; + tensor q_35_cast_fp16 = reshape(shape = var_3599, x = var_3598_cast_fp16)[name = string("q_35_cast_fp16")]; + tensor var_3601_begin_0 = const()[name = string("op_3601_begin_0"), val = tensor([0, 0, 2048])]; + tensor var_3601_end_0 = const()[name = string("op_3601_end_0"), val = tensor([1, 2048, 4096])]; + tensor var_3601_end_mask_0 = const()[name = string("op_3601_end_mask_0"), val = tensor([true, true, false])]; + tensor var_3601_cast_fp16 = slice_by_index(begin = var_3601_begin_0, end = var_3601_end_0, end_mask = var_3601_end_mask_0, x = qkv_7_cast_fp16)[name = string("op_3601_cast_fp16")]; + tensor var_3602 = const()[name = string("op_3602"), val = tensor([1, 2048, 16, 128])]; + tensor k_31_cast_fp16 = reshape(shape = var_3602, x = var_3601_cast_fp16)[name = string("k_31_cast_fp16")]; + tensor var_3604_begin_0 = const()[name = string("op_3604_begin_0"), val = tensor([0, 0, 4096])]; + tensor var_3604_end_0 = const()[name = string("op_3604_end_0"), val = tensor([1, 2048, 8192])]; + tensor var_3604_end_mask_0 = const()[name = string("op_3604_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3604_cast_fp16 = slice_by_index(begin = var_3604_begin_0, end = var_3604_end_0, end_mask = var_3604_end_mask_0, x = qkv_7_cast_fp16)[name = string("op_3604_cast_fp16")]; + tensor var_3605 = const()[name = string("op_3605"), val = tensor([1, 2048, 32, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_3605, x = var_3604_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor layers_4_linear_attn_in_proj_z_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482278208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492764032))))[name = string("layers_4_linear_attn_in_proj_z_weight_to_fp16_quantized")]; + tensor linear_32_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_linear_attn_in_proj_z_weight_to_fp16_quantized, x = x_103_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_3609 = const()[name = string("op_3609"), val = tensor([1, 2048, 32, 128])]; + tensor z_7_cast_fp16 = reshape(shape = var_3609, x = linear_32_cast_fp16)[name = string("z_7_cast_fp16")]; + tensor layers_4_linear_attn_in_proj_b_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492772288))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492854272))))[name = string("layers_4_linear_attn_in_proj_b_weight_to_fp16_quantized")]; + tensor linear_33_bias_0_to_fp16 = const()[name = string("linear_33_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492854400)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_33_bias_0_to_fp16, weight = layers_4_linear_attn_in_proj_b_weight_to_fp16_quantized, x = x_103_cast_fp16)[name = string("linear_33_cast_fp16")]; + tensor beta_13_cast_fp16 = sigmoid(x = linear_33_cast_fp16)[name = string("beta_13_cast_fp16")]; + tensor var_3618_weight_0_to_fp16 = const()[name = string("op_3618_weight_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492854528)))]; + tensor var_3618_bias_0_to_fp16 = const()[name = string("op_3618_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493018432)))]; + tensor var_3618_cast_fp16 = linear(bias = var_3618_bias_0_to_fp16, weight = var_3618_weight_0_to_fp16, x = x_103_cast_fp16)[name = string("op_3618_cast_fp16")]; + tensor var_3619_cast_fp16 = softplus(x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")]; + tensor var_3615_to_fp16 = const()[name = string("op_3615_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493018560)))]; + tensor g_13_cast_fp16 = mul(x = var_3615_to_fp16, y = var_3619_cast_fp16)[name = string("g_13_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = q_35_cast_fp16, y = q_35_cast_fp16)[name = string("op_3621_cast_fp16")]; + tensor var_3623_axes_0 = const()[name = string("op_3623_axes_0"), val = tensor([-1])]; + bool var_3623_keep_dims_0 = const()[name = string("op_3623_keep_dims_0"), val = bool(true)]; + tensor var_3623_cast_fp16 = reduce_sum(axes = var_3623_axes_0, keep_dims = var_3623_keep_dims_0, x = var_3621_cast_fp16)[name = string("op_3623_cast_fp16")]; + fp16 var_3624_to_fp16 = const()[name = string("op_3624_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3625_cast_fp16 = add(x = var_3623_cast_fp16, y = var_3624_to_fp16)[name = string("op_3625_cast_fp16")]; + fp32 var_3626_epsilon_0 = const()[name = string("op_3626_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3626_cast_fp16 = rsqrt(epsilon = var_3626_epsilon_0, x = var_3625_cast_fp16)[name = string("op_3626_cast_fp16")]; + tensor q_37_cast_fp16 = mul(x = q_35_cast_fp16, y = var_3626_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor var_3628_cast_fp16 = mul(x = k_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_3628_cast_fp16")]; + tensor var_3630_axes_0 = const()[name = string("op_3630_axes_0"), val = tensor([-1])]; + bool var_3630_keep_dims_0 = const()[name = string("op_3630_keep_dims_0"), val = bool(true)]; + tensor var_3630_cast_fp16 = reduce_sum(axes = var_3630_axes_0, keep_dims = var_3630_keep_dims_0, x = var_3628_cast_fp16)[name = string("op_3630_cast_fp16")]; + fp16 var_3631_to_fp16 = const()[name = string("op_3631_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_3632_cast_fp16 = add(x = var_3630_cast_fp16, y = var_3631_to_fp16)[name = string("op_3632_cast_fp16")]; + fp32 var_3633_epsilon_0 = const()[name = string("op_3633_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3633_cast_fp16 = rsqrt(epsilon = var_3633_epsilon_0, x = var_3632_cast_fp16)[name = string("op_3633_cast_fp16")]; + tensor k_33_cast_fp16 = mul(x = k_31_cast_fp16, y = var_3633_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor var_3638_axes_0 = const()[name = string("op_3638_axes_0"), val = tensor([3])]; + tensor var_3638_cast_fp16 = expand_dims(axes = var_3638_axes_0, x = q_37_cast_fp16)[name = string("op_3638_cast_fp16")]; + tensor var_3640_reps_0 = const()[name = string("op_3640_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_3640_cast_fp16 = tile(reps = var_3640_reps_0, x = var_3638_cast_fp16)[name = string("op_3640_cast_fp16")]; + tensor var_3641 = const()[name = string("op_3641"), val = tensor([1, 2048, 32, 128])]; + tensor q_39_cast_fp16 = reshape(shape = var_3641, x = var_3640_cast_fp16)[name = string("q_39_cast_fp16")]; + tensor var_3646_axes_0 = const()[name = string("op_3646_axes_0"), val = tensor([3])]; + tensor var_3646_cast_fp16 = expand_dims(axes = var_3646_axes_0, x = k_33_cast_fp16)[name = string("op_3646_cast_fp16")]; + tensor var_3648_reps_0 = const()[name = string("op_3648_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_3648_cast_fp16 = tile(reps = var_3648_reps_0, x = var_3646_cast_fp16)[name = string("op_3648_cast_fp16")]; + tensor var_3649 = const()[name = string("op_3649"), val = tensor([1, 2048, 32, 128])]; + tensor k_35_cast_fp16 = reshape(shape = var_3649, x = var_3648_cast_fp16)[name = string("k_35_cast_fp16")]; + fp16 var_3651_to_fp16 = const()[name = string("op_3651_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor q_41_cast_fp16 = mul(x = q_39_cast_fp16, y = var_3651_to_fp16)[name = string("q_41_cast_fp16")]; + tensor var_3653_begin_0 = const()[name = string("op_3653_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3653_end_0 = const()[name = string("op_3653_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_3653_end_mask_0 = const()[name = string("op_3653_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_3653_squeeze_mask_0 = const()[name = string("op_3653_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_3653_cast_fp16 = slice_by_index(begin = var_3653_begin_0, end = var_3653_end_0, end_mask = var_3653_end_mask_0, squeeze_mask = var_3653_squeeze_mask_0, x = q_41_cast_fp16)[name = string("op_3653_cast_fp16")]; + tensor var_3654 = const()[name = string("op_3654"), val = tensor([1, 0, 2])]; + tensor var_3656 = const()[name = string("op_3656"), val = tensor([32, 32, 64, 128])]; + tensor var_3655_cast_fp16 = transpose(perm = var_3654, x = var_3653_cast_fp16)[name = string("transpose_25")]; + tensor q_43_cast_fp16 = reshape(shape = var_3656, x = var_3655_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor var_3658_begin_0 = const()[name = string("op_3658_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3658_end_0 = const()[name = string("op_3658_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_3658_end_mask_0 = const()[name = string("op_3658_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_3658_squeeze_mask_0 = const()[name = string("op_3658_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_3658_cast_fp16 = slice_by_index(begin = var_3658_begin_0, end = var_3658_end_0, end_mask = var_3658_end_mask_0, squeeze_mask = var_3658_squeeze_mask_0, x = k_35_cast_fp16)[name = string("op_3658_cast_fp16")]; + tensor var_3659 = const()[name = string("op_3659"), val = tensor([1, 0, 2])]; + tensor var_3661 = const()[name = string("op_3661"), val = tensor([32, 32, 64, 128])]; + tensor var_3660_cast_fp16 = transpose(perm = var_3659, x = var_3658_cast_fp16)[name = string("transpose_24")]; + tensor k_37_cast_fp16 = reshape(shape = var_3661, x = var_3660_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor var_3663_begin_0 = const()[name = string("op_3663_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3663_end_0 = const()[name = string("op_3663_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_3663_end_mask_0 = const()[name = string("op_3663_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_3663_squeeze_mask_0 = const()[name = string("op_3663_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_3663_cast_fp16 = slice_by_index(begin = var_3663_begin_0, end = var_3663_end_0, end_mask = var_3663_end_mask_0, squeeze_mask = var_3663_squeeze_mask_0, x = v_17_cast_fp16)[name = string("op_3663_cast_fp16")]; + tensor var_3664 = const()[name = string("op_3664"), val = tensor([1, 0, 2])]; + tensor var_3666 = const()[name = string("op_3666"), val = tensor([32, 32, 64, 128])]; + tensor var_3665_cast_fp16 = transpose(perm = var_3664, x = var_3663_cast_fp16)[name = string("transpose_23")]; + tensor v_19_cast_fp16 = reshape(shape = var_3666, x = var_3665_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor var_3668_begin_0 = const()[name = string("op_3668_begin_0"), val = tensor([0, 0, 0])]; + tensor var_3668_end_0 = const()[name = string("op_3668_end_0"), val = tensor([1, 2048, 32])]; + tensor var_3668_end_mask_0 = const()[name = string("op_3668_end_mask_0"), val = tensor([false, true, true])]; + tensor var_3668_squeeze_mask_0 = const()[name = string("op_3668_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_3668_cast_fp16 = slice_by_index(begin = var_3668_begin_0, end = var_3668_end_0, end_mask = var_3668_end_mask_0, squeeze_mask = var_3668_squeeze_mask_0, x = beta_13_cast_fp16)[name = string("op_3668_cast_fp16")]; + tensor var_3669_perm_0 = const()[name = string("op_3669_perm_0"), val = tensor([1, 0])]; + tensor var_3670 = const()[name = string("op_3670"), val = tensor([32, 32, 64, 1])]; + tensor var_3669_cast_fp16 = transpose(perm = var_3669_perm_0, x = var_3668_cast_fp16)[name = string("transpose_22")]; + tensor beta_15_cast_fp16 = reshape(shape = var_3670, x = var_3669_cast_fp16)[name = string("beta_15_cast_fp16")]; + tensor var_3672_begin_0 = const()[name = string("op_3672_begin_0"), val = tensor([0, 0, 0])]; + tensor var_3672_end_0 = const()[name = string("op_3672_end_0"), val = tensor([1, 2048, 32])]; + tensor var_3672_end_mask_0 = const()[name = string("op_3672_end_mask_0"), val = tensor([false, true, true])]; + tensor var_3672_squeeze_mask_0 = const()[name = string("op_3672_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_3672_cast_fp16 = slice_by_index(begin = var_3672_begin_0, end = var_3672_end_0, end_mask = var_3672_end_mask_0, squeeze_mask = var_3672_squeeze_mask_0, x = g_13_cast_fp16)[name = string("op_3672_cast_fp16")]; + tensor var_3673_perm_0 = const()[name = string("op_3673_perm_0"), val = tensor([1, 0])]; + tensor var_3674 = const()[name = string("op_3674"), val = tensor([32, 32, 64])]; + tensor var_3673_cast_fp16 = transpose(perm = var_3673_perm_0, x = var_3672_cast_fp16)[name = string("transpose_21")]; + tensor g_15_cast_fp16 = reshape(shape = var_3674, x = var_3673_cast_fp16)[name = string("g_15_cast_fp16")]; + tensor g_col_7_axes_0 = const()[name = string("g_col_7_axes_0"), val = tensor([-1])]; + tensor g_col_7_cast_fp16 = expand_dims(axes = g_col_7_axes_0, x = g_15_cast_fp16)[name = string("g_col_7_cast_fp16")]; + bool var_3677_transpose_x_0 = const()[name = string("op_3677_transpose_x_0"), val = bool(false)]; + bool var_3677_transpose_y_0 = const()[name = string("op_3677_transpose_y_0"), val = bool(false)]; + tensor var_3677_cast_fp16 = matmul(transpose_x = var_3677_transpose_x_0, transpose_y = var_3677_transpose_y_0, x = layers_0_linear_attn_lower_incl_to_fp16, y = g_col_7_cast_fp16)[name = string("op_3677_cast_fp16")]; + tensor cum_7_axes_0 = const()[name = string("cum_7_axes_0"), val = tensor([-1])]; + tensor cum_7_cast_fp16 = squeeze(axes = cum_7_axes_0, x = var_3677_cast_fp16)[name = string("cum_7_cast_fp16")]; + bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)]; + bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)]; + tensor var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = layers_0_linear_attn_suffix_sum_t_to_fp16, y = g_col_7_cast_fp16)[name = string("op_3679_cast_fp16")]; + tensor to_end_7_axes_0 = const()[name = string("to_end_7_axes_0"), val = tensor([-1])]; + tensor to_end_7_cast_fp16 = squeeze(axes = to_end_7_axes_0, x = var_3679_cast_fp16)[name = string("to_end_7_cast_fp16")]; + bool var_3681_transpose_x_0 = const()[name = string("op_3681_transpose_x_0"), val = bool(false)]; + bool var_3681_transpose_y_0 = const()[name = string("op_3681_transpose_y_0"), val = bool(false)]; + tensor var_3681_cast_fp16 = matmul(transpose_x = var_3681_transpose_x_0, transpose_y = var_3681_transpose_y_0, x = layers_0_linear_attn_pair_sum_t_to_fp16, y = g_col_7_cast_fp16)[name = string("op_3681_cast_fp16")]; + tensor var_3682 = const()[name = string("op_3682"), val = tensor([32, 32, 64, 64])]; + tensor pair_43_cast_fp16 = reshape(shape = var_3682, x = var_3681_cast_fp16)[name = string("pair_43_cast_fp16")]; + tensor var_3684_cast_fp16 = exp(x = pair_43_cast_fp16)[name = string("op_3684_cast_fp16")]; + tensor pair_decay_7_cast_fp16 = mul(x = var_3684_cast_fp16, y = layers_0_linear_attn_lower_incl_to_fp16)[name = string("pair_decay_7_cast_fp16")]; + tensor k_beta_7_cast_fp16 = mul(x = k_37_cast_fp16, y = beta_15_cast_fp16)[name = string("k_beta_7_cast_fp16")]; + tensor v_beta_7_cast_fp16 = mul(x = v_19_cast_fp16, y = beta_15_cast_fp16)[name = string("v_beta_7_cast_fp16")]; + bool var_3689_transpose_x_1 = const()[name = string("op_3689_transpose_x_1"), val = bool(false)]; + bool var_3689_transpose_y_1 = const()[name = string("op_3689_transpose_y_1"), val = bool(true)]; + tensor var_3689_cast_fp16 = matmul(transpose_x = var_3689_transpose_x_1, transpose_y = var_3689_transpose_y_1, x = k_beta_7_cast_fp16, y = k_37_cast_fp16)[name = string("op_3689_cast_fp16")]; + tensor kkt_7_cast_fp16 = mul(x = var_3689_cast_fp16, y = pair_decay_7_cast_fp16)[name = string("kkt_7_cast_fp16")]; + bool var_3692_transpose_x_1 = const()[name = string("op_3692_transpose_x_1"), val = bool(false)]; + bool var_3692_transpose_y_1 = const()[name = string("op_3692_transpose_y_1"), val = bool(true)]; + tensor var_3692_cast_fp16 = matmul(transpose_x = var_3692_transpose_x_1, transpose_y = var_3692_transpose_y_1, x = q_43_cast_fp16, y = k_37_cast_fp16)[name = string("op_3692_cast_fp16")]; + tensor attn_intra_7_cast_fp16 = mul(x = var_3692_cast_fp16, y = pair_decay_7_cast_fp16)[name = string("attn_intra_7_cast_fp16")]; + tensor var_3694_cast_fp16 = mul(x = kkt_7_cast_fp16, y = layers_0_linear_attn_strict_lower_to_fp16)[name = string("op_3694_cast_fp16")]; + tensor t_13_cast_fp16 = add(x = layers_0_linear_attn_eye_to_fp16, y = var_3694_cast_fp16)[name = string("t_13_cast_fp16")]; + tensor var_3698 = const()[name = string("op_3698"), val = tensor([-1, 64, 64])]; + tensor t_15_cast_fp16 = reshape(shape = var_3698, x = t_13_cast_fp16)[name = string("t_15_cast_fp16")]; + tensor var_3703 = const()[name = string("op_3703"), val = tensor([1024, 32, 2, 32, 2])]; + tensor var_3704_cast_fp16 = reshape(shape = var_3703, x = t_15_cast_fp16)[name = string("op_3704_cast_fp16")]; + tensor var_3709_cast_fp16 = mul(x = var_3704_cast_fp16, y = var_227_to_fp16)[name = string("op_3709_cast_fp16")]; + tensor blocks_37_axes_0 = const()[name = string("blocks_37_axes_0"), val = tensor([-2])]; + bool blocks_37_keep_dims_0 = const()[name = string("blocks_37_keep_dims_0"), val = bool(false)]; + tensor blocks_37_cast_fp16 = reduce_sum(axes = blocks_37_axes_0, keep_dims = blocks_37_keep_dims_0, x = var_3709_cast_fp16)[name = string("blocks_37_cast_fp16")]; + tensor var_3712_begin_0 = const()[name = string("op_3712_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3712_end_0 = const()[name = string("op_3712_end_0"), val = tensor([1024, 32, 2, 2])]; + tensor var_3712_end_mask_0 = const()[name = string("op_3712_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3712_cast_fp16 = slice_by_index(begin = var_3712_begin_0, end = var_3712_end_0, end_mask = var_3712_end_mask_0, x = blocks_37_cast_fp16)[name = string("op_3712_cast_fp16")]; + tensor x_105_begin_0 = const()[name = string("x_105_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_105_end_0 = const()[name = string("x_105_end_0"), val = tensor([1024, 32, 1, 1])]; + tensor x_105_end_mask_0 = const()[name = string("x_105_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_105_cast_fp16 = slice_by_index(begin = x_105_begin_0, end = x_105_end_0, end_mask = x_105_end_mask_0, x = var_3712_cast_fp16)[name = string("x_105_cast_fp16")]; + bool var_3722_transpose_x_0 = const()[name = string("op_3722_transpose_x_0"), val = bool(false)]; + bool var_3722_transpose_y_0 = const()[name = string("op_3722_transpose_y_0"), val = bool(false)]; + tensor var_3722_cast_fp16 = matmul(transpose_x = var_3722_transpose_x_0, transpose_y = var_3722_transpose_y_0, x = x_105_cast_fp16, y = var_235_to_fp16)[name = string("op_3722_cast_fp16")]; + bool var_3723_transpose_x_0 = const()[name = string("op_3723_transpose_x_0"), val = bool(false)]; + bool var_3723_transpose_y_0 = const()[name = string("op_3723_transpose_y_0"), val = bool(false)]; + tensor var_3723_cast_fp16 = matmul(transpose_x = var_3723_transpose_x_0, transpose_y = var_3723_transpose_y_0, x = var_235_to_fp16, y = var_3722_cast_fp16)[name = string("op_3723_cast_fp16")]; + fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_37_cast_fp16 = mul(x = var_3723_cast_fp16, y = const_47_promoted_to_fp16)[name = string("lower_37_cast_fp16")]; + bool bottom_37_interleave_0 = const()[name = string("bottom_37_interleave_0"), val = bool(false)]; + tensor bottom_37_cast_fp16 = concat(axis = var_3544, interleave = bottom_37_interleave_0, values = (lower_37_cast_fp16, var_235_to_fp16))[name = string("bottom_37_cast_fp16")]; + bool inv_51_interleave_0 = const()[name = string("inv_51_interleave_0"), val = bool(false)]; + tensor inv_51_cast_fp16 = concat(axis = var_3515, interleave = inv_51_interleave_0, values = (top_1_to_fp16, bottom_37_cast_fp16))[name = string("inv_51_cast_fp16")]; + tensor var_3732 = const()[name = string("op_3732"), val = tensor([1024, 16, 4, 16, 4])]; + tensor var_3733_cast_fp16 = reshape(shape = var_3732, x = t_15_cast_fp16)[name = string("op_3733_cast_fp16")]; + tensor var_3738_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_256_to_fp16)[name = string("op_3738_cast_fp16")]; + tensor blocks_39_axes_0 = const()[name = string("blocks_39_axes_0"), val = tensor([-2])]; + bool blocks_39_keep_dims_0 = const()[name = string("blocks_39_keep_dims_0"), val = bool(false)]; + tensor blocks_39_cast_fp16 = reduce_sum(axes = blocks_39_axes_0, keep_dims = blocks_39_keep_dims_0, x = var_3738_cast_fp16)[name = string("blocks_39_cast_fp16")]; + tensor var_3741_begin_0 = const()[name = string("op_3741_begin_0"), val = tensor([0, 0, 2, 0])]; + tensor var_3741_end_0 = const()[name = string("op_3741_end_0"), val = tensor([1024, 16, 4, 4])]; + tensor var_3741_end_mask_0 = const()[name = string("op_3741_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3741_cast_fp16 = slice_by_index(begin = var_3741_begin_0, end = var_3741_end_0, end_mask = var_3741_end_mask_0, x = blocks_39_cast_fp16)[name = string("op_3741_cast_fp16")]; + tensor x_107_begin_0 = const()[name = string("x_107_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_107_end_0 = const()[name = string("x_107_end_0"), val = tensor([1024, 16, 2, 2])]; + tensor x_107_end_mask_0 = const()[name = string("x_107_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_107_cast_fp16 = slice_by_index(begin = x_107_begin_0, end = x_107_end_0, end_mask = x_107_end_mask_0, x = var_3741_cast_fp16)[name = string("x_107_cast_fp16")]; + tensor var_3743 = const()[name = string("op_3743"), val = tensor([1024, 16, 2, 2, 2])]; + tensor pair_47_cast_fp16 = reshape(shape = var_3743, x = inv_51_cast_fp16)[name = string("pair_47_cast_fp16")]; + tensor var_3745_begin_0 = const()[name = string("op_3745_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3745_end_0 = const()[name = string("op_3745_end_0"), val = tensor([1024, 16, 1, 2, 2])]; + tensor var_3745_end_mask_0 = const()[name = string("op_3745_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3745_squeeze_mask_0 = const()[name = string("op_3745_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3745_cast_fp16 = slice_by_index(begin = var_3745_begin_0, end = var_3745_end_0, end_mask = var_3745_end_mask_0, squeeze_mask = var_3745_squeeze_mask_0, x = pair_47_cast_fp16)[name = string("op_3745_cast_fp16")]; + tensor var_3748_begin_0 = const()[name = string("op_3748_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_3748_end_0 = const()[name = string("op_3748_end_0"), val = tensor([1024, 16, 2, 2, 2])]; + tensor var_3748_end_mask_0 = const()[name = string("op_3748_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3748_squeeze_mask_0 = const()[name = string("op_3748_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3748_cast_fp16 = slice_by_index(begin = var_3748_begin_0, end = var_3748_end_0, end_mask = var_3748_end_mask_0, squeeze_mask = var_3748_squeeze_mask_0, x = pair_47_cast_fp16)[name = string("op_3748_cast_fp16")]; + bool var_3751_transpose_x_0 = const()[name = string("op_3751_transpose_x_0"), val = bool(false)]; + bool var_3751_transpose_y_0 = const()[name = string("op_3751_transpose_y_0"), val = bool(false)]; + tensor var_3751_cast_fp16 = matmul(transpose_x = var_3751_transpose_x_0, transpose_y = var_3751_transpose_y_0, x = x_107_cast_fp16, y = var_3745_cast_fp16)[name = string("op_3751_cast_fp16")]; + bool var_3752_transpose_x_0 = const()[name = string("op_3752_transpose_x_0"), val = bool(false)]; + bool var_3752_transpose_y_0 = const()[name = string("op_3752_transpose_y_0"), val = bool(false)]; + tensor var_3752_cast_fp16 = matmul(transpose_x = var_3752_transpose_x_0, transpose_y = var_3752_transpose_y_0, x = var_3748_cast_fp16, y = var_3751_cast_fp16)[name = string("op_3752_cast_fp16")]; + fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_39_cast_fp16 = mul(x = var_3752_cast_fp16, y = const_48_promoted_to_fp16)[name = string("lower_39_cast_fp16")]; + bool top_39_interleave_0 = const()[name = string("top_39_interleave_0"), val = bool(false)]; + tensor top_39_cast_fp16 = concat(axis = var_3544, interleave = top_39_interleave_0, values = (var_3745_cast_fp16, var_273_to_fp16))[name = string("top_39_cast_fp16")]; + bool bottom_39_interleave_0 = const()[name = string("bottom_39_interleave_0"), val = bool(false)]; + tensor bottom_39_cast_fp16 = concat(axis = var_3544, interleave = bottom_39_interleave_0, values = (lower_39_cast_fp16, var_3748_cast_fp16))[name = string("bottom_39_cast_fp16")]; + bool inv_53_interleave_0 = const()[name = string("inv_53_interleave_0"), val = bool(false)]; + tensor inv_53_cast_fp16 = concat(axis = var_3515, interleave = inv_53_interleave_0, values = (top_39_cast_fp16, bottom_39_cast_fp16))[name = string("inv_53_cast_fp16")]; + tensor var_3761 = const()[name = string("op_3761"), val = tensor([1024, 8, 8, 8, 8])]; + tensor var_3762_cast_fp16 = reshape(shape = var_3761, x = t_15_cast_fp16)[name = string("op_3762_cast_fp16")]; + tensor var_3766_to_fp16 = const()[name = string("op_3766_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493018688)))]; + tensor var_3767_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3766_to_fp16)[name = string("op_3767_cast_fp16")]; + tensor blocks_41_axes_0 = const()[name = string("blocks_41_axes_0"), val = tensor([-2])]; + bool blocks_41_keep_dims_0 = const()[name = string("blocks_41_keep_dims_0"), val = bool(false)]; + tensor blocks_41_cast_fp16 = reduce_sum(axes = blocks_41_axes_0, keep_dims = blocks_41_keep_dims_0, x = var_3767_cast_fp16)[name = string("blocks_41_cast_fp16")]; + tensor var_3770_begin_0 = const()[name = string("op_3770_begin_0"), val = tensor([0, 0, 4, 0])]; + tensor var_3770_end_0 = const()[name = string("op_3770_end_0"), val = tensor([1024, 8, 8, 8])]; + tensor var_3770_end_mask_0 = const()[name = string("op_3770_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3770_cast_fp16 = slice_by_index(begin = var_3770_begin_0, end = var_3770_end_0, end_mask = var_3770_end_mask_0, x = blocks_41_cast_fp16)[name = string("op_3770_cast_fp16")]; + tensor x_109_begin_0 = const()[name = string("x_109_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_109_end_0 = const()[name = string("x_109_end_0"), val = tensor([1024, 8, 4, 4])]; + tensor x_109_end_mask_0 = const()[name = string("x_109_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_109_cast_fp16 = slice_by_index(begin = x_109_begin_0, end = x_109_end_0, end_mask = x_109_end_mask_0, x = var_3770_cast_fp16)[name = string("x_109_cast_fp16")]; + tensor var_3772 = const()[name = string("op_3772"), val = tensor([1024, 8, 2, 4, 4])]; + tensor pair_49_cast_fp16 = reshape(shape = var_3772, x = inv_53_cast_fp16)[name = string("pair_49_cast_fp16")]; + tensor var_3774_begin_0 = const()[name = string("op_3774_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3774_end_0 = const()[name = string("op_3774_end_0"), val = tensor([1024, 8, 1, 4, 4])]; + tensor var_3774_end_mask_0 = const()[name = string("op_3774_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3774_squeeze_mask_0 = const()[name = string("op_3774_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3774_cast_fp16 = slice_by_index(begin = var_3774_begin_0, end = var_3774_end_0, end_mask = var_3774_end_mask_0, squeeze_mask = var_3774_squeeze_mask_0, x = pair_49_cast_fp16)[name = string("op_3774_cast_fp16")]; + tensor var_3777_begin_0 = const()[name = string("op_3777_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_3777_end_0 = const()[name = string("op_3777_end_0"), val = tensor([1024, 8, 2, 4, 4])]; + tensor var_3777_end_mask_0 = const()[name = string("op_3777_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3777_squeeze_mask_0 = const()[name = string("op_3777_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3777_cast_fp16 = slice_by_index(begin = var_3777_begin_0, end = var_3777_end_0, end_mask = var_3777_end_mask_0, squeeze_mask = var_3777_squeeze_mask_0, x = pair_49_cast_fp16)[name = string("op_3777_cast_fp16")]; + bool var_3780_transpose_x_0 = const()[name = string("op_3780_transpose_x_0"), val = bool(false)]; + bool var_3780_transpose_y_0 = const()[name = string("op_3780_transpose_y_0"), val = bool(false)]; + tensor var_3780_cast_fp16 = matmul(transpose_x = var_3780_transpose_x_0, transpose_y = var_3780_transpose_y_0, x = x_109_cast_fp16, y = var_3774_cast_fp16)[name = string("op_3780_cast_fp16")]; + bool var_3781_transpose_x_0 = const()[name = string("op_3781_transpose_x_0"), val = bool(false)]; + bool var_3781_transpose_y_0 = const()[name = string("op_3781_transpose_y_0"), val = bool(false)]; + tensor var_3781_cast_fp16 = matmul(transpose_x = var_3781_transpose_x_0, transpose_y = var_3781_transpose_y_0, x = var_3777_cast_fp16, y = var_3780_cast_fp16)[name = string("op_3781_cast_fp16")]; + fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_41_cast_fp16 = mul(x = var_3781_cast_fp16, y = const_49_promoted_to_fp16)[name = string("lower_41_cast_fp16")]; + bool top_41_interleave_0 = const()[name = string("top_41_interleave_0"), val = bool(false)]; + tensor top_41_cast_fp16 = concat(axis = var_3544, interleave = top_41_interleave_0, values = (var_3774_cast_fp16, var_302_to_fp16))[name = string("top_41_cast_fp16")]; + bool bottom_41_interleave_0 = const()[name = string("bottom_41_interleave_0"), val = bool(false)]; + tensor bottom_41_cast_fp16 = concat(axis = var_3544, interleave = bottom_41_interleave_0, values = (lower_41_cast_fp16, var_3777_cast_fp16))[name = string("bottom_41_cast_fp16")]; + bool inv_55_interleave_0 = const()[name = string("inv_55_interleave_0"), val = bool(false)]; + tensor inv_55_cast_fp16 = concat(axis = var_3515, interleave = inv_55_interleave_0, values = (top_41_cast_fp16, bottom_41_cast_fp16))[name = string("inv_55_cast_fp16")]; + tensor var_3790 = const()[name = string("op_3790"), val = tensor([1024, 4, 16, 4, 16])]; + tensor var_3791_cast_fp16 = reshape(shape = var_3790, x = t_15_cast_fp16)[name = string("op_3791_cast_fp16")]; + tensor var_3795_to_fp16 = const()[name = string("op_3795_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493018880)))]; + tensor var_3796_cast_fp16 = mul(x = var_3791_cast_fp16, y = var_3795_to_fp16)[name = string("op_3796_cast_fp16")]; + tensor blocks_43_axes_0 = const()[name = string("blocks_43_axes_0"), val = tensor([-2])]; + bool blocks_43_keep_dims_0 = const()[name = string("blocks_43_keep_dims_0"), val = bool(false)]; + tensor blocks_43_cast_fp16 = reduce_sum(axes = blocks_43_axes_0, keep_dims = blocks_43_keep_dims_0, x = var_3796_cast_fp16)[name = string("blocks_43_cast_fp16")]; + tensor var_3799_begin_0 = const()[name = string("op_3799_begin_0"), val = tensor([0, 0, 8, 0])]; + tensor var_3799_end_0 = const()[name = string("op_3799_end_0"), val = tensor([1024, 4, 16, 16])]; + tensor var_3799_end_mask_0 = const()[name = string("op_3799_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3799_cast_fp16 = slice_by_index(begin = var_3799_begin_0, end = var_3799_end_0, end_mask = var_3799_end_mask_0, x = blocks_43_cast_fp16)[name = string("op_3799_cast_fp16")]; + tensor x_111_begin_0 = const()[name = string("x_111_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_111_end_0 = const()[name = string("x_111_end_0"), val = tensor([1024, 4, 8, 8])]; + tensor x_111_end_mask_0 = const()[name = string("x_111_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_111_cast_fp16 = slice_by_index(begin = x_111_begin_0, end = x_111_end_0, end_mask = x_111_end_mask_0, x = var_3799_cast_fp16)[name = string("x_111_cast_fp16")]; + tensor var_3801 = const()[name = string("op_3801"), val = tensor([1024, 4, 2, 8, 8])]; + tensor pair_51_cast_fp16 = reshape(shape = var_3801, x = inv_55_cast_fp16)[name = string("pair_51_cast_fp16")]; + tensor var_3803_begin_0 = const()[name = string("op_3803_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3803_end_0 = const()[name = string("op_3803_end_0"), val = tensor([1024, 4, 1, 8, 8])]; + tensor var_3803_end_mask_0 = const()[name = string("op_3803_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3803_squeeze_mask_0 = const()[name = string("op_3803_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3803_cast_fp16 = slice_by_index(begin = var_3803_begin_0, end = var_3803_end_0, end_mask = var_3803_end_mask_0, squeeze_mask = var_3803_squeeze_mask_0, x = pair_51_cast_fp16)[name = string("op_3803_cast_fp16")]; + tensor var_3806_begin_0 = const()[name = string("op_3806_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_3806_end_0 = const()[name = string("op_3806_end_0"), val = tensor([1024, 4, 2, 8, 8])]; + tensor var_3806_end_mask_0 = const()[name = string("op_3806_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3806_squeeze_mask_0 = const()[name = string("op_3806_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3806_cast_fp16 = slice_by_index(begin = var_3806_begin_0, end = var_3806_end_0, end_mask = var_3806_end_mask_0, squeeze_mask = var_3806_squeeze_mask_0, x = pair_51_cast_fp16)[name = string("op_3806_cast_fp16")]; + bool var_3809_transpose_x_0 = const()[name = string("op_3809_transpose_x_0"), val = bool(false)]; + bool var_3809_transpose_y_0 = const()[name = string("op_3809_transpose_y_0"), val = bool(false)]; + tensor var_3809_cast_fp16 = matmul(transpose_x = var_3809_transpose_x_0, transpose_y = var_3809_transpose_y_0, x = x_111_cast_fp16, y = var_3803_cast_fp16)[name = string("op_3809_cast_fp16")]; + bool var_3810_transpose_x_0 = const()[name = string("op_3810_transpose_x_0"), val = bool(false)]; + bool var_3810_transpose_y_0 = const()[name = string("op_3810_transpose_y_0"), val = bool(false)]; + tensor var_3810_cast_fp16 = matmul(transpose_x = var_3810_transpose_x_0, transpose_y = var_3810_transpose_y_0, x = var_3806_cast_fp16, y = var_3809_cast_fp16)[name = string("op_3810_cast_fp16")]; + fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_43_cast_fp16 = mul(x = var_3810_cast_fp16, y = const_50_promoted_to_fp16)[name = string("lower_43_cast_fp16")]; + bool top_43_interleave_0 = const()[name = string("top_43_interleave_0"), val = bool(false)]; + tensor top_43_cast_fp16 = concat(axis = var_3544, interleave = top_43_interleave_0, values = (var_3803_cast_fp16, var_331_to_fp16))[name = string("top_43_cast_fp16")]; + bool bottom_43_interleave_0 = const()[name = string("bottom_43_interleave_0"), val = bool(false)]; + tensor bottom_43_cast_fp16 = concat(axis = var_3544, interleave = bottom_43_interleave_0, values = (lower_43_cast_fp16, var_3806_cast_fp16))[name = string("bottom_43_cast_fp16")]; + bool inv_57_interleave_0 = const()[name = string("inv_57_interleave_0"), val = bool(false)]; + tensor inv_57_cast_fp16 = concat(axis = var_3515, interleave = inv_57_interleave_0, values = (top_43_cast_fp16, bottom_43_cast_fp16))[name = string("inv_57_cast_fp16")]; + tensor var_3819 = const()[name = string("op_3819"), val = tensor([1024, 2, 32, 2, 32])]; + tensor var_3820_cast_fp16 = reshape(shape = var_3819, x = t_15_cast_fp16)[name = string("op_3820_cast_fp16")]; + tensor var_3824_to_fp16 = const()[name = string("op_3824_to_fp16"), val = tensor([[[[0x1p+0], [0x0p+0]]], [[[0x0p+0], [0x1p+0]]]])]; + tensor var_3825_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_3824_to_fp16)[name = string("op_3825_cast_fp16")]; + tensor blocks_45_axes_0 = const()[name = string("blocks_45_axes_0"), val = tensor([-2])]; + bool blocks_45_keep_dims_0 = const()[name = string("blocks_45_keep_dims_0"), val = bool(false)]; + tensor blocks_45_cast_fp16 = reduce_sum(axes = blocks_45_axes_0, keep_dims = blocks_45_keep_dims_0, x = var_3825_cast_fp16)[name = string("blocks_45_cast_fp16")]; + tensor var_3828_begin_0 = const()[name = string("op_3828_begin_0"), val = tensor([0, 0, 16, 0])]; + tensor var_3828_end_0 = const()[name = string("op_3828_end_0"), val = tensor([1024, 2, 32, 32])]; + tensor var_3828_end_mask_0 = const()[name = string("op_3828_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3828_cast_fp16 = slice_by_index(begin = var_3828_begin_0, end = var_3828_end_0, end_mask = var_3828_end_mask_0, x = blocks_45_cast_fp16)[name = string("op_3828_cast_fp16")]; + tensor x_113_begin_0 = const()[name = string("x_113_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_113_end_0 = const()[name = string("x_113_end_0"), val = tensor([1024, 2, 16, 16])]; + tensor x_113_end_mask_0 = const()[name = string("x_113_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_113_cast_fp16 = slice_by_index(begin = x_113_begin_0, end = x_113_end_0, end_mask = x_113_end_mask_0, x = var_3828_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_3830 = const()[name = string("op_3830"), val = tensor([1024, 2, 2, 16, 16])]; + tensor pair_53_cast_fp16 = reshape(shape = var_3830, x = inv_57_cast_fp16)[name = string("pair_53_cast_fp16")]; + tensor var_3832_begin_0 = const()[name = string("op_3832_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3832_end_0 = const()[name = string("op_3832_end_0"), val = tensor([1024, 2, 1, 16, 16])]; + tensor var_3832_end_mask_0 = const()[name = string("op_3832_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3832_squeeze_mask_0 = const()[name = string("op_3832_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3832_cast_fp16 = slice_by_index(begin = var_3832_begin_0, end = var_3832_end_0, end_mask = var_3832_end_mask_0, squeeze_mask = var_3832_squeeze_mask_0, x = pair_53_cast_fp16)[name = string("op_3832_cast_fp16")]; + tensor var_3835_begin_0 = const()[name = string("op_3835_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_3835_end_0 = const()[name = string("op_3835_end_0"), val = tensor([1024, 2, 2, 16, 16])]; + tensor var_3835_end_mask_0 = const()[name = string("op_3835_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3835_squeeze_mask_0 = const()[name = string("op_3835_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3835_cast_fp16 = slice_by_index(begin = var_3835_begin_0, end = var_3835_end_0, end_mask = var_3835_end_mask_0, squeeze_mask = var_3835_squeeze_mask_0, x = pair_53_cast_fp16)[name = string("op_3835_cast_fp16")]; + bool var_3838_transpose_x_0 = const()[name = string("op_3838_transpose_x_0"), val = bool(false)]; + bool var_3838_transpose_y_0 = const()[name = string("op_3838_transpose_y_0"), val = bool(false)]; + tensor var_3838_cast_fp16 = matmul(transpose_x = var_3838_transpose_x_0, transpose_y = var_3838_transpose_y_0, x = x_113_cast_fp16, y = var_3832_cast_fp16)[name = string("op_3838_cast_fp16")]; + bool var_3839_transpose_x_0 = const()[name = string("op_3839_transpose_x_0"), val = bool(false)]; + bool var_3839_transpose_y_0 = const()[name = string("op_3839_transpose_y_0"), val = bool(false)]; + tensor var_3839_cast_fp16 = matmul(transpose_x = var_3839_transpose_x_0, transpose_y = var_3839_transpose_y_0, x = var_3835_cast_fp16, y = var_3838_cast_fp16)[name = string("op_3839_cast_fp16")]; + fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_45_cast_fp16 = mul(x = var_3839_cast_fp16, y = const_51_promoted_to_fp16)[name = string("lower_45_cast_fp16")]; + bool top_45_interleave_0 = const()[name = string("top_45_interleave_0"), val = bool(false)]; + tensor top_45_cast_fp16 = concat(axis = var_3544, interleave = top_45_interleave_0, values = (var_3832_cast_fp16, var_360_to_fp16))[name = string("top_45_cast_fp16")]; + bool bottom_45_interleave_0 = const()[name = string("bottom_45_interleave_0"), val = bool(false)]; + tensor bottom_45_cast_fp16 = concat(axis = var_3544, interleave = bottom_45_interleave_0, values = (lower_45_cast_fp16, var_3835_cast_fp16))[name = string("bottom_45_cast_fp16")]; + bool inv_59_interleave_0 = const()[name = string("inv_59_interleave_0"), val = bool(false)]; + tensor inv_59_cast_fp16 = concat(axis = var_3515, interleave = inv_59_interleave_0, values = (top_45_cast_fp16, bottom_45_cast_fp16))[name = string("inv_59_cast_fp16")]; + tensor var_3848 = const()[name = string("op_3848"), val = tensor([1024, 1, 64, 1, 64])]; + tensor var_3849_cast_fp16 = reshape(shape = var_3848, x = t_15_cast_fp16)[name = string("op_3849_cast_fp16")]; + tensor blocks_47_axes_0 = const()[name = string("blocks_47_axes_0"), val = tensor([-2])]; + bool blocks_47_keep_dims_0 = const()[name = string("blocks_47_keep_dims_0"), val = bool(false)]; + tensor blocks_47_cast_fp16 = reduce_sum(axes = blocks_47_axes_0, keep_dims = blocks_47_keep_dims_0, x = var_3849_cast_fp16)[name = string("blocks_47_cast_fp16")]; + tensor var_3857_begin_0 = const()[name = string("op_3857_begin_0"), val = tensor([0, 0, 32, 0])]; + tensor var_3857_end_0 = const()[name = string("op_3857_end_0"), val = tensor([1024, 1, 64, 64])]; + tensor var_3857_end_mask_0 = const()[name = string("op_3857_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3857_cast_fp16 = slice_by_index(begin = var_3857_begin_0, end = var_3857_end_0, end_mask = var_3857_end_mask_0, x = blocks_47_cast_fp16)[name = string("op_3857_cast_fp16")]; + tensor x_115_begin_0 = const()[name = string("x_115_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_115_end_0 = const()[name = string("x_115_end_0"), val = tensor([1024, 1, 32, 32])]; + tensor x_115_end_mask_0 = const()[name = string("x_115_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_115_cast_fp16 = slice_by_index(begin = x_115_begin_0, end = x_115_end_0, end_mask = x_115_end_mask_0, x = var_3857_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_3859 = const()[name = string("op_3859"), val = tensor([1024, 1, 2, 32, 32])]; + tensor pair_55_cast_fp16 = reshape(shape = var_3859, x = inv_59_cast_fp16)[name = string("pair_55_cast_fp16")]; + tensor var_3861_begin_0 = const()[name = string("op_3861_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3861_end_0 = const()[name = string("op_3861_end_0"), val = tensor([1024, 1, 1, 32, 32])]; + tensor var_3861_end_mask_0 = const()[name = string("op_3861_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3861_squeeze_mask_0 = const()[name = string("op_3861_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3861_cast_fp16 = slice_by_index(begin = var_3861_begin_0, end = var_3861_end_0, end_mask = var_3861_end_mask_0, squeeze_mask = var_3861_squeeze_mask_0, x = pair_55_cast_fp16)[name = string("op_3861_cast_fp16")]; + tensor var_3864_begin_0 = const()[name = string("op_3864_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_3864_end_0 = const()[name = string("op_3864_end_0"), val = tensor([1024, 1, 2, 32, 32])]; + tensor var_3864_end_mask_0 = const()[name = string("op_3864_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_3864_squeeze_mask_0 = const()[name = string("op_3864_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_3864_cast_fp16 = slice_by_index(begin = var_3864_begin_0, end = var_3864_end_0, end_mask = var_3864_end_mask_0, squeeze_mask = var_3864_squeeze_mask_0, x = pair_55_cast_fp16)[name = string("op_3864_cast_fp16")]; + bool var_3867_transpose_x_0 = const()[name = string("op_3867_transpose_x_0"), val = bool(false)]; + bool var_3867_transpose_y_0 = const()[name = string("op_3867_transpose_y_0"), val = bool(false)]; + tensor var_3867_cast_fp16 = matmul(transpose_x = var_3867_transpose_x_0, transpose_y = var_3867_transpose_y_0, x = x_115_cast_fp16, y = var_3861_cast_fp16)[name = string("op_3867_cast_fp16")]; + bool var_3868_transpose_x_0 = const()[name = string("op_3868_transpose_x_0"), val = bool(false)]; + bool var_3868_transpose_y_0 = const()[name = string("op_3868_transpose_y_0"), val = bool(false)]; + tensor var_3868_cast_fp16 = matmul(transpose_x = var_3868_transpose_x_0, transpose_y = var_3868_transpose_y_0, x = var_3864_cast_fp16, y = var_3867_cast_fp16)[name = string("op_3868_cast_fp16")]; + fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_47_cast_fp16 = mul(x = var_3868_cast_fp16, y = const_52_promoted_to_fp16)[name = string("lower_47_cast_fp16")]; + bool top_47_interleave_0 = const()[name = string("top_47_interleave_0"), val = bool(false)]; + tensor top_47_cast_fp16 = concat(axis = var_3544, interleave = top_47_interleave_0, values = (var_3861_cast_fp16, var_389_to_fp16))[name = string("top_47_cast_fp16")]; + bool bottom_47_interleave_0 = const()[name = string("bottom_47_interleave_0"), val = bool(false)]; + tensor bottom_47_cast_fp16 = concat(axis = var_3544, interleave = bottom_47_interleave_0, values = (lower_47_cast_fp16, var_3864_cast_fp16))[name = string("bottom_47_cast_fp16")]; + bool inv_61_interleave_0 = const()[name = string("inv_61_interleave_0"), val = bool(false)]; + tensor inv_61_cast_fp16 = concat(axis = var_3515, interleave = inv_61_interleave_0, values = (top_47_cast_fp16, bottom_47_cast_fp16))[name = string("inv_61_cast_fp16")]; + tensor var_3877 = const()[name = string("op_3877"), val = tensor([32, 32, 64, 64])]; + tensor inv_63_cast_fp16 = reshape(shape = var_3877, x = inv_61_cast_fp16)[name = string("inv_63_cast_fp16")]; + bool new_v_7_transpose_x_0 = const()[name = string("new_v_7_transpose_x_0"), val = bool(false)]; + bool new_v_7_transpose_y_0 = const()[name = string("new_v_7_transpose_y_0"), val = bool(false)]; + tensor new_v_7_cast_fp16 = matmul(transpose_x = new_v_7_transpose_x_0, transpose_y = new_v_7_transpose_y_0, x = inv_63_cast_fp16, y = v_beta_7_cast_fp16)[name = string("new_v_7_cast_fp16")]; + tensor var_3880_cast_fp16 = exp(x = cum_7_cast_fp16)[name = string("op_3880_cast_fp16")]; + tensor var_3881_axes_0 = const()[name = string("op_3881_axes_0"), val = tensor([3])]; + tensor var_3881_cast_fp16 = expand_dims(axes = var_3881_axes_0, x = var_3880_cast_fp16)[name = string("op_3881_cast_fp16")]; + tensor var_3882_cast_fp16 = mul(x = k_beta_7_cast_fp16, y = var_3881_cast_fp16)[name = string("op_3882_cast_fp16")]; + bool k_cumdecay_7_transpose_x_0 = const()[name = string("k_cumdecay_7_transpose_x_0"), val = bool(false)]; + bool k_cumdecay_7_transpose_y_0 = const()[name = string("k_cumdecay_7_transpose_y_0"), val = bool(false)]; + tensor k_cumdecay_7_cast_fp16 = matmul(transpose_x = k_cumdecay_7_transpose_x_0, transpose_y = k_cumdecay_7_transpose_y_0, x = inv_63_cast_fp16, y = var_3882_cast_fp16)[name = string("k_cumdecay_7_cast_fp16")]; + tensor q_dec_7_cast_fp16 = mul(x = q_43_cast_fp16, y = var_3881_cast_fp16)[name = string("q_dec_7_cast_fp16")]; + tensor var_3887_cast_fp16 = exp(x = to_end_7_cast_fp16)[name = string("op_3887_cast_fp16")]; + tensor var_3888_axes_0 = const()[name = string("op_3888_axes_0"), val = tensor([3])]; + tensor var_3888_cast_fp16 = expand_dims(axes = var_3888_axes_0, x = var_3887_cast_fp16)[name = string("op_3888_cast_fp16")]; + tensor k_dec_7_cast_fp16 = mul(x = k_37_cast_fp16, y = var_3888_cast_fp16)[name = string("k_dec_7_cast_fp16")]; + tensor var_3890_begin_0 = const()[name = string("op_3890_begin_0"), val = tensor([0, 0, -1])]; + tensor var_3890_end_0 = const()[name = string("op_3890_end_0"), val = tensor([32, 32, 64])]; + tensor var_3890_end_mask_0 = const()[name = string("op_3890_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3890_cast_fp16 = slice_by_index(begin = var_3890_begin_0, end = var_3890_end_0, end_mask = var_3890_end_mask_0, x = cum_7_cast_fp16)[name = string("op_3890_cast_fp16")]; + tensor var_3891_cast_fp16 = exp(x = var_3890_cast_fp16)[name = string("op_3891_cast_fp16")]; + tensor chunk_decay_7_axes_0 = const()[name = string("chunk_decay_7_axes_0"), val = tensor([3])]; + tensor chunk_decay_7_cast_fp16 = expand_dims(axes = chunk_decay_7_axes_0, x = var_3891_cast_fp16)[name = string("chunk_decay_7_cast_fp16")]; + tensor var_3896_begin_0 = const()[name = string("op_3896_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3896_end_0 = const()[name = string("op_3896_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_3896_end_mask_0 = const()[name = string("op_3896_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3896_squeeze_mask_0 = const()[name = string("op_3896_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3896_cast_fp16 = slice_by_index(begin = var_3896_begin_0, end = var_3896_end_0, end_mask = var_3896_end_mask_0, squeeze_mask = var_3896_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_3896_cast_fp16")]; + tensor var_3898_begin_0 = const()[name = string("op_3898_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3898_end_0 = const()[name = string("op_3898_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_3898_end_mask_0 = const()[name = string("op_3898_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3898_squeeze_mask_0 = const()[name = string("op_3898_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3898_cast_fp16 = slice_by_index(begin = var_3898_begin_0, end = var_3898_end_0, end_mask = var_3898_end_mask_0, squeeze_mask = var_3898_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_3898_cast_fp16")]; + bool var_3899_transpose_x_0 = const()[name = string("op_3899_transpose_x_0"), val = bool(false)]; + bool var_3899_transpose_y_0 = const()[name = string("op_3899_transpose_y_0"), val = bool(false)]; + tensor var_3899_cast_fp16 = matmul(transpose_x = var_3899_transpose_x_0, transpose_y = var_3899_transpose_y_0, x = var_3898_cast_fp16, y = state_1_to_fp16)[name = string("op_3899_cast_fp16")]; + tensor v_new_193_cast_fp16 = sub(x = var_3896_cast_fp16, y = var_3899_cast_fp16)[name = string("v_new_193_cast_fp16")]; + tensor var_3902_begin_0 = const()[name = string("op_3902_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3902_end_0 = const()[name = string("op_3902_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_3902_end_mask_0 = const()[name = string("op_3902_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3902_squeeze_mask_0 = const()[name = string("op_3902_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3902_cast_fp16 = slice_by_index(begin = var_3902_begin_0, end = var_3902_end_0, end_mask = var_3902_end_mask_0, squeeze_mask = var_3902_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_3902_cast_fp16")]; + bool var_3903_transpose_x_0 = const()[name = string("op_3903_transpose_x_0"), val = bool(false)]; + bool var_3903_transpose_y_0 = const()[name = string("op_3903_transpose_y_0"), val = bool(false)]; + tensor var_3903_cast_fp16 = matmul(transpose_x = var_3903_transpose_x_0, transpose_y = var_3903_transpose_y_0, x = var_3902_cast_fp16, y = state_1_to_fp16)[name = string("op_3903_cast_fp16")]; + tensor var_3905_begin_0 = const()[name = string("op_3905_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3905_end_0 = const()[name = string("op_3905_end_0"), val = tensor([32, 1, 64, 64])]; + tensor var_3905_end_mask_0 = const()[name = string("op_3905_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3905_squeeze_mask_0 = const()[name = string("op_3905_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3905_cast_fp16 = slice_by_index(begin = var_3905_begin_0, end = var_3905_end_0, end_mask = var_3905_end_mask_0, squeeze_mask = var_3905_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_3905_cast_fp16")]; + bool var_3906_transpose_x_0 = const()[name = string("op_3906_transpose_x_0"), val = bool(false)]; + bool var_3906_transpose_y_0 = const()[name = string("op_3906_transpose_y_0"), val = bool(false)]; + tensor var_3906_cast_fp16 = matmul(transpose_x = var_3906_transpose_x_0, transpose_y = var_3906_transpose_y_0, x = var_3905_cast_fp16, y = v_new_193_cast_fp16)[name = string("op_3906_cast_fp16")]; + tensor var_3907_cast_fp16 = add(x = var_3903_cast_fp16, y = var_3906_cast_fp16)[name = string("op_3907_cast_fp16")]; + tensor var_3909_begin_0 = const()[name = string("op_3909_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3909_end_0 = const()[name = string("op_3909_end_0"), val = tensor([32, 1, 1, 1])]; + tensor var_3909_end_mask_0 = const()[name = string("op_3909_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3909_squeeze_mask_0 = const()[name = string("op_3909_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3909_cast_fp16 = slice_by_index(begin = var_3909_begin_0, end = var_3909_end_0, end_mask = var_3909_end_mask_0, squeeze_mask = var_3909_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_3909_cast_fp16")]; + tensor var_3910_cast_fp16 = mul(x = state_1_to_fp16, y = var_3909_cast_fp16)[name = string("op_3910_cast_fp16")]; + tensor var_3912_begin_0 = const()[name = string("op_3912_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3912_end_0 = const()[name = string("op_3912_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_3912_end_mask_0 = const()[name = string("op_3912_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3912_squeeze_mask_0 = const()[name = string("op_3912_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3912_cast_fp16 = slice_by_index(begin = var_3912_begin_0, end = var_3912_end_0, end_mask = var_3912_end_mask_0, squeeze_mask = var_3912_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_3912_cast_fp16")]; + bool var_3914_transpose_x_1 = const()[name = string("op_3914_transpose_x_1"), val = bool(true)]; + bool var_3914_transpose_y_1 = const()[name = string("op_3914_transpose_y_1"), val = bool(false)]; + tensor var_3914_cast_fp16 = matmul(transpose_x = var_3914_transpose_x_1, transpose_y = var_3914_transpose_y_1, x = var_3912_cast_fp16, y = v_new_193_cast_fp16)[name = string("op_3914_cast_fp16")]; + tensor state_195_cast_fp16 = add(x = var_3910_cast_fp16, y = var_3914_cast_fp16)[name = string("state_195_cast_fp16")]; + tensor var_3917_begin_0 = const()[name = string("op_3917_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_3917_end_0 = const()[name = string("op_3917_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_3917_end_mask_0 = const()[name = string("op_3917_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3917_squeeze_mask_0 = const()[name = string("op_3917_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3917_cast_fp16 = slice_by_index(begin = var_3917_begin_0, end = var_3917_end_0, end_mask = var_3917_end_mask_0, squeeze_mask = var_3917_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_3917_cast_fp16")]; + tensor var_3919_begin_0 = const()[name = string("op_3919_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_3919_end_0 = const()[name = string("op_3919_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_3919_end_mask_0 = const()[name = string("op_3919_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3919_squeeze_mask_0 = const()[name = string("op_3919_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3919_cast_fp16 = slice_by_index(begin = var_3919_begin_0, end = var_3919_end_0, end_mask = var_3919_end_mask_0, squeeze_mask = var_3919_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_3919_cast_fp16")]; + bool var_3920_transpose_x_0 = const()[name = string("op_3920_transpose_x_0"), val = bool(false)]; + bool var_3920_transpose_y_0 = const()[name = string("op_3920_transpose_y_0"), val = bool(false)]; + tensor var_3920_cast_fp16 = matmul(transpose_x = var_3920_transpose_x_0, transpose_y = var_3920_transpose_y_0, x = var_3919_cast_fp16, y = state_195_cast_fp16)[name = string("op_3920_cast_fp16")]; + tensor v_new_195_cast_fp16 = sub(x = var_3917_cast_fp16, y = var_3920_cast_fp16)[name = string("v_new_195_cast_fp16")]; + tensor var_3923_begin_0 = const()[name = string("op_3923_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_3923_end_0 = const()[name = string("op_3923_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_3923_end_mask_0 = const()[name = string("op_3923_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3923_squeeze_mask_0 = const()[name = string("op_3923_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3923_cast_fp16 = slice_by_index(begin = var_3923_begin_0, end = var_3923_end_0, end_mask = var_3923_end_mask_0, squeeze_mask = var_3923_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_3923_cast_fp16")]; + bool var_3924_transpose_x_0 = const()[name = string("op_3924_transpose_x_0"), val = bool(false)]; + bool var_3924_transpose_y_0 = const()[name = string("op_3924_transpose_y_0"), val = bool(false)]; + tensor var_3924_cast_fp16 = matmul(transpose_x = var_3924_transpose_x_0, transpose_y = var_3924_transpose_y_0, x = var_3923_cast_fp16, y = state_195_cast_fp16)[name = string("op_3924_cast_fp16")]; + tensor var_3926_begin_0 = const()[name = string("op_3926_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_3926_end_0 = const()[name = string("op_3926_end_0"), val = tensor([32, 2, 64, 64])]; + tensor var_3926_end_mask_0 = const()[name = string("op_3926_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3926_squeeze_mask_0 = const()[name = string("op_3926_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3926_cast_fp16 = slice_by_index(begin = var_3926_begin_0, end = var_3926_end_0, end_mask = var_3926_end_mask_0, squeeze_mask = var_3926_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_3926_cast_fp16")]; + bool var_3927_transpose_x_0 = const()[name = string("op_3927_transpose_x_0"), val = bool(false)]; + bool var_3927_transpose_y_0 = const()[name = string("op_3927_transpose_y_0"), val = bool(false)]; + tensor var_3927_cast_fp16 = matmul(transpose_x = var_3927_transpose_x_0, transpose_y = var_3927_transpose_y_0, x = var_3926_cast_fp16, y = v_new_195_cast_fp16)[name = string("op_3927_cast_fp16")]; + tensor var_3928_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3927_cast_fp16)[name = string("op_3928_cast_fp16")]; + tensor var_3930_begin_0 = const()[name = string("op_3930_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_3930_end_0 = const()[name = string("op_3930_end_0"), val = tensor([32, 2, 1, 1])]; + tensor var_3930_end_mask_0 = const()[name = string("op_3930_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3930_squeeze_mask_0 = const()[name = string("op_3930_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3930_cast_fp16 = slice_by_index(begin = var_3930_begin_0, end = var_3930_end_0, end_mask = var_3930_end_mask_0, squeeze_mask = var_3930_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_3930_cast_fp16")]; + tensor var_3931_cast_fp16 = mul(x = state_195_cast_fp16, y = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")]; + tensor var_3933_begin_0 = const()[name = string("op_3933_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_3933_end_0 = const()[name = string("op_3933_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_3933_end_mask_0 = const()[name = string("op_3933_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3933_squeeze_mask_0 = const()[name = string("op_3933_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3933_cast_fp16 = slice_by_index(begin = var_3933_begin_0, end = var_3933_end_0, end_mask = var_3933_end_mask_0, squeeze_mask = var_3933_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_3933_cast_fp16")]; + bool var_3935_transpose_x_1 = const()[name = string("op_3935_transpose_x_1"), val = bool(true)]; + bool var_3935_transpose_y_1 = const()[name = string("op_3935_transpose_y_1"), val = bool(false)]; + tensor var_3935_cast_fp16 = matmul(transpose_x = var_3935_transpose_x_1, transpose_y = var_3935_transpose_y_1, x = var_3933_cast_fp16, y = v_new_195_cast_fp16)[name = string("op_3935_cast_fp16")]; + tensor state_197_cast_fp16 = add(x = var_3931_cast_fp16, y = var_3935_cast_fp16)[name = string("state_197_cast_fp16")]; + tensor var_3938_begin_0 = const()[name = string("op_3938_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_3938_end_0 = const()[name = string("op_3938_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_3938_end_mask_0 = const()[name = string("op_3938_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3938_squeeze_mask_0 = const()[name = string("op_3938_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3938_cast_fp16 = slice_by_index(begin = var_3938_begin_0, end = var_3938_end_0, end_mask = var_3938_end_mask_0, squeeze_mask = var_3938_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_3938_cast_fp16")]; + tensor var_3940_begin_0 = const()[name = string("op_3940_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_3940_end_0 = const()[name = string("op_3940_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_3940_end_mask_0 = const()[name = string("op_3940_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3940_squeeze_mask_0 = const()[name = string("op_3940_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3940_cast_fp16 = slice_by_index(begin = var_3940_begin_0, end = var_3940_end_0, end_mask = var_3940_end_mask_0, squeeze_mask = var_3940_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_3940_cast_fp16")]; + bool var_3941_transpose_x_0 = const()[name = string("op_3941_transpose_x_0"), val = bool(false)]; + bool var_3941_transpose_y_0 = const()[name = string("op_3941_transpose_y_0"), val = bool(false)]; + tensor var_3941_cast_fp16 = matmul(transpose_x = var_3941_transpose_x_0, transpose_y = var_3941_transpose_y_0, x = var_3940_cast_fp16, y = state_197_cast_fp16)[name = string("op_3941_cast_fp16")]; + tensor v_new_197_cast_fp16 = sub(x = var_3938_cast_fp16, y = var_3941_cast_fp16)[name = string("v_new_197_cast_fp16")]; + tensor var_3944_begin_0 = const()[name = string("op_3944_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_3944_end_0 = const()[name = string("op_3944_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_3944_end_mask_0 = const()[name = string("op_3944_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3944_squeeze_mask_0 = const()[name = string("op_3944_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3944_cast_fp16 = slice_by_index(begin = var_3944_begin_0, end = var_3944_end_0, end_mask = var_3944_end_mask_0, squeeze_mask = var_3944_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_3944_cast_fp16")]; + bool var_3945_transpose_x_0 = const()[name = string("op_3945_transpose_x_0"), val = bool(false)]; + bool var_3945_transpose_y_0 = const()[name = string("op_3945_transpose_y_0"), val = bool(false)]; + tensor var_3945_cast_fp16 = matmul(transpose_x = var_3945_transpose_x_0, transpose_y = var_3945_transpose_y_0, x = var_3944_cast_fp16, y = state_197_cast_fp16)[name = string("op_3945_cast_fp16")]; + tensor var_3947_begin_0 = const()[name = string("op_3947_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_3947_end_0 = const()[name = string("op_3947_end_0"), val = tensor([32, 3, 64, 64])]; + tensor var_3947_end_mask_0 = const()[name = string("op_3947_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3947_squeeze_mask_0 = const()[name = string("op_3947_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3947_cast_fp16 = slice_by_index(begin = var_3947_begin_0, end = var_3947_end_0, end_mask = var_3947_end_mask_0, squeeze_mask = var_3947_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_3947_cast_fp16")]; + bool var_3948_transpose_x_0 = const()[name = string("op_3948_transpose_x_0"), val = bool(false)]; + bool var_3948_transpose_y_0 = const()[name = string("op_3948_transpose_y_0"), val = bool(false)]; + tensor var_3948_cast_fp16 = matmul(transpose_x = var_3948_transpose_x_0, transpose_y = var_3948_transpose_y_0, x = var_3947_cast_fp16, y = v_new_197_cast_fp16)[name = string("op_3948_cast_fp16")]; + tensor var_3949_cast_fp16 = add(x = var_3945_cast_fp16, y = var_3948_cast_fp16)[name = string("op_3949_cast_fp16")]; + tensor var_3951_begin_0 = const()[name = string("op_3951_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_3951_end_0 = const()[name = string("op_3951_end_0"), val = tensor([32, 3, 1, 1])]; + tensor var_3951_end_mask_0 = const()[name = string("op_3951_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3951_squeeze_mask_0 = const()[name = string("op_3951_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3951_cast_fp16 = slice_by_index(begin = var_3951_begin_0, end = var_3951_end_0, end_mask = var_3951_end_mask_0, squeeze_mask = var_3951_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_3951_cast_fp16")]; + tensor var_3952_cast_fp16 = mul(x = state_197_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3952_cast_fp16")]; + tensor var_3954_begin_0 = const()[name = string("op_3954_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_3954_end_0 = const()[name = string("op_3954_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_3954_end_mask_0 = const()[name = string("op_3954_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3954_squeeze_mask_0 = const()[name = string("op_3954_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3954_cast_fp16 = slice_by_index(begin = var_3954_begin_0, end = var_3954_end_0, end_mask = var_3954_end_mask_0, squeeze_mask = var_3954_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_3954_cast_fp16")]; + bool var_3956_transpose_x_1 = const()[name = string("op_3956_transpose_x_1"), val = bool(true)]; + bool var_3956_transpose_y_1 = const()[name = string("op_3956_transpose_y_1"), val = bool(false)]; + tensor var_3956_cast_fp16 = matmul(transpose_x = var_3956_transpose_x_1, transpose_y = var_3956_transpose_y_1, x = var_3954_cast_fp16, y = v_new_197_cast_fp16)[name = string("op_3956_cast_fp16")]; + tensor state_199_cast_fp16 = add(x = var_3952_cast_fp16, y = var_3956_cast_fp16)[name = string("state_199_cast_fp16")]; + tensor var_3959_begin_0 = const()[name = string("op_3959_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_3959_end_0 = const()[name = string("op_3959_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_3959_end_mask_0 = const()[name = string("op_3959_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3959_squeeze_mask_0 = const()[name = string("op_3959_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3959_cast_fp16 = slice_by_index(begin = var_3959_begin_0, end = var_3959_end_0, end_mask = var_3959_end_mask_0, squeeze_mask = var_3959_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_3959_cast_fp16")]; + tensor var_3961_begin_0 = const()[name = string("op_3961_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_3961_end_0 = const()[name = string("op_3961_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_3961_end_mask_0 = const()[name = string("op_3961_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3961_squeeze_mask_0 = const()[name = string("op_3961_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3961_cast_fp16 = slice_by_index(begin = var_3961_begin_0, end = var_3961_end_0, end_mask = var_3961_end_mask_0, squeeze_mask = var_3961_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_3961_cast_fp16")]; + bool var_3962_transpose_x_0 = const()[name = string("op_3962_transpose_x_0"), val = bool(false)]; + bool var_3962_transpose_y_0 = const()[name = string("op_3962_transpose_y_0"), val = bool(false)]; + tensor var_3962_cast_fp16 = matmul(transpose_x = var_3962_transpose_x_0, transpose_y = var_3962_transpose_y_0, x = var_3961_cast_fp16, y = state_199_cast_fp16)[name = string("op_3962_cast_fp16")]; + tensor v_new_199_cast_fp16 = sub(x = var_3959_cast_fp16, y = var_3962_cast_fp16)[name = string("v_new_199_cast_fp16")]; + tensor var_3965_begin_0 = const()[name = string("op_3965_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_3965_end_0 = const()[name = string("op_3965_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_3965_end_mask_0 = const()[name = string("op_3965_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3965_squeeze_mask_0 = const()[name = string("op_3965_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3965_cast_fp16 = slice_by_index(begin = var_3965_begin_0, end = var_3965_end_0, end_mask = var_3965_end_mask_0, squeeze_mask = var_3965_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_3965_cast_fp16")]; + bool var_3966_transpose_x_0 = const()[name = string("op_3966_transpose_x_0"), val = bool(false)]; + bool var_3966_transpose_y_0 = const()[name = string("op_3966_transpose_y_0"), val = bool(false)]; + tensor var_3966_cast_fp16 = matmul(transpose_x = var_3966_transpose_x_0, transpose_y = var_3966_transpose_y_0, x = var_3965_cast_fp16, y = state_199_cast_fp16)[name = string("op_3966_cast_fp16")]; + tensor var_3968_begin_0 = const()[name = string("op_3968_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_3968_end_0 = const()[name = string("op_3968_end_0"), val = tensor([32, 4, 64, 64])]; + tensor var_3968_end_mask_0 = const()[name = string("op_3968_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3968_squeeze_mask_0 = const()[name = string("op_3968_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3968_cast_fp16 = slice_by_index(begin = var_3968_begin_0, end = var_3968_end_0, end_mask = var_3968_end_mask_0, squeeze_mask = var_3968_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_3968_cast_fp16")]; + bool var_3969_transpose_x_0 = const()[name = string("op_3969_transpose_x_0"), val = bool(false)]; + bool var_3969_transpose_y_0 = const()[name = string("op_3969_transpose_y_0"), val = bool(false)]; + tensor var_3969_cast_fp16 = matmul(transpose_x = var_3969_transpose_x_0, transpose_y = var_3969_transpose_y_0, x = var_3968_cast_fp16, y = v_new_199_cast_fp16)[name = string("op_3969_cast_fp16")]; + tensor var_3970_cast_fp16 = add(x = var_3966_cast_fp16, y = var_3969_cast_fp16)[name = string("op_3970_cast_fp16")]; + tensor var_3972_begin_0 = const()[name = string("op_3972_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_3972_end_0 = const()[name = string("op_3972_end_0"), val = tensor([32, 4, 1, 1])]; + tensor var_3972_end_mask_0 = const()[name = string("op_3972_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3972_squeeze_mask_0 = const()[name = string("op_3972_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3972_cast_fp16 = slice_by_index(begin = var_3972_begin_0, end = var_3972_end_0, end_mask = var_3972_end_mask_0, squeeze_mask = var_3972_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_3972_cast_fp16")]; + tensor var_3973_cast_fp16 = mul(x = state_199_cast_fp16, y = var_3972_cast_fp16)[name = string("op_3973_cast_fp16")]; + tensor var_3975_begin_0 = const()[name = string("op_3975_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_3975_end_0 = const()[name = string("op_3975_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_3975_end_mask_0 = const()[name = string("op_3975_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3975_squeeze_mask_0 = const()[name = string("op_3975_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3975_cast_fp16 = slice_by_index(begin = var_3975_begin_0, end = var_3975_end_0, end_mask = var_3975_end_mask_0, squeeze_mask = var_3975_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_3975_cast_fp16")]; + bool var_3977_transpose_x_1 = const()[name = string("op_3977_transpose_x_1"), val = bool(true)]; + bool var_3977_transpose_y_1 = const()[name = string("op_3977_transpose_y_1"), val = bool(false)]; + tensor var_3977_cast_fp16 = matmul(transpose_x = var_3977_transpose_x_1, transpose_y = var_3977_transpose_y_1, x = var_3975_cast_fp16, y = v_new_199_cast_fp16)[name = string("op_3977_cast_fp16")]; + tensor state_201_cast_fp16 = add(x = var_3973_cast_fp16, y = var_3977_cast_fp16)[name = string("state_201_cast_fp16")]; + tensor var_3980_begin_0 = const()[name = string("op_3980_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_3980_end_0 = const()[name = string("op_3980_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_3980_end_mask_0 = const()[name = string("op_3980_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3980_squeeze_mask_0 = const()[name = string("op_3980_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3980_cast_fp16 = slice_by_index(begin = var_3980_begin_0, end = var_3980_end_0, end_mask = var_3980_end_mask_0, squeeze_mask = var_3980_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_3980_cast_fp16")]; + tensor var_3982_begin_0 = const()[name = string("op_3982_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_3982_end_0 = const()[name = string("op_3982_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_3982_end_mask_0 = const()[name = string("op_3982_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3982_squeeze_mask_0 = const()[name = string("op_3982_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3982_cast_fp16 = slice_by_index(begin = var_3982_begin_0, end = var_3982_end_0, end_mask = var_3982_end_mask_0, squeeze_mask = var_3982_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_3982_cast_fp16")]; + bool var_3983_transpose_x_0 = const()[name = string("op_3983_transpose_x_0"), val = bool(false)]; + bool var_3983_transpose_y_0 = const()[name = string("op_3983_transpose_y_0"), val = bool(false)]; + tensor var_3983_cast_fp16 = matmul(transpose_x = var_3983_transpose_x_0, transpose_y = var_3983_transpose_y_0, x = var_3982_cast_fp16, y = state_201_cast_fp16)[name = string("op_3983_cast_fp16")]; + tensor v_new_201_cast_fp16 = sub(x = var_3980_cast_fp16, y = var_3983_cast_fp16)[name = string("v_new_201_cast_fp16")]; + tensor var_3986_begin_0 = const()[name = string("op_3986_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_3986_end_0 = const()[name = string("op_3986_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_3986_end_mask_0 = const()[name = string("op_3986_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3986_squeeze_mask_0 = const()[name = string("op_3986_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3986_cast_fp16 = slice_by_index(begin = var_3986_begin_0, end = var_3986_end_0, end_mask = var_3986_end_mask_0, squeeze_mask = var_3986_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_3986_cast_fp16")]; + bool var_3987_transpose_x_0 = const()[name = string("op_3987_transpose_x_0"), val = bool(false)]; + bool var_3987_transpose_y_0 = const()[name = string("op_3987_transpose_y_0"), val = bool(false)]; + tensor var_3987_cast_fp16 = matmul(transpose_x = var_3987_transpose_x_0, transpose_y = var_3987_transpose_y_0, x = var_3986_cast_fp16, y = state_201_cast_fp16)[name = string("op_3987_cast_fp16")]; + tensor var_3989_begin_0 = const()[name = string("op_3989_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_3989_end_0 = const()[name = string("op_3989_end_0"), val = tensor([32, 5, 64, 64])]; + tensor var_3989_end_mask_0 = const()[name = string("op_3989_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3989_squeeze_mask_0 = const()[name = string("op_3989_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3989_cast_fp16 = slice_by_index(begin = var_3989_begin_0, end = var_3989_end_0, end_mask = var_3989_end_mask_0, squeeze_mask = var_3989_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_3989_cast_fp16")]; + bool var_3990_transpose_x_0 = const()[name = string("op_3990_transpose_x_0"), val = bool(false)]; + bool var_3990_transpose_y_0 = const()[name = string("op_3990_transpose_y_0"), val = bool(false)]; + tensor var_3990_cast_fp16 = matmul(transpose_x = var_3990_transpose_x_0, transpose_y = var_3990_transpose_y_0, x = var_3989_cast_fp16, y = v_new_201_cast_fp16)[name = string("op_3990_cast_fp16")]; + tensor var_3991_cast_fp16 = add(x = var_3987_cast_fp16, y = var_3990_cast_fp16)[name = string("op_3991_cast_fp16")]; + tensor var_3993_begin_0 = const()[name = string("op_3993_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_3993_end_0 = const()[name = string("op_3993_end_0"), val = tensor([32, 5, 1, 1])]; + tensor var_3993_end_mask_0 = const()[name = string("op_3993_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3993_squeeze_mask_0 = const()[name = string("op_3993_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3993_cast_fp16 = slice_by_index(begin = var_3993_begin_0, end = var_3993_end_0, end_mask = var_3993_end_mask_0, squeeze_mask = var_3993_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_3993_cast_fp16")]; + tensor var_3994_cast_fp16 = mul(x = state_201_cast_fp16, y = var_3993_cast_fp16)[name = string("op_3994_cast_fp16")]; + tensor var_3996_begin_0 = const()[name = string("op_3996_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_3996_end_0 = const()[name = string("op_3996_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_3996_end_mask_0 = const()[name = string("op_3996_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3996_squeeze_mask_0 = const()[name = string("op_3996_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, squeeze_mask = var_3996_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_3996_cast_fp16")]; + bool var_3998_transpose_x_1 = const()[name = string("op_3998_transpose_x_1"), val = bool(true)]; + bool var_3998_transpose_y_1 = const()[name = string("op_3998_transpose_y_1"), val = bool(false)]; + tensor var_3998_cast_fp16 = matmul(transpose_x = var_3998_transpose_x_1, transpose_y = var_3998_transpose_y_1, x = var_3996_cast_fp16, y = v_new_201_cast_fp16)[name = string("op_3998_cast_fp16")]; + tensor state_203_cast_fp16 = add(x = var_3994_cast_fp16, y = var_3998_cast_fp16)[name = string("state_203_cast_fp16")]; + tensor var_4001_begin_0 = const()[name = string("op_4001_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_4001_end_0 = const()[name = string("op_4001_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_4001_end_mask_0 = const()[name = string("op_4001_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4001_squeeze_mask_0 = const()[name = string("op_4001_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4001_cast_fp16 = slice_by_index(begin = var_4001_begin_0, end = var_4001_end_0, end_mask = var_4001_end_mask_0, squeeze_mask = var_4001_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4001_cast_fp16")]; + tensor var_4003_begin_0 = const()[name = string("op_4003_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_4003_end_0 = const()[name = string("op_4003_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_4003_end_mask_0 = const()[name = string("op_4003_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4003_squeeze_mask_0 = const()[name = string("op_4003_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4003_cast_fp16 = slice_by_index(begin = var_4003_begin_0, end = var_4003_end_0, end_mask = var_4003_end_mask_0, squeeze_mask = var_4003_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4003_cast_fp16")]; + bool var_4004_transpose_x_0 = const()[name = string("op_4004_transpose_x_0"), val = bool(false)]; + bool var_4004_transpose_y_0 = const()[name = string("op_4004_transpose_y_0"), val = bool(false)]; + tensor var_4004_cast_fp16 = matmul(transpose_x = var_4004_transpose_x_0, transpose_y = var_4004_transpose_y_0, x = var_4003_cast_fp16, y = state_203_cast_fp16)[name = string("op_4004_cast_fp16")]; + tensor v_new_203_cast_fp16 = sub(x = var_4001_cast_fp16, y = var_4004_cast_fp16)[name = string("v_new_203_cast_fp16")]; + tensor var_4007_begin_0 = const()[name = string("op_4007_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_4007_end_0 = const()[name = string("op_4007_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_4007_end_mask_0 = const()[name = string("op_4007_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4007_squeeze_mask_0 = const()[name = string("op_4007_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4007_cast_fp16 = slice_by_index(begin = var_4007_begin_0, end = var_4007_end_0, end_mask = var_4007_end_mask_0, squeeze_mask = var_4007_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4007_cast_fp16")]; + bool var_4008_transpose_x_0 = const()[name = string("op_4008_transpose_x_0"), val = bool(false)]; + bool var_4008_transpose_y_0 = const()[name = string("op_4008_transpose_y_0"), val = bool(false)]; + tensor var_4008_cast_fp16 = matmul(transpose_x = var_4008_transpose_x_0, transpose_y = var_4008_transpose_y_0, x = var_4007_cast_fp16, y = state_203_cast_fp16)[name = string("op_4008_cast_fp16")]; + tensor var_4010_begin_0 = const()[name = string("op_4010_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_4010_end_0 = const()[name = string("op_4010_end_0"), val = tensor([32, 6, 64, 64])]; + tensor var_4010_end_mask_0 = const()[name = string("op_4010_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4010_squeeze_mask_0 = const()[name = string("op_4010_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4010_cast_fp16 = slice_by_index(begin = var_4010_begin_0, end = var_4010_end_0, end_mask = var_4010_end_mask_0, squeeze_mask = var_4010_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4010_cast_fp16")]; + bool var_4011_transpose_x_0 = const()[name = string("op_4011_transpose_x_0"), val = bool(false)]; + bool var_4011_transpose_y_0 = const()[name = string("op_4011_transpose_y_0"), val = bool(false)]; + tensor var_4011_cast_fp16 = matmul(transpose_x = var_4011_transpose_x_0, transpose_y = var_4011_transpose_y_0, x = var_4010_cast_fp16, y = v_new_203_cast_fp16)[name = string("op_4011_cast_fp16")]; + tensor var_4012_cast_fp16 = add(x = var_4008_cast_fp16, y = var_4011_cast_fp16)[name = string("op_4012_cast_fp16")]; + tensor var_4014_begin_0 = const()[name = string("op_4014_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_4014_end_0 = const()[name = string("op_4014_end_0"), val = tensor([32, 6, 1, 1])]; + tensor var_4014_end_mask_0 = const()[name = string("op_4014_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4014_squeeze_mask_0 = const()[name = string("op_4014_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4014_cast_fp16 = slice_by_index(begin = var_4014_begin_0, end = var_4014_end_0, end_mask = var_4014_end_mask_0, squeeze_mask = var_4014_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4014_cast_fp16")]; + tensor var_4015_cast_fp16 = mul(x = state_203_cast_fp16, y = var_4014_cast_fp16)[name = string("op_4015_cast_fp16")]; + tensor var_4017_begin_0 = const()[name = string("op_4017_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_4017_end_0 = const()[name = string("op_4017_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_4017_end_mask_0 = const()[name = string("op_4017_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4017_squeeze_mask_0 = const()[name = string("op_4017_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4017_cast_fp16 = slice_by_index(begin = var_4017_begin_0, end = var_4017_end_0, end_mask = var_4017_end_mask_0, squeeze_mask = var_4017_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4017_cast_fp16")]; + bool var_4019_transpose_x_1 = const()[name = string("op_4019_transpose_x_1"), val = bool(true)]; + bool var_4019_transpose_y_1 = const()[name = string("op_4019_transpose_y_1"), val = bool(false)]; + tensor var_4019_cast_fp16 = matmul(transpose_x = var_4019_transpose_x_1, transpose_y = var_4019_transpose_y_1, x = var_4017_cast_fp16, y = v_new_203_cast_fp16)[name = string("op_4019_cast_fp16")]; + tensor state_205_cast_fp16 = add(x = var_4015_cast_fp16, y = var_4019_cast_fp16)[name = string("state_205_cast_fp16")]; + tensor var_4022_begin_0 = const()[name = string("op_4022_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_4022_end_0 = const()[name = string("op_4022_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_4022_end_mask_0 = const()[name = string("op_4022_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4022_squeeze_mask_0 = const()[name = string("op_4022_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4022_cast_fp16 = slice_by_index(begin = var_4022_begin_0, end = var_4022_end_0, end_mask = var_4022_end_mask_0, squeeze_mask = var_4022_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4022_cast_fp16")]; + tensor var_4024_begin_0 = const()[name = string("op_4024_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_4024_end_0 = const()[name = string("op_4024_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_4024_end_mask_0 = const()[name = string("op_4024_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4024_squeeze_mask_0 = const()[name = string("op_4024_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4024_cast_fp16 = slice_by_index(begin = var_4024_begin_0, end = var_4024_end_0, end_mask = var_4024_end_mask_0, squeeze_mask = var_4024_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4024_cast_fp16")]; + bool var_4025_transpose_x_0 = const()[name = string("op_4025_transpose_x_0"), val = bool(false)]; + bool var_4025_transpose_y_0 = const()[name = string("op_4025_transpose_y_0"), val = bool(false)]; + tensor var_4025_cast_fp16 = matmul(transpose_x = var_4025_transpose_x_0, transpose_y = var_4025_transpose_y_0, x = var_4024_cast_fp16, y = state_205_cast_fp16)[name = string("op_4025_cast_fp16")]; + tensor v_new_205_cast_fp16 = sub(x = var_4022_cast_fp16, y = var_4025_cast_fp16)[name = string("v_new_205_cast_fp16")]; + tensor var_4028_begin_0 = const()[name = string("op_4028_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_4028_end_0 = const()[name = string("op_4028_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_4028_end_mask_0 = const()[name = string("op_4028_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4028_squeeze_mask_0 = const()[name = string("op_4028_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4028_cast_fp16 = slice_by_index(begin = var_4028_begin_0, end = var_4028_end_0, end_mask = var_4028_end_mask_0, squeeze_mask = var_4028_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4028_cast_fp16")]; + bool var_4029_transpose_x_0 = const()[name = string("op_4029_transpose_x_0"), val = bool(false)]; + bool var_4029_transpose_y_0 = const()[name = string("op_4029_transpose_y_0"), val = bool(false)]; + tensor var_4029_cast_fp16 = matmul(transpose_x = var_4029_transpose_x_0, transpose_y = var_4029_transpose_y_0, x = var_4028_cast_fp16, y = state_205_cast_fp16)[name = string("op_4029_cast_fp16")]; + tensor var_4031_begin_0 = const()[name = string("op_4031_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_4031_end_0 = const()[name = string("op_4031_end_0"), val = tensor([32, 7, 64, 64])]; + tensor var_4031_end_mask_0 = const()[name = string("op_4031_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4031_squeeze_mask_0 = const()[name = string("op_4031_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4031_cast_fp16 = slice_by_index(begin = var_4031_begin_0, end = var_4031_end_0, end_mask = var_4031_end_mask_0, squeeze_mask = var_4031_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4031_cast_fp16")]; + bool var_4032_transpose_x_0 = const()[name = string("op_4032_transpose_x_0"), val = bool(false)]; + bool var_4032_transpose_y_0 = const()[name = string("op_4032_transpose_y_0"), val = bool(false)]; + tensor var_4032_cast_fp16 = matmul(transpose_x = var_4032_transpose_x_0, transpose_y = var_4032_transpose_y_0, x = var_4031_cast_fp16, y = v_new_205_cast_fp16)[name = string("op_4032_cast_fp16")]; + tensor var_4033_cast_fp16 = add(x = var_4029_cast_fp16, y = var_4032_cast_fp16)[name = string("op_4033_cast_fp16")]; + tensor var_4035_begin_0 = const()[name = string("op_4035_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_4035_end_0 = const()[name = string("op_4035_end_0"), val = tensor([32, 7, 1, 1])]; + tensor var_4035_end_mask_0 = const()[name = string("op_4035_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4035_squeeze_mask_0 = const()[name = string("op_4035_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4035_cast_fp16 = slice_by_index(begin = var_4035_begin_0, end = var_4035_end_0, end_mask = var_4035_end_mask_0, squeeze_mask = var_4035_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4035_cast_fp16")]; + tensor var_4036_cast_fp16 = mul(x = state_205_cast_fp16, y = var_4035_cast_fp16)[name = string("op_4036_cast_fp16")]; + tensor var_4038_begin_0 = const()[name = string("op_4038_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_4038_end_0 = const()[name = string("op_4038_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_4038_end_mask_0 = const()[name = string("op_4038_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4038_squeeze_mask_0 = const()[name = string("op_4038_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4038_cast_fp16 = slice_by_index(begin = var_4038_begin_0, end = var_4038_end_0, end_mask = var_4038_end_mask_0, squeeze_mask = var_4038_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4038_cast_fp16")]; + bool var_4040_transpose_x_1 = const()[name = string("op_4040_transpose_x_1"), val = bool(true)]; + bool var_4040_transpose_y_1 = const()[name = string("op_4040_transpose_y_1"), val = bool(false)]; + tensor var_4040_cast_fp16 = matmul(transpose_x = var_4040_transpose_x_1, transpose_y = var_4040_transpose_y_1, x = var_4038_cast_fp16, y = v_new_205_cast_fp16)[name = string("op_4040_cast_fp16")]; + tensor state_207_cast_fp16 = add(x = var_4036_cast_fp16, y = var_4040_cast_fp16)[name = string("state_207_cast_fp16")]; + tensor var_4043_begin_0 = const()[name = string("op_4043_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_4043_end_0 = const()[name = string("op_4043_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_4043_end_mask_0 = const()[name = string("op_4043_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4043_squeeze_mask_0 = const()[name = string("op_4043_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4043_cast_fp16 = slice_by_index(begin = var_4043_begin_0, end = var_4043_end_0, end_mask = var_4043_end_mask_0, squeeze_mask = var_4043_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4043_cast_fp16")]; + tensor var_4045_begin_0 = const()[name = string("op_4045_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_4045_end_0 = const()[name = string("op_4045_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_4045_end_mask_0 = const()[name = string("op_4045_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4045_squeeze_mask_0 = const()[name = string("op_4045_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4045_cast_fp16 = slice_by_index(begin = var_4045_begin_0, end = var_4045_end_0, end_mask = var_4045_end_mask_0, squeeze_mask = var_4045_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4045_cast_fp16")]; + bool var_4046_transpose_x_0 = const()[name = string("op_4046_transpose_x_0"), val = bool(false)]; + bool var_4046_transpose_y_0 = const()[name = string("op_4046_transpose_y_0"), val = bool(false)]; + tensor var_4046_cast_fp16 = matmul(transpose_x = var_4046_transpose_x_0, transpose_y = var_4046_transpose_y_0, x = var_4045_cast_fp16, y = state_207_cast_fp16)[name = string("op_4046_cast_fp16")]; + tensor v_new_207_cast_fp16 = sub(x = var_4043_cast_fp16, y = var_4046_cast_fp16)[name = string("v_new_207_cast_fp16")]; + tensor var_4049_begin_0 = const()[name = string("op_4049_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_4049_end_0 = const()[name = string("op_4049_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_4049_end_mask_0 = const()[name = string("op_4049_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4049_squeeze_mask_0 = const()[name = string("op_4049_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4049_cast_fp16 = slice_by_index(begin = var_4049_begin_0, end = var_4049_end_0, end_mask = var_4049_end_mask_0, squeeze_mask = var_4049_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4049_cast_fp16")]; + bool var_4050_transpose_x_0 = const()[name = string("op_4050_transpose_x_0"), val = bool(false)]; + bool var_4050_transpose_y_0 = const()[name = string("op_4050_transpose_y_0"), val = bool(false)]; + tensor var_4050_cast_fp16 = matmul(transpose_x = var_4050_transpose_x_0, transpose_y = var_4050_transpose_y_0, x = var_4049_cast_fp16, y = state_207_cast_fp16)[name = string("op_4050_cast_fp16")]; + tensor var_4052_begin_0 = const()[name = string("op_4052_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_4052_end_0 = const()[name = string("op_4052_end_0"), val = tensor([32, 8, 64, 64])]; + tensor var_4052_end_mask_0 = const()[name = string("op_4052_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4052_squeeze_mask_0 = const()[name = string("op_4052_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4052_cast_fp16 = slice_by_index(begin = var_4052_begin_0, end = var_4052_end_0, end_mask = var_4052_end_mask_0, squeeze_mask = var_4052_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4052_cast_fp16")]; + bool var_4053_transpose_x_0 = const()[name = string("op_4053_transpose_x_0"), val = bool(false)]; + bool var_4053_transpose_y_0 = const()[name = string("op_4053_transpose_y_0"), val = bool(false)]; + tensor var_4053_cast_fp16 = matmul(transpose_x = var_4053_transpose_x_0, transpose_y = var_4053_transpose_y_0, x = var_4052_cast_fp16, y = v_new_207_cast_fp16)[name = string("op_4053_cast_fp16")]; + tensor var_4054_cast_fp16 = add(x = var_4050_cast_fp16, y = var_4053_cast_fp16)[name = string("op_4054_cast_fp16")]; + tensor var_4056_begin_0 = const()[name = string("op_4056_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_4056_end_0 = const()[name = string("op_4056_end_0"), val = tensor([32, 8, 1, 1])]; + tensor var_4056_end_mask_0 = const()[name = string("op_4056_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4056_squeeze_mask_0 = const()[name = string("op_4056_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4056_cast_fp16 = slice_by_index(begin = var_4056_begin_0, end = var_4056_end_0, end_mask = var_4056_end_mask_0, squeeze_mask = var_4056_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4056_cast_fp16")]; + tensor var_4057_cast_fp16 = mul(x = state_207_cast_fp16, y = var_4056_cast_fp16)[name = string("op_4057_cast_fp16")]; + tensor var_4059_begin_0 = const()[name = string("op_4059_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_4059_end_0 = const()[name = string("op_4059_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_4059_end_mask_0 = const()[name = string("op_4059_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4059_squeeze_mask_0 = const()[name = string("op_4059_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4059_cast_fp16 = slice_by_index(begin = var_4059_begin_0, end = var_4059_end_0, end_mask = var_4059_end_mask_0, squeeze_mask = var_4059_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4059_cast_fp16")]; + bool var_4061_transpose_x_1 = const()[name = string("op_4061_transpose_x_1"), val = bool(true)]; + bool var_4061_transpose_y_1 = const()[name = string("op_4061_transpose_y_1"), val = bool(false)]; + tensor var_4061_cast_fp16 = matmul(transpose_x = var_4061_transpose_x_1, transpose_y = var_4061_transpose_y_1, x = var_4059_cast_fp16, y = v_new_207_cast_fp16)[name = string("op_4061_cast_fp16")]; + tensor state_209_cast_fp16 = add(x = var_4057_cast_fp16, y = var_4061_cast_fp16)[name = string("state_209_cast_fp16")]; + tensor var_4064_begin_0 = const()[name = string("op_4064_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_4064_end_0 = const()[name = string("op_4064_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_4064_end_mask_0 = const()[name = string("op_4064_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4064_squeeze_mask_0 = const()[name = string("op_4064_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4064_cast_fp16 = slice_by_index(begin = var_4064_begin_0, end = var_4064_end_0, end_mask = var_4064_end_mask_0, squeeze_mask = var_4064_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4064_cast_fp16")]; + tensor var_4066_begin_0 = const()[name = string("op_4066_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_4066_end_0 = const()[name = string("op_4066_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_4066_end_mask_0 = const()[name = string("op_4066_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4066_squeeze_mask_0 = const()[name = string("op_4066_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4066_cast_fp16 = slice_by_index(begin = var_4066_begin_0, end = var_4066_end_0, end_mask = var_4066_end_mask_0, squeeze_mask = var_4066_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4066_cast_fp16")]; + bool var_4067_transpose_x_0 = const()[name = string("op_4067_transpose_x_0"), val = bool(false)]; + bool var_4067_transpose_y_0 = const()[name = string("op_4067_transpose_y_0"), val = bool(false)]; + tensor var_4067_cast_fp16 = matmul(transpose_x = var_4067_transpose_x_0, transpose_y = var_4067_transpose_y_0, x = var_4066_cast_fp16, y = state_209_cast_fp16)[name = string("op_4067_cast_fp16")]; + tensor v_new_209_cast_fp16 = sub(x = var_4064_cast_fp16, y = var_4067_cast_fp16)[name = string("v_new_209_cast_fp16")]; + tensor var_4070_begin_0 = const()[name = string("op_4070_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_4070_end_0 = const()[name = string("op_4070_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_4070_end_mask_0 = const()[name = string("op_4070_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4070_squeeze_mask_0 = const()[name = string("op_4070_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4070_cast_fp16 = slice_by_index(begin = var_4070_begin_0, end = var_4070_end_0, end_mask = var_4070_end_mask_0, squeeze_mask = var_4070_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4070_cast_fp16")]; + bool var_4071_transpose_x_0 = const()[name = string("op_4071_transpose_x_0"), val = bool(false)]; + bool var_4071_transpose_y_0 = const()[name = string("op_4071_transpose_y_0"), val = bool(false)]; + tensor var_4071_cast_fp16 = matmul(transpose_x = var_4071_transpose_x_0, transpose_y = var_4071_transpose_y_0, x = var_4070_cast_fp16, y = state_209_cast_fp16)[name = string("op_4071_cast_fp16")]; + tensor var_4073_begin_0 = const()[name = string("op_4073_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_4073_end_0 = const()[name = string("op_4073_end_0"), val = tensor([32, 9, 64, 64])]; + tensor var_4073_end_mask_0 = const()[name = string("op_4073_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4073_squeeze_mask_0 = const()[name = string("op_4073_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4073_cast_fp16 = slice_by_index(begin = var_4073_begin_0, end = var_4073_end_0, end_mask = var_4073_end_mask_0, squeeze_mask = var_4073_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4073_cast_fp16")]; + bool var_4074_transpose_x_0 = const()[name = string("op_4074_transpose_x_0"), val = bool(false)]; + bool var_4074_transpose_y_0 = const()[name = string("op_4074_transpose_y_0"), val = bool(false)]; + tensor var_4074_cast_fp16 = matmul(transpose_x = var_4074_transpose_x_0, transpose_y = var_4074_transpose_y_0, x = var_4073_cast_fp16, y = v_new_209_cast_fp16)[name = string("op_4074_cast_fp16")]; + tensor var_4075_cast_fp16 = add(x = var_4071_cast_fp16, y = var_4074_cast_fp16)[name = string("op_4075_cast_fp16")]; + tensor var_4077_begin_0 = const()[name = string("op_4077_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_4077_end_0 = const()[name = string("op_4077_end_0"), val = tensor([32, 9, 1, 1])]; + tensor var_4077_end_mask_0 = const()[name = string("op_4077_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4077_squeeze_mask_0 = const()[name = string("op_4077_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4077_cast_fp16 = slice_by_index(begin = var_4077_begin_0, end = var_4077_end_0, end_mask = var_4077_end_mask_0, squeeze_mask = var_4077_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4077_cast_fp16")]; + tensor var_4078_cast_fp16 = mul(x = state_209_cast_fp16, y = var_4077_cast_fp16)[name = string("op_4078_cast_fp16")]; + tensor var_4080_begin_0 = const()[name = string("op_4080_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_4080_end_0 = const()[name = string("op_4080_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_4080_end_mask_0 = const()[name = string("op_4080_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4080_squeeze_mask_0 = const()[name = string("op_4080_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4080_cast_fp16 = slice_by_index(begin = var_4080_begin_0, end = var_4080_end_0, end_mask = var_4080_end_mask_0, squeeze_mask = var_4080_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4080_cast_fp16")]; + bool var_4082_transpose_x_1 = const()[name = string("op_4082_transpose_x_1"), val = bool(true)]; + bool var_4082_transpose_y_1 = const()[name = string("op_4082_transpose_y_1"), val = bool(false)]; + tensor var_4082_cast_fp16 = matmul(transpose_x = var_4082_transpose_x_1, transpose_y = var_4082_transpose_y_1, x = var_4080_cast_fp16, y = v_new_209_cast_fp16)[name = string("op_4082_cast_fp16")]; + tensor state_211_cast_fp16 = add(x = var_4078_cast_fp16, y = var_4082_cast_fp16)[name = string("state_211_cast_fp16")]; + tensor var_4085_begin_0 = const()[name = string("op_4085_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_4085_end_0 = const()[name = string("op_4085_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_4085_end_mask_0 = const()[name = string("op_4085_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4085_squeeze_mask_0 = const()[name = string("op_4085_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4085_cast_fp16 = slice_by_index(begin = var_4085_begin_0, end = var_4085_end_0, end_mask = var_4085_end_mask_0, squeeze_mask = var_4085_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4085_cast_fp16")]; + tensor var_4087_begin_0 = const()[name = string("op_4087_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_4087_end_0 = const()[name = string("op_4087_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_4087_end_mask_0 = const()[name = string("op_4087_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4087_squeeze_mask_0 = const()[name = string("op_4087_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4087_cast_fp16 = slice_by_index(begin = var_4087_begin_0, end = var_4087_end_0, end_mask = var_4087_end_mask_0, squeeze_mask = var_4087_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4087_cast_fp16")]; + bool var_4088_transpose_x_0 = const()[name = string("op_4088_transpose_x_0"), val = bool(false)]; + bool var_4088_transpose_y_0 = const()[name = string("op_4088_transpose_y_0"), val = bool(false)]; + tensor var_4088_cast_fp16 = matmul(transpose_x = var_4088_transpose_x_0, transpose_y = var_4088_transpose_y_0, x = var_4087_cast_fp16, y = state_211_cast_fp16)[name = string("op_4088_cast_fp16")]; + tensor v_new_211_cast_fp16 = sub(x = var_4085_cast_fp16, y = var_4088_cast_fp16)[name = string("v_new_211_cast_fp16")]; + tensor var_4091_begin_0 = const()[name = string("op_4091_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_4091_end_0 = const()[name = string("op_4091_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_4091_end_mask_0 = const()[name = string("op_4091_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4091_squeeze_mask_0 = const()[name = string("op_4091_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4091_cast_fp16 = slice_by_index(begin = var_4091_begin_0, end = var_4091_end_0, end_mask = var_4091_end_mask_0, squeeze_mask = var_4091_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4091_cast_fp16")]; + bool var_4092_transpose_x_0 = const()[name = string("op_4092_transpose_x_0"), val = bool(false)]; + bool var_4092_transpose_y_0 = const()[name = string("op_4092_transpose_y_0"), val = bool(false)]; + tensor var_4092_cast_fp16 = matmul(transpose_x = var_4092_transpose_x_0, transpose_y = var_4092_transpose_y_0, x = var_4091_cast_fp16, y = state_211_cast_fp16)[name = string("op_4092_cast_fp16")]; + tensor var_4094_begin_0 = const()[name = string("op_4094_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_4094_end_0 = const()[name = string("op_4094_end_0"), val = tensor([32, 10, 64, 64])]; + tensor var_4094_end_mask_0 = const()[name = string("op_4094_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4094_squeeze_mask_0 = const()[name = string("op_4094_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4094_cast_fp16 = slice_by_index(begin = var_4094_begin_0, end = var_4094_end_0, end_mask = var_4094_end_mask_0, squeeze_mask = var_4094_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4094_cast_fp16")]; + bool var_4095_transpose_x_0 = const()[name = string("op_4095_transpose_x_0"), val = bool(false)]; + bool var_4095_transpose_y_0 = const()[name = string("op_4095_transpose_y_0"), val = bool(false)]; + tensor var_4095_cast_fp16 = matmul(transpose_x = var_4095_transpose_x_0, transpose_y = var_4095_transpose_y_0, x = var_4094_cast_fp16, y = v_new_211_cast_fp16)[name = string("op_4095_cast_fp16")]; + tensor var_4096_cast_fp16 = add(x = var_4092_cast_fp16, y = var_4095_cast_fp16)[name = string("op_4096_cast_fp16")]; + tensor var_4098_begin_0 = const()[name = string("op_4098_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_4098_end_0 = const()[name = string("op_4098_end_0"), val = tensor([32, 10, 1, 1])]; + tensor var_4098_end_mask_0 = const()[name = string("op_4098_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4098_squeeze_mask_0 = const()[name = string("op_4098_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4098_cast_fp16 = slice_by_index(begin = var_4098_begin_0, end = var_4098_end_0, end_mask = var_4098_end_mask_0, squeeze_mask = var_4098_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4098_cast_fp16")]; + tensor var_4099_cast_fp16 = mul(x = state_211_cast_fp16, y = var_4098_cast_fp16)[name = string("op_4099_cast_fp16")]; + tensor var_4101_begin_0 = const()[name = string("op_4101_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_4101_end_0 = const()[name = string("op_4101_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_4101_end_mask_0 = const()[name = string("op_4101_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4101_squeeze_mask_0 = const()[name = string("op_4101_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4101_cast_fp16 = slice_by_index(begin = var_4101_begin_0, end = var_4101_end_0, end_mask = var_4101_end_mask_0, squeeze_mask = var_4101_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4101_cast_fp16")]; + bool var_4103_transpose_x_1 = const()[name = string("op_4103_transpose_x_1"), val = bool(true)]; + bool var_4103_transpose_y_1 = const()[name = string("op_4103_transpose_y_1"), val = bool(false)]; + tensor var_4103_cast_fp16 = matmul(transpose_x = var_4103_transpose_x_1, transpose_y = var_4103_transpose_y_1, x = var_4101_cast_fp16, y = v_new_211_cast_fp16)[name = string("op_4103_cast_fp16")]; + tensor state_213_cast_fp16 = add(x = var_4099_cast_fp16, y = var_4103_cast_fp16)[name = string("state_213_cast_fp16")]; + tensor var_4106_begin_0 = const()[name = string("op_4106_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_4106_end_0 = const()[name = string("op_4106_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_4106_end_mask_0 = const()[name = string("op_4106_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4106_squeeze_mask_0 = const()[name = string("op_4106_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4106_cast_fp16 = slice_by_index(begin = var_4106_begin_0, end = var_4106_end_0, end_mask = var_4106_end_mask_0, squeeze_mask = var_4106_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4106_cast_fp16")]; + tensor var_4108_begin_0 = const()[name = string("op_4108_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_4108_end_0 = const()[name = string("op_4108_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_4108_end_mask_0 = const()[name = string("op_4108_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4108_squeeze_mask_0 = const()[name = string("op_4108_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4108_cast_fp16 = slice_by_index(begin = var_4108_begin_0, end = var_4108_end_0, end_mask = var_4108_end_mask_0, squeeze_mask = var_4108_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4108_cast_fp16")]; + bool var_4109_transpose_x_0 = const()[name = string("op_4109_transpose_x_0"), val = bool(false)]; + bool var_4109_transpose_y_0 = const()[name = string("op_4109_transpose_y_0"), val = bool(false)]; + tensor var_4109_cast_fp16 = matmul(transpose_x = var_4109_transpose_x_0, transpose_y = var_4109_transpose_y_0, x = var_4108_cast_fp16, y = state_213_cast_fp16)[name = string("op_4109_cast_fp16")]; + tensor v_new_213_cast_fp16 = sub(x = var_4106_cast_fp16, y = var_4109_cast_fp16)[name = string("v_new_213_cast_fp16")]; + tensor var_4112_begin_0 = const()[name = string("op_4112_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_4112_end_0 = const()[name = string("op_4112_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_4112_end_mask_0 = const()[name = string("op_4112_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4112_squeeze_mask_0 = const()[name = string("op_4112_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4112_cast_fp16 = slice_by_index(begin = var_4112_begin_0, end = var_4112_end_0, end_mask = var_4112_end_mask_0, squeeze_mask = var_4112_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4112_cast_fp16")]; + bool var_4113_transpose_x_0 = const()[name = string("op_4113_transpose_x_0"), val = bool(false)]; + bool var_4113_transpose_y_0 = const()[name = string("op_4113_transpose_y_0"), val = bool(false)]; + tensor var_4113_cast_fp16 = matmul(transpose_x = var_4113_transpose_x_0, transpose_y = var_4113_transpose_y_0, x = var_4112_cast_fp16, y = state_213_cast_fp16)[name = string("op_4113_cast_fp16")]; + tensor var_4115_begin_0 = const()[name = string("op_4115_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_4115_end_0 = const()[name = string("op_4115_end_0"), val = tensor([32, 11, 64, 64])]; + tensor var_4115_end_mask_0 = const()[name = string("op_4115_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4115_squeeze_mask_0 = const()[name = string("op_4115_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4115_cast_fp16 = slice_by_index(begin = var_4115_begin_0, end = var_4115_end_0, end_mask = var_4115_end_mask_0, squeeze_mask = var_4115_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4115_cast_fp16")]; + bool var_4116_transpose_x_0 = const()[name = string("op_4116_transpose_x_0"), val = bool(false)]; + bool var_4116_transpose_y_0 = const()[name = string("op_4116_transpose_y_0"), val = bool(false)]; + tensor var_4116_cast_fp16 = matmul(transpose_x = var_4116_transpose_x_0, transpose_y = var_4116_transpose_y_0, x = var_4115_cast_fp16, y = v_new_213_cast_fp16)[name = string("op_4116_cast_fp16")]; + tensor var_4117_cast_fp16 = add(x = var_4113_cast_fp16, y = var_4116_cast_fp16)[name = string("op_4117_cast_fp16")]; + tensor var_4119_begin_0 = const()[name = string("op_4119_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_4119_end_0 = const()[name = string("op_4119_end_0"), val = tensor([32, 11, 1, 1])]; + tensor var_4119_end_mask_0 = const()[name = string("op_4119_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4119_squeeze_mask_0 = const()[name = string("op_4119_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4119_cast_fp16 = slice_by_index(begin = var_4119_begin_0, end = var_4119_end_0, end_mask = var_4119_end_mask_0, squeeze_mask = var_4119_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4119_cast_fp16")]; + tensor var_4120_cast_fp16 = mul(x = state_213_cast_fp16, y = var_4119_cast_fp16)[name = string("op_4120_cast_fp16")]; + tensor var_4122_begin_0 = const()[name = string("op_4122_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_4122_end_0 = const()[name = string("op_4122_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_4122_end_mask_0 = const()[name = string("op_4122_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4122_squeeze_mask_0 = const()[name = string("op_4122_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4122_cast_fp16 = slice_by_index(begin = var_4122_begin_0, end = var_4122_end_0, end_mask = var_4122_end_mask_0, squeeze_mask = var_4122_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4122_cast_fp16")]; + bool var_4124_transpose_x_1 = const()[name = string("op_4124_transpose_x_1"), val = bool(true)]; + bool var_4124_transpose_y_1 = const()[name = string("op_4124_transpose_y_1"), val = bool(false)]; + tensor var_4124_cast_fp16 = matmul(transpose_x = var_4124_transpose_x_1, transpose_y = var_4124_transpose_y_1, x = var_4122_cast_fp16, y = v_new_213_cast_fp16)[name = string("op_4124_cast_fp16")]; + tensor state_215_cast_fp16 = add(x = var_4120_cast_fp16, y = var_4124_cast_fp16)[name = string("state_215_cast_fp16")]; + tensor var_4127_begin_0 = const()[name = string("op_4127_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_4127_end_0 = const()[name = string("op_4127_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_4127_end_mask_0 = const()[name = string("op_4127_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4127_squeeze_mask_0 = const()[name = string("op_4127_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4127_cast_fp16 = slice_by_index(begin = var_4127_begin_0, end = var_4127_end_0, end_mask = var_4127_end_mask_0, squeeze_mask = var_4127_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4127_cast_fp16")]; + tensor var_4129_begin_0 = const()[name = string("op_4129_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_4129_end_0 = const()[name = string("op_4129_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_4129_end_mask_0 = const()[name = string("op_4129_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4129_squeeze_mask_0 = const()[name = string("op_4129_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4129_cast_fp16 = slice_by_index(begin = var_4129_begin_0, end = var_4129_end_0, end_mask = var_4129_end_mask_0, squeeze_mask = var_4129_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4129_cast_fp16")]; + bool var_4130_transpose_x_0 = const()[name = string("op_4130_transpose_x_0"), val = bool(false)]; + bool var_4130_transpose_y_0 = const()[name = string("op_4130_transpose_y_0"), val = bool(false)]; + tensor var_4130_cast_fp16 = matmul(transpose_x = var_4130_transpose_x_0, transpose_y = var_4130_transpose_y_0, x = var_4129_cast_fp16, y = state_215_cast_fp16)[name = string("op_4130_cast_fp16")]; + tensor v_new_215_cast_fp16 = sub(x = var_4127_cast_fp16, y = var_4130_cast_fp16)[name = string("v_new_215_cast_fp16")]; + tensor var_4133_begin_0 = const()[name = string("op_4133_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_4133_end_0 = const()[name = string("op_4133_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_4133_end_mask_0 = const()[name = string("op_4133_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4133_squeeze_mask_0 = const()[name = string("op_4133_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4133_cast_fp16 = slice_by_index(begin = var_4133_begin_0, end = var_4133_end_0, end_mask = var_4133_end_mask_0, squeeze_mask = var_4133_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4133_cast_fp16")]; + bool var_4134_transpose_x_0 = const()[name = string("op_4134_transpose_x_0"), val = bool(false)]; + bool var_4134_transpose_y_0 = const()[name = string("op_4134_transpose_y_0"), val = bool(false)]; + tensor var_4134_cast_fp16 = matmul(transpose_x = var_4134_transpose_x_0, transpose_y = var_4134_transpose_y_0, x = var_4133_cast_fp16, y = state_215_cast_fp16)[name = string("op_4134_cast_fp16")]; + tensor var_4136_begin_0 = const()[name = string("op_4136_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_4136_end_0 = const()[name = string("op_4136_end_0"), val = tensor([32, 12, 64, 64])]; + tensor var_4136_end_mask_0 = const()[name = string("op_4136_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4136_squeeze_mask_0 = const()[name = string("op_4136_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4136_cast_fp16 = slice_by_index(begin = var_4136_begin_0, end = var_4136_end_0, end_mask = var_4136_end_mask_0, squeeze_mask = var_4136_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4136_cast_fp16")]; + bool var_4137_transpose_x_0 = const()[name = string("op_4137_transpose_x_0"), val = bool(false)]; + bool var_4137_transpose_y_0 = const()[name = string("op_4137_transpose_y_0"), val = bool(false)]; + tensor var_4137_cast_fp16 = matmul(transpose_x = var_4137_transpose_x_0, transpose_y = var_4137_transpose_y_0, x = var_4136_cast_fp16, y = v_new_215_cast_fp16)[name = string("op_4137_cast_fp16")]; + tensor var_4138_cast_fp16 = add(x = var_4134_cast_fp16, y = var_4137_cast_fp16)[name = string("op_4138_cast_fp16")]; + tensor var_4140_begin_0 = const()[name = string("op_4140_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_4140_end_0 = const()[name = string("op_4140_end_0"), val = tensor([32, 12, 1, 1])]; + tensor var_4140_end_mask_0 = const()[name = string("op_4140_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4140_squeeze_mask_0 = const()[name = string("op_4140_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4140_cast_fp16 = slice_by_index(begin = var_4140_begin_0, end = var_4140_end_0, end_mask = var_4140_end_mask_0, squeeze_mask = var_4140_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4140_cast_fp16")]; + tensor var_4141_cast_fp16 = mul(x = state_215_cast_fp16, y = var_4140_cast_fp16)[name = string("op_4141_cast_fp16")]; + tensor var_4143_begin_0 = const()[name = string("op_4143_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_4143_end_0 = const()[name = string("op_4143_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_4143_end_mask_0 = const()[name = string("op_4143_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4143_squeeze_mask_0 = const()[name = string("op_4143_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4143_cast_fp16 = slice_by_index(begin = var_4143_begin_0, end = var_4143_end_0, end_mask = var_4143_end_mask_0, squeeze_mask = var_4143_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4143_cast_fp16")]; + bool var_4145_transpose_x_1 = const()[name = string("op_4145_transpose_x_1"), val = bool(true)]; + bool var_4145_transpose_y_1 = const()[name = string("op_4145_transpose_y_1"), val = bool(false)]; + tensor var_4145_cast_fp16 = matmul(transpose_x = var_4145_transpose_x_1, transpose_y = var_4145_transpose_y_1, x = var_4143_cast_fp16, y = v_new_215_cast_fp16)[name = string("op_4145_cast_fp16")]; + tensor state_217_cast_fp16 = add(x = var_4141_cast_fp16, y = var_4145_cast_fp16)[name = string("state_217_cast_fp16")]; + tensor var_4148_begin_0 = const()[name = string("op_4148_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_4148_end_0 = const()[name = string("op_4148_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_4148_end_mask_0 = const()[name = string("op_4148_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4148_squeeze_mask_0 = const()[name = string("op_4148_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4148_cast_fp16 = slice_by_index(begin = var_4148_begin_0, end = var_4148_end_0, end_mask = var_4148_end_mask_0, squeeze_mask = var_4148_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4148_cast_fp16")]; + tensor var_4150_begin_0 = const()[name = string("op_4150_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_4150_end_0 = const()[name = string("op_4150_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_4150_end_mask_0 = const()[name = string("op_4150_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4150_squeeze_mask_0 = const()[name = string("op_4150_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4150_cast_fp16 = slice_by_index(begin = var_4150_begin_0, end = var_4150_end_0, end_mask = var_4150_end_mask_0, squeeze_mask = var_4150_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4150_cast_fp16")]; + bool var_4151_transpose_x_0 = const()[name = string("op_4151_transpose_x_0"), val = bool(false)]; + bool var_4151_transpose_y_0 = const()[name = string("op_4151_transpose_y_0"), val = bool(false)]; + tensor var_4151_cast_fp16 = matmul(transpose_x = var_4151_transpose_x_0, transpose_y = var_4151_transpose_y_0, x = var_4150_cast_fp16, y = state_217_cast_fp16)[name = string("op_4151_cast_fp16")]; + tensor v_new_217_cast_fp16 = sub(x = var_4148_cast_fp16, y = var_4151_cast_fp16)[name = string("v_new_217_cast_fp16")]; + tensor var_4154_begin_0 = const()[name = string("op_4154_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_4154_end_0 = const()[name = string("op_4154_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_4154_end_mask_0 = const()[name = string("op_4154_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4154_squeeze_mask_0 = const()[name = string("op_4154_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4154_cast_fp16 = slice_by_index(begin = var_4154_begin_0, end = var_4154_end_0, end_mask = var_4154_end_mask_0, squeeze_mask = var_4154_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4154_cast_fp16")]; + bool var_4155_transpose_x_0 = const()[name = string("op_4155_transpose_x_0"), val = bool(false)]; + bool var_4155_transpose_y_0 = const()[name = string("op_4155_transpose_y_0"), val = bool(false)]; + tensor var_4155_cast_fp16 = matmul(transpose_x = var_4155_transpose_x_0, transpose_y = var_4155_transpose_y_0, x = var_4154_cast_fp16, y = state_217_cast_fp16)[name = string("op_4155_cast_fp16")]; + tensor var_4157_begin_0 = const()[name = string("op_4157_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_4157_end_0 = const()[name = string("op_4157_end_0"), val = tensor([32, 13, 64, 64])]; + tensor var_4157_end_mask_0 = const()[name = string("op_4157_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4157_squeeze_mask_0 = const()[name = string("op_4157_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4157_cast_fp16 = slice_by_index(begin = var_4157_begin_0, end = var_4157_end_0, end_mask = var_4157_end_mask_0, squeeze_mask = var_4157_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4157_cast_fp16")]; + bool var_4158_transpose_x_0 = const()[name = string("op_4158_transpose_x_0"), val = bool(false)]; + bool var_4158_transpose_y_0 = const()[name = string("op_4158_transpose_y_0"), val = bool(false)]; + tensor var_4158_cast_fp16 = matmul(transpose_x = var_4158_transpose_x_0, transpose_y = var_4158_transpose_y_0, x = var_4157_cast_fp16, y = v_new_217_cast_fp16)[name = string("op_4158_cast_fp16")]; + tensor var_4159_cast_fp16 = add(x = var_4155_cast_fp16, y = var_4158_cast_fp16)[name = string("op_4159_cast_fp16")]; + tensor var_4161_begin_0 = const()[name = string("op_4161_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_4161_end_0 = const()[name = string("op_4161_end_0"), val = tensor([32, 13, 1, 1])]; + tensor var_4161_end_mask_0 = const()[name = string("op_4161_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4161_squeeze_mask_0 = const()[name = string("op_4161_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4161_cast_fp16 = slice_by_index(begin = var_4161_begin_0, end = var_4161_end_0, end_mask = var_4161_end_mask_0, squeeze_mask = var_4161_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4161_cast_fp16")]; + tensor var_4162_cast_fp16 = mul(x = state_217_cast_fp16, y = var_4161_cast_fp16)[name = string("op_4162_cast_fp16")]; + tensor var_4164_begin_0 = const()[name = string("op_4164_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_4164_end_0 = const()[name = string("op_4164_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_4164_end_mask_0 = const()[name = string("op_4164_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4164_squeeze_mask_0 = const()[name = string("op_4164_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4164_cast_fp16 = slice_by_index(begin = var_4164_begin_0, end = var_4164_end_0, end_mask = var_4164_end_mask_0, squeeze_mask = var_4164_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4164_cast_fp16")]; + bool var_4166_transpose_x_1 = const()[name = string("op_4166_transpose_x_1"), val = bool(true)]; + bool var_4166_transpose_y_1 = const()[name = string("op_4166_transpose_y_1"), val = bool(false)]; + tensor var_4166_cast_fp16 = matmul(transpose_x = var_4166_transpose_x_1, transpose_y = var_4166_transpose_y_1, x = var_4164_cast_fp16, y = v_new_217_cast_fp16)[name = string("op_4166_cast_fp16")]; + tensor state_219_cast_fp16 = add(x = var_4162_cast_fp16, y = var_4166_cast_fp16)[name = string("state_219_cast_fp16")]; + tensor var_4169_begin_0 = const()[name = string("op_4169_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_4169_end_0 = const()[name = string("op_4169_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_4169_end_mask_0 = const()[name = string("op_4169_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4169_squeeze_mask_0 = const()[name = string("op_4169_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4169_cast_fp16 = slice_by_index(begin = var_4169_begin_0, end = var_4169_end_0, end_mask = var_4169_end_mask_0, squeeze_mask = var_4169_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4169_cast_fp16")]; + tensor var_4171_begin_0 = const()[name = string("op_4171_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_4171_end_0 = const()[name = string("op_4171_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_4171_end_mask_0 = const()[name = string("op_4171_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4171_squeeze_mask_0 = const()[name = string("op_4171_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4171_cast_fp16 = slice_by_index(begin = var_4171_begin_0, end = var_4171_end_0, end_mask = var_4171_end_mask_0, squeeze_mask = var_4171_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4171_cast_fp16")]; + bool var_4172_transpose_x_0 = const()[name = string("op_4172_transpose_x_0"), val = bool(false)]; + bool var_4172_transpose_y_0 = const()[name = string("op_4172_transpose_y_0"), val = bool(false)]; + tensor var_4172_cast_fp16 = matmul(transpose_x = var_4172_transpose_x_0, transpose_y = var_4172_transpose_y_0, x = var_4171_cast_fp16, y = state_219_cast_fp16)[name = string("op_4172_cast_fp16")]; + tensor v_new_219_cast_fp16 = sub(x = var_4169_cast_fp16, y = var_4172_cast_fp16)[name = string("v_new_219_cast_fp16")]; + tensor var_4175_begin_0 = const()[name = string("op_4175_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_4175_end_0 = const()[name = string("op_4175_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_4175_end_mask_0 = const()[name = string("op_4175_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4175_squeeze_mask_0 = const()[name = string("op_4175_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4175_cast_fp16 = slice_by_index(begin = var_4175_begin_0, end = var_4175_end_0, end_mask = var_4175_end_mask_0, squeeze_mask = var_4175_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4175_cast_fp16")]; + bool var_4176_transpose_x_0 = const()[name = string("op_4176_transpose_x_0"), val = bool(false)]; + bool var_4176_transpose_y_0 = const()[name = string("op_4176_transpose_y_0"), val = bool(false)]; + tensor var_4176_cast_fp16 = matmul(transpose_x = var_4176_transpose_x_0, transpose_y = var_4176_transpose_y_0, x = var_4175_cast_fp16, y = state_219_cast_fp16)[name = string("op_4176_cast_fp16")]; + tensor var_4178_begin_0 = const()[name = string("op_4178_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_4178_end_0 = const()[name = string("op_4178_end_0"), val = tensor([32, 14, 64, 64])]; + tensor var_4178_end_mask_0 = const()[name = string("op_4178_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4178_squeeze_mask_0 = const()[name = string("op_4178_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4178_cast_fp16 = slice_by_index(begin = var_4178_begin_0, end = var_4178_end_0, end_mask = var_4178_end_mask_0, squeeze_mask = var_4178_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4178_cast_fp16")]; + bool var_4179_transpose_x_0 = const()[name = string("op_4179_transpose_x_0"), val = bool(false)]; + bool var_4179_transpose_y_0 = const()[name = string("op_4179_transpose_y_0"), val = bool(false)]; + tensor var_4179_cast_fp16 = matmul(transpose_x = var_4179_transpose_x_0, transpose_y = var_4179_transpose_y_0, x = var_4178_cast_fp16, y = v_new_219_cast_fp16)[name = string("op_4179_cast_fp16")]; + tensor var_4180_cast_fp16 = add(x = var_4176_cast_fp16, y = var_4179_cast_fp16)[name = string("op_4180_cast_fp16")]; + tensor var_4182_begin_0 = const()[name = string("op_4182_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_4182_end_0 = const()[name = string("op_4182_end_0"), val = tensor([32, 14, 1, 1])]; + tensor var_4182_end_mask_0 = const()[name = string("op_4182_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4182_squeeze_mask_0 = const()[name = string("op_4182_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4182_cast_fp16 = slice_by_index(begin = var_4182_begin_0, end = var_4182_end_0, end_mask = var_4182_end_mask_0, squeeze_mask = var_4182_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4182_cast_fp16")]; + tensor var_4183_cast_fp16 = mul(x = state_219_cast_fp16, y = var_4182_cast_fp16)[name = string("op_4183_cast_fp16")]; + tensor var_4185_begin_0 = const()[name = string("op_4185_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_4185_end_0 = const()[name = string("op_4185_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_4185_end_mask_0 = const()[name = string("op_4185_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4185_squeeze_mask_0 = const()[name = string("op_4185_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4185_cast_fp16 = slice_by_index(begin = var_4185_begin_0, end = var_4185_end_0, end_mask = var_4185_end_mask_0, squeeze_mask = var_4185_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4185_cast_fp16")]; + bool var_4187_transpose_x_1 = const()[name = string("op_4187_transpose_x_1"), val = bool(true)]; + bool var_4187_transpose_y_1 = const()[name = string("op_4187_transpose_y_1"), val = bool(false)]; + tensor var_4187_cast_fp16 = matmul(transpose_x = var_4187_transpose_x_1, transpose_y = var_4187_transpose_y_1, x = var_4185_cast_fp16, y = v_new_219_cast_fp16)[name = string("op_4187_cast_fp16")]; + tensor state_221_cast_fp16 = add(x = var_4183_cast_fp16, y = var_4187_cast_fp16)[name = string("state_221_cast_fp16")]; + tensor var_4190_begin_0 = const()[name = string("op_4190_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_4190_end_0 = const()[name = string("op_4190_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_4190_end_mask_0 = const()[name = string("op_4190_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4190_squeeze_mask_0 = const()[name = string("op_4190_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4190_cast_fp16 = slice_by_index(begin = var_4190_begin_0, end = var_4190_end_0, end_mask = var_4190_end_mask_0, squeeze_mask = var_4190_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4190_cast_fp16")]; + tensor var_4192_begin_0 = const()[name = string("op_4192_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_4192_end_0 = const()[name = string("op_4192_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_4192_end_mask_0 = const()[name = string("op_4192_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4192_squeeze_mask_0 = const()[name = string("op_4192_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4192_cast_fp16 = slice_by_index(begin = var_4192_begin_0, end = var_4192_end_0, end_mask = var_4192_end_mask_0, squeeze_mask = var_4192_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4192_cast_fp16")]; + bool var_4193_transpose_x_0 = const()[name = string("op_4193_transpose_x_0"), val = bool(false)]; + bool var_4193_transpose_y_0 = const()[name = string("op_4193_transpose_y_0"), val = bool(false)]; + tensor var_4193_cast_fp16 = matmul(transpose_x = var_4193_transpose_x_0, transpose_y = var_4193_transpose_y_0, x = var_4192_cast_fp16, y = state_221_cast_fp16)[name = string("op_4193_cast_fp16")]; + tensor v_new_221_cast_fp16 = sub(x = var_4190_cast_fp16, y = var_4193_cast_fp16)[name = string("v_new_221_cast_fp16")]; + tensor var_4196_begin_0 = const()[name = string("op_4196_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_4196_end_0 = const()[name = string("op_4196_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_4196_end_mask_0 = const()[name = string("op_4196_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4196_squeeze_mask_0 = const()[name = string("op_4196_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4196_cast_fp16 = slice_by_index(begin = var_4196_begin_0, end = var_4196_end_0, end_mask = var_4196_end_mask_0, squeeze_mask = var_4196_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4196_cast_fp16")]; + bool var_4197_transpose_x_0 = const()[name = string("op_4197_transpose_x_0"), val = bool(false)]; + bool var_4197_transpose_y_0 = const()[name = string("op_4197_transpose_y_0"), val = bool(false)]; + tensor var_4197_cast_fp16 = matmul(transpose_x = var_4197_transpose_x_0, transpose_y = var_4197_transpose_y_0, x = var_4196_cast_fp16, y = state_221_cast_fp16)[name = string("op_4197_cast_fp16")]; + tensor var_4199_begin_0 = const()[name = string("op_4199_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_4199_end_0 = const()[name = string("op_4199_end_0"), val = tensor([32, 15, 64, 64])]; + tensor var_4199_end_mask_0 = const()[name = string("op_4199_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4199_squeeze_mask_0 = const()[name = string("op_4199_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4199_cast_fp16 = slice_by_index(begin = var_4199_begin_0, end = var_4199_end_0, end_mask = var_4199_end_mask_0, squeeze_mask = var_4199_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4199_cast_fp16")]; + bool var_4200_transpose_x_0 = const()[name = string("op_4200_transpose_x_0"), val = bool(false)]; + bool var_4200_transpose_y_0 = const()[name = string("op_4200_transpose_y_0"), val = bool(false)]; + tensor var_4200_cast_fp16 = matmul(transpose_x = var_4200_transpose_x_0, transpose_y = var_4200_transpose_y_0, x = var_4199_cast_fp16, y = v_new_221_cast_fp16)[name = string("op_4200_cast_fp16")]; + tensor var_4201_cast_fp16 = add(x = var_4197_cast_fp16, y = var_4200_cast_fp16)[name = string("op_4201_cast_fp16")]; + tensor var_4203_begin_0 = const()[name = string("op_4203_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_4203_end_0 = const()[name = string("op_4203_end_0"), val = tensor([32, 15, 1, 1])]; + tensor var_4203_end_mask_0 = const()[name = string("op_4203_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4203_squeeze_mask_0 = const()[name = string("op_4203_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4203_cast_fp16 = slice_by_index(begin = var_4203_begin_0, end = var_4203_end_0, end_mask = var_4203_end_mask_0, squeeze_mask = var_4203_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4203_cast_fp16")]; + tensor var_4204_cast_fp16 = mul(x = state_221_cast_fp16, y = var_4203_cast_fp16)[name = string("op_4204_cast_fp16")]; + tensor var_4206_begin_0 = const()[name = string("op_4206_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_4206_end_0 = const()[name = string("op_4206_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_4206_end_mask_0 = const()[name = string("op_4206_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4206_squeeze_mask_0 = const()[name = string("op_4206_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4206_cast_fp16 = slice_by_index(begin = var_4206_begin_0, end = var_4206_end_0, end_mask = var_4206_end_mask_0, squeeze_mask = var_4206_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4206_cast_fp16")]; + bool var_4208_transpose_x_1 = const()[name = string("op_4208_transpose_x_1"), val = bool(true)]; + bool var_4208_transpose_y_1 = const()[name = string("op_4208_transpose_y_1"), val = bool(false)]; + tensor var_4208_cast_fp16 = matmul(transpose_x = var_4208_transpose_x_1, transpose_y = var_4208_transpose_y_1, x = var_4206_cast_fp16, y = v_new_221_cast_fp16)[name = string("op_4208_cast_fp16")]; + tensor state_223_cast_fp16 = add(x = var_4204_cast_fp16, y = var_4208_cast_fp16)[name = string("state_223_cast_fp16")]; + tensor var_4211_begin_0 = const()[name = string("op_4211_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_4211_end_0 = const()[name = string("op_4211_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_4211_end_mask_0 = const()[name = string("op_4211_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4211_squeeze_mask_0 = const()[name = string("op_4211_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4211_cast_fp16 = slice_by_index(begin = var_4211_begin_0, end = var_4211_end_0, end_mask = var_4211_end_mask_0, squeeze_mask = var_4211_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4211_cast_fp16")]; + tensor var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_4213_end_0 = const()[name = string("op_4213_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4213_squeeze_mask_0 = const()[name = string("op_4213_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = var_4213_end_0, end_mask = var_4213_end_mask_0, squeeze_mask = var_4213_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4213_cast_fp16")]; + bool var_4214_transpose_x_0 = const()[name = string("op_4214_transpose_x_0"), val = bool(false)]; + bool var_4214_transpose_y_0 = const()[name = string("op_4214_transpose_y_0"), val = bool(false)]; + tensor var_4214_cast_fp16 = matmul(transpose_x = var_4214_transpose_x_0, transpose_y = var_4214_transpose_y_0, x = var_4213_cast_fp16, y = state_223_cast_fp16)[name = string("op_4214_cast_fp16")]; + tensor v_new_223_cast_fp16 = sub(x = var_4211_cast_fp16, y = var_4214_cast_fp16)[name = string("v_new_223_cast_fp16")]; + tensor var_4217_begin_0 = const()[name = string("op_4217_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_4217_end_0 = const()[name = string("op_4217_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_4217_end_mask_0 = const()[name = string("op_4217_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4217_squeeze_mask_0 = const()[name = string("op_4217_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4217_cast_fp16 = slice_by_index(begin = var_4217_begin_0, end = var_4217_end_0, end_mask = var_4217_end_mask_0, squeeze_mask = var_4217_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4217_cast_fp16")]; + bool var_4218_transpose_x_0 = const()[name = string("op_4218_transpose_x_0"), val = bool(false)]; + bool var_4218_transpose_y_0 = const()[name = string("op_4218_transpose_y_0"), val = bool(false)]; + tensor var_4218_cast_fp16 = matmul(transpose_x = var_4218_transpose_x_0, transpose_y = var_4218_transpose_y_0, x = var_4217_cast_fp16, y = state_223_cast_fp16)[name = string("op_4218_cast_fp16")]; + tensor var_4220_begin_0 = const()[name = string("op_4220_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_4220_end_0 = const()[name = string("op_4220_end_0"), val = tensor([32, 16, 64, 64])]; + tensor var_4220_end_mask_0 = const()[name = string("op_4220_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4220_squeeze_mask_0 = const()[name = string("op_4220_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4220_cast_fp16 = slice_by_index(begin = var_4220_begin_0, end = var_4220_end_0, end_mask = var_4220_end_mask_0, squeeze_mask = var_4220_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4220_cast_fp16")]; + bool var_4221_transpose_x_0 = const()[name = string("op_4221_transpose_x_0"), val = bool(false)]; + bool var_4221_transpose_y_0 = const()[name = string("op_4221_transpose_y_0"), val = bool(false)]; + tensor var_4221_cast_fp16 = matmul(transpose_x = var_4221_transpose_x_0, transpose_y = var_4221_transpose_y_0, x = var_4220_cast_fp16, y = v_new_223_cast_fp16)[name = string("op_4221_cast_fp16")]; + tensor var_4222_cast_fp16 = add(x = var_4218_cast_fp16, y = var_4221_cast_fp16)[name = string("op_4222_cast_fp16")]; + tensor var_4224_begin_0 = const()[name = string("op_4224_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_4224_end_0 = const()[name = string("op_4224_end_0"), val = tensor([32, 16, 1, 1])]; + tensor var_4224_end_mask_0 = const()[name = string("op_4224_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4224_squeeze_mask_0 = const()[name = string("op_4224_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4224_cast_fp16 = slice_by_index(begin = var_4224_begin_0, end = var_4224_end_0, end_mask = var_4224_end_mask_0, squeeze_mask = var_4224_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4224_cast_fp16")]; + tensor var_4225_cast_fp16 = mul(x = state_223_cast_fp16, y = var_4224_cast_fp16)[name = string("op_4225_cast_fp16")]; + tensor var_4227_begin_0 = const()[name = string("op_4227_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_4227_end_0 = const()[name = string("op_4227_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_4227_end_mask_0 = const()[name = string("op_4227_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4227_squeeze_mask_0 = const()[name = string("op_4227_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4227_cast_fp16 = slice_by_index(begin = var_4227_begin_0, end = var_4227_end_0, end_mask = var_4227_end_mask_0, squeeze_mask = var_4227_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4227_cast_fp16")]; + bool var_4229_transpose_x_1 = const()[name = string("op_4229_transpose_x_1"), val = bool(true)]; + bool var_4229_transpose_y_1 = const()[name = string("op_4229_transpose_y_1"), val = bool(false)]; + tensor var_4229_cast_fp16 = matmul(transpose_x = var_4229_transpose_x_1, transpose_y = var_4229_transpose_y_1, x = var_4227_cast_fp16, y = v_new_223_cast_fp16)[name = string("op_4229_cast_fp16")]; + tensor state_225_cast_fp16 = add(x = var_4225_cast_fp16, y = var_4229_cast_fp16)[name = string("state_225_cast_fp16")]; + tensor var_4232_begin_0 = const()[name = string("op_4232_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_4232_end_0 = const()[name = string("op_4232_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_4232_end_mask_0 = const()[name = string("op_4232_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4232_squeeze_mask_0 = const()[name = string("op_4232_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4232_cast_fp16 = slice_by_index(begin = var_4232_begin_0, end = var_4232_end_0, end_mask = var_4232_end_mask_0, squeeze_mask = var_4232_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4232_cast_fp16")]; + tensor var_4234_begin_0 = const()[name = string("op_4234_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_4234_end_0 = const()[name = string("op_4234_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_4234_end_mask_0 = const()[name = string("op_4234_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4234_squeeze_mask_0 = const()[name = string("op_4234_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4234_cast_fp16 = slice_by_index(begin = var_4234_begin_0, end = var_4234_end_0, end_mask = var_4234_end_mask_0, squeeze_mask = var_4234_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4234_cast_fp16")]; + bool var_4235_transpose_x_0 = const()[name = string("op_4235_transpose_x_0"), val = bool(false)]; + bool var_4235_transpose_y_0 = const()[name = string("op_4235_transpose_y_0"), val = bool(false)]; + tensor var_4235_cast_fp16 = matmul(transpose_x = var_4235_transpose_x_0, transpose_y = var_4235_transpose_y_0, x = var_4234_cast_fp16, y = state_225_cast_fp16)[name = string("op_4235_cast_fp16")]; + tensor v_new_225_cast_fp16 = sub(x = var_4232_cast_fp16, y = var_4235_cast_fp16)[name = string("v_new_225_cast_fp16")]; + tensor var_4238_begin_0 = const()[name = string("op_4238_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_4238_end_0 = const()[name = string("op_4238_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_4238_end_mask_0 = const()[name = string("op_4238_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4238_squeeze_mask_0 = const()[name = string("op_4238_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4238_cast_fp16 = slice_by_index(begin = var_4238_begin_0, end = var_4238_end_0, end_mask = var_4238_end_mask_0, squeeze_mask = var_4238_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4238_cast_fp16")]; + bool var_4239_transpose_x_0 = const()[name = string("op_4239_transpose_x_0"), val = bool(false)]; + bool var_4239_transpose_y_0 = const()[name = string("op_4239_transpose_y_0"), val = bool(false)]; + tensor var_4239_cast_fp16 = matmul(transpose_x = var_4239_transpose_x_0, transpose_y = var_4239_transpose_y_0, x = var_4238_cast_fp16, y = state_225_cast_fp16)[name = string("op_4239_cast_fp16")]; + tensor var_4241_begin_0 = const()[name = string("op_4241_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_4241_end_0 = const()[name = string("op_4241_end_0"), val = tensor([32, 17, 64, 64])]; + tensor var_4241_end_mask_0 = const()[name = string("op_4241_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4241_squeeze_mask_0 = const()[name = string("op_4241_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4241_cast_fp16 = slice_by_index(begin = var_4241_begin_0, end = var_4241_end_0, end_mask = var_4241_end_mask_0, squeeze_mask = var_4241_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4241_cast_fp16")]; + bool var_4242_transpose_x_0 = const()[name = string("op_4242_transpose_x_0"), val = bool(false)]; + bool var_4242_transpose_y_0 = const()[name = string("op_4242_transpose_y_0"), val = bool(false)]; + tensor var_4242_cast_fp16 = matmul(transpose_x = var_4242_transpose_x_0, transpose_y = var_4242_transpose_y_0, x = var_4241_cast_fp16, y = v_new_225_cast_fp16)[name = string("op_4242_cast_fp16")]; + tensor var_4243_cast_fp16 = add(x = var_4239_cast_fp16, y = var_4242_cast_fp16)[name = string("op_4243_cast_fp16")]; + tensor var_4245_begin_0 = const()[name = string("op_4245_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_4245_end_0 = const()[name = string("op_4245_end_0"), val = tensor([32, 17, 1, 1])]; + tensor var_4245_end_mask_0 = const()[name = string("op_4245_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4245_squeeze_mask_0 = const()[name = string("op_4245_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4245_cast_fp16 = slice_by_index(begin = var_4245_begin_0, end = var_4245_end_0, end_mask = var_4245_end_mask_0, squeeze_mask = var_4245_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4245_cast_fp16")]; + tensor var_4246_cast_fp16 = mul(x = state_225_cast_fp16, y = var_4245_cast_fp16)[name = string("op_4246_cast_fp16")]; + tensor var_4248_begin_0 = const()[name = string("op_4248_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_4248_end_0 = const()[name = string("op_4248_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_4248_end_mask_0 = const()[name = string("op_4248_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4248_squeeze_mask_0 = const()[name = string("op_4248_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4248_cast_fp16 = slice_by_index(begin = var_4248_begin_0, end = var_4248_end_0, end_mask = var_4248_end_mask_0, squeeze_mask = var_4248_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4248_cast_fp16")]; + bool var_4250_transpose_x_1 = const()[name = string("op_4250_transpose_x_1"), val = bool(true)]; + bool var_4250_transpose_y_1 = const()[name = string("op_4250_transpose_y_1"), val = bool(false)]; + tensor var_4250_cast_fp16 = matmul(transpose_x = var_4250_transpose_x_1, transpose_y = var_4250_transpose_y_1, x = var_4248_cast_fp16, y = v_new_225_cast_fp16)[name = string("op_4250_cast_fp16")]; + tensor state_227_cast_fp16 = add(x = var_4246_cast_fp16, y = var_4250_cast_fp16)[name = string("state_227_cast_fp16")]; + tensor var_4253_begin_0 = const()[name = string("op_4253_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_4253_end_0 = const()[name = string("op_4253_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_4253_end_mask_0 = const()[name = string("op_4253_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4253_squeeze_mask_0 = const()[name = string("op_4253_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4253_cast_fp16 = slice_by_index(begin = var_4253_begin_0, end = var_4253_end_0, end_mask = var_4253_end_mask_0, squeeze_mask = var_4253_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4253_cast_fp16")]; + tensor var_4255_begin_0 = const()[name = string("op_4255_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_4255_end_0 = const()[name = string("op_4255_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_4255_end_mask_0 = const()[name = string("op_4255_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4255_squeeze_mask_0 = const()[name = string("op_4255_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4255_cast_fp16 = slice_by_index(begin = var_4255_begin_0, end = var_4255_end_0, end_mask = var_4255_end_mask_0, squeeze_mask = var_4255_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4255_cast_fp16")]; + bool var_4256_transpose_x_0 = const()[name = string("op_4256_transpose_x_0"), val = bool(false)]; + bool var_4256_transpose_y_0 = const()[name = string("op_4256_transpose_y_0"), val = bool(false)]; + tensor var_4256_cast_fp16 = matmul(transpose_x = var_4256_transpose_x_0, transpose_y = var_4256_transpose_y_0, x = var_4255_cast_fp16, y = state_227_cast_fp16)[name = string("op_4256_cast_fp16")]; + tensor v_new_227_cast_fp16 = sub(x = var_4253_cast_fp16, y = var_4256_cast_fp16)[name = string("v_new_227_cast_fp16")]; + tensor var_4259_begin_0 = const()[name = string("op_4259_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_4259_end_0 = const()[name = string("op_4259_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_4259_end_mask_0 = const()[name = string("op_4259_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4259_squeeze_mask_0 = const()[name = string("op_4259_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4259_cast_fp16 = slice_by_index(begin = var_4259_begin_0, end = var_4259_end_0, end_mask = var_4259_end_mask_0, squeeze_mask = var_4259_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4259_cast_fp16")]; + bool var_4260_transpose_x_0 = const()[name = string("op_4260_transpose_x_0"), val = bool(false)]; + bool var_4260_transpose_y_0 = const()[name = string("op_4260_transpose_y_0"), val = bool(false)]; + tensor var_4260_cast_fp16 = matmul(transpose_x = var_4260_transpose_x_0, transpose_y = var_4260_transpose_y_0, x = var_4259_cast_fp16, y = state_227_cast_fp16)[name = string("op_4260_cast_fp16")]; + tensor var_4262_begin_0 = const()[name = string("op_4262_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_4262_end_0 = const()[name = string("op_4262_end_0"), val = tensor([32, 18, 64, 64])]; + tensor var_4262_end_mask_0 = const()[name = string("op_4262_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4262_squeeze_mask_0 = const()[name = string("op_4262_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4262_cast_fp16 = slice_by_index(begin = var_4262_begin_0, end = var_4262_end_0, end_mask = var_4262_end_mask_0, squeeze_mask = var_4262_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4262_cast_fp16")]; + bool var_4263_transpose_x_0 = const()[name = string("op_4263_transpose_x_0"), val = bool(false)]; + bool var_4263_transpose_y_0 = const()[name = string("op_4263_transpose_y_0"), val = bool(false)]; + tensor var_4263_cast_fp16 = matmul(transpose_x = var_4263_transpose_x_0, transpose_y = var_4263_transpose_y_0, x = var_4262_cast_fp16, y = v_new_227_cast_fp16)[name = string("op_4263_cast_fp16")]; + tensor var_4264_cast_fp16 = add(x = var_4260_cast_fp16, y = var_4263_cast_fp16)[name = string("op_4264_cast_fp16")]; + tensor var_4266_begin_0 = const()[name = string("op_4266_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_4266_end_0 = const()[name = string("op_4266_end_0"), val = tensor([32, 18, 1, 1])]; + tensor var_4266_end_mask_0 = const()[name = string("op_4266_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4266_squeeze_mask_0 = const()[name = string("op_4266_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4266_cast_fp16 = slice_by_index(begin = var_4266_begin_0, end = var_4266_end_0, end_mask = var_4266_end_mask_0, squeeze_mask = var_4266_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4266_cast_fp16")]; + tensor var_4267_cast_fp16 = mul(x = state_227_cast_fp16, y = var_4266_cast_fp16)[name = string("op_4267_cast_fp16")]; + tensor var_4269_begin_0 = const()[name = string("op_4269_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_4269_end_0 = const()[name = string("op_4269_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_4269_end_mask_0 = const()[name = string("op_4269_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4269_squeeze_mask_0 = const()[name = string("op_4269_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4269_cast_fp16 = slice_by_index(begin = var_4269_begin_0, end = var_4269_end_0, end_mask = var_4269_end_mask_0, squeeze_mask = var_4269_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4269_cast_fp16")]; + bool var_4271_transpose_x_1 = const()[name = string("op_4271_transpose_x_1"), val = bool(true)]; + bool var_4271_transpose_y_1 = const()[name = string("op_4271_transpose_y_1"), val = bool(false)]; + tensor var_4271_cast_fp16 = matmul(transpose_x = var_4271_transpose_x_1, transpose_y = var_4271_transpose_y_1, x = var_4269_cast_fp16, y = v_new_227_cast_fp16)[name = string("op_4271_cast_fp16")]; + tensor state_229_cast_fp16 = add(x = var_4267_cast_fp16, y = var_4271_cast_fp16)[name = string("state_229_cast_fp16")]; + tensor var_4274_begin_0 = const()[name = string("op_4274_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_4274_end_0 = const()[name = string("op_4274_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_4274_end_mask_0 = const()[name = string("op_4274_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4274_squeeze_mask_0 = const()[name = string("op_4274_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4274_cast_fp16 = slice_by_index(begin = var_4274_begin_0, end = var_4274_end_0, end_mask = var_4274_end_mask_0, squeeze_mask = var_4274_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4274_cast_fp16")]; + tensor var_4276_begin_0 = const()[name = string("op_4276_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_4276_end_0 = const()[name = string("op_4276_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_4276_end_mask_0 = const()[name = string("op_4276_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4276_squeeze_mask_0 = const()[name = string("op_4276_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4276_cast_fp16 = slice_by_index(begin = var_4276_begin_0, end = var_4276_end_0, end_mask = var_4276_end_mask_0, squeeze_mask = var_4276_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4276_cast_fp16")]; + bool var_4277_transpose_x_0 = const()[name = string("op_4277_transpose_x_0"), val = bool(false)]; + bool var_4277_transpose_y_0 = const()[name = string("op_4277_transpose_y_0"), val = bool(false)]; + tensor var_4277_cast_fp16 = matmul(transpose_x = var_4277_transpose_x_0, transpose_y = var_4277_transpose_y_0, x = var_4276_cast_fp16, y = state_229_cast_fp16)[name = string("op_4277_cast_fp16")]; + tensor v_new_229_cast_fp16 = sub(x = var_4274_cast_fp16, y = var_4277_cast_fp16)[name = string("v_new_229_cast_fp16")]; + tensor var_4280_begin_0 = const()[name = string("op_4280_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_4280_end_0 = const()[name = string("op_4280_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_4280_end_mask_0 = const()[name = string("op_4280_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4280_squeeze_mask_0 = const()[name = string("op_4280_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4280_cast_fp16 = slice_by_index(begin = var_4280_begin_0, end = var_4280_end_0, end_mask = var_4280_end_mask_0, squeeze_mask = var_4280_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4280_cast_fp16")]; + bool var_4281_transpose_x_0 = const()[name = string("op_4281_transpose_x_0"), val = bool(false)]; + bool var_4281_transpose_y_0 = const()[name = string("op_4281_transpose_y_0"), val = bool(false)]; + tensor var_4281_cast_fp16 = matmul(transpose_x = var_4281_transpose_x_0, transpose_y = var_4281_transpose_y_0, x = var_4280_cast_fp16, y = state_229_cast_fp16)[name = string("op_4281_cast_fp16")]; + tensor var_4283_begin_0 = const()[name = string("op_4283_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_4283_end_0 = const()[name = string("op_4283_end_0"), val = tensor([32, 19, 64, 64])]; + tensor var_4283_end_mask_0 = const()[name = string("op_4283_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4283_squeeze_mask_0 = const()[name = string("op_4283_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4283_cast_fp16 = slice_by_index(begin = var_4283_begin_0, end = var_4283_end_0, end_mask = var_4283_end_mask_0, squeeze_mask = var_4283_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4283_cast_fp16")]; + bool var_4284_transpose_x_0 = const()[name = string("op_4284_transpose_x_0"), val = bool(false)]; + bool var_4284_transpose_y_0 = const()[name = string("op_4284_transpose_y_0"), val = bool(false)]; + tensor var_4284_cast_fp16 = matmul(transpose_x = var_4284_transpose_x_0, transpose_y = var_4284_transpose_y_0, x = var_4283_cast_fp16, y = v_new_229_cast_fp16)[name = string("op_4284_cast_fp16")]; + tensor var_4285_cast_fp16 = add(x = var_4281_cast_fp16, y = var_4284_cast_fp16)[name = string("op_4285_cast_fp16")]; + tensor var_4287_begin_0 = const()[name = string("op_4287_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_4287_end_0 = const()[name = string("op_4287_end_0"), val = tensor([32, 19, 1, 1])]; + tensor var_4287_end_mask_0 = const()[name = string("op_4287_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4287_squeeze_mask_0 = const()[name = string("op_4287_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4287_cast_fp16 = slice_by_index(begin = var_4287_begin_0, end = var_4287_end_0, end_mask = var_4287_end_mask_0, squeeze_mask = var_4287_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4287_cast_fp16")]; + tensor var_4288_cast_fp16 = mul(x = state_229_cast_fp16, y = var_4287_cast_fp16)[name = string("op_4288_cast_fp16")]; + tensor var_4290_begin_0 = const()[name = string("op_4290_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_4290_end_0 = const()[name = string("op_4290_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_4290_end_mask_0 = const()[name = string("op_4290_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4290_squeeze_mask_0 = const()[name = string("op_4290_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4290_cast_fp16 = slice_by_index(begin = var_4290_begin_0, end = var_4290_end_0, end_mask = var_4290_end_mask_0, squeeze_mask = var_4290_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4290_cast_fp16")]; + bool var_4292_transpose_x_1 = const()[name = string("op_4292_transpose_x_1"), val = bool(true)]; + bool var_4292_transpose_y_1 = const()[name = string("op_4292_transpose_y_1"), val = bool(false)]; + tensor var_4292_cast_fp16 = matmul(transpose_x = var_4292_transpose_x_1, transpose_y = var_4292_transpose_y_1, x = var_4290_cast_fp16, y = v_new_229_cast_fp16)[name = string("op_4292_cast_fp16")]; + tensor state_231_cast_fp16 = add(x = var_4288_cast_fp16, y = var_4292_cast_fp16)[name = string("state_231_cast_fp16")]; + tensor var_4295_begin_0 = const()[name = string("op_4295_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_4295_end_0 = const()[name = string("op_4295_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_4295_end_mask_0 = const()[name = string("op_4295_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4295_squeeze_mask_0 = const()[name = string("op_4295_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4295_cast_fp16 = slice_by_index(begin = var_4295_begin_0, end = var_4295_end_0, end_mask = var_4295_end_mask_0, squeeze_mask = var_4295_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4295_cast_fp16")]; + tensor var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4297_squeeze_mask_0 = const()[name = string("op_4297_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, squeeze_mask = var_4297_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4297_cast_fp16")]; + bool var_4298_transpose_x_0 = const()[name = string("op_4298_transpose_x_0"), val = bool(false)]; + bool var_4298_transpose_y_0 = const()[name = string("op_4298_transpose_y_0"), val = bool(false)]; + tensor var_4298_cast_fp16 = matmul(transpose_x = var_4298_transpose_x_0, transpose_y = var_4298_transpose_y_0, x = var_4297_cast_fp16, y = state_231_cast_fp16)[name = string("op_4298_cast_fp16")]; + tensor v_new_231_cast_fp16 = sub(x = var_4295_cast_fp16, y = var_4298_cast_fp16)[name = string("v_new_231_cast_fp16")]; + tensor var_4301_begin_0 = const()[name = string("op_4301_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_4301_end_0 = const()[name = string("op_4301_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_4301_end_mask_0 = const()[name = string("op_4301_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4301_squeeze_mask_0 = const()[name = string("op_4301_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4301_cast_fp16 = slice_by_index(begin = var_4301_begin_0, end = var_4301_end_0, end_mask = var_4301_end_mask_0, squeeze_mask = var_4301_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4301_cast_fp16")]; + bool var_4302_transpose_x_0 = const()[name = string("op_4302_transpose_x_0"), val = bool(false)]; + bool var_4302_transpose_y_0 = const()[name = string("op_4302_transpose_y_0"), val = bool(false)]; + tensor var_4302_cast_fp16 = matmul(transpose_x = var_4302_transpose_x_0, transpose_y = var_4302_transpose_y_0, x = var_4301_cast_fp16, y = state_231_cast_fp16)[name = string("op_4302_cast_fp16")]; + tensor var_4304_begin_0 = const()[name = string("op_4304_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_4304_end_0 = const()[name = string("op_4304_end_0"), val = tensor([32, 20, 64, 64])]; + tensor var_4304_end_mask_0 = const()[name = string("op_4304_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4304_squeeze_mask_0 = const()[name = string("op_4304_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4304_cast_fp16 = slice_by_index(begin = var_4304_begin_0, end = var_4304_end_0, end_mask = var_4304_end_mask_0, squeeze_mask = var_4304_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4304_cast_fp16")]; + bool var_4305_transpose_x_0 = const()[name = string("op_4305_transpose_x_0"), val = bool(false)]; + bool var_4305_transpose_y_0 = const()[name = string("op_4305_transpose_y_0"), val = bool(false)]; + tensor var_4305_cast_fp16 = matmul(transpose_x = var_4305_transpose_x_0, transpose_y = var_4305_transpose_y_0, x = var_4304_cast_fp16, y = v_new_231_cast_fp16)[name = string("op_4305_cast_fp16")]; + tensor var_4306_cast_fp16 = add(x = var_4302_cast_fp16, y = var_4305_cast_fp16)[name = string("op_4306_cast_fp16")]; + tensor var_4308_begin_0 = const()[name = string("op_4308_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_4308_end_0 = const()[name = string("op_4308_end_0"), val = tensor([32, 20, 1, 1])]; + tensor var_4308_end_mask_0 = const()[name = string("op_4308_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4308_squeeze_mask_0 = const()[name = string("op_4308_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4308_cast_fp16 = slice_by_index(begin = var_4308_begin_0, end = var_4308_end_0, end_mask = var_4308_end_mask_0, squeeze_mask = var_4308_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4308_cast_fp16")]; + tensor var_4309_cast_fp16 = mul(x = state_231_cast_fp16, y = var_4308_cast_fp16)[name = string("op_4309_cast_fp16")]; + tensor var_4311_begin_0 = const()[name = string("op_4311_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_4311_end_0 = const()[name = string("op_4311_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_4311_end_mask_0 = const()[name = string("op_4311_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4311_squeeze_mask_0 = const()[name = string("op_4311_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4311_cast_fp16 = slice_by_index(begin = var_4311_begin_0, end = var_4311_end_0, end_mask = var_4311_end_mask_0, squeeze_mask = var_4311_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4311_cast_fp16")]; + bool var_4313_transpose_x_1 = const()[name = string("op_4313_transpose_x_1"), val = bool(true)]; + bool var_4313_transpose_y_1 = const()[name = string("op_4313_transpose_y_1"), val = bool(false)]; + tensor var_4313_cast_fp16 = matmul(transpose_x = var_4313_transpose_x_1, transpose_y = var_4313_transpose_y_1, x = var_4311_cast_fp16, y = v_new_231_cast_fp16)[name = string("op_4313_cast_fp16")]; + tensor state_233_cast_fp16 = add(x = var_4309_cast_fp16, y = var_4313_cast_fp16)[name = string("state_233_cast_fp16")]; + tensor var_4316_begin_0 = const()[name = string("op_4316_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_4316_end_0 = const()[name = string("op_4316_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_4316_end_mask_0 = const()[name = string("op_4316_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4316_squeeze_mask_0 = const()[name = string("op_4316_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4316_cast_fp16 = slice_by_index(begin = var_4316_begin_0, end = var_4316_end_0, end_mask = var_4316_end_mask_0, squeeze_mask = var_4316_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4316_cast_fp16")]; + tensor var_4318_begin_0 = const()[name = string("op_4318_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_4318_end_0 = const()[name = string("op_4318_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_4318_end_mask_0 = const()[name = string("op_4318_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4318_squeeze_mask_0 = const()[name = string("op_4318_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4318_cast_fp16 = slice_by_index(begin = var_4318_begin_0, end = var_4318_end_0, end_mask = var_4318_end_mask_0, squeeze_mask = var_4318_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4318_cast_fp16")]; + bool var_4319_transpose_x_0 = const()[name = string("op_4319_transpose_x_0"), val = bool(false)]; + bool var_4319_transpose_y_0 = const()[name = string("op_4319_transpose_y_0"), val = bool(false)]; + tensor var_4319_cast_fp16 = matmul(transpose_x = var_4319_transpose_x_0, transpose_y = var_4319_transpose_y_0, x = var_4318_cast_fp16, y = state_233_cast_fp16)[name = string("op_4319_cast_fp16")]; + tensor v_new_233_cast_fp16 = sub(x = var_4316_cast_fp16, y = var_4319_cast_fp16)[name = string("v_new_233_cast_fp16")]; + tensor var_4322_begin_0 = const()[name = string("op_4322_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_4322_end_0 = const()[name = string("op_4322_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_4322_end_mask_0 = const()[name = string("op_4322_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4322_squeeze_mask_0 = const()[name = string("op_4322_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4322_cast_fp16 = slice_by_index(begin = var_4322_begin_0, end = var_4322_end_0, end_mask = var_4322_end_mask_0, squeeze_mask = var_4322_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4322_cast_fp16")]; + bool var_4323_transpose_x_0 = const()[name = string("op_4323_transpose_x_0"), val = bool(false)]; + bool var_4323_transpose_y_0 = const()[name = string("op_4323_transpose_y_0"), val = bool(false)]; + tensor var_4323_cast_fp16 = matmul(transpose_x = var_4323_transpose_x_0, transpose_y = var_4323_transpose_y_0, x = var_4322_cast_fp16, y = state_233_cast_fp16)[name = string("op_4323_cast_fp16")]; + tensor var_4325_begin_0 = const()[name = string("op_4325_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_4325_end_0 = const()[name = string("op_4325_end_0"), val = tensor([32, 21, 64, 64])]; + tensor var_4325_end_mask_0 = const()[name = string("op_4325_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4325_squeeze_mask_0 = const()[name = string("op_4325_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4325_cast_fp16 = slice_by_index(begin = var_4325_begin_0, end = var_4325_end_0, end_mask = var_4325_end_mask_0, squeeze_mask = var_4325_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4325_cast_fp16")]; + bool var_4326_transpose_x_0 = const()[name = string("op_4326_transpose_x_0"), val = bool(false)]; + bool var_4326_transpose_y_0 = const()[name = string("op_4326_transpose_y_0"), val = bool(false)]; + tensor var_4326_cast_fp16 = matmul(transpose_x = var_4326_transpose_x_0, transpose_y = var_4326_transpose_y_0, x = var_4325_cast_fp16, y = v_new_233_cast_fp16)[name = string("op_4326_cast_fp16")]; + tensor var_4327_cast_fp16 = add(x = var_4323_cast_fp16, y = var_4326_cast_fp16)[name = string("op_4327_cast_fp16")]; + tensor var_4329_begin_0 = const()[name = string("op_4329_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_4329_end_0 = const()[name = string("op_4329_end_0"), val = tensor([32, 21, 1, 1])]; + tensor var_4329_end_mask_0 = const()[name = string("op_4329_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4329_squeeze_mask_0 = const()[name = string("op_4329_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4329_cast_fp16 = slice_by_index(begin = var_4329_begin_0, end = var_4329_end_0, end_mask = var_4329_end_mask_0, squeeze_mask = var_4329_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4329_cast_fp16")]; + tensor var_4330_cast_fp16 = mul(x = state_233_cast_fp16, y = var_4329_cast_fp16)[name = string("op_4330_cast_fp16")]; + tensor var_4332_begin_0 = const()[name = string("op_4332_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_4332_end_0 = const()[name = string("op_4332_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_4332_end_mask_0 = const()[name = string("op_4332_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4332_squeeze_mask_0 = const()[name = string("op_4332_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4332_cast_fp16 = slice_by_index(begin = var_4332_begin_0, end = var_4332_end_0, end_mask = var_4332_end_mask_0, squeeze_mask = var_4332_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4332_cast_fp16")]; + bool var_4334_transpose_x_1 = const()[name = string("op_4334_transpose_x_1"), val = bool(true)]; + bool var_4334_transpose_y_1 = const()[name = string("op_4334_transpose_y_1"), val = bool(false)]; + tensor var_4334_cast_fp16 = matmul(transpose_x = var_4334_transpose_x_1, transpose_y = var_4334_transpose_y_1, x = var_4332_cast_fp16, y = v_new_233_cast_fp16)[name = string("op_4334_cast_fp16")]; + tensor state_235_cast_fp16 = add(x = var_4330_cast_fp16, y = var_4334_cast_fp16)[name = string("state_235_cast_fp16")]; + tensor var_4337_begin_0 = const()[name = string("op_4337_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_4337_end_0 = const()[name = string("op_4337_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_4337_end_mask_0 = const()[name = string("op_4337_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4337_squeeze_mask_0 = const()[name = string("op_4337_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4337_cast_fp16 = slice_by_index(begin = var_4337_begin_0, end = var_4337_end_0, end_mask = var_4337_end_mask_0, squeeze_mask = var_4337_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4337_cast_fp16")]; + tensor var_4339_begin_0 = const()[name = string("op_4339_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_4339_end_0 = const()[name = string("op_4339_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_4339_end_mask_0 = const()[name = string("op_4339_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4339_squeeze_mask_0 = const()[name = string("op_4339_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4339_cast_fp16 = slice_by_index(begin = var_4339_begin_0, end = var_4339_end_0, end_mask = var_4339_end_mask_0, squeeze_mask = var_4339_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4339_cast_fp16")]; + bool var_4340_transpose_x_0 = const()[name = string("op_4340_transpose_x_0"), val = bool(false)]; + bool var_4340_transpose_y_0 = const()[name = string("op_4340_transpose_y_0"), val = bool(false)]; + tensor var_4340_cast_fp16 = matmul(transpose_x = var_4340_transpose_x_0, transpose_y = var_4340_transpose_y_0, x = var_4339_cast_fp16, y = state_235_cast_fp16)[name = string("op_4340_cast_fp16")]; + tensor v_new_235_cast_fp16 = sub(x = var_4337_cast_fp16, y = var_4340_cast_fp16)[name = string("v_new_235_cast_fp16")]; + tensor var_4343_begin_0 = const()[name = string("op_4343_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_4343_end_0 = const()[name = string("op_4343_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_4343_end_mask_0 = const()[name = string("op_4343_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4343_squeeze_mask_0 = const()[name = string("op_4343_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4343_cast_fp16 = slice_by_index(begin = var_4343_begin_0, end = var_4343_end_0, end_mask = var_4343_end_mask_0, squeeze_mask = var_4343_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4343_cast_fp16")]; + bool var_4344_transpose_x_0 = const()[name = string("op_4344_transpose_x_0"), val = bool(false)]; + bool var_4344_transpose_y_0 = const()[name = string("op_4344_transpose_y_0"), val = bool(false)]; + tensor var_4344_cast_fp16 = matmul(transpose_x = var_4344_transpose_x_0, transpose_y = var_4344_transpose_y_0, x = var_4343_cast_fp16, y = state_235_cast_fp16)[name = string("op_4344_cast_fp16")]; + tensor var_4346_begin_0 = const()[name = string("op_4346_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_4346_end_0 = const()[name = string("op_4346_end_0"), val = tensor([32, 22, 64, 64])]; + tensor var_4346_end_mask_0 = const()[name = string("op_4346_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4346_squeeze_mask_0 = const()[name = string("op_4346_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4346_cast_fp16 = slice_by_index(begin = var_4346_begin_0, end = var_4346_end_0, end_mask = var_4346_end_mask_0, squeeze_mask = var_4346_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4346_cast_fp16")]; + bool var_4347_transpose_x_0 = const()[name = string("op_4347_transpose_x_0"), val = bool(false)]; + bool var_4347_transpose_y_0 = const()[name = string("op_4347_transpose_y_0"), val = bool(false)]; + tensor var_4347_cast_fp16 = matmul(transpose_x = var_4347_transpose_x_0, transpose_y = var_4347_transpose_y_0, x = var_4346_cast_fp16, y = v_new_235_cast_fp16)[name = string("op_4347_cast_fp16")]; + tensor var_4348_cast_fp16 = add(x = var_4344_cast_fp16, y = var_4347_cast_fp16)[name = string("op_4348_cast_fp16")]; + tensor var_4350_begin_0 = const()[name = string("op_4350_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_4350_end_0 = const()[name = string("op_4350_end_0"), val = tensor([32, 22, 1, 1])]; + tensor var_4350_end_mask_0 = const()[name = string("op_4350_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4350_squeeze_mask_0 = const()[name = string("op_4350_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4350_cast_fp16 = slice_by_index(begin = var_4350_begin_0, end = var_4350_end_0, end_mask = var_4350_end_mask_0, squeeze_mask = var_4350_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4350_cast_fp16")]; + tensor var_4351_cast_fp16 = mul(x = state_235_cast_fp16, y = var_4350_cast_fp16)[name = string("op_4351_cast_fp16")]; + tensor var_4353_begin_0 = const()[name = string("op_4353_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_4353_end_0 = const()[name = string("op_4353_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_4353_end_mask_0 = const()[name = string("op_4353_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4353_squeeze_mask_0 = const()[name = string("op_4353_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4353_cast_fp16 = slice_by_index(begin = var_4353_begin_0, end = var_4353_end_0, end_mask = var_4353_end_mask_0, squeeze_mask = var_4353_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4353_cast_fp16")]; + bool var_4355_transpose_x_1 = const()[name = string("op_4355_transpose_x_1"), val = bool(true)]; + bool var_4355_transpose_y_1 = const()[name = string("op_4355_transpose_y_1"), val = bool(false)]; + tensor var_4355_cast_fp16 = matmul(transpose_x = var_4355_transpose_x_1, transpose_y = var_4355_transpose_y_1, x = var_4353_cast_fp16, y = v_new_235_cast_fp16)[name = string("op_4355_cast_fp16")]; + tensor state_237_cast_fp16 = add(x = var_4351_cast_fp16, y = var_4355_cast_fp16)[name = string("state_237_cast_fp16")]; + tensor var_4358_begin_0 = const()[name = string("op_4358_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_4358_end_0 = const()[name = string("op_4358_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_4358_end_mask_0 = const()[name = string("op_4358_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4358_squeeze_mask_0 = const()[name = string("op_4358_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4358_cast_fp16 = slice_by_index(begin = var_4358_begin_0, end = var_4358_end_0, end_mask = var_4358_end_mask_0, squeeze_mask = var_4358_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4358_cast_fp16")]; + tensor var_4360_begin_0 = const()[name = string("op_4360_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_4360_end_0 = const()[name = string("op_4360_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_4360_end_mask_0 = const()[name = string("op_4360_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4360_squeeze_mask_0 = const()[name = string("op_4360_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4360_cast_fp16 = slice_by_index(begin = var_4360_begin_0, end = var_4360_end_0, end_mask = var_4360_end_mask_0, squeeze_mask = var_4360_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4360_cast_fp16")]; + bool var_4361_transpose_x_0 = const()[name = string("op_4361_transpose_x_0"), val = bool(false)]; + bool var_4361_transpose_y_0 = const()[name = string("op_4361_transpose_y_0"), val = bool(false)]; + tensor var_4361_cast_fp16 = matmul(transpose_x = var_4361_transpose_x_0, transpose_y = var_4361_transpose_y_0, x = var_4360_cast_fp16, y = state_237_cast_fp16)[name = string("op_4361_cast_fp16")]; + tensor v_new_237_cast_fp16 = sub(x = var_4358_cast_fp16, y = var_4361_cast_fp16)[name = string("v_new_237_cast_fp16")]; + tensor var_4364_begin_0 = const()[name = string("op_4364_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_4364_end_0 = const()[name = string("op_4364_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_4364_end_mask_0 = const()[name = string("op_4364_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4364_squeeze_mask_0 = const()[name = string("op_4364_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4364_cast_fp16 = slice_by_index(begin = var_4364_begin_0, end = var_4364_end_0, end_mask = var_4364_end_mask_0, squeeze_mask = var_4364_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4364_cast_fp16")]; + bool var_4365_transpose_x_0 = const()[name = string("op_4365_transpose_x_0"), val = bool(false)]; + bool var_4365_transpose_y_0 = const()[name = string("op_4365_transpose_y_0"), val = bool(false)]; + tensor var_4365_cast_fp16 = matmul(transpose_x = var_4365_transpose_x_0, transpose_y = var_4365_transpose_y_0, x = var_4364_cast_fp16, y = state_237_cast_fp16)[name = string("op_4365_cast_fp16")]; + tensor var_4367_begin_0 = const()[name = string("op_4367_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_4367_end_0 = const()[name = string("op_4367_end_0"), val = tensor([32, 23, 64, 64])]; + tensor var_4367_end_mask_0 = const()[name = string("op_4367_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4367_squeeze_mask_0 = const()[name = string("op_4367_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4367_cast_fp16 = slice_by_index(begin = var_4367_begin_0, end = var_4367_end_0, end_mask = var_4367_end_mask_0, squeeze_mask = var_4367_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4367_cast_fp16")]; + bool var_4368_transpose_x_0 = const()[name = string("op_4368_transpose_x_0"), val = bool(false)]; + bool var_4368_transpose_y_0 = const()[name = string("op_4368_transpose_y_0"), val = bool(false)]; + tensor var_4368_cast_fp16 = matmul(transpose_x = var_4368_transpose_x_0, transpose_y = var_4368_transpose_y_0, x = var_4367_cast_fp16, y = v_new_237_cast_fp16)[name = string("op_4368_cast_fp16")]; + tensor var_4369_cast_fp16 = add(x = var_4365_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4369_cast_fp16")]; + tensor var_4371_begin_0 = const()[name = string("op_4371_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_4371_end_0 = const()[name = string("op_4371_end_0"), val = tensor([32, 23, 1, 1])]; + tensor var_4371_end_mask_0 = const()[name = string("op_4371_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4371_squeeze_mask_0 = const()[name = string("op_4371_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4371_cast_fp16 = slice_by_index(begin = var_4371_begin_0, end = var_4371_end_0, end_mask = var_4371_end_mask_0, squeeze_mask = var_4371_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4371_cast_fp16")]; + tensor var_4372_cast_fp16 = mul(x = state_237_cast_fp16, y = var_4371_cast_fp16)[name = string("op_4372_cast_fp16")]; + tensor var_4374_begin_0 = const()[name = string("op_4374_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_4374_end_0 = const()[name = string("op_4374_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_4374_end_mask_0 = const()[name = string("op_4374_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4374_squeeze_mask_0 = const()[name = string("op_4374_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4374_cast_fp16 = slice_by_index(begin = var_4374_begin_0, end = var_4374_end_0, end_mask = var_4374_end_mask_0, squeeze_mask = var_4374_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4374_cast_fp16")]; + bool var_4376_transpose_x_1 = const()[name = string("op_4376_transpose_x_1"), val = bool(true)]; + bool var_4376_transpose_y_1 = const()[name = string("op_4376_transpose_y_1"), val = bool(false)]; + tensor var_4376_cast_fp16 = matmul(transpose_x = var_4376_transpose_x_1, transpose_y = var_4376_transpose_y_1, x = var_4374_cast_fp16, y = v_new_237_cast_fp16)[name = string("op_4376_cast_fp16")]; + tensor state_239_cast_fp16 = add(x = var_4372_cast_fp16, y = var_4376_cast_fp16)[name = string("state_239_cast_fp16")]; + tensor var_4379_begin_0 = const()[name = string("op_4379_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_4379_end_0 = const()[name = string("op_4379_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_4379_end_mask_0 = const()[name = string("op_4379_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4379_squeeze_mask_0 = const()[name = string("op_4379_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4379_cast_fp16 = slice_by_index(begin = var_4379_begin_0, end = var_4379_end_0, end_mask = var_4379_end_mask_0, squeeze_mask = var_4379_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4379_cast_fp16")]; + tensor var_4381_begin_0 = const()[name = string("op_4381_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_4381_end_0 = const()[name = string("op_4381_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_4381_end_mask_0 = const()[name = string("op_4381_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4381_squeeze_mask_0 = const()[name = string("op_4381_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4381_cast_fp16 = slice_by_index(begin = var_4381_begin_0, end = var_4381_end_0, end_mask = var_4381_end_mask_0, squeeze_mask = var_4381_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4381_cast_fp16")]; + bool var_4382_transpose_x_0 = const()[name = string("op_4382_transpose_x_0"), val = bool(false)]; + bool var_4382_transpose_y_0 = const()[name = string("op_4382_transpose_y_0"), val = bool(false)]; + tensor var_4382_cast_fp16 = matmul(transpose_x = var_4382_transpose_x_0, transpose_y = var_4382_transpose_y_0, x = var_4381_cast_fp16, y = state_239_cast_fp16)[name = string("op_4382_cast_fp16")]; + tensor v_new_239_cast_fp16 = sub(x = var_4379_cast_fp16, y = var_4382_cast_fp16)[name = string("v_new_239_cast_fp16")]; + tensor var_4385_begin_0 = const()[name = string("op_4385_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_4385_end_0 = const()[name = string("op_4385_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_4385_end_mask_0 = const()[name = string("op_4385_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4385_squeeze_mask_0 = const()[name = string("op_4385_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4385_cast_fp16 = slice_by_index(begin = var_4385_begin_0, end = var_4385_end_0, end_mask = var_4385_end_mask_0, squeeze_mask = var_4385_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4385_cast_fp16")]; + bool var_4386_transpose_x_0 = const()[name = string("op_4386_transpose_x_0"), val = bool(false)]; + bool var_4386_transpose_y_0 = const()[name = string("op_4386_transpose_y_0"), val = bool(false)]; + tensor var_4386_cast_fp16 = matmul(transpose_x = var_4386_transpose_x_0, transpose_y = var_4386_transpose_y_0, x = var_4385_cast_fp16, y = state_239_cast_fp16)[name = string("op_4386_cast_fp16")]; + tensor var_4388_begin_0 = const()[name = string("op_4388_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_4388_end_0 = const()[name = string("op_4388_end_0"), val = tensor([32, 24, 64, 64])]; + tensor var_4388_end_mask_0 = const()[name = string("op_4388_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4388_squeeze_mask_0 = const()[name = string("op_4388_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4388_cast_fp16 = slice_by_index(begin = var_4388_begin_0, end = var_4388_end_0, end_mask = var_4388_end_mask_0, squeeze_mask = var_4388_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4388_cast_fp16")]; + bool var_4389_transpose_x_0 = const()[name = string("op_4389_transpose_x_0"), val = bool(false)]; + bool var_4389_transpose_y_0 = const()[name = string("op_4389_transpose_y_0"), val = bool(false)]; + tensor var_4389_cast_fp16 = matmul(transpose_x = var_4389_transpose_x_0, transpose_y = var_4389_transpose_y_0, x = var_4388_cast_fp16, y = v_new_239_cast_fp16)[name = string("op_4389_cast_fp16")]; + tensor var_4390_cast_fp16 = add(x = var_4386_cast_fp16, y = var_4389_cast_fp16)[name = string("op_4390_cast_fp16")]; + tensor var_4392_begin_0 = const()[name = string("op_4392_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_4392_end_0 = const()[name = string("op_4392_end_0"), val = tensor([32, 24, 1, 1])]; + tensor var_4392_end_mask_0 = const()[name = string("op_4392_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4392_squeeze_mask_0 = const()[name = string("op_4392_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4392_cast_fp16 = slice_by_index(begin = var_4392_begin_0, end = var_4392_end_0, end_mask = var_4392_end_mask_0, squeeze_mask = var_4392_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4392_cast_fp16")]; + tensor var_4393_cast_fp16 = mul(x = state_239_cast_fp16, y = var_4392_cast_fp16)[name = string("op_4393_cast_fp16")]; + tensor var_4395_begin_0 = const()[name = string("op_4395_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_4395_end_0 = const()[name = string("op_4395_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_4395_end_mask_0 = const()[name = string("op_4395_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4395_squeeze_mask_0 = const()[name = string("op_4395_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4395_cast_fp16 = slice_by_index(begin = var_4395_begin_0, end = var_4395_end_0, end_mask = var_4395_end_mask_0, squeeze_mask = var_4395_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4395_cast_fp16")]; + bool var_4397_transpose_x_1 = const()[name = string("op_4397_transpose_x_1"), val = bool(true)]; + bool var_4397_transpose_y_1 = const()[name = string("op_4397_transpose_y_1"), val = bool(false)]; + tensor var_4397_cast_fp16 = matmul(transpose_x = var_4397_transpose_x_1, transpose_y = var_4397_transpose_y_1, x = var_4395_cast_fp16, y = v_new_239_cast_fp16)[name = string("op_4397_cast_fp16")]; + tensor state_241_cast_fp16 = add(x = var_4393_cast_fp16, y = var_4397_cast_fp16)[name = string("state_241_cast_fp16")]; + tensor var_4400_begin_0 = const()[name = string("op_4400_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_4400_end_0 = const()[name = string("op_4400_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_4400_end_mask_0 = const()[name = string("op_4400_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4400_squeeze_mask_0 = const()[name = string("op_4400_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4400_cast_fp16 = slice_by_index(begin = var_4400_begin_0, end = var_4400_end_0, end_mask = var_4400_end_mask_0, squeeze_mask = var_4400_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4400_cast_fp16")]; + tensor var_4402_begin_0 = const()[name = string("op_4402_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_4402_end_0 = const()[name = string("op_4402_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_4402_end_mask_0 = const()[name = string("op_4402_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4402_squeeze_mask_0 = const()[name = string("op_4402_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4402_cast_fp16 = slice_by_index(begin = var_4402_begin_0, end = var_4402_end_0, end_mask = var_4402_end_mask_0, squeeze_mask = var_4402_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4402_cast_fp16")]; + bool var_4403_transpose_x_0 = const()[name = string("op_4403_transpose_x_0"), val = bool(false)]; + bool var_4403_transpose_y_0 = const()[name = string("op_4403_transpose_y_0"), val = bool(false)]; + tensor var_4403_cast_fp16 = matmul(transpose_x = var_4403_transpose_x_0, transpose_y = var_4403_transpose_y_0, x = var_4402_cast_fp16, y = state_241_cast_fp16)[name = string("op_4403_cast_fp16")]; + tensor v_new_241_cast_fp16 = sub(x = var_4400_cast_fp16, y = var_4403_cast_fp16)[name = string("v_new_241_cast_fp16")]; + tensor var_4406_begin_0 = const()[name = string("op_4406_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_4406_end_0 = const()[name = string("op_4406_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_4406_end_mask_0 = const()[name = string("op_4406_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4406_squeeze_mask_0 = const()[name = string("op_4406_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4406_cast_fp16 = slice_by_index(begin = var_4406_begin_0, end = var_4406_end_0, end_mask = var_4406_end_mask_0, squeeze_mask = var_4406_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4406_cast_fp16")]; + bool var_4407_transpose_x_0 = const()[name = string("op_4407_transpose_x_0"), val = bool(false)]; + bool var_4407_transpose_y_0 = const()[name = string("op_4407_transpose_y_0"), val = bool(false)]; + tensor var_4407_cast_fp16 = matmul(transpose_x = var_4407_transpose_x_0, transpose_y = var_4407_transpose_y_0, x = var_4406_cast_fp16, y = state_241_cast_fp16)[name = string("op_4407_cast_fp16")]; + tensor var_4409_begin_0 = const()[name = string("op_4409_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_4409_end_0 = const()[name = string("op_4409_end_0"), val = tensor([32, 25, 64, 64])]; + tensor var_4409_end_mask_0 = const()[name = string("op_4409_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4409_squeeze_mask_0 = const()[name = string("op_4409_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4409_cast_fp16 = slice_by_index(begin = var_4409_begin_0, end = var_4409_end_0, end_mask = var_4409_end_mask_0, squeeze_mask = var_4409_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4409_cast_fp16")]; + bool var_4410_transpose_x_0 = const()[name = string("op_4410_transpose_x_0"), val = bool(false)]; + bool var_4410_transpose_y_0 = const()[name = string("op_4410_transpose_y_0"), val = bool(false)]; + tensor var_4410_cast_fp16 = matmul(transpose_x = var_4410_transpose_x_0, transpose_y = var_4410_transpose_y_0, x = var_4409_cast_fp16, y = v_new_241_cast_fp16)[name = string("op_4410_cast_fp16")]; + tensor var_4411_cast_fp16 = add(x = var_4407_cast_fp16, y = var_4410_cast_fp16)[name = string("op_4411_cast_fp16")]; + tensor var_4413_begin_0 = const()[name = string("op_4413_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_4413_end_0 = const()[name = string("op_4413_end_0"), val = tensor([32, 25, 1, 1])]; + tensor var_4413_end_mask_0 = const()[name = string("op_4413_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4413_squeeze_mask_0 = const()[name = string("op_4413_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4413_cast_fp16 = slice_by_index(begin = var_4413_begin_0, end = var_4413_end_0, end_mask = var_4413_end_mask_0, squeeze_mask = var_4413_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4413_cast_fp16")]; + tensor var_4414_cast_fp16 = mul(x = state_241_cast_fp16, y = var_4413_cast_fp16)[name = string("op_4414_cast_fp16")]; + tensor var_4416_begin_0 = const()[name = string("op_4416_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_4416_end_0 = const()[name = string("op_4416_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_4416_end_mask_0 = const()[name = string("op_4416_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4416_squeeze_mask_0 = const()[name = string("op_4416_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4416_cast_fp16 = slice_by_index(begin = var_4416_begin_0, end = var_4416_end_0, end_mask = var_4416_end_mask_0, squeeze_mask = var_4416_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4416_cast_fp16")]; + bool var_4418_transpose_x_1 = const()[name = string("op_4418_transpose_x_1"), val = bool(true)]; + bool var_4418_transpose_y_1 = const()[name = string("op_4418_transpose_y_1"), val = bool(false)]; + tensor var_4418_cast_fp16 = matmul(transpose_x = var_4418_transpose_x_1, transpose_y = var_4418_transpose_y_1, x = var_4416_cast_fp16, y = v_new_241_cast_fp16)[name = string("op_4418_cast_fp16")]; + tensor state_243_cast_fp16 = add(x = var_4414_cast_fp16, y = var_4418_cast_fp16)[name = string("state_243_cast_fp16")]; + tensor var_4421_begin_0 = const()[name = string("op_4421_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_4421_end_0 = const()[name = string("op_4421_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_4421_end_mask_0 = const()[name = string("op_4421_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4421_squeeze_mask_0 = const()[name = string("op_4421_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4421_cast_fp16 = slice_by_index(begin = var_4421_begin_0, end = var_4421_end_0, end_mask = var_4421_end_mask_0, squeeze_mask = var_4421_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4421_cast_fp16")]; + tensor var_4423_begin_0 = const()[name = string("op_4423_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_4423_end_0 = const()[name = string("op_4423_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_4423_end_mask_0 = const()[name = string("op_4423_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4423_squeeze_mask_0 = const()[name = string("op_4423_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4423_cast_fp16 = slice_by_index(begin = var_4423_begin_0, end = var_4423_end_0, end_mask = var_4423_end_mask_0, squeeze_mask = var_4423_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4423_cast_fp16")]; + bool var_4424_transpose_x_0 = const()[name = string("op_4424_transpose_x_0"), val = bool(false)]; + bool var_4424_transpose_y_0 = const()[name = string("op_4424_transpose_y_0"), val = bool(false)]; + tensor var_4424_cast_fp16 = matmul(transpose_x = var_4424_transpose_x_0, transpose_y = var_4424_transpose_y_0, x = var_4423_cast_fp16, y = state_243_cast_fp16)[name = string("op_4424_cast_fp16")]; + tensor v_new_243_cast_fp16 = sub(x = var_4421_cast_fp16, y = var_4424_cast_fp16)[name = string("v_new_243_cast_fp16")]; + tensor var_4427_begin_0 = const()[name = string("op_4427_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_4427_end_0 = const()[name = string("op_4427_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_4427_end_mask_0 = const()[name = string("op_4427_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4427_squeeze_mask_0 = const()[name = string("op_4427_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4427_cast_fp16 = slice_by_index(begin = var_4427_begin_0, end = var_4427_end_0, end_mask = var_4427_end_mask_0, squeeze_mask = var_4427_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4427_cast_fp16")]; + bool var_4428_transpose_x_0 = const()[name = string("op_4428_transpose_x_0"), val = bool(false)]; + bool var_4428_transpose_y_0 = const()[name = string("op_4428_transpose_y_0"), val = bool(false)]; + tensor var_4428_cast_fp16 = matmul(transpose_x = var_4428_transpose_x_0, transpose_y = var_4428_transpose_y_0, x = var_4427_cast_fp16, y = state_243_cast_fp16)[name = string("op_4428_cast_fp16")]; + tensor var_4430_begin_0 = const()[name = string("op_4430_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_4430_end_0 = const()[name = string("op_4430_end_0"), val = tensor([32, 26, 64, 64])]; + tensor var_4430_end_mask_0 = const()[name = string("op_4430_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4430_squeeze_mask_0 = const()[name = string("op_4430_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4430_cast_fp16 = slice_by_index(begin = var_4430_begin_0, end = var_4430_end_0, end_mask = var_4430_end_mask_0, squeeze_mask = var_4430_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4430_cast_fp16")]; + bool var_4431_transpose_x_0 = const()[name = string("op_4431_transpose_x_0"), val = bool(false)]; + bool var_4431_transpose_y_0 = const()[name = string("op_4431_transpose_y_0"), val = bool(false)]; + tensor var_4431_cast_fp16 = matmul(transpose_x = var_4431_transpose_x_0, transpose_y = var_4431_transpose_y_0, x = var_4430_cast_fp16, y = v_new_243_cast_fp16)[name = string("op_4431_cast_fp16")]; + tensor var_4432_cast_fp16 = add(x = var_4428_cast_fp16, y = var_4431_cast_fp16)[name = string("op_4432_cast_fp16")]; + tensor var_4434_begin_0 = const()[name = string("op_4434_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_4434_end_0 = const()[name = string("op_4434_end_0"), val = tensor([32, 26, 1, 1])]; + tensor var_4434_end_mask_0 = const()[name = string("op_4434_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4434_squeeze_mask_0 = const()[name = string("op_4434_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4434_cast_fp16 = slice_by_index(begin = var_4434_begin_0, end = var_4434_end_0, end_mask = var_4434_end_mask_0, squeeze_mask = var_4434_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = state_243_cast_fp16, y = var_4434_cast_fp16)[name = string("op_4435_cast_fp16")]; + tensor var_4437_begin_0 = const()[name = string("op_4437_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_4437_end_0 = const()[name = string("op_4437_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_4437_end_mask_0 = const()[name = string("op_4437_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4437_squeeze_mask_0 = const()[name = string("op_4437_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4437_cast_fp16 = slice_by_index(begin = var_4437_begin_0, end = var_4437_end_0, end_mask = var_4437_end_mask_0, squeeze_mask = var_4437_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4437_cast_fp16")]; + bool var_4439_transpose_x_1 = const()[name = string("op_4439_transpose_x_1"), val = bool(true)]; + bool var_4439_transpose_y_1 = const()[name = string("op_4439_transpose_y_1"), val = bool(false)]; + tensor var_4439_cast_fp16 = matmul(transpose_x = var_4439_transpose_x_1, transpose_y = var_4439_transpose_y_1, x = var_4437_cast_fp16, y = v_new_243_cast_fp16)[name = string("op_4439_cast_fp16")]; + tensor state_245_cast_fp16 = add(x = var_4435_cast_fp16, y = var_4439_cast_fp16)[name = string("state_245_cast_fp16")]; + tensor var_4442_begin_0 = const()[name = string("op_4442_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_4442_end_0 = const()[name = string("op_4442_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_4442_end_mask_0 = const()[name = string("op_4442_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4442_squeeze_mask_0 = const()[name = string("op_4442_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4442_cast_fp16 = slice_by_index(begin = var_4442_begin_0, end = var_4442_end_0, end_mask = var_4442_end_mask_0, squeeze_mask = var_4442_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4442_cast_fp16")]; + tensor var_4444_begin_0 = const()[name = string("op_4444_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_4444_end_0 = const()[name = string("op_4444_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_4444_end_mask_0 = const()[name = string("op_4444_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4444_squeeze_mask_0 = const()[name = string("op_4444_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4444_cast_fp16 = slice_by_index(begin = var_4444_begin_0, end = var_4444_end_0, end_mask = var_4444_end_mask_0, squeeze_mask = var_4444_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4444_cast_fp16")]; + bool var_4445_transpose_x_0 = const()[name = string("op_4445_transpose_x_0"), val = bool(false)]; + bool var_4445_transpose_y_0 = const()[name = string("op_4445_transpose_y_0"), val = bool(false)]; + tensor var_4445_cast_fp16 = matmul(transpose_x = var_4445_transpose_x_0, transpose_y = var_4445_transpose_y_0, x = var_4444_cast_fp16, y = state_245_cast_fp16)[name = string("op_4445_cast_fp16")]; + tensor v_new_245_cast_fp16 = sub(x = var_4442_cast_fp16, y = var_4445_cast_fp16)[name = string("v_new_245_cast_fp16")]; + tensor var_4448_begin_0 = const()[name = string("op_4448_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_4448_end_0 = const()[name = string("op_4448_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_4448_end_mask_0 = const()[name = string("op_4448_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4448_squeeze_mask_0 = const()[name = string("op_4448_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4448_cast_fp16 = slice_by_index(begin = var_4448_begin_0, end = var_4448_end_0, end_mask = var_4448_end_mask_0, squeeze_mask = var_4448_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4448_cast_fp16")]; + bool var_4449_transpose_x_0 = const()[name = string("op_4449_transpose_x_0"), val = bool(false)]; + bool var_4449_transpose_y_0 = const()[name = string("op_4449_transpose_y_0"), val = bool(false)]; + tensor var_4449_cast_fp16 = matmul(transpose_x = var_4449_transpose_x_0, transpose_y = var_4449_transpose_y_0, x = var_4448_cast_fp16, y = state_245_cast_fp16)[name = string("op_4449_cast_fp16")]; + tensor var_4451_begin_0 = const()[name = string("op_4451_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_4451_end_0 = const()[name = string("op_4451_end_0"), val = tensor([32, 27, 64, 64])]; + tensor var_4451_end_mask_0 = const()[name = string("op_4451_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4451_squeeze_mask_0 = const()[name = string("op_4451_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4451_cast_fp16 = slice_by_index(begin = var_4451_begin_0, end = var_4451_end_0, end_mask = var_4451_end_mask_0, squeeze_mask = var_4451_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4451_cast_fp16")]; + bool var_4452_transpose_x_0 = const()[name = string("op_4452_transpose_x_0"), val = bool(false)]; + bool var_4452_transpose_y_0 = const()[name = string("op_4452_transpose_y_0"), val = bool(false)]; + tensor var_4452_cast_fp16 = matmul(transpose_x = var_4452_transpose_x_0, transpose_y = var_4452_transpose_y_0, x = var_4451_cast_fp16, y = v_new_245_cast_fp16)[name = string("op_4452_cast_fp16")]; + tensor var_4453_cast_fp16 = add(x = var_4449_cast_fp16, y = var_4452_cast_fp16)[name = string("op_4453_cast_fp16")]; + tensor var_4455_begin_0 = const()[name = string("op_4455_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_4455_end_0 = const()[name = string("op_4455_end_0"), val = tensor([32, 27, 1, 1])]; + tensor var_4455_end_mask_0 = const()[name = string("op_4455_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4455_squeeze_mask_0 = const()[name = string("op_4455_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4455_cast_fp16 = slice_by_index(begin = var_4455_begin_0, end = var_4455_end_0, end_mask = var_4455_end_mask_0, squeeze_mask = var_4455_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4455_cast_fp16")]; + tensor var_4456_cast_fp16 = mul(x = state_245_cast_fp16, y = var_4455_cast_fp16)[name = string("op_4456_cast_fp16")]; + tensor var_4458_begin_0 = const()[name = string("op_4458_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_4458_end_0 = const()[name = string("op_4458_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_4458_end_mask_0 = const()[name = string("op_4458_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4458_squeeze_mask_0 = const()[name = string("op_4458_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4458_cast_fp16 = slice_by_index(begin = var_4458_begin_0, end = var_4458_end_0, end_mask = var_4458_end_mask_0, squeeze_mask = var_4458_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4458_cast_fp16")]; + bool var_4460_transpose_x_1 = const()[name = string("op_4460_transpose_x_1"), val = bool(true)]; + bool var_4460_transpose_y_1 = const()[name = string("op_4460_transpose_y_1"), val = bool(false)]; + tensor var_4460_cast_fp16 = matmul(transpose_x = var_4460_transpose_x_1, transpose_y = var_4460_transpose_y_1, x = var_4458_cast_fp16, y = v_new_245_cast_fp16)[name = string("op_4460_cast_fp16")]; + tensor state_247_cast_fp16 = add(x = var_4456_cast_fp16, y = var_4460_cast_fp16)[name = string("state_247_cast_fp16")]; + tensor var_4463_begin_0 = const()[name = string("op_4463_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_4463_end_0 = const()[name = string("op_4463_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_4463_end_mask_0 = const()[name = string("op_4463_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4463_squeeze_mask_0 = const()[name = string("op_4463_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4463_cast_fp16 = slice_by_index(begin = var_4463_begin_0, end = var_4463_end_0, end_mask = var_4463_end_mask_0, squeeze_mask = var_4463_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4463_cast_fp16")]; + tensor var_4465_begin_0 = const()[name = string("op_4465_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_4465_end_0 = const()[name = string("op_4465_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_4465_end_mask_0 = const()[name = string("op_4465_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4465_squeeze_mask_0 = const()[name = string("op_4465_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4465_cast_fp16 = slice_by_index(begin = var_4465_begin_0, end = var_4465_end_0, end_mask = var_4465_end_mask_0, squeeze_mask = var_4465_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4465_cast_fp16")]; + bool var_4466_transpose_x_0 = const()[name = string("op_4466_transpose_x_0"), val = bool(false)]; + bool var_4466_transpose_y_0 = const()[name = string("op_4466_transpose_y_0"), val = bool(false)]; + tensor var_4466_cast_fp16 = matmul(transpose_x = var_4466_transpose_x_0, transpose_y = var_4466_transpose_y_0, x = var_4465_cast_fp16, y = state_247_cast_fp16)[name = string("op_4466_cast_fp16")]; + tensor v_new_247_cast_fp16 = sub(x = var_4463_cast_fp16, y = var_4466_cast_fp16)[name = string("v_new_247_cast_fp16")]; + tensor var_4469_begin_0 = const()[name = string("op_4469_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_4469_end_0 = const()[name = string("op_4469_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_4469_end_mask_0 = const()[name = string("op_4469_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4469_squeeze_mask_0 = const()[name = string("op_4469_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4469_cast_fp16 = slice_by_index(begin = var_4469_begin_0, end = var_4469_end_0, end_mask = var_4469_end_mask_0, squeeze_mask = var_4469_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4469_cast_fp16")]; + bool var_4470_transpose_x_0 = const()[name = string("op_4470_transpose_x_0"), val = bool(false)]; + bool var_4470_transpose_y_0 = const()[name = string("op_4470_transpose_y_0"), val = bool(false)]; + tensor var_4470_cast_fp16 = matmul(transpose_x = var_4470_transpose_x_0, transpose_y = var_4470_transpose_y_0, x = var_4469_cast_fp16, y = state_247_cast_fp16)[name = string("op_4470_cast_fp16")]; + tensor var_4472_begin_0 = const()[name = string("op_4472_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_4472_end_0 = const()[name = string("op_4472_end_0"), val = tensor([32, 28, 64, 64])]; + tensor var_4472_end_mask_0 = const()[name = string("op_4472_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4472_squeeze_mask_0 = const()[name = string("op_4472_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4472_cast_fp16 = slice_by_index(begin = var_4472_begin_0, end = var_4472_end_0, end_mask = var_4472_end_mask_0, squeeze_mask = var_4472_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4472_cast_fp16")]; + bool var_4473_transpose_x_0 = const()[name = string("op_4473_transpose_x_0"), val = bool(false)]; + bool var_4473_transpose_y_0 = const()[name = string("op_4473_transpose_y_0"), val = bool(false)]; + tensor var_4473_cast_fp16 = matmul(transpose_x = var_4473_transpose_x_0, transpose_y = var_4473_transpose_y_0, x = var_4472_cast_fp16, y = v_new_247_cast_fp16)[name = string("op_4473_cast_fp16")]; + tensor var_4474_cast_fp16 = add(x = var_4470_cast_fp16, y = var_4473_cast_fp16)[name = string("op_4474_cast_fp16")]; + tensor var_4476_begin_0 = const()[name = string("op_4476_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_4476_end_0 = const()[name = string("op_4476_end_0"), val = tensor([32, 28, 1, 1])]; + tensor var_4476_end_mask_0 = const()[name = string("op_4476_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4476_squeeze_mask_0 = const()[name = string("op_4476_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4476_cast_fp16 = slice_by_index(begin = var_4476_begin_0, end = var_4476_end_0, end_mask = var_4476_end_mask_0, squeeze_mask = var_4476_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4476_cast_fp16")]; + tensor var_4477_cast_fp16 = mul(x = state_247_cast_fp16, y = var_4476_cast_fp16)[name = string("op_4477_cast_fp16")]; + tensor var_4479_begin_0 = const()[name = string("op_4479_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_4479_end_0 = const()[name = string("op_4479_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_4479_end_mask_0 = const()[name = string("op_4479_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4479_squeeze_mask_0 = const()[name = string("op_4479_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4479_cast_fp16 = slice_by_index(begin = var_4479_begin_0, end = var_4479_end_0, end_mask = var_4479_end_mask_0, squeeze_mask = var_4479_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4479_cast_fp16")]; + bool var_4481_transpose_x_1 = const()[name = string("op_4481_transpose_x_1"), val = bool(true)]; + bool var_4481_transpose_y_1 = const()[name = string("op_4481_transpose_y_1"), val = bool(false)]; + tensor var_4481_cast_fp16 = matmul(transpose_x = var_4481_transpose_x_1, transpose_y = var_4481_transpose_y_1, x = var_4479_cast_fp16, y = v_new_247_cast_fp16)[name = string("op_4481_cast_fp16")]; + tensor state_249_cast_fp16 = add(x = var_4477_cast_fp16, y = var_4481_cast_fp16)[name = string("state_249_cast_fp16")]; + tensor var_4484_begin_0 = const()[name = string("op_4484_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_4484_end_0 = const()[name = string("op_4484_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_4484_end_mask_0 = const()[name = string("op_4484_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4484_squeeze_mask_0 = const()[name = string("op_4484_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4484_cast_fp16 = slice_by_index(begin = var_4484_begin_0, end = var_4484_end_0, end_mask = var_4484_end_mask_0, squeeze_mask = var_4484_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4484_cast_fp16")]; + tensor var_4486_begin_0 = const()[name = string("op_4486_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_4486_end_0 = const()[name = string("op_4486_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_4486_end_mask_0 = const()[name = string("op_4486_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4486_squeeze_mask_0 = const()[name = string("op_4486_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4486_cast_fp16 = slice_by_index(begin = var_4486_begin_0, end = var_4486_end_0, end_mask = var_4486_end_mask_0, squeeze_mask = var_4486_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4486_cast_fp16")]; + bool var_4487_transpose_x_0 = const()[name = string("op_4487_transpose_x_0"), val = bool(false)]; + bool var_4487_transpose_y_0 = const()[name = string("op_4487_transpose_y_0"), val = bool(false)]; + tensor var_4487_cast_fp16 = matmul(transpose_x = var_4487_transpose_x_0, transpose_y = var_4487_transpose_y_0, x = var_4486_cast_fp16, y = state_249_cast_fp16)[name = string("op_4487_cast_fp16")]; + tensor v_new_249_cast_fp16 = sub(x = var_4484_cast_fp16, y = var_4487_cast_fp16)[name = string("v_new_249_cast_fp16")]; + tensor var_4490_begin_0 = const()[name = string("op_4490_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_4490_end_0 = const()[name = string("op_4490_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_4490_end_mask_0 = const()[name = string("op_4490_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4490_squeeze_mask_0 = const()[name = string("op_4490_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4490_cast_fp16 = slice_by_index(begin = var_4490_begin_0, end = var_4490_end_0, end_mask = var_4490_end_mask_0, squeeze_mask = var_4490_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4490_cast_fp16")]; + bool var_4491_transpose_x_0 = const()[name = string("op_4491_transpose_x_0"), val = bool(false)]; + bool var_4491_transpose_y_0 = const()[name = string("op_4491_transpose_y_0"), val = bool(false)]; + tensor var_4491_cast_fp16 = matmul(transpose_x = var_4491_transpose_x_0, transpose_y = var_4491_transpose_y_0, x = var_4490_cast_fp16, y = state_249_cast_fp16)[name = string("op_4491_cast_fp16")]; + tensor var_4493_begin_0 = const()[name = string("op_4493_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_4493_end_0 = const()[name = string("op_4493_end_0"), val = tensor([32, 29, 64, 64])]; + tensor var_4493_end_mask_0 = const()[name = string("op_4493_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4493_squeeze_mask_0 = const()[name = string("op_4493_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4493_cast_fp16 = slice_by_index(begin = var_4493_begin_0, end = var_4493_end_0, end_mask = var_4493_end_mask_0, squeeze_mask = var_4493_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4493_cast_fp16")]; + bool var_4494_transpose_x_0 = const()[name = string("op_4494_transpose_x_0"), val = bool(false)]; + bool var_4494_transpose_y_0 = const()[name = string("op_4494_transpose_y_0"), val = bool(false)]; + tensor var_4494_cast_fp16 = matmul(transpose_x = var_4494_transpose_x_0, transpose_y = var_4494_transpose_y_0, x = var_4493_cast_fp16, y = v_new_249_cast_fp16)[name = string("op_4494_cast_fp16")]; + tensor var_4495_cast_fp16 = add(x = var_4491_cast_fp16, y = var_4494_cast_fp16)[name = string("op_4495_cast_fp16")]; + tensor var_4497_begin_0 = const()[name = string("op_4497_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_4497_end_0 = const()[name = string("op_4497_end_0"), val = tensor([32, 29, 1, 1])]; + tensor var_4497_end_mask_0 = const()[name = string("op_4497_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4497_squeeze_mask_0 = const()[name = string("op_4497_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4497_cast_fp16 = slice_by_index(begin = var_4497_begin_0, end = var_4497_end_0, end_mask = var_4497_end_mask_0, squeeze_mask = var_4497_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4497_cast_fp16")]; + tensor var_4498_cast_fp16 = mul(x = state_249_cast_fp16, y = var_4497_cast_fp16)[name = string("op_4498_cast_fp16")]; + tensor var_4500_begin_0 = const()[name = string("op_4500_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_4500_end_0 = const()[name = string("op_4500_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_4500_end_mask_0 = const()[name = string("op_4500_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4500_squeeze_mask_0 = const()[name = string("op_4500_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4500_cast_fp16 = slice_by_index(begin = var_4500_begin_0, end = var_4500_end_0, end_mask = var_4500_end_mask_0, squeeze_mask = var_4500_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4500_cast_fp16")]; + bool var_4502_transpose_x_1 = const()[name = string("op_4502_transpose_x_1"), val = bool(true)]; + bool var_4502_transpose_y_1 = const()[name = string("op_4502_transpose_y_1"), val = bool(false)]; + tensor var_4502_cast_fp16 = matmul(transpose_x = var_4502_transpose_x_1, transpose_y = var_4502_transpose_y_1, x = var_4500_cast_fp16, y = v_new_249_cast_fp16)[name = string("op_4502_cast_fp16")]; + tensor state_251_cast_fp16 = add(x = var_4498_cast_fp16, y = var_4502_cast_fp16)[name = string("state_251_cast_fp16")]; + tensor var_4505_begin_0 = const()[name = string("op_4505_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_4505_end_0 = const()[name = string("op_4505_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_4505_end_mask_0 = const()[name = string("op_4505_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4505_squeeze_mask_0 = const()[name = string("op_4505_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4505_cast_fp16 = slice_by_index(begin = var_4505_begin_0, end = var_4505_end_0, end_mask = var_4505_end_mask_0, squeeze_mask = var_4505_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4505_cast_fp16")]; + tensor var_4507_begin_0 = const()[name = string("op_4507_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_4507_end_0 = const()[name = string("op_4507_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_4507_end_mask_0 = const()[name = string("op_4507_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4507_squeeze_mask_0 = const()[name = string("op_4507_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4507_cast_fp16 = slice_by_index(begin = var_4507_begin_0, end = var_4507_end_0, end_mask = var_4507_end_mask_0, squeeze_mask = var_4507_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4507_cast_fp16")]; + bool var_4508_transpose_x_0 = const()[name = string("op_4508_transpose_x_0"), val = bool(false)]; + bool var_4508_transpose_y_0 = const()[name = string("op_4508_transpose_y_0"), val = bool(false)]; + tensor var_4508_cast_fp16 = matmul(transpose_x = var_4508_transpose_x_0, transpose_y = var_4508_transpose_y_0, x = var_4507_cast_fp16, y = state_251_cast_fp16)[name = string("op_4508_cast_fp16")]; + tensor v_new_251_cast_fp16 = sub(x = var_4505_cast_fp16, y = var_4508_cast_fp16)[name = string("v_new_251_cast_fp16")]; + tensor var_4511_begin_0 = const()[name = string("op_4511_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_4511_end_0 = const()[name = string("op_4511_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_4511_end_mask_0 = const()[name = string("op_4511_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4511_squeeze_mask_0 = const()[name = string("op_4511_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4511_cast_fp16 = slice_by_index(begin = var_4511_begin_0, end = var_4511_end_0, end_mask = var_4511_end_mask_0, squeeze_mask = var_4511_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4511_cast_fp16")]; + bool var_4512_transpose_x_0 = const()[name = string("op_4512_transpose_x_0"), val = bool(false)]; + bool var_4512_transpose_y_0 = const()[name = string("op_4512_transpose_y_0"), val = bool(false)]; + tensor var_4512_cast_fp16 = matmul(transpose_x = var_4512_transpose_x_0, transpose_y = var_4512_transpose_y_0, x = var_4511_cast_fp16, y = state_251_cast_fp16)[name = string("op_4512_cast_fp16")]; + tensor var_4514_begin_0 = const()[name = string("op_4514_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_4514_end_0 = const()[name = string("op_4514_end_0"), val = tensor([32, 30, 64, 64])]; + tensor var_4514_end_mask_0 = const()[name = string("op_4514_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4514_squeeze_mask_0 = const()[name = string("op_4514_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4514_cast_fp16 = slice_by_index(begin = var_4514_begin_0, end = var_4514_end_0, end_mask = var_4514_end_mask_0, squeeze_mask = var_4514_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4514_cast_fp16")]; + bool var_4515_transpose_x_0 = const()[name = string("op_4515_transpose_x_0"), val = bool(false)]; + bool var_4515_transpose_y_0 = const()[name = string("op_4515_transpose_y_0"), val = bool(false)]; + tensor var_4515_cast_fp16 = matmul(transpose_x = var_4515_transpose_x_0, transpose_y = var_4515_transpose_y_0, x = var_4514_cast_fp16, y = v_new_251_cast_fp16)[name = string("op_4515_cast_fp16")]; + tensor var_4516_cast_fp16 = add(x = var_4512_cast_fp16, y = var_4515_cast_fp16)[name = string("op_4516_cast_fp16")]; + tensor var_4518_begin_0 = const()[name = string("op_4518_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_4518_end_0 = const()[name = string("op_4518_end_0"), val = tensor([32, 30, 1, 1])]; + tensor var_4518_end_mask_0 = const()[name = string("op_4518_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4518_squeeze_mask_0 = const()[name = string("op_4518_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4518_cast_fp16 = slice_by_index(begin = var_4518_begin_0, end = var_4518_end_0, end_mask = var_4518_end_mask_0, squeeze_mask = var_4518_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4518_cast_fp16")]; + tensor var_4519_cast_fp16 = mul(x = state_251_cast_fp16, y = var_4518_cast_fp16)[name = string("op_4519_cast_fp16")]; + tensor var_4521_begin_0 = const()[name = string("op_4521_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_4521_end_0 = const()[name = string("op_4521_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_4521_end_mask_0 = const()[name = string("op_4521_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4521_squeeze_mask_0 = const()[name = string("op_4521_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4521_cast_fp16 = slice_by_index(begin = var_4521_begin_0, end = var_4521_end_0, end_mask = var_4521_end_mask_0, squeeze_mask = var_4521_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4521_cast_fp16")]; + bool var_4523_transpose_x_1 = const()[name = string("op_4523_transpose_x_1"), val = bool(true)]; + bool var_4523_transpose_y_1 = const()[name = string("op_4523_transpose_y_1"), val = bool(false)]; + tensor var_4523_cast_fp16 = matmul(transpose_x = var_4523_transpose_x_1, transpose_y = var_4523_transpose_y_1, x = var_4521_cast_fp16, y = v_new_251_cast_fp16)[name = string("op_4523_cast_fp16")]; + tensor state_253_cast_fp16 = add(x = var_4519_cast_fp16, y = var_4523_cast_fp16)[name = string("state_253_cast_fp16")]; + tensor var_4526_begin_0 = const()[name = string("op_4526_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_4526_end_0 = const()[name = string("op_4526_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_4526_end_mask_0 = const()[name = string("op_4526_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4526_squeeze_mask_0 = const()[name = string("op_4526_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4526_cast_fp16 = slice_by_index(begin = var_4526_begin_0, end = var_4526_end_0, end_mask = var_4526_end_mask_0, squeeze_mask = var_4526_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4526_cast_fp16")]; + tensor var_4528_begin_0 = const()[name = string("op_4528_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_4528_end_0 = const()[name = string("op_4528_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_4528_end_mask_0 = const()[name = string("op_4528_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4528_squeeze_mask_0 = const()[name = string("op_4528_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4528_cast_fp16 = slice_by_index(begin = var_4528_begin_0, end = var_4528_end_0, end_mask = var_4528_end_mask_0, squeeze_mask = var_4528_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4528_cast_fp16")]; + bool var_4529_transpose_x_0 = const()[name = string("op_4529_transpose_x_0"), val = bool(false)]; + bool var_4529_transpose_y_0 = const()[name = string("op_4529_transpose_y_0"), val = bool(false)]; + tensor var_4529_cast_fp16 = matmul(transpose_x = var_4529_transpose_x_0, transpose_y = var_4529_transpose_y_0, x = var_4528_cast_fp16, y = state_253_cast_fp16)[name = string("op_4529_cast_fp16")]; + tensor v_new_253_cast_fp16 = sub(x = var_4526_cast_fp16, y = var_4529_cast_fp16)[name = string("v_new_253_cast_fp16")]; + tensor var_4532_begin_0 = const()[name = string("op_4532_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_4532_end_0 = const()[name = string("op_4532_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_4532_end_mask_0 = const()[name = string("op_4532_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4532_squeeze_mask_0 = const()[name = string("op_4532_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4532_cast_fp16 = slice_by_index(begin = var_4532_begin_0, end = var_4532_end_0, end_mask = var_4532_end_mask_0, squeeze_mask = var_4532_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4532_cast_fp16")]; + bool var_4533_transpose_x_0 = const()[name = string("op_4533_transpose_x_0"), val = bool(false)]; + bool var_4533_transpose_y_0 = const()[name = string("op_4533_transpose_y_0"), val = bool(false)]; + tensor var_4533_cast_fp16 = matmul(transpose_x = var_4533_transpose_x_0, transpose_y = var_4533_transpose_y_0, x = var_4532_cast_fp16, y = state_253_cast_fp16)[name = string("op_4533_cast_fp16")]; + tensor var_4535_begin_0 = const()[name = string("op_4535_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_4535_end_0 = const()[name = string("op_4535_end_0"), val = tensor([32, 31, 64, 64])]; + tensor var_4535_end_mask_0 = const()[name = string("op_4535_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4535_squeeze_mask_0 = const()[name = string("op_4535_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4535_cast_fp16 = slice_by_index(begin = var_4535_begin_0, end = var_4535_end_0, end_mask = var_4535_end_mask_0, squeeze_mask = var_4535_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4535_cast_fp16")]; + bool var_4536_transpose_x_0 = const()[name = string("op_4536_transpose_x_0"), val = bool(false)]; + bool var_4536_transpose_y_0 = const()[name = string("op_4536_transpose_y_0"), val = bool(false)]; + tensor var_4536_cast_fp16 = matmul(transpose_x = var_4536_transpose_x_0, transpose_y = var_4536_transpose_y_0, x = var_4535_cast_fp16, y = v_new_253_cast_fp16)[name = string("op_4536_cast_fp16")]; + tensor var_4537_cast_fp16 = add(x = var_4533_cast_fp16, y = var_4536_cast_fp16)[name = string("op_4537_cast_fp16")]; + tensor var_4539_begin_0 = const()[name = string("op_4539_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_4539_end_0 = const()[name = string("op_4539_end_0"), val = tensor([32, 31, 1, 1])]; + tensor var_4539_end_mask_0 = const()[name = string("op_4539_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4539_squeeze_mask_0 = const()[name = string("op_4539_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4539_cast_fp16 = slice_by_index(begin = var_4539_begin_0, end = var_4539_end_0, end_mask = var_4539_end_mask_0, squeeze_mask = var_4539_squeeze_mask_0, x = chunk_decay_7_cast_fp16)[name = string("op_4539_cast_fp16")]; + tensor var_4540_cast_fp16 = mul(x = state_253_cast_fp16, y = var_4539_cast_fp16)[name = string("op_4540_cast_fp16")]; + tensor var_4542_begin_0 = const()[name = string("op_4542_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_4542_end_0 = const()[name = string("op_4542_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_4542_end_mask_0 = const()[name = string("op_4542_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4542_squeeze_mask_0 = const()[name = string("op_4542_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4542_cast_fp16 = slice_by_index(begin = var_4542_begin_0, end = var_4542_end_0, end_mask = var_4542_end_mask_0, squeeze_mask = var_4542_squeeze_mask_0, x = k_dec_7_cast_fp16)[name = string("op_4542_cast_fp16")]; + bool var_4544_transpose_x_1 = const()[name = string("op_4544_transpose_x_1"), val = bool(true)]; + bool var_4544_transpose_y_1 = const()[name = string("op_4544_transpose_y_1"), val = bool(false)]; + tensor var_4544_cast_fp16 = matmul(transpose_x = var_4544_transpose_x_1, transpose_y = var_4544_transpose_y_1, x = var_4542_cast_fp16, y = v_new_253_cast_fp16)[name = string("op_4544_cast_fp16")]; + tensor state_255_cast_fp16 = add(x = var_4540_cast_fp16, y = var_4544_cast_fp16)[name = string("state_255_cast_fp16")]; + tensor var_4547_begin_0 = const()[name = string("op_4547_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_4547_end_0 = const()[name = string("op_4547_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_4547_end_mask_0 = const()[name = string("op_4547_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4547_squeeze_mask_0 = const()[name = string("op_4547_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4547_cast_fp16 = slice_by_index(begin = var_4547_begin_0, end = var_4547_end_0, end_mask = var_4547_end_mask_0, squeeze_mask = var_4547_squeeze_mask_0, x = new_v_7_cast_fp16)[name = string("op_4547_cast_fp16")]; + tensor var_4549_begin_0 = const()[name = string("op_4549_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_4549_end_0 = const()[name = string("op_4549_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_4549_end_mask_0 = const()[name = string("op_4549_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4549_squeeze_mask_0 = const()[name = string("op_4549_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4549_cast_fp16 = slice_by_index(begin = var_4549_begin_0, end = var_4549_end_0, end_mask = var_4549_end_mask_0, squeeze_mask = var_4549_squeeze_mask_0, x = k_cumdecay_7_cast_fp16)[name = string("op_4549_cast_fp16")]; + bool var_4550_transpose_x_0 = const()[name = string("op_4550_transpose_x_0"), val = bool(false)]; + bool var_4550_transpose_y_0 = const()[name = string("op_4550_transpose_y_0"), val = bool(false)]; + tensor var_4550_cast_fp16 = matmul(transpose_x = var_4550_transpose_x_0, transpose_y = var_4550_transpose_y_0, x = var_4549_cast_fp16, y = state_255_cast_fp16)[name = string("op_4550_cast_fp16")]; + tensor v_new_255_cast_fp16 = sub(x = var_4547_cast_fp16, y = var_4550_cast_fp16)[name = string("v_new_255_cast_fp16")]; + tensor var_4553_begin_0 = const()[name = string("op_4553_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_4553_end_0 = const()[name = string("op_4553_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_4553_end_mask_0 = const()[name = string("op_4553_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4553_squeeze_mask_0 = const()[name = string("op_4553_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4553_cast_fp16 = slice_by_index(begin = var_4553_begin_0, end = var_4553_end_0, end_mask = var_4553_end_mask_0, squeeze_mask = var_4553_squeeze_mask_0, x = q_dec_7_cast_fp16)[name = string("op_4553_cast_fp16")]; + bool var_4554_transpose_x_0 = const()[name = string("op_4554_transpose_x_0"), val = bool(false)]; + bool var_4554_transpose_y_0 = const()[name = string("op_4554_transpose_y_0"), val = bool(false)]; + tensor var_4554_cast_fp16 = matmul(transpose_x = var_4554_transpose_x_0, transpose_y = var_4554_transpose_y_0, x = var_4553_cast_fp16, y = state_255_cast_fp16)[name = string("op_4554_cast_fp16")]; + tensor var_4556_begin_0 = const()[name = string("op_4556_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_4556_end_0 = const()[name = string("op_4556_end_0"), val = tensor([32, 32, 64, 64])]; + tensor var_4556_end_mask_0 = const()[name = string("op_4556_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4556_squeeze_mask_0 = const()[name = string("op_4556_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4556_cast_fp16 = slice_by_index(begin = var_4556_begin_0, end = var_4556_end_0, end_mask = var_4556_end_mask_0, squeeze_mask = var_4556_squeeze_mask_0, x = attn_intra_7_cast_fp16)[name = string("op_4556_cast_fp16")]; + bool var_4557_transpose_x_0 = const()[name = string("op_4557_transpose_x_0"), val = bool(false)]; + bool var_4557_transpose_y_0 = const()[name = string("op_4557_transpose_y_0"), val = bool(false)]; + tensor var_4557_cast_fp16 = matmul(transpose_x = var_4557_transpose_x_0, transpose_y = var_4557_transpose_y_0, x = var_4556_cast_fp16, y = v_new_255_cast_fp16)[name = string("op_4557_cast_fp16")]; + tensor var_4558_cast_fp16 = add(x = var_4554_cast_fp16, y = var_4557_cast_fp16)[name = string("op_4558_cast_fp16")]; + int32 var_4560_axis_0 = const()[name = string("op_4560_axis_0"), val = int32(1)]; + tensor var_4560_cast_fp16 = stack(axis = var_4560_axis_0, values = (var_3907_cast_fp16, var_3928_cast_fp16, var_3949_cast_fp16, var_3970_cast_fp16, var_3991_cast_fp16, var_4012_cast_fp16, var_4033_cast_fp16, var_4054_cast_fp16, var_4075_cast_fp16, var_4096_cast_fp16, var_4117_cast_fp16, var_4138_cast_fp16, var_4159_cast_fp16, var_4180_cast_fp16, var_4201_cast_fp16, var_4222_cast_fp16, var_4243_cast_fp16, var_4264_cast_fp16, var_4285_cast_fp16, var_4306_cast_fp16, var_4327_cast_fp16, var_4348_cast_fp16, var_4369_cast_fp16, var_4390_cast_fp16, var_4411_cast_fp16, var_4432_cast_fp16, var_4453_cast_fp16, var_4474_cast_fp16, var_4495_cast_fp16, var_4516_cast_fp16, var_4537_cast_fp16, var_4558_cast_fp16))[name = string("op_4560_cast_fp16")]; + tensor var_4561 = const()[name = string("op_4561"), val = tensor([32, 2048, 128])]; + tensor var_4562_cast_fp16 = reshape(shape = var_4561, x = var_4560_cast_fp16)[name = string("op_4562_cast_fp16")]; + tensor var_4563 = const()[name = string("op_4563"), val = tensor([1, 0, 2])]; + tensor input_63_begin_0 = const()[name = string("input_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor input_63_end_0 = const()[name = string("input_63_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor input_63_end_mask_0 = const()[name = string("input_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor input_63_squeeze_mask_0 = const()[name = string("input_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor input_63_cast_fp16 = slice_by_index(begin = input_63_begin_0, end = input_63_end_0, end_mask = input_63_end_mask_0, squeeze_mask = input_63_squeeze_mask_0, x = z_7_cast_fp16)[name = string("input_63_cast_fp16")]; + fp16 var_3543_promoted_1_to_fp16 = const()[name = string("op_3543_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor x_117_cast_fp16 = transpose(perm = var_4563, x = var_4562_cast_fp16)[name = string("transpose_20")]; + tensor var_4567_cast_fp16 = pow(x = x_117_cast_fp16, y = var_3543_promoted_1_to_fp16)[name = string("op_4567_cast_fp16")]; + tensor var_4569_axes_0 = const()[name = string("op_4569_axes_0"), val = tensor([-1])]; + bool var_4569_keep_dims_0 = const()[name = string("op_4569_keep_dims_0"), val = bool(true)]; + tensor var_4569_cast_fp16 = reduce_mean(axes = var_4569_axes_0, keep_dims = var_4569_keep_dims_0, x = var_4567_cast_fp16)[name = string("op_4569_cast_fp16")]; + fp16 var_4570_to_fp16 = const()[name = string("op_4570_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_4571_cast_fp16 = add(x = var_4569_cast_fp16, y = var_4570_to_fp16)[name = string("op_4571_cast_fp16")]; + fp32 var_4572_epsilon_0 = const()[name = string("op_4572_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4572_cast_fp16 = rsqrt(epsilon = var_4572_epsilon_0, x = var_4571_cast_fp16)[name = string("op_4572_cast_fp16")]; + tensor x_119_cast_fp16 = mul(x = x_117_cast_fp16, y = var_4572_cast_fp16)[name = string("x_119_cast_fp16")]; + tensor layers_4_linear_attn_norm_weight_to_fp16 = const()[name = string("layers_4_linear_attn_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493019008)))]; + tensor var_4574_cast_fp16 = mul(x = layers_4_linear_attn_norm_weight_to_fp16, y = x_119_cast_fp16)[name = string("op_4574_cast_fp16")]; + tensor var_4575_cast_fp16 = silu(x = input_63_cast_fp16)[name = string("op_4575_cast_fp16")]; + tensor core_7_cast_fp16 = mul(x = var_4574_cast_fp16, y = var_4575_cast_fp16)[name = string("core_7_cast_fp16")]; + tensor var_4577 = const()[name = string("op_4577"), val = tensor([1, 2048, 4096])]; + tensor input_65_cast_fp16 = reshape(shape = var_4577, x = core_7_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor layers_4_linear_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493019328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503505152))))[name = string("layers_4_linear_attn_out_proj_weight_to_fp16_quantized")]; + tensor linear_35_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_linear_attn_out_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = x_99_cast_fp16, y = linear_35_cast_fp16)[name = string("x_121_cast_fp16")]; + fp16 var_3543_promoted_2_to_fp16 = const()[name = string("op_3543_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor var_4583_cast_fp16 = pow(x = x_121_cast_fp16, y = var_3543_promoted_2_to_fp16)[name = string("op_4583_cast_fp16")]; + tensor var_4585_axes_0 = const()[name = string("op_4585_axes_0"), val = tensor([-1])]; + bool var_4585_keep_dims_0 = const()[name = string("op_4585_keep_dims_0"), val = bool(true)]; + tensor var_4585_cast_fp16 = reduce_mean(axes = var_4585_axes_0, keep_dims = var_4585_keep_dims_0, x = var_4583_cast_fp16)[name = string("op_4585_cast_fp16")]; + fp16 var_4586_to_fp16 = const()[name = string("op_4586_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_4587_cast_fp16 = add(x = var_4585_cast_fp16, y = var_4586_to_fp16)[name = string("op_4587_cast_fp16")]; + fp32 var_4588_epsilon_0 = const()[name = string("op_4588_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4588_cast_fp16 = rsqrt(epsilon = var_4588_epsilon_0, x = var_4587_cast_fp16)[name = string("op_4588_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_121_cast_fp16, y = var_4588_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor var_4591_to_fp16 = const()[name = string("op_4591_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503510336)))]; + tensor input_67_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4591_to_fp16)[name = string("input_67_cast_fp16")]; + tensor layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503515520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527108544))))[name = string("layers_4_mlp_gate_proj_weight_to_fp16_quantized")]; + tensor linear_36_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_36_cast_fp16")]; + tensor var_4598_cast_fp16 = silu(x = linear_36_cast_fp16)[name = string("op_4598_cast_fp16")]; + tensor layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527127040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550720064))))[name = string("layers_4_mlp_up_proj_weight_to_fp16_quantized")]; + tensor linear_37_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_71_cast_fp16 = mul(x = var_4598_cast_fp16, y = linear_37_cast_fp16)[name = string("input_71_cast_fp16")]; + tensor layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550738560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574331584))))[name = string("layers_4_mlp_down_proj_weight_to_fp16_quantized")]; + tensor linear_38_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_38_cast_fp16")]; + tensor x_125_cast_fp16 = add(x = x_121_cast_fp16, y = linear_38_cast_fp16)[name = string("x_125_cast_fp16")]; + int32 var_4618 = const()[name = string("op_4618"), val = int32(-2)]; + int32 var_4647 = const()[name = string("op_4647"), val = int32(-1)]; + fp16 var_4646_promoted_to_fp16 = const()[name = string("op_4646_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_4656_cast_fp16 = pow(x = x_125_cast_fp16, y = var_4646_promoted_to_fp16)[name = string("op_4656_cast_fp16")]; + tensor var_4658_axes_0 = const()[name = string("op_4658_axes_0"), val = tensor([-1])]; + bool var_4658_keep_dims_0 = const()[name = string("op_4658_keep_dims_0"), val = bool(true)]; + tensor var_4658_cast_fp16 = reduce_mean(axes = var_4658_axes_0, keep_dims = var_4658_keep_dims_0, x = var_4656_cast_fp16)[name = string("op_4658_cast_fp16")]; + fp16 var_4659_to_fp16 = const()[name = string("op_4659_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_4660_cast_fp16 = add(x = var_4658_cast_fp16, y = var_4659_to_fp16)[name = string("op_4660_cast_fp16")]; + fp32 var_4661_epsilon_0 = const()[name = string("op_4661_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4661_cast_fp16 = rsqrt(epsilon = var_4661_epsilon_0, x = var_4660_cast_fp16)[name = string("op_4661_cast_fp16")]; + tensor x_127_cast_fp16 = mul(x = x_125_cast_fp16, y = var_4661_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor var_4664_to_fp16 = const()[name = string("op_4664_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574336768)))]; + tensor x_129_cast_fp16 = mul(x = x_127_cast_fp16, y = var_4664_to_fp16)[name = string("x_129_cast_fp16")]; + tensor layers_5_linear_attn_in_proj_qkv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574341952))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595313536))))[name = string("layers_5_linear_attn_in_proj_qkv_weight_to_fp16_quantized")]; + tensor linear_39_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_linear_attn_in_proj_qkv_weight_to_fp16_quantized, x = x_129_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor input_73_perm_0 = const()[name = string("input_73_perm_0"), val = tensor([0, 2, 1])]; + tensor input_75_pad_0 = const()[name = string("input_75_pad_0"), val = tensor([0, 0, 0, 0, 3, 0])]; + string input_75_mode_0 = const()[name = string("input_75_mode_0"), val = string("constant")]; + fp16 const_54_to_fp16 = const()[name = string("const_54_to_fp16"), val = fp16(0x0p+0)]; + tensor input_73_cast_fp16 = transpose(perm = input_73_perm_0, x = linear_39_cast_fp16)[name = string("transpose_19")]; + tensor input_75_cast_fp16 = pad(constant_val = const_54_to_fp16, mode = input_75_mode_0, pad = input_75_pad_0, x = input_73_cast_fp16)[name = string("input_75_cast_fp16")]; + string input_77_pad_type_0 = const()[name = string("input_77_pad_type_0"), val = string("valid")]; + int32 input_77_groups_0 = const()[name = string("input_77_groups_0"), val = int32(8192)]; + tensor input_77_strides_0 = const()[name = string("input_77_strides_0"), val = tensor([1])]; + tensor input_77_pad_0 = const()[name = string("input_77_pad_0"), val = tensor([0, 0])]; + tensor input_77_dilations_0 = const()[name = string("input_77_dilations_0"), val = tensor([1])]; + tensor layers_5_linear_attn_conv1d_weight_to_fp16 = const()[name = string("layers_5_linear_attn_conv1d_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595329984)))]; + tensor input_77_cast_fp16 = conv(dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_5_linear_attn_conv1d_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor var_4699_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("op_4699_cast_fp16")]; + tensor qkv_9_perm_0 = const()[name = string("qkv_9_perm_0"), val = tensor([0, 2, 1])]; + tensor var_4701_begin_0 = const()[name = string("op_4701_begin_0"), val = tensor([0, 0, 0])]; + tensor var_4701_end_0 = const()[name = string("op_4701_end_0"), val = tensor([1, 2048, 2048])]; + tensor var_4701_end_mask_0 = const()[name = string("op_4701_end_mask_0"), val = tensor([true, true, false])]; + tensor qkv_9_cast_fp16 = transpose(perm = qkv_9_perm_0, x = var_4699_cast_fp16)[name = string("transpose_18")]; + tensor var_4701_cast_fp16 = slice_by_index(begin = var_4701_begin_0, end = var_4701_end_0, end_mask = var_4701_end_mask_0, x = qkv_9_cast_fp16)[name = string("op_4701_cast_fp16")]; + tensor var_4702 = const()[name = string("op_4702"), val = tensor([1, 2048, 16, 128])]; + tensor q_45_cast_fp16 = reshape(shape = var_4702, x = var_4701_cast_fp16)[name = string("q_45_cast_fp16")]; + tensor var_4704_begin_0 = const()[name = string("op_4704_begin_0"), val = tensor([0, 0, 2048])]; + tensor var_4704_end_0 = const()[name = string("op_4704_end_0"), val = tensor([1, 2048, 4096])]; + tensor var_4704_end_mask_0 = const()[name = string("op_4704_end_mask_0"), val = tensor([true, true, false])]; + tensor var_4704_cast_fp16 = slice_by_index(begin = var_4704_begin_0, end = var_4704_end_0, end_mask = var_4704_end_mask_0, x = qkv_9_cast_fp16)[name = string("op_4704_cast_fp16")]; + tensor var_4705 = const()[name = string("op_4705"), val = tensor([1, 2048, 16, 128])]; + tensor k_39_cast_fp16 = reshape(shape = var_4705, x = var_4704_cast_fp16)[name = string("k_39_cast_fp16")]; + tensor var_4707_begin_0 = const()[name = string("op_4707_begin_0"), val = tensor([0, 0, 4096])]; + tensor var_4707_end_0 = const()[name = string("op_4707_end_0"), val = tensor([1, 2048, 8192])]; + tensor var_4707_end_mask_0 = const()[name = string("op_4707_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4707_cast_fp16 = slice_by_index(begin = var_4707_begin_0, end = var_4707_end_0, end_mask = var_4707_end_mask_0, x = qkv_9_cast_fp16)[name = string("op_4707_cast_fp16")]; + tensor var_4708 = const()[name = string("op_4708"), val = tensor([1, 2048, 32, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_4708, x = var_4707_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor layers_5_linear_attn_in_proj_z_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595395584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605881408))))[name = string("layers_5_linear_attn_in_proj_z_weight_to_fp16_quantized")]; + tensor linear_40_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_linear_attn_in_proj_z_weight_to_fp16_quantized, x = x_129_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_4712 = const()[name = string("op_4712"), val = tensor([1, 2048, 32, 128])]; + tensor z_9_cast_fp16 = reshape(shape = var_4712, x = linear_40_cast_fp16)[name = string("z_9_cast_fp16")]; + tensor layers_5_linear_attn_in_proj_b_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605889664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605971648))))[name = string("layers_5_linear_attn_in_proj_b_weight_to_fp16_quantized")]; + tensor linear_41_bias_0_to_fp16 = const()[name = string("linear_41_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605971776)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_41_bias_0_to_fp16, weight = layers_5_linear_attn_in_proj_b_weight_to_fp16_quantized, x = x_129_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor beta_17_cast_fp16 = sigmoid(x = linear_41_cast_fp16)[name = string("beta_17_cast_fp16")]; + tensor var_4721_weight_0_to_fp16 = const()[name = string("op_4721_weight_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605971904)))]; + tensor var_4721_bias_0_to_fp16 = const()[name = string("op_4721_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606135808)))]; + tensor var_4721_cast_fp16 = linear(bias = var_4721_bias_0_to_fp16, weight = var_4721_weight_0_to_fp16, x = x_129_cast_fp16)[name = string("op_4721_cast_fp16")]; + tensor var_4722_cast_fp16 = softplus(x = var_4721_cast_fp16)[name = string("op_4722_cast_fp16")]; + tensor var_4718_to_fp16 = const()[name = string("op_4718_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606135936)))]; + tensor g_17_cast_fp16 = mul(x = var_4718_to_fp16, y = var_4722_cast_fp16)[name = string("g_17_cast_fp16")]; + tensor var_4724_cast_fp16 = mul(x = q_45_cast_fp16, y = q_45_cast_fp16)[name = string("op_4724_cast_fp16")]; + tensor var_4726_axes_0 = const()[name = string("op_4726_axes_0"), val = tensor([-1])]; + bool var_4726_keep_dims_0 = const()[name = string("op_4726_keep_dims_0"), val = bool(true)]; + tensor var_4726_cast_fp16 = reduce_sum(axes = var_4726_axes_0, keep_dims = var_4726_keep_dims_0, x = var_4724_cast_fp16)[name = string("op_4726_cast_fp16")]; + fp16 var_4727_to_fp16 = const()[name = string("op_4727_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_4728_cast_fp16 = add(x = var_4726_cast_fp16, y = var_4727_to_fp16)[name = string("op_4728_cast_fp16")]; + fp32 var_4729_epsilon_0 = const()[name = string("op_4729_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4729_cast_fp16 = rsqrt(epsilon = var_4729_epsilon_0, x = var_4728_cast_fp16)[name = string("op_4729_cast_fp16")]; + tensor q_47_cast_fp16 = mul(x = q_45_cast_fp16, y = var_4729_cast_fp16)[name = string("q_47_cast_fp16")]; + tensor var_4731_cast_fp16 = mul(x = k_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_4731_cast_fp16")]; + tensor var_4733_axes_0 = const()[name = string("op_4733_axes_0"), val = tensor([-1])]; + bool var_4733_keep_dims_0 = const()[name = string("op_4733_keep_dims_0"), val = bool(true)]; + tensor var_4733_cast_fp16 = reduce_sum(axes = var_4733_axes_0, keep_dims = var_4733_keep_dims_0, x = var_4731_cast_fp16)[name = string("op_4733_cast_fp16")]; + fp16 var_4734_to_fp16 = const()[name = string("op_4734_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_4735_cast_fp16 = add(x = var_4733_cast_fp16, y = var_4734_to_fp16)[name = string("op_4735_cast_fp16")]; + fp32 var_4736_epsilon_0 = const()[name = string("op_4736_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4736_cast_fp16 = rsqrt(epsilon = var_4736_epsilon_0, x = var_4735_cast_fp16)[name = string("op_4736_cast_fp16")]; + tensor k_41_cast_fp16 = mul(x = k_39_cast_fp16, y = var_4736_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor var_4741_axes_0 = const()[name = string("op_4741_axes_0"), val = tensor([3])]; + tensor var_4741_cast_fp16 = expand_dims(axes = var_4741_axes_0, x = q_47_cast_fp16)[name = string("op_4741_cast_fp16")]; + tensor var_4743_reps_0 = const()[name = string("op_4743_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_4743_cast_fp16 = tile(reps = var_4743_reps_0, x = var_4741_cast_fp16)[name = string("op_4743_cast_fp16")]; + tensor var_4744 = const()[name = string("op_4744"), val = tensor([1, 2048, 32, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_4744, x = var_4743_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor var_4749_axes_0 = const()[name = string("op_4749_axes_0"), val = tensor([3])]; + tensor var_4749_cast_fp16 = expand_dims(axes = var_4749_axes_0, x = k_41_cast_fp16)[name = string("op_4749_cast_fp16")]; + tensor var_4751_reps_0 = const()[name = string("op_4751_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_4751_cast_fp16 = tile(reps = var_4751_reps_0, x = var_4749_cast_fp16)[name = string("op_4751_cast_fp16")]; + tensor var_4752 = const()[name = string("op_4752"), val = tensor([1, 2048, 32, 128])]; + tensor k_43_cast_fp16 = reshape(shape = var_4752, x = var_4751_cast_fp16)[name = string("k_43_cast_fp16")]; + fp16 var_4754_to_fp16 = const()[name = string("op_4754_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor q_51_cast_fp16 = mul(x = q_49_cast_fp16, y = var_4754_to_fp16)[name = string("q_51_cast_fp16")]; + tensor var_4756_begin_0 = const()[name = string("op_4756_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4756_end_0 = const()[name = string("op_4756_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_4756_end_mask_0 = const()[name = string("op_4756_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_4756_squeeze_mask_0 = const()[name = string("op_4756_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_4756_cast_fp16 = slice_by_index(begin = var_4756_begin_0, end = var_4756_end_0, end_mask = var_4756_end_mask_0, squeeze_mask = var_4756_squeeze_mask_0, x = q_51_cast_fp16)[name = string("op_4756_cast_fp16")]; + tensor var_4757 = const()[name = string("op_4757"), val = tensor([1, 0, 2])]; + tensor var_4759 = const()[name = string("op_4759"), val = tensor([32, 32, 64, 128])]; + tensor var_4758_cast_fp16 = transpose(perm = var_4757, x = var_4756_cast_fp16)[name = string("transpose_17")]; + tensor q_53_cast_fp16 = reshape(shape = var_4759, x = var_4758_cast_fp16)[name = string("q_53_cast_fp16")]; + tensor var_4761_begin_0 = const()[name = string("op_4761_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4761_end_0 = const()[name = string("op_4761_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_4761_end_mask_0 = const()[name = string("op_4761_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_4761_squeeze_mask_0 = const()[name = string("op_4761_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_4761_cast_fp16 = slice_by_index(begin = var_4761_begin_0, end = var_4761_end_0, end_mask = var_4761_end_mask_0, squeeze_mask = var_4761_squeeze_mask_0, x = k_43_cast_fp16)[name = string("op_4761_cast_fp16")]; + tensor var_4762 = const()[name = string("op_4762"), val = tensor([1, 0, 2])]; + tensor var_4764 = const()[name = string("op_4764"), val = tensor([32, 32, 64, 128])]; + tensor var_4763_cast_fp16 = transpose(perm = var_4762, x = var_4761_cast_fp16)[name = string("transpose_16")]; + tensor k_45_cast_fp16 = reshape(shape = var_4764, x = var_4763_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor var_4766_begin_0 = const()[name = string("op_4766_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4766_end_0 = const()[name = string("op_4766_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_4766_end_mask_0 = const()[name = string("op_4766_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_4766_squeeze_mask_0 = const()[name = string("op_4766_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_4766_cast_fp16 = slice_by_index(begin = var_4766_begin_0, end = var_4766_end_0, end_mask = var_4766_end_mask_0, squeeze_mask = var_4766_squeeze_mask_0, x = v_21_cast_fp16)[name = string("op_4766_cast_fp16")]; + tensor var_4767 = const()[name = string("op_4767"), val = tensor([1, 0, 2])]; + tensor var_4769 = const()[name = string("op_4769"), val = tensor([32, 32, 64, 128])]; + tensor var_4768_cast_fp16 = transpose(perm = var_4767, x = var_4766_cast_fp16)[name = string("transpose_15")]; + tensor v_23_cast_fp16 = reshape(shape = var_4769, x = var_4768_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor var_4771_begin_0 = const()[name = string("op_4771_begin_0"), val = tensor([0, 0, 0])]; + tensor var_4771_end_0 = const()[name = string("op_4771_end_0"), val = tensor([1, 2048, 32])]; + tensor var_4771_end_mask_0 = const()[name = string("op_4771_end_mask_0"), val = tensor([false, true, true])]; + tensor var_4771_squeeze_mask_0 = const()[name = string("op_4771_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_4771_cast_fp16 = slice_by_index(begin = var_4771_begin_0, end = var_4771_end_0, end_mask = var_4771_end_mask_0, squeeze_mask = var_4771_squeeze_mask_0, x = beta_17_cast_fp16)[name = string("op_4771_cast_fp16")]; + tensor var_4772_perm_0 = const()[name = string("op_4772_perm_0"), val = tensor([1, 0])]; + tensor var_4773 = const()[name = string("op_4773"), val = tensor([32, 32, 64, 1])]; + tensor var_4772_cast_fp16 = transpose(perm = var_4772_perm_0, x = var_4771_cast_fp16)[name = string("transpose_14")]; + tensor beta_19_cast_fp16 = reshape(shape = var_4773, x = var_4772_cast_fp16)[name = string("beta_19_cast_fp16")]; + tensor var_4775_begin_0 = const()[name = string("op_4775_begin_0"), val = tensor([0, 0, 0])]; + tensor var_4775_end_0 = const()[name = string("op_4775_end_0"), val = tensor([1, 2048, 32])]; + tensor var_4775_end_mask_0 = const()[name = string("op_4775_end_mask_0"), val = tensor([false, true, true])]; + tensor var_4775_squeeze_mask_0 = const()[name = string("op_4775_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_4775_cast_fp16 = slice_by_index(begin = var_4775_begin_0, end = var_4775_end_0, end_mask = var_4775_end_mask_0, squeeze_mask = var_4775_squeeze_mask_0, x = g_17_cast_fp16)[name = string("op_4775_cast_fp16")]; + tensor var_4776_perm_0 = const()[name = string("op_4776_perm_0"), val = tensor([1, 0])]; + tensor var_4777 = const()[name = string("op_4777"), val = tensor([32, 32, 64])]; + tensor var_4776_cast_fp16 = transpose(perm = var_4776_perm_0, x = var_4775_cast_fp16)[name = string("transpose_13")]; + tensor g_19_cast_fp16 = reshape(shape = var_4777, x = var_4776_cast_fp16)[name = string("g_19_cast_fp16")]; + tensor g_col_9_axes_0 = const()[name = string("g_col_9_axes_0"), val = tensor([-1])]; + tensor g_col_9_cast_fp16 = expand_dims(axes = g_col_9_axes_0, x = g_19_cast_fp16)[name = string("g_col_9_cast_fp16")]; + bool var_4780_transpose_x_0 = const()[name = string("op_4780_transpose_x_0"), val = bool(false)]; + bool var_4780_transpose_y_0 = const()[name = string("op_4780_transpose_y_0"), val = bool(false)]; + tensor var_4780_cast_fp16 = matmul(transpose_x = var_4780_transpose_x_0, transpose_y = var_4780_transpose_y_0, x = layers_0_linear_attn_lower_incl_to_fp16, y = g_col_9_cast_fp16)[name = string("op_4780_cast_fp16")]; + tensor cum_9_axes_0 = const()[name = string("cum_9_axes_0"), val = tensor([-1])]; + tensor cum_9_cast_fp16 = squeeze(axes = cum_9_axes_0, x = var_4780_cast_fp16)[name = string("cum_9_cast_fp16")]; + bool var_4782_transpose_x_0 = const()[name = string("op_4782_transpose_x_0"), val = bool(false)]; + bool var_4782_transpose_y_0 = const()[name = string("op_4782_transpose_y_0"), val = bool(false)]; + tensor var_4782_cast_fp16 = matmul(transpose_x = var_4782_transpose_x_0, transpose_y = var_4782_transpose_y_0, x = layers_0_linear_attn_suffix_sum_t_to_fp16, y = g_col_9_cast_fp16)[name = string("op_4782_cast_fp16")]; + tensor to_end_9_axes_0 = const()[name = string("to_end_9_axes_0"), val = tensor([-1])]; + tensor to_end_9_cast_fp16 = squeeze(axes = to_end_9_axes_0, x = var_4782_cast_fp16)[name = string("to_end_9_cast_fp16")]; + bool var_4784_transpose_x_0 = const()[name = string("op_4784_transpose_x_0"), val = bool(false)]; + bool var_4784_transpose_y_0 = const()[name = string("op_4784_transpose_y_0"), val = bool(false)]; + tensor var_4784_cast_fp16 = matmul(transpose_x = var_4784_transpose_x_0, transpose_y = var_4784_transpose_y_0, x = layers_0_linear_attn_pair_sum_t_to_fp16, y = g_col_9_cast_fp16)[name = string("op_4784_cast_fp16")]; + tensor var_4785 = const()[name = string("op_4785"), val = tensor([32, 32, 64, 64])]; + tensor pair_57_cast_fp16 = reshape(shape = var_4785, x = var_4784_cast_fp16)[name = string("pair_57_cast_fp16")]; + tensor var_4787_cast_fp16 = exp(x = pair_57_cast_fp16)[name = string("op_4787_cast_fp16")]; + tensor pair_decay_9_cast_fp16 = mul(x = var_4787_cast_fp16, y = layers_0_linear_attn_lower_incl_to_fp16)[name = string("pair_decay_9_cast_fp16")]; + tensor k_beta_9_cast_fp16 = mul(x = k_45_cast_fp16, y = beta_19_cast_fp16)[name = string("k_beta_9_cast_fp16")]; + tensor v_beta_9_cast_fp16 = mul(x = v_23_cast_fp16, y = beta_19_cast_fp16)[name = string("v_beta_9_cast_fp16")]; + bool var_4792_transpose_x_1 = const()[name = string("op_4792_transpose_x_1"), val = bool(false)]; + bool var_4792_transpose_y_1 = const()[name = string("op_4792_transpose_y_1"), val = bool(true)]; + tensor var_4792_cast_fp16 = matmul(transpose_x = var_4792_transpose_x_1, transpose_y = var_4792_transpose_y_1, x = k_beta_9_cast_fp16, y = k_45_cast_fp16)[name = string("op_4792_cast_fp16")]; + tensor kkt_9_cast_fp16 = mul(x = var_4792_cast_fp16, y = pair_decay_9_cast_fp16)[name = string("kkt_9_cast_fp16")]; + bool var_4795_transpose_x_1 = const()[name = string("op_4795_transpose_x_1"), val = bool(false)]; + bool var_4795_transpose_y_1 = const()[name = string("op_4795_transpose_y_1"), val = bool(true)]; + tensor var_4795_cast_fp16 = matmul(transpose_x = var_4795_transpose_x_1, transpose_y = var_4795_transpose_y_1, x = q_53_cast_fp16, y = k_45_cast_fp16)[name = string("op_4795_cast_fp16")]; + tensor attn_intra_9_cast_fp16 = mul(x = var_4795_cast_fp16, y = pair_decay_9_cast_fp16)[name = string("attn_intra_9_cast_fp16")]; + tensor var_4797_cast_fp16 = mul(x = kkt_9_cast_fp16, y = layers_0_linear_attn_strict_lower_to_fp16)[name = string("op_4797_cast_fp16")]; + tensor t_17_cast_fp16 = add(x = layers_0_linear_attn_eye_to_fp16, y = var_4797_cast_fp16)[name = string("t_17_cast_fp16")]; + tensor var_4801 = const()[name = string("op_4801"), val = tensor([-1, 64, 64])]; + tensor t_19_cast_fp16 = reshape(shape = var_4801, x = t_17_cast_fp16)[name = string("t_19_cast_fp16")]; + tensor var_4806 = const()[name = string("op_4806"), val = tensor([1024, 32, 2, 32, 2])]; + tensor var_4807_cast_fp16 = reshape(shape = var_4806, x = t_19_cast_fp16)[name = string("op_4807_cast_fp16")]; + tensor var_4812_cast_fp16 = mul(x = var_4807_cast_fp16, y = var_227_to_fp16)[name = string("op_4812_cast_fp16")]; + tensor blocks_49_axes_0 = const()[name = string("blocks_49_axes_0"), val = tensor([-2])]; + bool blocks_49_keep_dims_0 = const()[name = string("blocks_49_keep_dims_0"), val = bool(false)]; + tensor blocks_49_cast_fp16 = reduce_sum(axes = blocks_49_axes_0, keep_dims = blocks_49_keep_dims_0, x = var_4812_cast_fp16)[name = string("blocks_49_cast_fp16")]; + tensor var_4815_begin_0 = const()[name = string("op_4815_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4815_end_0 = const()[name = string("op_4815_end_0"), val = tensor([1024, 32, 2, 2])]; + tensor var_4815_end_mask_0 = const()[name = string("op_4815_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4815_cast_fp16 = slice_by_index(begin = var_4815_begin_0, end = var_4815_end_0, end_mask = var_4815_end_mask_0, x = blocks_49_cast_fp16)[name = string("op_4815_cast_fp16")]; + tensor x_131_begin_0 = const()[name = string("x_131_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_131_end_0 = const()[name = string("x_131_end_0"), val = tensor([1024, 32, 1, 1])]; + tensor x_131_end_mask_0 = const()[name = string("x_131_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_131_cast_fp16 = slice_by_index(begin = x_131_begin_0, end = x_131_end_0, end_mask = x_131_end_mask_0, x = var_4815_cast_fp16)[name = string("x_131_cast_fp16")]; + bool var_4825_transpose_x_0 = const()[name = string("op_4825_transpose_x_0"), val = bool(false)]; + bool var_4825_transpose_y_0 = const()[name = string("op_4825_transpose_y_0"), val = bool(false)]; + tensor var_4825_cast_fp16 = matmul(transpose_x = var_4825_transpose_x_0, transpose_y = var_4825_transpose_y_0, x = x_131_cast_fp16, y = var_235_to_fp16)[name = string("op_4825_cast_fp16")]; + bool var_4826_transpose_x_0 = const()[name = string("op_4826_transpose_x_0"), val = bool(false)]; + bool var_4826_transpose_y_0 = const()[name = string("op_4826_transpose_y_0"), val = bool(false)]; + tensor var_4826_cast_fp16 = matmul(transpose_x = var_4826_transpose_x_0, transpose_y = var_4826_transpose_y_0, x = var_235_to_fp16, y = var_4825_cast_fp16)[name = string("op_4826_cast_fp16")]; + fp16 const_59_promoted_to_fp16 = const()[name = string("const_59_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_49_cast_fp16 = mul(x = var_4826_cast_fp16, y = const_59_promoted_to_fp16)[name = string("lower_49_cast_fp16")]; + bool bottom_49_interleave_0 = const()[name = string("bottom_49_interleave_0"), val = bool(false)]; + tensor bottom_49_cast_fp16 = concat(axis = var_4647, interleave = bottom_49_interleave_0, values = (lower_49_cast_fp16, var_235_to_fp16))[name = string("bottom_49_cast_fp16")]; + bool inv_67_interleave_0 = const()[name = string("inv_67_interleave_0"), val = bool(false)]; + tensor inv_67_cast_fp16 = concat(axis = var_4618, interleave = inv_67_interleave_0, values = (top_1_to_fp16, bottom_49_cast_fp16))[name = string("inv_67_cast_fp16")]; + tensor var_4835 = const()[name = string("op_4835"), val = tensor([1024, 16, 4, 16, 4])]; + tensor var_4836_cast_fp16 = reshape(shape = var_4835, x = t_19_cast_fp16)[name = string("op_4836_cast_fp16")]; + tensor var_4841_cast_fp16 = mul(x = var_4836_cast_fp16, y = var_256_to_fp16)[name = string("op_4841_cast_fp16")]; + tensor blocks_51_axes_0 = const()[name = string("blocks_51_axes_0"), val = tensor([-2])]; + bool blocks_51_keep_dims_0 = const()[name = string("blocks_51_keep_dims_0"), val = bool(false)]; + tensor blocks_51_cast_fp16 = reduce_sum(axes = blocks_51_axes_0, keep_dims = blocks_51_keep_dims_0, x = var_4841_cast_fp16)[name = string("blocks_51_cast_fp16")]; + tensor var_4844_begin_0 = const()[name = string("op_4844_begin_0"), val = tensor([0, 0, 2, 0])]; + tensor var_4844_end_0 = const()[name = string("op_4844_end_0"), val = tensor([1024, 16, 4, 4])]; + tensor var_4844_end_mask_0 = const()[name = string("op_4844_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4844_cast_fp16 = slice_by_index(begin = var_4844_begin_0, end = var_4844_end_0, end_mask = var_4844_end_mask_0, x = blocks_51_cast_fp16)[name = string("op_4844_cast_fp16")]; + tensor x_133_begin_0 = const()[name = string("x_133_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_133_end_0 = const()[name = string("x_133_end_0"), val = tensor([1024, 16, 2, 2])]; + tensor x_133_end_mask_0 = const()[name = string("x_133_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_133_cast_fp16 = slice_by_index(begin = x_133_begin_0, end = x_133_end_0, end_mask = x_133_end_mask_0, x = var_4844_cast_fp16)[name = string("x_133_cast_fp16")]; + tensor var_4846 = const()[name = string("op_4846"), val = tensor([1024, 16, 2, 2, 2])]; + tensor pair_61_cast_fp16 = reshape(shape = var_4846, x = inv_67_cast_fp16)[name = string("pair_61_cast_fp16")]; + tensor var_4848_begin_0 = const()[name = string("op_4848_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4848_end_0 = const()[name = string("op_4848_end_0"), val = tensor([1024, 16, 1, 2, 2])]; + tensor var_4848_end_mask_0 = const()[name = string("op_4848_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4848_squeeze_mask_0 = const()[name = string("op_4848_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, squeeze_mask = var_4848_squeeze_mask_0, x = pair_61_cast_fp16)[name = string("op_4848_cast_fp16")]; + tensor var_4851_begin_0 = const()[name = string("op_4851_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_4851_end_0 = const()[name = string("op_4851_end_0"), val = tensor([1024, 16, 2, 2, 2])]; + tensor var_4851_end_mask_0 = const()[name = string("op_4851_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4851_squeeze_mask_0 = const()[name = string("op_4851_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4851_cast_fp16 = slice_by_index(begin = var_4851_begin_0, end = var_4851_end_0, end_mask = var_4851_end_mask_0, squeeze_mask = var_4851_squeeze_mask_0, x = pair_61_cast_fp16)[name = string("op_4851_cast_fp16")]; + bool var_4854_transpose_x_0 = const()[name = string("op_4854_transpose_x_0"), val = bool(false)]; + bool var_4854_transpose_y_0 = const()[name = string("op_4854_transpose_y_0"), val = bool(false)]; + tensor var_4854_cast_fp16 = matmul(transpose_x = var_4854_transpose_x_0, transpose_y = var_4854_transpose_y_0, x = x_133_cast_fp16, y = var_4848_cast_fp16)[name = string("op_4854_cast_fp16")]; + bool var_4855_transpose_x_0 = const()[name = string("op_4855_transpose_x_0"), val = bool(false)]; + bool var_4855_transpose_y_0 = const()[name = string("op_4855_transpose_y_0"), val = bool(false)]; + tensor var_4855_cast_fp16 = matmul(transpose_x = var_4855_transpose_x_0, transpose_y = var_4855_transpose_y_0, x = var_4851_cast_fp16, y = var_4854_cast_fp16)[name = string("op_4855_cast_fp16")]; + fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_51_cast_fp16 = mul(x = var_4855_cast_fp16, y = const_60_promoted_to_fp16)[name = string("lower_51_cast_fp16")]; + bool top_51_interleave_0 = const()[name = string("top_51_interleave_0"), val = bool(false)]; + tensor top_51_cast_fp16 = concat(axis = var_4647, interleave = top_51_interleave_0, values = (var_4848_cast_fp16, var_273_to_fp16))[name = string("top_51_cast_fp16")]; + bool bottom_51_interleave_0 = const()[name = string("bottom_51_interleave_0"), val = bool(false)]; + tensor bottom_51_cast_fp16 = concat(axis = var_4647, interleave = bottom_51_interleave_0, values = (lower_51_cast_fp16, var_4851_cast_fp16))[name = string("bottom_51_cast_fp16")]; + bool inv_69_interleave_0 = const()[name = string("inv_69_interleave_0"), val = bool(false)]; + tensor inv_69_cast_fp16 = concat(axis = var_4618, interleave = inv_69_interleave_0, values = (top_51_cast_fp16, bottom_51_cast_fp16))[name = string("inv_69_cast_fp16")]; + tensor var_4864 = const()[name = string("op_4864"), val = tensor([1024, 8, 8, 8, 8])]; + tensor var_4865_cast_fp16 = reshape(shape = var_4864, x = t_19_cast_fp16)[name = string("op_4865_cast_fp16")]; + tensor var_4869_to_fp16 = const()[name = string("op_4869_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606136064)))]; + tensor var_4870_cast_fp16 = mul(x = var_4865_cast_fp16, y = var_4869_to_fp16)[name = string("op_4870_cast_fp16")]; + tensor blocks_53_axes_0 = const()[name = string("blocks_53_axes_0"), val = tensor([-2])]; + bool blocks_53_keep_dims_0 = const()[name = string("blocks_53_keep_dims_0"), val = bool(false)]; + tensor blocks_53_cast_fp16 = reduce_sum(axes = blocks_53_axes_0, keep_dims = blocks_53_keep_dims_0, x = var_4870_cast_fp16)[name = string("blocks_53_cast_fp16")]; + tensor var_4873_begin_0 = const()[name = string("op_4873_begin_0"), val = tensor([0, 0, 4, 0])]; + tensor var_4873_end_0 = const()[name = string("op_4873_end_0"), val = tensor([1024, 8, 8, 8])]; + tensor var_4873_end_mask_0 = const()[name = string("op_4873_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4873_cast_fp16 = slice_by_index(begin = var_4873_begin_0, end = var_4873_end_0, end_mask = var_4873_end_mask_0, x = blocks_53_cast_fp16)[name = string("op_4873_cast_fp16")]; + tensor x_135_begin_0 = const()[name = string("x_135_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_135_end_0 = const()[name = string("x_135_end_0"), val = tensor([1024, 8, 4, 4])]; + tensor x_135_end_mask_0 = const()[name = string("x_135_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_135_cast_fp16 = slice_by_index(begin = x_135_begin_0, end = x_135_end_0, end_mask = x_135_end_mask_0, x = var_4873_cast_fp16)[name = string("x_135_cast_fp16")]; + tensor var_4875 = const()[name = string("op_4875"), val = tensor([1024, 8, 2, 4, 4])]; + tensor pair_63_cast_fp16 = reshape(shape = var_4875, x = inv_69_cast_fp16)[name = string("pair_63_cast_fp16")]; + tensor var_4877_begin_0 = const()[name = string("op_4877_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4877_end_0 = const()[name = string("op_4877_end_0"), val = tensor([1024, 8, 1, 4, 4])]; + tensor var_4877_end_mask_0 = const()[name = string("op_4877_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4877_squeeze_mask_0 = const()[name = string("op_4877_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4877_cast_fp16 = slice_by_index(begin = var_4877_begin_0, end = var_4877_end_0, end_mask = var_4877_end_mask_0, squeeze_mask = var_4877_squeeze_mask_0, x = pair_63_cast_fp16)[name = string("op_4877_cast_fp16")]; + tensor var_4880_begin_0 = const()[name = string("op_4880_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_4880_end_0 = const()[name = string("op_4880_end_0"), val = tensor([1024, 8, 2, 4, 4])]; + tensor var_4880_end_mask_0 = const()[name = string("op_4880_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4880_squeeze_mask_0 = const()[name = string("op_4880_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4880_cast_fp16 = slice_by_index(begin = var_4880_begin_0, end = var_4880_end_0, end_mask = var_4880_end_mask_0, squeeze_mask = var_4880_squeeze_mask_0, x = pair_63_cast_fp16)[name = string("op_4880_cast_fp16")]; + bool var_4883_transpose_x_0 = const()[name = string("op_4883_transpose_x_0"), val = bool(false)]; + bool var_4883_transpose_y_0 = const()[name = string("op_4883_transpose_y_0"), val = bool(false)]; + tensor var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_0, transpose_y = var_4883_transpose_y_0, x = x_135_cast_fp16, y = var_4877_cast_fp16)[name = string("op_4883_cast_fp16")]; + bool var_4884_transpose_x_0 = const()[name = string("op_4884_transpose_x_0"), val = bool(false)]; + bool var_4884_transpose_y_0 = const()[name = string("op_4884_transpose_y_0"), val = bool(false)]; + tensor var_4884_cast_fp16 = matmul(transpose_x = var_4884_transpose_x_0, transpose_y = var_4884_transpose_y_0, x = var_4880_cast_fp16, y = var_4883_cast_fp16)[name = string("op_4884_cast_fp16")]; + fp16 const_61_promoted_to_fp16 = const()[name = string("const_61_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_53_cast_fp16 = mul(x = var_4884_cast_fp16, y = const_61_promoted_to_fp16)[name = string("lower_53_cast_fp16")]; + bool top_53_interleave_0 = const()[name = string("top_53_interleave_0"), val = bool(false)]; + tensor top_53_cast_fp16 = concat(axis = var_4647, interleave = top_53_interleave_0, values = (var_4877_cast_fp16, var_302_to_fp16))[name = string("top_53_cast_fp16")]; + bool bottom_53_interleave_0 = const()[name = string("bottom_53_interleave_0"), val = bool(false)]; + tensor bottom_53_cast_fp16 = concat(axis = var_4647, interleave = bottom_53_interleave_0, values = (lower_53_cast_fp16, var_4880_cast_fp16))[name = string("bottom_53_cast_fp16")]; + bool inv_71_interleave_0 = const()[name = string("inv_71_interleave_0"), val = bool(false)]; + tensor inv_71_cast_fp16 = concat(axis = var_4618, interleave = inv_71_interleave_0, values = (top_53_cast_fp16, bottom_53_cast_fp16))[name = string("inv_71_cast_fp16")]; + tensor var_4893 = const()[name = string("op_4893"), val = tensor([1024, 4, 16, 4, 16])]; + tensor var_4894_cast_fp16 = reshape(shape = var_4893, x = t_19_cast_fp16)[name = string("op_4894_cast_fp16")]; + tensor var_4898_to_fp16 = const()[name = string("op_4898_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606136256)))]; + tensor var_4899_cast_fp16 = mul(x = var_4894_cast_fp16, y = var_4898_to_fp16)[name = string("op_4899_cast_fp16")]; + tensor blocks_55_axes_0 = const()[name = string("blocks_55_axes_0"), val = tensor([-2])]; + bool blocks_55_keep_dims_0 = const()[name = string("blocks_55_keep_dims_0"), val = bool(false)]; + tensor blocks_55_cast_fp16 = reduce_sum(axes = blocks_55_axes_0, keep_dims = blocks_55_keep_dims_0, x = var_4899_cast_fp16)[name = string("blocks_55_cast_fp16")]; + tensor var_4902_begin_0 = const()[name = string("op_4902_begin_0"), val = tensor([0, 0, 8, 0])]; + tensor var_4902_end_0 = const()[name = string("op_4902_end_0"), val = tensor([1024, 4, 16, 16])]; + tensor var_4902_end_mask_0 = const()[name = string("op_4902_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4902_cast_fp16 = slice_by_index(begin = var_4902_begin_0, end = var_4902_end_0, end_mask = var_4902_end_mask_0, x = blocks_55_cast_fp16)[name = string("op_4902_cast_fp16")]; + tensor x_137_begin_0 = const()[name = string("x_137_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_137_end_0 = const()[name = string("x_137_end_0"), val = tensor([1024, 4, 8, 8])]; + tensor x_137_end_mask_0 = const()[name = string("x_137_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_137_cast_fp16 = slice_by_index(begin = x_137_begin_0, end = x_137_end_0, end_mask = x_137_end_mask_0, x = var_4902_cast_fp16)[name = string("x_137_cast_fp16")]; + tensor var_4904 = const()[name = string("op_4904"), val = tensor([1024, 4, 2, 8, 8])]; + tensor pair_65_cast_fp16 = reshape(shape = var_4904, x = inv_71_cast_fp16)[name = string("pair_65_cast_fp16")]; + tensor var_4906_begin_0 = const()[name = string("op_4906_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4906_end_0 = const()[name = string("op_4906_end_0"), val = tensor([1024, 4, 1, 8, 8])]; + tensor var_4906_end_mask_0 = const()[name = string("op_4906_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4906_squeeze_mask_0 = const()[name = string("op_4906_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4906_cast_fp16 = slice_by_index(begin = var_4906_begin_0, end = var_4906_end_0, end_mask = var_4906_end_mask_0, squeeze_mask = var_4906_squeeze_mask_0, x = pair_65_cast_fp16)[name = string("op_4906_cast_fp16")]; + tensor var_4909_begin_0 = const()[name = string("op_4909_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_4909_end_0 = const()[name = string("op_4909_end_0"), val = tensor([1024, 4, 2, 8, 8])]; + tensor var_4909_end_mask_0 = const()[name = string("op_4909_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4909_squeeze_mask_0 = const()[name = string("op_4909_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4909_cast_fp16 = slice_by_index(begin = var_4909_begin_0, end = var_4909_end_0, end_mask = var_4909_end_mask_0, squeeze_mask = var_4909_squeeze_mask_0, x = pair_65_cast_fp16)[name = string("op_4909_cast_fp16")]; + bool var_4912_transpose_x_0 = const()[name = string("op_4912_transpose_x_0"), val = bool(false)]; + bool var_4912_transpose_y_0 = const()[name = string("op_4912_transpose_y_0"), val = bool(false)]; + tensor var_4912_cast_fp16 = matmul(transpose_x = var_4912_transpose_x_0, transpose_y = var_4912_transpose_y_0, x = x_137_cast_fp16, y = var_4906_cast_fp16)[name = string("op_4912_cast_fp16")]; + bool var_4913_transpose_x_0 = const()[name = string("op_4913_transpose_x_0"), val = bool(false)]; + bool var_4913_transpose_y_0 = const()[name = string("op_4913_transpose_y_0"), val = bool(false)]; + tensor var_4913_cast_fp16 = matmul(transpose_x = var_4913_transpose_x_0, transpose_y = var_4913_transpose_y_0, x = var_4909_cast_fp16, y = var_4912_cast_fp16)[name = string("op_4913_cast_fp16")]; + fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_55_cast_fp16 = mul(x = var_4913_cast_fp16, y = const_62_promoted_to_fp16)[name = string("lower_55_cast_fp16")]; + bool top_55_interleave_0 = const()[name = string("top_55_interleave_0"), val = bool(false)]; + tensor top_55_cast_fp16 = concat(axis = var_4647, interleave = top_55_interleave_0, values = (var_4906_cast_fp16, var_331_to_fp16))[name = string("top_55_cast_fp16")]; + bool bottom_55_interleave_0 = const()[name = string("bottom_55_interleave_0"), val = bool(false)]; + tensor bottom_55_cast_fp16 = concat(axis = var_4647, interleave = bottom_55_interleave_0, values = (lower_55_cast_fp16, var_4909_cast_fp16))[name = string("bottom_55_cast_fp16")]; + bool inv_73_interleave_0 = const()[name = string("inv_73_interleave_0"), val = bool(false)]; + tensor inv_73_cast_fp16 = concat(axis = var_4618, interleave = inv_73_interleave_0, values = (top_55_cast_fp16, bottom_55_cast_fp16))[name = string("inv_73_cast_fp16")]; + tensor var_4922 = const()[name = string("op_4922"), val = tensor([1024, 2, 32, 2, 32])]; + tensor var_4923_cast_fp16 = reshape(shape = var_4922, x = t_19_cast_fp16)[name = string("op_4923_cast_fp16")]; + tensor var_4927_to_fp16 = const()[name = string("op_4927_to_fp16"), val = tensor([[[[0x1p+0], [0x0p+0]]], [[[0x0p+0], [0x1p+0]]]])]; + tensor var_4928_cast_fp16 = mul(x = var_4923_cast_fp16, y = var_4927_to_fp16)[name = string("op_4928_cast_fp16")]; + tensor blocks_57_axes_0 = const()[name = string("blocks_57_axes_0"), val = tensor([-2])]; + bool blocks_57_keep_dims_0 = const()[name = string("blocks_57_keep_dims_0"), val = bool(false)]; + tensor blocks_57_cast_fp16 = reduce_sum(axes = blocks_57_axes_0, keep_dims = blocks_57_keep_dims_0, x = var_4928_cast_fp16)[name = string("blocks_57_cast_fp16")]; + tensor var_4931_begin_0 = const()[name = string("op_4931_begin_0"), val = tensor([0, 0, 16, 0])]; + tensor var_4931_end_0 = const()[name = string("op_4931_end_0"), val = tensor([1024, 2, 32, 32])]; + tensor var_4931_end_mask_0 = const()[name = string("op_4931_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4931_cast_fp16 = slice_by_index(begin = var_4931_begin_0, end = var_4931_end_0, end_mask = var_4931_end_mask_0, x = blocks_57_cast_fp16)[name = string("op_4931_cast_fp16")]; + tensor x_139_begin_0 = const()[name = string("x_139_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_139_end_0 = const()[name = string("x_139_end_0"), val = tensor([1024, 2, 16, 16])]; + tensor x_139_end_mask_0 = const()[name = string("x_139_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_139_cast_fp16 = slice_by_index(begin = x_139_begin_0, end = x_139_end_0, end_mask = x_139_end_mask_0, x = var_4931_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_4933 = const()[name = string("op_4933"), val = tensor([1024, 2, 2, 16, 16])]; + tensor pair_67_cast_fp16 = reshape(shape = var_4933, x = inv_73_cast_fp16)[name = string("pair_67_cast_fp16")]; + tensor var_4935_begin_0 = const()[name = string("op_4935_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4935_end_0 = const()[name = string("op_4935_end_0"), val = tensor([1024, 2, 1, 16, 16])]; + tensor var_4935_end_mask_0 = const()[name = string("op_4935_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4935_squeeze_mask_0 = const()[name = string("op_4935_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4935_cast_fp16 = slice_by_index(begin = var_4935_begin_0, end = var_4935_end_0, end_mask = var_4935_end_mask_0, squeeze_mask = var_4935_squeeze_mask_0, x = pair_67_cast_fp16)[name = string("op_4935_cast_fp16")]; + tensor var_4938_begin_0 = const()[name = string("op_4938_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_4938_end_0 = const()[name = string("op_4938_end_0"), val = tensor([1024, 2, 2, 16, 16])]; + tensor var_4938_end_mask_0 = const()[name = string("op_4938_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4938_squeeze_mask_0 = const()[name = string("op_4938_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4938_cast_fp16 = slice_by_index(begin = var_4938_begin_0, end = var_4938_end_0, end_mask = var_4938_end_mask_0, squeeze_mask = var_4938_squeeze_mask_0, x = pair_67_cast_fp16)[name = string("op_4938_cast_fp16")]; + bool var_4941_transpose_x_0 = const()[name = string("op_4941_transpose_x_0"), val = bool(false)]; + bool var_4941_transpose_y_0 = const()[name = string("op_4941_transpose_y_0"), val = bool(false)]; + tensor var_4941_cast_fp16 = matmul(transpose_x = var_4941_transpose_x_0, transpose_y = var_4941_transpose_y_0, x = x_139_cast_fp16, y = var_4935_cast_fp16)[name = string("op_4941_cast_fp16")]; + bool var_4942_transpose_x_0 = const()[name = string("op_4942_transpose_x_0"), val = bool(false)]; + bool var_4942_transpose_y_0 = const()[name = string("op_4942_transpose_y_0"), val = bool(false)]; + tensor var_4942_cast_fp16 = matmul(transpose_x = var_4942_transpose_x_0, transpose_y = var_4942_transpose_y_0, x = var_4938_cast_fp16, y = var_4941_cast_fp16)[name = string("op_4942_cast_fp16")]; + fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_57_cast_fp16 = mul(x = var_4942_cast_fp16, y = const_63_promoted_to_fp16)[name = string("lower_57_cast_fp16")]; + bool top_57_interleave_0 = const()[name = string("top_57_interleave_0"), val = bool(false)]; + tensor top_57_cast_fp16 = concat(axis = var_4647, interleave = top_57_interleave_0, values = (var_4935_cast_fp16, var_360_to_fp16))[name = string("top_57_cast_fp16")]; + bool bottom_57_interleave_0 = const()[name = string("bottom_57_interleave_0"), val = bool(false)]; + tensor bottom_57_cast_fp16 = concat(axis = var_4647, interleave = bottom_57_interleave_0, values = (lower_57_cast_fp16, var_4938_cast_fp16))[name = string("bottom_57_cast_fp16")]; + bool inv_75_interleave_0 = const()[name = string("inv_75_interleave_0"), val = bool(false)]; + tensor inv_75_cast_fp16 = concat(axis = var_4618, interleave = inv_75_interleave_0, values = (top_57_cast_fp16, bottom_57_cast_fp16))[name = string("inv_75_cast_fp16")]; + tensor var_4951 = const()[name = string("op_4951"), val = tensor([1024, 1, 64, 1, 64])]; + tensor var_4952_cast_fp16 = reshape(shape = var_4951, x = t_19_cast_fp16)[name = string("op_4952_cast_fp16")]; + tensor blocks_59_axes_0 = const()[name = string("blocks_59_axes_0"), val = tensor([-2])]; + bool blocks_59_keep_dims_0 = const()[name = string("blocks_59_keep_dims_0"), val = bool(false)]; + tensor blocks_59_cast_fp16 = reduce_sum(axes = blocks_59_axes_0, keep_dims = blocks_59_keep_dims_0, x = var_4952_cast_fp16)[name = string("blocks_59_cast_fp16")]; + tensor var_4960_begin_0 = const()[name = string("op_4960_begin_0"), val = tensor([0, 0, 32, 0])]; + tensor var_4960_end_0 = const()[name = string("op_4960_end_0"), val = tensor([1024, 1, 64, 64])]; + tensor var_4960_end_mask_0 = const()[name = string("op_4960_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4960_cast_fp16 = slice_by_index(begin = var_4960_begin_0, end = var_4960_end_0, end_mask = var_4960_end_mask_0, x = blocks_59_cast_fp16)[name = string("op_4960_cast_fp16")]; + tensor x_141_begin_0 = const()[name = string("x_141_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_141_end_0 = const()[name = string("x_141_end_0"), val = tensor([1024, 1, 32, 32])]; + tensor x_141_end_mask_0 = const()[name = string("x_141_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_141_cast_fp16 = slice_by_index(begin = x_141_begin_0, end = x_141_end_0, end_mask = x_141_end_mask_0, x = var_4960_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_4962 = const()[name = string("op_4962"), val = tensor([1024, 1, 2, 32, 32])]; + tensor pair_69_cast_fp16 = reshape(shape = var_4962, x = inv_75_cast_fp16)[name = string("pair_69_cast_fp16")]; + tensor var_4964_begin_0 = const()[name = string("op_4964_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4964_end_0 = const()[name = string("op_4964_end_0"), val = tensor([1024, 1, 1, 32, 32])]; + tensor var_4964_end_mask_0 = const()[name = string("op_4964_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4964_squeeze_mask_0 = const()[name = string("op_4964_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4964_cast_fp16 = slice_by_index(begin = var_4964_begin_0, end = var_4964_end_0, end_mask = var_4964_end_mask_0, squeeze_mask = var_4964_squeeze_mask_0, x = pair_69_cast_fp16)[name = string("op_4964_cast_fp16")]; + tensor var_4967_begin_0 = const()[name = string("op_4967_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_4967_end_0 = const()[name = string("op_4967_end_0"), val = tensor([1024, 1, 2, 32, 32])]; + tensor var_4967_end_mask_0 = const()[name = string("op_4967_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_4967_squeeze_mask_0 = const()[name = string("op_4967_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_4967_cast_fp16 = slice_by_index(begin = var_4967_begin_0, end = var_4967_end_0, end_mask = var_4967_end_mask_0, squeeze_mask = var_4967_squeeze_mask_0, x = pair_69_cast_fp16)[name = string("op_4967_cast_fp16")]; + bool var_4970_transpose_x_0 = const()[name = string("op_4970_transpose_x_0"), val = bool(false)]; + bool var_4970_transpose_y_0 = const()[name = string("op_4970_transpose_y_0"), val = bool(false)]; + tensor var_4970_cast_fp16 = matmul(transpose_x = var_4970_transpose_x_0, transpose_y = var_4970_transpose_y_0, x = x_141_cast_fp16, y = var_4964_cast_fp16)[name = string("op_4970_cast_fp16")]; + bool var_4971_transpose_x_0 = const()[name = string("op_4971_transpose_x_0"), val = bool(false)]; + bool var_4971_transpose_y_0 = const()[name = string("op_4971_transpose_y_0"), val = bool(false)]; + tensor var_4971_cast_fp16 = matmul(transpose_x = var_4971_transpose_x_0, transpose_y = var_4971_transpose_y_0, x = var_4967_cast_fp16, y = var_4970_cast_fp16)[name = string("op_4971_cast_fp16")]; + fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_59_cast_fp16 = mul(x = var_4971_cast_fp16, y = const_64_promoted_to_fp16)[name = string("lower_59_cast_fp16")]; + bool top_59_interleave_0 = const()[name = string("top_59_interleave_0"), val = bool(false)]; + tensor top_59_cast_fp16 = concat(axis = var_4647, interleave = top_59_interleave_0, values = (var_4964_cast_fp16, var_389_to_fp16))[name = string("top_59_cast_fp16")]; + bool bottom_59_interleave_0 = const()[name = string("bottom_59_interleave_0"), val = bool(false)]; + tensor bottom_59_cast_fp16 = concat(axis = var_4647, interleave = bottom_59_interleave_0, values = (lower_59_cast_fp16, var_4967_cast_fp16))[name = string("bottom_59_cast_fp16")]; + bool inv_77_interleave_0 = const()[name = string("inv_77_interleave_0"), val = bool(false)]; + tensor inv_77_cast_fp16 = concat(axis = var_4618, interleave = inv_77_interleave_0, values = (top_59_cast_fp16, bottom_59_cast_fp16))[name = string("inv_77_cast_fp16")]; + tensor var_4980 = const()[name = string("op_4980"), val = tensor([32, 32, 64, 64])]; + tensor inv_79_cast_fp16 = reshape(shape = var_4980, x = inv_77_cast_fp16)[name = string("inv_79_cast_fp16")]; + bool new_v_9_transpose_x_0 = const()[name = string("new_v_9_transpose_x_0"), val = bool(false)]; + bool new_v_9_transpose_y_0 = const()[name = string("new_v_9_transpose_y_0"), val = bool(false)]; + tensor new_v_9_cast_fp16 = matmul(transpose_x = new_v_9_transpose_x_0, transpose_y = new_v_9_transpose_y_0, x = inv_79_cast_fp16, y = v_beta_9_cast_fp16)[name = string("new_v_9_cast_fp16")]; + tensor var_4983_cast_fp16 = exp(x = cum_9_cast_fp16)[name = string("op_4983_cast_fp16")]; + tensor var_4984_axes_0 = const()[name = string("op_4984_axes_0"), val = tensor([3])]; + tensor var_4984_cast_fp16 = expand_dims(axes = var_4984_axes_0, x = var_4983_cast_fp16)[name = string("op_4984_cast_fp16")]; + tensor var_4985_cast_fp16 = mul(x = k_beta_9_cast_fp16, y = var_4984_cast_fp16)[name = string("op_4985_cast_fp16")]; + bool k_cumdecay_9_transpose_x_0 = const()[name = string("k_cumdecay_9_transpose_x_0"), val = bool(false)]; + bool k_cumdecay_9_transpose_y_0 = const()[name = string("k_cumdecay_9_transpose_y_0"), val = bool(false)]; + tensor k_cumdecay_9_cast_fp16 = matmul(transpose_x = k_cumdecay_9_transpose_x_0, transpose_y = k_cumdecay_9_transpose_y_0, x = inv_79_cast_fp16, y = var_4985_cast_fp16)[name = string("k_cumdecay_9_cast_fp16")]; + tensor q_dec_9_cast_fp16 = mul(x = q_53_cast_fp16, y = var_4984_cast_fp16)[name = string("q_dec_9_cast_fp16")]; + tensor var_4990_cast_fp16 = exp(x = to_end_9_cast_fp16)[name = string("op_4990_cast_fp16")]; + tensor var_4991_axes_0 = const()[name = string("op_4991_axes_0"), val = tensor([3])]; + tensor var_4991_cast_fp16 = expand_dims(axes = var_4991_axes_0, x = var_4990_cast_fp16)[name = string("op_4991_cast_fp16")]; + tensor k_dec_9_cast_fp16 = mul(x = k_45_cast_fp16, y = var_4991_cast_fp16)[name = string("k_dec_9_cast_fp16")]; + tensor var_4993_begin_0 = const()[name = string("op_4993_begin_0"), val = tensor([0, 0, -1])]; + tensor var_4993_end_0 = const()[name = string("op_4993_end_0"), val = tensor([32, 32, 64])]; + tensor var_4993_end_mask_0 = const()[name = string("op_4993_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4993_cast_fp16 = slice_by_index(begin = var_4993_begin_0, end = var_4993_end_0, end_mask = var_4993_end_mask_0, x = cum_9_cast_fp16)[name = string("op_4993_cast_fp16")]; + tensor var_4994_cast_fp16 = exp(x = var_4993_cast_fp16)[name = string("op_4994_cast_fp16")]; + tensor chunk_decay_9_axes_0 = const()[name = string("chunk_decay_9_axes_0"), val = tensor([3])]; + tensor chunk_decay_9_cast_fp16 = expand_dims(axes = chunk_decay_9_axes_0, x = var_4994_cast_fp16)[name = string("chunk_decay_9_cast_fp16")]; + tensor var_4999_begin_0 = const()[name = string("op_4999_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4999_end_0 = const()[name = string("op_4999_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_4999_end_mask_0 = const()[name = string("op_4999_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4999_squeeze_mask_0 = const()[name = string("op_4999_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_4999_cast_fp16 = slice_by_index(begin = var_4999_begin_0, end = var_4999_end_0, end_mask = var_4999_end_mask_0, squeeze_mask = var_4999_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_4999_cast_fp16")]; + tensor var_5001_begin_0 = const()[name = string("op_5001_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5001_end_0 = const()[name = string("op_5001_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_5001_end_mask_0 = const()[name = string("op_5001_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5001_squeeze_mask_0 = const()[name = string("op_5001_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, squeeze_mask = var_5001_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5001_cast_fp16")]; + bool var_5002_transpose_x_0 = const()[name = string("op_5002_transpose_x_0"), val = bool(false)]; + bool var_5002_transpose_y_0 = const()[name = string("op_5002_transpose_y_0"), val = bool(false)]; + tensor var_5002_cast_fp16 = matmul(transpose_x = var_5002_transpose_x_0, transpose_y = var_5002_transpose_y_0, x = var_5001_cast_fp16, y = state_1_to_fp16)[name = string("op_5002_cast_fp16")]; + tensor v_new_257_cast_fp16 = sub(x = var_4999_cast_fp16, y = var_5002_cast_fp16)[name = string("v_new_257_cast_fp16")]; + tensor var_5005_begin_0 = const()[name = string("op_5005_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5005_end_0 = const()[name = string("op_5005_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_5005_end_mask_0 = const()[name = string("op_5005_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5005_squeeze_mask_0 = const()[name = string("op_5005_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5005_cast_fp16 = slice_by_index(begin = var_5005_begin_0, end = var_5005_end_0, end_mask = var_5005_end_mask_0, squeeze_mask = var_5005_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5005_cast_fp16")]; + bool var_5006_transpose_x_0 = const()[name = string("op_5006_transpose_x_0"), val = bool(false)]; + bool var_5006_transpose_y_0 = const()[name = string("op_5006_transpose_y_0"), val = bool(false)]; + tensor var_5006_cast_fp16 = matmul(transpose_x = var_5006_transpose_x_0, transpose_y = var_5006_transpose_y_0, x = var_5005_cast_fp16, y = state_1_to_fp16)[name = string("op_5006_cast_fp16")]; + tensor var_5008_begin_0 = const()[name = string("op_5008_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5008_end_0 = const()[name = string("op_5008_end_0"), val = tensor([32, 1, 64, 64])]; + tensor var_5008_end_mask_0 = const()[name = string("op_5008_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5008_squeeze_mask_0 = const()[name = string("op_5008_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5008_cast_fp16 = slice_by_index(begin = var_5008_begin_0, end = var_5008_end_0, end_mask = var_5008_end_mask_0, squeeze_mask = var_5008_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5008_cast_fp16")]; + bool var_5009_transpose_x_0 = const()[name = string("op_5009_transpose_x_0"), val = bool(false)]; + bool var_5009_transpose_y_0 = const()[name = string("op_5009_transpose_y_0"), val = bool(false)]; + tensor var_5009_cast_fp16 = matmul(transpose_x = var_5009_transpose_x_0, transpose_y = var_5009_transpose_y_0, x = var_5008_cast_fp16, y = v_new_257_cast_fp16)[name = string("op_5009_cast_fp16")]; + tensor var_5010_cast_fp16 = add(x = var_5006_cast_fp16, y = var_5009_cast_fp16)[name = string("op_5010_cast_fp16")]; + tensor var_5012_begin_0 = const()[name = string("op_5012_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5012_end_0 = const()[name = string("op_5012_end_0"), val = tensor([32, 1, 1, 1])]; + tensor var_5012_end_mask_0 = const()[name = string("op_5012_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5012_squeeze_mask_0 = const()[name = string("op_5012_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5012_cast_fp16 = slice_by_index(begin = var_5012_begin_0, end = var_5012_end_0, end_mask = var_5012_end_mask_0, squeeze_mask = var_5012_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5012_cast_fp16")]; + tensor var_5013_cast_fp16 = mul(x = state_1_to_fp16, y = var_5012_cast_fp16)[name = string("op_5013_cast_fp16")]; + tensor var_5015_begin_0 = const()[name = string("op_5015_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5015_end_0 = const()[name = string("op_5015_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_5015_end_mask_0 = const()[name = string("op_5015_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5015_squeeze_mask_0 = const()[name = string("op_5015_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5015_cast_fp16 = slice_by_index(begin = var_5015_begin_0, end = var_5015_end_0, end_mask = var_5015_end_mask_0, squeeze_mask = var_5015_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5015_cast_fp16")]; + bool var_5017_transpose_x_1 = const()[name = string("op_5017_transpose_x_1"), val = bool(true)]; + bool var_5017_transpose_y_1 = const()[name = string("op_5017_transpose_y_1"), val = bool(false)]; + tensor var_5017_cast_fp16 = matmul(transpose_x = var_5017_transpose_x_1, transpose_y = var_5017_transpose_y_1, x = var_5015_cast_fp16, y = v_new_257_cast_fp16)[name = string("op_5017_cast_fp16")]; + tensor state_259_cast_fp16 = add(x = var_5013_cast_fp16, y = var_5017_cast_fp16)[name = string("state_259_cast_fp16")]; + tensor var_5020_begin_0 = const()[name = string("op_5020_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_5020_end_0 = const()[name = string("op_5020_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_5020_end_mask_0 = const()[name = string("op_5020_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5020_squeeze_mask_0 = const()[name = string("op_5020_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5020_cast_fp16 = slice_by_index(begin = var_5020_begin_0, end = var_5020_end_0, end_mask = var_5020_end_mask_0, squeeze_mask = var_5020_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5020_cast_fp16")]; + tensor var_5022_begin_0 = const()[name = string("op_5022_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_5022_end_0 = const()[name = string("op_5022_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_5022_end_mask_0 = const()[name = string("op_5022_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5022_squeeze_mask_0 = const()[name = string("op_5022_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5022_cast_fp16 = slice_by_index(begin = var_5022_begin_0, end = var_5022_end_0, end_mask = var_5022_end_mask_0, squeeze_mask = var_5022_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5022_cast_fp16")]; + bool var_5023_transpose_x_0 = const()[name = string("op_5023_transpose_x_0"), val = bool(false)]; + bool var_5023_transpose_y_0 = const()[name = string("op_5023_transpose_y_0"), val = bool(false)]; + tensor var_5023_cast_fp16 = matmul(transpose_x = var_5023_transpose_x_0, transpose_y = var_5023_transpose_y_0, x = var_5022_cast_fp16, y = state_259_cast_fp16)[name = string("op_5023_cast_fp16")]; + tensor v_new_259_cast_fp16 = sub(x = var_5020_cast_fp16, y = var_5023_cast_fp16)[name = string("v_new_259_cast_fp16")]; + tensor var_5026_begin_0 = const()[name = string("op_5026_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_5026_end_0 = const()[name = string("op_5026_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_5026_end_mask_0 = const()[name = string("op_5026_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5026_squeeze_mask_0 = const()[name = string("op_5026_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5026_cast_fp16 = slice_by_index(begin = var_5026_begin_0, end = var_5026_end_0, end_mask = var_5026_end_mask_0, squeeze_mask = var_5026_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5026_cast_fp16")]; + bool var_5027_transpose_x_0 = const()[name = string("op_5027_transpose_x_0"), val = bool(false)]; + bool var_5027_transpose_y_0 = const()[name = string("op_5027_transpose_y_0"), val = bool(false)]; + tensor var_5027_cast_fp16 = matmul(transpose_x = var_5027_transpose_x_0, transpose_y = var_5027_transpose_y_0, x = var_5026_cast_fp16, y = state_259_cast_fp16)[name = string("op_5027_cast_fp16")]; + tensor var_5029_begin_0 = const()[name = string("op_5029_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_5029_end_0 = const()[name = string("op_5029_end_0"), val = tensor([32, 2, 64, 64])]; + tensor var_5029_end_mask_0 = const()[name = string("op_5029_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5029_squeeze_mask_0 = const()[name = string("op_5029_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5029_cast_fp16 = slice_by_index(begin = var_5029_begin_0, end = var_5029_end_0, end_mask = var_5029_end_mask_0, squeeze_mask = var_5029_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5029_cast_fp16")]; + bool var_5030_transpose_x_0 = const()[name = string("op_5030_transpose_x_0"), val = bool(false)]; + bool var_5030_transpose_y_0 = const()[name = string("op_5030_transpose_y_0"), val = bool(false)]; + tensor var_5030_cast_fp16 = matmul(transpose_x = var_5030_transpose_x_0, transpose_y = var_5030_transpose_y_0, x = var_5029_cast_fp16, y = v_new_259_cast_fp16)[name = string("op_5030_cast_fp16")]; + tensor var_5031_cast_fp16 = add(x = var_5027_cast_fp16, y = var_5030_cast_fp16)[name = string("op_5031_cast_fp16")]; + tensor var_5033_begin_0 = const()[name = string("op_5033_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_5033_end_0 = const()[name = string("op_5033_end_0"), val = tensor([32, 2, 1, 1])]; + tensor var_5033_end_mask_0 = const()[name = string("op_5033_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5033_squeeze_mask_0 = const()[name = string("op_5033_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5033_cast_fp16 = slice_by_index(begin = var_5033_begin_0, end = var_5033_end_0, end_mask = var_5033_end_mask_0, squeeze_mask = var_5033_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5033_cast_fp16")]; + tensor var_5034_cast_fp16 = mul(x = state_259_cast_fp16, y = var_5033_cast_fp16)[name = string("op_5034_cast_fp16")]; + tensor var_5036_begin_0 = const()[name = string("op_5036_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_5036_end_0 = const()[name = string("op_5036_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_5036_end_mask_0 = const()[name = string("op_5036_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5036_squeeze_mask_0 = const()[name = string("op_5036_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5036_cast_fp16 = slice_by_index(begin = var_5036_begin_0, end = var_5036_end_0, end_mask = var_5036_end_mask_0, squeeze_mask = var_5036_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5036_cast_fp16")]; + bool var_5038_transpose_x_1 = const()[name = string("op_5038_transpose_x_1"), val = bool(true)]; + bool var_5038_transpose_y_1 = const()[name = string("op_5038_transpose_y_1"), val = bool(false)]; + tensor var_5038_cast_fp16 = matmul(transpose_x = var_5038_transpose_x_1, transpose_y = var_5038_transpose_y_1, x = var_5036_cast_fp16, y = v_new_259_cast_fp16)[name = string("op_5038_cast_fp16")]; + tensor state_261_cast_fp16 = add(x = var_5034_cast_fp16, y = var_5038_cast_fp16)[name = string("state_261_cast_fp16")]; + tensor var_5041_begin_0 = const()[name = string("op_5041_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_5041_end_0 = const()[name = string("op_5041_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_5041_end_mask_0 = const()[name = string("op_5041_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5041_squeeze_mask_0 = const()[name = string("op_5041_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5041_cast_fp16 = slice_by_index(begin = var_5041_begin_0, end = var_5041_end_0, end_mask = var_5041_end_mask_0, squeeze_mask = var_5041_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5041_cast_fp16")]; + tensor var_5043_begin_0 = const()[name = string("op_5043_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_5043_end_0 = const()[name = string("op_5043_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_5043_end_mask_0 = const()[name = string("op_5043_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5043_squeeze_mask_0 = const()[name = string("op_5043_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5043_cast_fp16 = slice_by_index(begin = var_5043_begin_0, end = var_5043_end_0, end_mask = var_5043_end_mask_0, squeeze_mask = var_5043_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5043_cast_fp16")]; + bool var_5044_transpose_x_0 = const()[name = string("op_5044_transpose_x_0"), val = bool(false)]; + bool var_5044_transpose_y_0 = const()[name = string("op_5044_transpose_y_0"), val = bool(false)]; + tensor var_5044_cast_fp16 = matmul(transpose_x = var_5044_transpose_x_0, transpose_y = var_5044_transpose_y_0, x = var_5043_cast_fp16, y = state_261_cast_fp16)[name = string("op_5044_cast_fp16")]; + tensor v_new_261_cast_fp16 = sub(x = var_5041_cast_fp16, y = var_5044_cast_fp16)[name = string("v_new_261_cast_fp16")]; + tensor var_5047_begin_0 = const()[name = string("op_5047_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_5047_end_0 = const()[name = string("op_5047_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_5047_end_mask_0 = const()[name = string("op_5047_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5047_squeeze_mask_0 = const()[name = string("op_5047_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5047_cast_fp16 = slice_by_index(begin = var_5047_begin_0, end = var_5047_end_0, end_mask = var_5047_end_mask_0, squeeze_mask = var_5047_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5047_cast_fp16")]; + bool var_5048_transpose_x_0 = const()[name = string("op_5048_transpose_x_0"), val = bool(false)]; + bool var_5048_transpose_y_0 = const()[name = string("op_5048_transpose_y_0"), val = bool(false)]; + tensor var_5048_cast_fp16 = matmul(transpose_x = var_5048_transpose_x_0, transpose_y = var_5048_transpose_y_0, x = var_5047_cast_fp16, y = state_261_cast_fp16)[name = string("op_5048_cast_fp16")]; + tensor var_5050_begin_0 = const()[name = string("op_5050_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_5050_end_0 = const()[name = string("op_5050_end_0"), val = tensor([32, 3, 64, 64])]; + tensor var_5050_end_mask_0 = const()[name = string("op_5050_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5050_squeeze_mask_0 = const()[name = string("op_5050_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5050_cast_fp16 = slice_by_index(begin = var_5050_begin_0, end = var_5050_end_0, end_mask = var_5050_end_mask_0, squeeze_mask = var_5050_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5050_cast_fp16")]; + bool var_5051_transpose_x_0 = const()[name = string("op_5051_transpose_x_0"), val = bool(false)]; + bool var_5051_transpose_y_0 = const()[name = string("op_5051_transpose_y_0"), val = bool(false)]; + tensor var_5051_cast_fp16 = matmul(transpose_x = var_5051_transpose_x_0, transpose_y = var_5051_transpose_y_0, x = var_5050_cast_fp16, y = v_new_261_cast_fp16)[name = string("op_5051_cast_fp16")]; + tensor var_5052_cast_fp16 = add(x = var_5048_cast_fp16, y = var_5051_cast_fp16)[name = string("op_5052_cast_fp16")]; + tensor var_5054_begin_0 = const()[name = string("op_5054_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_5054_end_0 = const()[name = string("op_5054_end_0"), val = tensor([32, 3, 1, 1])]; + tensor var_5054_end_mask_0 = const()[name = string("op_5054_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5054_squeeze_mask_0 = const()[name = string("op_5054_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5054_cast_fp16 = slice_by_index(begin = var_5054_begin_0, end = var_5054_end_0, end_mask = var_5054_end_mask_0, squeeze_mask = var_5054_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5054_cast_fp16")]; + tensor var_5055_cast_fp16 = mul(x = state_261_cast_fp16, y = var_5054_cast_fp16)[name = string("op_5055_cast_fp16")]; + tensor var_5057_begin_0 = const()[name = string("op_5057_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_5057_end_0 = const()[name = string("op_5057_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_5057_end_mask_0 = const()[name = string("op_5057_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5057_squeeze_mask_0 = const()[name = string("op_5057_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5057_cast_fp16 = slice_by_index(begin = var_5057_begin_0, end = var_5057_end_0, end_mask = var_5057_end_mask_0, squeeze_mask = var_5057_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5057_cast_fp16")]; + bool var_5059_transpose_x_1 = const()[name = string("op_5059_transpose_x_1"), val = bool(true)]; + bool var_5059_transpose_y_1 = const()[name = string("op_5059_transpose_y_1"), val = bool(false)]; + tensor var_5059_cast_fp16 = matmul(transpose_x = var_5059_transpose_x_1, transpose_y = var_5059_transpose_y_1, x = var_5057_cast_fp16, y = v_new_261_cast_fp16)[name = string("op_5059_cast_fp16")]; + tensor state_263_cast_fp16 = add(x = var_5055_cast_fp16, y = var_5059_cast_fp16)[name = string("state_263_cast_fp16")]; + tensor var_5062_begin_0 = const()[name = string("op_5062_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_5062_end_0 = const()[name = string("op_5062_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_5062_end_mask_0 = const()[name = string("op_5062_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5062_squeeze_mask_0 = const()[name = string("op_5062_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5062_cast_fp16 = slice_by_index(begin = var_5062_begin_0, end = var_5062_end_0, end_mask = var_5062_end_mask_0, squeeze_mask = var_5062_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5062_cast_fp16")]; + tensor var_5064_begin_0 = const()[name = string("op_5064_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_5064_end_0 = const()[name = string("op_5064_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_5064_end_mask_0 = const()[name = string("op_5064_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5064_squeeze_mask_0 = const()[name = string("op_5064_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5064_cast_fp16 = slice_by_index(begin = var_5064_begin_0, end = var_5064_end_0, end_mask = var_5064_end_mask_0, squeeze_mask = var_5064_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5064_cast_fp16")]; + bool var_5065_transpose_x_0 = const()[name = string("op_5065_transpose_x_0"), val = bool(false)]; + bool var_5065_transpose_y_0 = const()[name = string("op_5065_transpose_y_0"), val = bool(false)]; + tensor var_5065_cast_fp16 = matmul(transpose_x = var_5065_transpose_x_0, transpose_y = var_5065_transpose_y_0, x = var_5064_cast_fp16, y = state_263_cast_fp16)[name = string("op_5065_cast_fp16")]; + tensor v_new_263_cast_fp16 = sub(x = var_5062_cast_fp16, y = var_5065_cast_fp16)[name = string("v_new_263_cast_fp16")]; + tensor var_5068_begin_0 = const()[name = string("op_5068_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_5068_end_0 = const()[name = string("op_5068_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_5068_end_mask_0 = const()[name = string("op_5068_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5068_squeeze_mask_0 = const()[name = string("op_5068_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5068_cast_fp16 = slice_by_index(begin = var_5068_begin_0, end = var_5068_end_0, end_mask = var_5068_end_mask_0, squeeze_mask = var_5068_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5068_cast_fp16")]; + bool var_5069_transpose_x_0 = const()[name = string("op_5069_transpose_x_0"), val = bool(false)]; + bool var_5069_transpose_y_0 = const()[name = string("op_5069_transpose_y_0"), val = bool(false)]; + tensor var_5069_cast_fp16 = matmul(transpose_x = var_5069_transpose_x_0, transpose_y = var_5069_transpose_y_0, x = var_5068_cast_fp16, y = state_263_cast_fp16)[name = string("op_5069_cast_fp16")]; + tensor var_5071_begin_0 = const()[name = string("op_5071_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_5071_end_0 = const()[name = string("op_5071_end_0"), val = tensor([32, 4, 64, 64])]; + tensor var_5071_end_mask_0 = const()[name = string("op_5071_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5071_squeeze_mask_0 = const()[name = string("op_5071_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5071_cast_fp16 = slice_by_index(begin = var_5071_begin_0, end = var_5071_end_0, end_mask = var_5071_end_mask_0, squeeze_mask = var_5071_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5071_cast_fp16")]; + bool var_5072_transpose_x_0 = const()[name = string("op_5072_transpose_x_0"), val = bool(false)]; + bool var_5072_transpose_y_0 = const()[name = string("op_5072_transpose_y_0"), val = bool(false)]; + tensor var_5072_cast_fp16 = matmul(transpose_x = var_5072_transpose_x_0, transpose_y = var_5072_transpose_y_0, x = var_5071_cast_fp16, y = v_new_263_cast_fp16)[name = string("op_5072_cast_fp16")]; + tensor var_5073_cast_fp16 = add(x = var_5069_cast_fp16, y = var_5072_cast_fp16)[name = string("op_5073_cast_fp16")]; + tensor var_5075_begin_0 = const()[name = string("op_5075_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_5075_end_0 = const()[name = string("op_5075_end_0"), val = tensor([32, 4, 1, 1])]; + tensor var_5075_end_mask_0 = const()[name = string("op_5075_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5075_squeeze_mask_0 = const()[name = string("op_5075_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5075_cast_fp16 = slice_by_index(begin = var_5075_begin_0, end = var_5075_end_0, end_mask = var_5075_end_mask_0, squeeze_mask = var_5075_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5075_cast_fp16")]; + tensor var_5076_cast_fp16 = mul(x = state_263_cast_fp16, y = var_5075_cast_fp16)[name = string("op_5076_cast_fp16")]; + tensor var_5078_begin_0 = const()[name = string("op_5078_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_5078_end_0 = const()[name = string("op_5078_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_5078_end_mask_0 = const()[name = string("op_5078_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5078_squeeze_mask_0 = const()[name = string("op_5078_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5078_cast_fp16 = slice_by_index(begin = var_5078_begin_0, end = var_5078_end_0, end_mask = var_5078_end_mask_0, squeeze_mask = var_5078_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5078_cast_fp16")]; + bool var_5080_transpose_x_1 = const()[name = string("op_5080_transpose_x_1"), val = bool(true)]; + bool var_5080_transpose_y_1 = const()[name = string("op_5080_transpose_y_1"), val = bool(false)]; + tensor var_5080_cast_fp16 = matmul(transpose_x = var_5080_transpose_x_1, transpose_y = var_5080_transpose_y_1, x = var_5078_cast_fp16, y = v_new_263_cast_fp16)[name = string("op_5080_cast_fp16")]; + tensor state_265_cast_fp16 = add(x = var_5076_cast_fp16, y = var_5080_cast_fp16)[name = string("state_265_cast_fp16")]; + tensor var_5083_begin_0 = const()[name = string("op_5083_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5083_end_0 = const()[name = string("op_5083_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_5083_end_mask_0 = const()[name = string("op_5083_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5083_squeeze_mask_0 = const()[name = string("op_5083_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5083_cast_fp16 = slice_by_index(begin = var_5083_begin_0, end = var_5083_end_0, end_mask = var_5083_end_mask_0, squeeze_mask = var_5083_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5083_cast_fp16")]; + tensor var_5085_begin_0 = const()[name = string("op_5085_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5085_end_0 = const()[name = string("op_5085_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_5085_end_mask_0 = const()[name = string("op_5085_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5085_squeeze_mask_0 = const()[name = string("op_5085_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5085_cast_fp16 = slice_by_index(begin = var_5085_begin_0, end = var_5085_end_0, end_mask = var_5085_end_mask_0, squeeze_mask = var_5085_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5085_cast_fp16")]; + bool var_5086_transpose_x_0 = const()[name = string("op_5086_transpose_x_0"), val = bool(false)]; + bool var_5086_transpose_y_0 = const()[name = string("op_5086_transpose_y_0"), val = bool(false)]; + tensor var_5086_cast_fp16 = matmul(transpose_x = var_5086_transpose_x_0, transpose_y = var_5086_transpose_y_0, x = var_5085_cast_fp16, y = state_265_cast_fp16)[name = string("op_5086_cast_fp16")]; + tensor v_new_265_cast_fp16 = sub(x = var_5083_cast_fp16, y = var_5086_cast_fp16)[name = string("v_new_265_cast_fp16")]; + tensor var_5089_begin_0 = const()[name = string("op_5089_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5089_end_0 = const()[name = string("op_5089_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_5089_end_mask_0 = const()[name = string("op_5089_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5089_squeeze_mask_0 = const()[name = string("op_5089_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5089_cast_fp16 = slice_by_index(begin = var_5089_begin_0, end = var_5089_end_0, end_mask = var_5089_end_mask_0, squeeze_mask = var_5089_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5089_cast_fp16")]; + bool var_5090_transpose_x_0 = const()[name = string("op_5090_transpose_x_0"), val = bool(false)]; + bool var_5090_transpose_y_0 = const()[name = string("op_5090_transpose_y_0"), val = bool(false)]; + tensor var_5090_cast_fp16 = matmul(transpose_x = var_5090_transpose_x_0, transpose_y = var_5090_transpose_y_0, x = var_5089_cast_fp16, y = state_265_cast_fp16)[name = string("op_5090_cast_fp16")]; + tensor var_5092_begin_0 = const()[name = string("op_5092_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5092_end_0 = const()[name = string("op_5092_end_0"), val = tensor([32, 5, 64, 64])]; + tensor var_5092_end_mask_0 = const()[name = string("op_5092_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5092_squeeze_mask_0 = const()[name = string("op_5092_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5092_cast_fp16 = slice_by_index(begin = var_5092_begin_0, end = var_5092_end_0, end_mask = var_5092_end_mask_0, squeeze_mask = var_5092_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5092_cast_fp16")]; + bool var_5093_transpose_x_0 = const()[name = string("op_5093_transpose_x_0"), val = bool(false)]; + bool var_5093_transpose_y_0 = const()[name = string("op_5093_transpose_y_0"), val = bool(false)]; + tensor var_5093_cast_fp16 = matmul(transpose_x = var_5093_transpose_x_0, transpose_y = var_5093_transpose_y_0, x = var_5092_cast_fp16, y = v_new_265_cast_fp16)[name = string("op_5093_cast_fp16")]; + tensor var_5094_cast_fp16 = add(x = var_5090_cast_fp16, y = var_5093_cast_fp16)[name = string("op_5094_cast_fp16")]; + tensor var_5096_begin_0 = const()[name = string("op_5096_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5096_end_0 = const()[name = string("op_5096_end_0"), val = tensor([32, 5, 1, 1])]; + tensor var_5096_end_mask_0 = const()[name = string("op_5096_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5096_squeeze_mask_0 = const()[name = string("op_5096_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5096_cast_fp16 = slice_by_index(begin = var_5096_begin_0, end = var_5096_end_0, end_mask = var_5096_end_mask_0, squeeze_mask = var_5096_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5096_cast_fp16")]; + tensor var_5097_cast_fp16 = mul(x = state_265_cast_fp16, y = var_5096_cast_fp16)[name = string("op_5097_cast_fp16")]; + tensor var_5099_begin_0 = const()[name = string("op_5099_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5099_end_0 = const()[name = string("op_5099_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_5099_end_mask_0 = const()[name = string("op_5099_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5099_squeeze_mask_0 = const()[name = string("op_5099_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5099_cast_fp16 = slice_by_index(begin = var_5099_begin_0, end = var_5099_end_0, end_mask = var_5099_end_mask_0, squeeze_mask = var_5099_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5099_cast_fp16")]; + bool var_5101_transpose_x_1 = const()[name = string("op_5101_transpose_x_1"), val = bool(true)]; + bool var_5101_transpose_y_1 = const()[name = string("op_5101_transpose_y_1"), val = bool(false)]; + tensor var_5101_cast_fp16 = matmul(transpose_x = var_5101_transpose_x_1, transpose_y = var_5101_transpose_y_1, x = var_5099_cast_fp16, y = v_new_265_cast_fp16)[name = string("op_5101_cast_fp16")]; + tensor state_267_cast_fp16 = add(x = var_5097_cast_fp16, y = var_5101_cast_fp16)[name = string("state_267_cast_fp16")]; + tensor var_5104_begin_0 = const()[name = string("op_5104_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_5104_end_0 = const()[name = string("op_5104_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_5104_end_mask_0 = const()[name = string("op_5104_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5104_squeeze_mask_0 = const()[name = string("op_5104_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5104_cast_fp16 = slice_by_index(begin = var_5104_begin_0, end = var_5104_end_0, end_mask = var_5104_end_mask_0, squeeze_mask = var_5104_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5104_cast_fp16")]; + tensor var_5106_begin_0 = const()[name = string("op_5106_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_5106_end_0 = const()[name = string("op_5106_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_5106_end_mask_0 = const()[name = string("op_5106_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5106_squeeze_mask_0 = const()[name = string("op_5106_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5106_cast_fp16 = slice_by_index(begin = var_5106_begin_0, end = var_5106_end_0, end_mask = var_5106_end_mask_0, squeeze_mask = var_5106_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5106_cast_fp16")]; + bool var_5107_transpose_x_0 = const()[name = string("op_5107_transpose_x_0"), val = bool(false)]; + bool var_5107_transpose_y_0 = const()[name = string("op_5107_transpose_y_0"), val = bool(false)]; + tensor var_5107_cast_fp16 = matmul(transpose_x = var_5107_transpose_x_0, transpose_y = var_5107_transpose_y_0, x = var_5106_cast_fp16, y = state_267_cast_fp16)[name = string("op_5107_cast_fp16")]; + tensor v_new_267_cast_fp16 = sub(x = var_5104_cast_fp16, y = var_5107_cast_fp16)[name = string("v_new_267_cast_fp16")]; + tensor var_5110_begin_0 = const()[name = string("op_5110_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_5110_end_0 = const()[name = string("op_5110_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_5110_end_mask_0 = const()[name = string("op_5110_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5110_squeeze_mask_0 = const()[name = string("op_5110_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5110_cast_fp16 = slice_by_index(begin = var_5110_begin_0, end = var_5110_end_0, end_mask = var_5110_end_mask_0, squeeze_mask = var_5110_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5110_cast_fp16")]; + bool var_5111_transpose_x_0 = const()[name = string("op_5111_transpose_x_0"), val = bool(false)]; + bool var_5111_transpose_y_0 = const()[name = string("op_5111_transpose_y_0"), val = bool(false)]; + tensor var_5111_cast_fp16 = matmul(transpose_x = var_5111_transpose_x_0, transpose_y = var_5111_transpose_y_0, x = var_5110_cast_fp16, y = state_267_cast_fp16)[name = string("op_5111_cast_fp16")]; + tensor var_5113_begin_0 = const()[name = string("op_5113_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_5113_end_0 = const()[name = string("op_5113_end_0"), val = tensor([32, 6, 64, 64])]; + tensor var_5113_end_mask_0 = const()[name = string("op_5113_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5113_squeeze_mask_0 = const()[name = string("op_5113_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5113_cast_fp16 = slice_by_index(begin = var_5113_begin_0, end = var_5113_end_0, end_mask = var_5113_end_mask_0, squeeze_mask = var_5113_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5113_cast_fp16")]; + bool var_5114_transpose_x_0 = const()[name = string("op_5114_transpose_x_0"), val = bool(false)]; + bool var_5114_transpose_y_0 = const()[name = string("op_5114_transpose_y_0"), val = bool(false)]; + tensor var_5114_cast_fp16 = matmul(transpose_x = var_5114_transpose_x_0, transpose_y = var_5114_transpose_y_0, x = var_5113_cast_fp16, y = v_new_267_cast_fp16)[name = string("op_5114_cast_fp16")]; + tensor var_5115_cast_fp16 = add(x = var_5111_cast_fp16, y = var_5114_cast_fp16)[name = string("op_5115_cast_fp16")]; + tensor var_5117_begin_0 = const()[name = string("op_5117_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_5117_end_0 = const()[name = string("op_5117_end_0"), val = tensor([32, 6, 1, 1])]; + tensor var_5117_end_mask_0 = const()[name = string("op_5117_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5117_squeeze_mask_0 = const()[name = string("op_5117_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5117_cast_fp16 = slice_by_index(begin = var_5117_begin_0, end = var_5117_end_0, end_mask = var_5117_end_mask_0, squeeze_mask = var_5117_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5117_cast_fp16")]; + tensor var_5118_cast_fp16 = mul(x = state_267_cast_fp16, y = var_5117_cast_fp16)[name = string("op_5118_cast_fp16")]; + tensor var_5120_begin_0 = const()[name = string("op_5120_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_5120_end_0 = const()[name = string("op_5120_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_5120_end_mask_0 = const()[name = string("op_5120_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5120_squeeze_mask_0 = const()[name = string("op_5120_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5120_cast_fp16 = slice_by_index(begin = var_5120_begin_0, end = var_5120_end_0, end_mask = var_5120_end_mask_0, squeeze_mask = var_5120_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5120_cast_fp16")]; + bool var_5122_transpose_x_1 = const()[name = string("op_5122_transpose_x_1"), val = bool(true)]; + bool var_5122_transpose_y_1 = const()[name = string("op_5122_transpose_y_1"), val = bool(false)]; + tensor var_5122_cast_fp16 = matmul(transpose_x = var_5122_transpose_x_1, transpose_y = var_5122_transpose_y_1, x = var_5120_cast_fp16, y = v_new_267_cast_fp16)[name = string("op_5122_cast_fp16")]; + tensor state_269_cast_fp16 = add(x = var_5118_cast_fp16, y = var_5122_cast_fp16)[name = string("state_269_cast_fp16")]; + tensor var_5125_begin_0 = const()[name = string("op_5125_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_5125_end_0 = const()[name = string("op_5125_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_5125_end_mask_0 = const()[name = string("op_5125_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5125_squeeze_mask_0 = const()[name = string("op_5125_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5125_cast_fp16 = slice_by_index(begin = var_5125_begin_0, end = var_5125_end_0, end_mask = var_5125_end_mask_0, squeeze_mask = var_5125_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5125_cast_fp16")]; + tensor var_5127_begin_0 = const()[name = string("op_5127_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_5127_end_0 = const()[name = string("op_5127_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_5127_end_mask_0 = const()[name = string("op_5127_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5127_squeeze_mask_0 = const()[name = string("op_5127_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5127_cast_fp16 = slice_by_index(begin = var_5127_begin_0, end = var_5127_end_0, end_mask = var_5127_end_mask_0, squeeze_mask = var_5127_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5127_cast_fp16")]; + bool var_5128_transpose_x_0 = const()[name = string("op_5128_transpose_x_0"), val = bool(false)]; + bool var_5128_transpose_y_0 = const()[name = string("op_5128_transpose_y_0"), val = bool(false)]; + tensor var_5128_cast_fp16 = matmul(transpose_x = var_5128_transpose_x_0, transpose_y = var_5128_transpose_y_0, x = var_5127_cast_fp16, y = state_269_cast_fp16)[name = string("op_5128_cast_fp16")]; + tensor v_new_269_cast_fp16 = sub(x = var_5125_cast_fp16, y = var_5128_cast_fp16)[name = string("v_new_269_cast_fp16")]; + tensor var_5131_begin_0 = const()[name = string("op_5131_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_5131_end_0 = const()[name = string("op_5131_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_5131_end_mask_0 = const()[name = string("op_5131_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5131_squeeze_mask_0 = const()[name = string("op_5131_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5131_cast_fp16 = slice_by_index(begin = var_5131_begin_0, end = var_5131_end_0, end_mask = var_5131_end_mask_0, squeeze_mask = var_5131_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5131_cast_fp16")]; + bool var_5132_transpose_x_0 = const()[name = string("op_5132_transpose_x_0"), val = bool(false)]; + bool var_5132_transpose_y_0 = const()[name = string("op_5132_transpose_y_0"), val = bool(false)]; + tensor var_5132_cast_fp16 = matmul(transpose_x = var_5132_transpose_x_0, transpose_y = var_5132_transpose_y_0, x = var_5131_cast_fp16, y = state_269_cast_fp16)[name = string("op_5132_cast_fp16")]; + tensor var_5134_begin_0 = const()[name = string("op_5134_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_5134_end_0 = const()[name = string("op_5134_end_0"), val = tensor([32, 7, 64, 64])]; + tensor var_5134_end_mask_0 = const()[name = string("op_5134_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5134_squeeze_mask_0 = const()[name = string("op_5134_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5134_cast_fp16 = slice_by_index(begin = var_5134_begin_0, end = var_5134_end_0, end_mask = var_5134_end_mask_0, squeeze_mask = var_5134_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5134_cast_fp16")]; + bool var_5135_transpose_x_0 = const()[name = string("op_5135_transpose_x_0"), val = bool(false)]; + bool var_5135_transpose_y_0 = const()[name = string("op_5135_transpose_y_0"), val = bool(false)]; + tensor var_5135_cast_fp16 = matmul(transpose_x = var_5135_transpose_x_0, transpose_y = var_5135_transpose_y_0, x = var_5134_cast_fp16, y = v_new_269_cast_fp16)[name = string("op_5135_cast_fp16")]; + tensor var_5136_cast_fp16 = add(x = var_5132_cast_fp16, y = var_5135_cast_fp16)[name = string("op_5136_cast_fp16")]; + tensor var_5138_begin_0 = const()[name = string("op_5138_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_5138_end_0 = const()[name = string("op_5138_end_0"), val = tensor([32, 7, 1, 1])]; + tensor var_5138_end_mask_0 = const()[name = string("op_5138_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5138_squeeze_mask_0 = const()[name = string("op_5138_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5138_cast_fp16 = slice_by_index(begin = var_5138_begin_0, end = var_5138_end_0, end_mask = var_5138_end_mask_0, squeeze_mask = var_5138_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5138_cast_fp16")]; + tensor var_5139_cast_fp16 = mul(x = state_269_cast_fp16, y = var_5138_cast_fp16)[name = string("op_5139_cast_fp16")]; + tensor var_5141_begin_0 = const()[name = string("op_5141_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_5141_end_0 = const()[name = string("op_5141_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_5141_end_mask_0 = const()[name = string("op_5141_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5141_squeeze_mask_0 = const()[name = string("op_5141_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5141_cast_fp16 = slice_by_index(begin = var_5141_begin_0, end = var_5141_end_0, end_mask = var_5141_end_mask_0, squeeze_mask = var_5141_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5141_cast_fp16")]; + bool var_5143_transpose_x_1 = const()[name = string("op_5143_transpose_x_1"), val = bool(true)]; + bool var_5143_transpose_y_1 = const()[name = string("op_5143_transpose_y_1"), val = bool(false)]; + tensor var_5143_cast_fp16 = matmul(transpose_x = var_5143_transpose_x_1, transpose_y = var_5143_transpose_y_1, x = var_5141_cast_fp16, y = v_new_269_cast_fp16)[name = string("op_5143_cast_fp16")]; + tensor state_271_cast_fp16 = add(x = var_5139_cast_fp16, y = var_5143_cast_fp16)[name = string("state_271_cast_fp16")]; + tensor var_5146_begin_0 = const()[name = string("op_5146_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_5146_end_0 = const()[name = string("op_5146_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_5146_end_mask_0 = const()[name = string("op_5146_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5146_squeeze_mask_0 = const()[name = string("op_5146_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5146_cast_fp16 = slice_by_index(begin = var_5146_begin_0, end = var_5146_end_0, end_mask = var_5146_end_mask_0, squeeze_mask = var_5146_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5146_cast_fp16")]; + tensor var_5148_begin_0 = const()[name = string("op_5148_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_5148_end_0 = const()[name = string("op_5148_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_5148_end_mask_0 = const()[name = string("op_5148_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5148_squeeze_mask_0 = const()[name = string("op_5148_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5148_cast_fp16 = slice_by_index(begin = var_5148_begin_0, end = var_5148_end_0, end_mask = var_5148_end_mask_0, squeeze_mask = var_5148_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5148_cast_fp16")]; + bool var_5149_transpose_x_0 = const()[name = string("op_5149_transpose_x_0"), val = bool(false)]; + bool var_5149_transpose_y_0 = const()[name = string("op_5149_transpose_y_0"), val = bool(false)]; + tensor var_5149_cast_fp16 = matmul(transpose_x = var_5149_transpose_x_0, transpose_y = var_5149_transpose_y_0, x = var_5148_cast_fp16, y = state_271_cast_fp16)[name = string("op_5149_cast_fp16")]; + tensor v_new_271_cast_fp16 = sub(x = var_5146_cast_fp16, y = var_5149_cast_fp16)[name = string("v_new_271_cast_fp16")]; + tensor var_5152_begin_0 = const()[name = string("op_5152_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_5152_end_0 = const()[name = string("op_5152_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_5152_end_mask_0 = const()[name = string("op_5152_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5152_squeeze_mask_0 = const()[name = string("op_5152_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5152_cast_fp16 = slice_by_index(begin = var_5152_begin_0, end = var_5152_end_0, end_mask = var_5152_end_mask_0, squeeze_mask = var_5152_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5152_cast_fp16")]; + bool var_5153_transpose_x_0 = const()[name = string("op_5153_transpose_x_0"), val = bool(false)]; + bool var_5153_transpose_y_0 = const()[name = string("op_5153_transpose_y_0"), val = bool(false)]; + tensor var_5153_cast_fp16 = matmul(transpose_x = var_5153_transpose_x_0, transpose_y = var_5153_transpose_y_0, x = var_5152_cast_fp16, y = state_271_cast_fp16)[name = string("op_5153_cast_fp16")]; + tensor var_5155_begin_0 = const()[name = string("op_5155_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_5155_end_0 = const()[name = string("op_5155_end_0"), val = tensor([32, 8, 64, 64])]; + tensor var_5155_end_mask_0 = const()[name = string("op_5155_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5155_squeeze_mask_0 = const()[name = string("op_5155_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5155_cast_fp16 = slice_by_index(begin = var_5155_begin_0, end = var_5155_end_0, end_mask = var_5155_end_mask_0, squeeze_mask = var_5155_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5155_cast_fp16")]; + bool var_5156_transpose_x_0 = const()[name = string("op_5156_transpose_x_0"), val = bool(false)]; + bool var_5156_transpose_y_0 = const()[name = string("op_5156_transpose_y_0"), val = bool(false)]; + tensor var_5156_cast_fp16 = matmul(transpose_x = var_5156_transpose_x_0, transpose_y = var_5156_transpose_y_0, x = var_5155_cast_fp16, y = v_new_271_cast_fp16)[name = string("op_5156_cast_fp16")]; + tensor var_5157_cast_fp16 = add(x = var_5153_cast_fp16, y = var_5156_cast_fp16)[name = string("op_5157_cast_fp16")]; + tensor var_5159_begin_0 = const()[name = string("op_5159_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_5159_end_0 = const()[name = string("op_5159_end_0"), val = tensor([32, 8, 1, 1])]; + tensor var_5159_end_mask_0 = const()[name = string("op_5159_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5159_squeeze_mask_0 = const()[name = string("op_5159_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5159_cast_fp16 = slice_by_index(begin = var_5159_begin_0, end = var_5159_end_0, end_mask = var_5159_end_mask_0, squeeze_mask = var_5159_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5159_cast_fp16")]; + tensor var_5160_cast_fp16 = mul(x = state_271_cast_fp16, y = var_5159_cast_fp16)[name = string("op_5160_cast_fp16")]; + tensor var_5162_begin_0 = const()[name = string("op_5162_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_5162_end_0 = const()[name = string("op_5162_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_5162_end_mask_0 = const()[name = string("op_5162_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5162_squeeze_mask_0 = const()[name = string("op_5162_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5162_cast_fp16 = slice_by_index(begin = var_5162_begin_0, end = var_5162_end_0, end_mask = var_5162_end_mask_0, squeeze_mask = var_5162_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5162_cast_fp16")]; + bool var_5164_transpose_x_1 = const()[name = string("op_5164_transpose_x_1"), val = bool(true)]; + bool var_5164_transpose_y_1 = const()[name = string("op_5164_transpose_y_1"), val = bool(false)]; + tensor var_5164_cast_fp16 = matmul(transpose_x = var_5164_transpose_x_1, transpose_y = var_5164_transpose_y_1, x = var_5162_cast_fp16, y = v_new_271_cast_fp16)[name = string("op_5164_cast_fp16")]; + tensor state_273_cast_fp16 = add(x = var_5160_cast_fp16, y = var_5164_cast_fp16)[name = string("state_273_cast_fp16")]; + tensor var_5167_begin_0 = const()[name = string("op_5167_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_5167_end_0 = const()[name = string("op_5167_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_5167_end_mask_0 = const()[name = string("op_5167_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5167_squeeze_mask_0 = const()[name = string("op_5167_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5167_cast_fp16 = slice_by_index(begin = var_5167_begin_0, end = var_5167_end_0, end_mask = var_5167_end_mask_0, squeeze_mask = var_5167_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5167_cast_fp16")]; + tensor var_5169_begin_0 = const()[name = string("op_5169_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_5169_end_0 = const()[name = string("op_5169_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_5169_end_mask_0 = const()[name = string("op_5169_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5169_squeeze_mask_0 = const()[name = string("op_5169_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5169_cast_fp16 = slice_by_index(begin = var_5169_begin_0, end = var_5169_end_0, end_mask = var_5169_end_mask_0, squeeze_mask = var_5169_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5169_cast_fp16")]; + bool var_5170_transpose_x_0 = const()[name = string("op_5170_transpose_x_0"), val = bool(false)]; + bool var_5170_transpose_y_0 = const()[name = string("op_5170_transpose_y_0"), val = bool(false)]; + tensor var_5170_cast_fp16 = matmul(transpose_x = var_5170_transpose_x_0, transpose_y = var_5170_transpose_y_0, x = var_5169_cast_fp16, y = state_273_cast_fp16)[name = string("op_5170_cast_fp16")]; + tensor v_new_273_cast_fp16 = sub(x = var_5167_cast_fp16, y = var_5170_cast_fp16)[name = string("v_new_273_cast_fp16")]; + tensor var_5173_begin_0 = const()[name = string("op_5173_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_5173_end_0 = const()[name = string("op_5173_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_5173_end_mask_0 = const()[name = string("op_5173_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5173_squeeze_mask_0 = const()[name = string("op_5173_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5173_cast_fp16 = slice_by_index(begin = var_5173_begin_0, end = var_5173_end_0, end_mask = var_5173_end_mask_0, squeeze_mask = var_5173_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5173_cast_fp16")]; + bool var_5174_transpose_x_0 = const()[name = string("op_5174_transpose_x_0"), val = bool(false)]; + bool var_5174_transpose_y_0 = const()[name = string("op_5174_transpose_y_0"), val = bool(false)]; + tensor var_5174_cast_fp16 = matmul(transpose_x = var_5174_transpose_x_0, transpose_y = var_5174_transpose_y_0, x = var_5173_cast_fp16, y = state_273_cast_fp16)[name = string("op_5174_cast_fp16")]; + tensor var_5176_begin_0 = const()[name = string("op_5176_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_5176_end_0 = const()[name = string("op_5176_end_0"), val = tensor([32, 9, 64, 64])]; + tensor var_5176_end_mask_0 = const()[name = string("op_5176_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5176_squeeze_mask_0 = const()[name = string("op_5176_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5176_cast_fp16 = slice_by_index(begin = var_5176_begin_0, end = var_5176_end_0, end_mask = var_5176_end_mask_0, squeeze_mask = var_5176_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5176_cast_fp16")]; + bool var_5177_transpose_x_0 = const()[name = string("op_5177_transpose_x_0"), val = bool(false)]; + bool var_5177_transpose_y_0 = const()[name = string("op_5177_transpose_y_0"), val = bool(false)]; + tensor var_5177_cast_fp16 = matmul(transpose_x = var_5177_transpose_x_0, transpose_y = var_5177_transpose_y_0, x = var_5176_cast_fp16, y = v_new_273_cast_fp16)[name = string("op_5177_cast_fp16")]; + tensor var_5178_cast_fp16 = add(x = var_5174_cast_fp16, y = var_5177_cast_fp16)[name = string("op_5178_cast_fp16")]; + tensor var_5180_begin_0 = const()[name = string("op_5180_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_5180_end_0 = const()[name = string("op_5180_end_0"), val = tensor([32, 9, 1, 1])]; + tensor var_5180_end_mask_0 = const()[name = string("op_5180_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5180_squeeze_mask_0 = const()[name = string("op_5180_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5180_cast_fp16 = slice_by_index(begin = var_5180_begin_0, end = var_5180_end_0, end_mask = var_5180_end_mask_0, squeeze_mask = var_5180_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5180_cast_fp16")]; + tensor var_5181_cast_fp16 = mul(x = state_273_cast_fp16, y = var_5180_cast_fp16)[name = string("op_5181_cast_fp16")]; + tensor var_5183_begin_0 = const()[name = string("op_5183_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_5183_end_0 = const()[name = string("op_5183_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_5183_end_mask_0 = const()[name = string("op_5183_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5183_squeeze_mask_0 = const()[name = string("op_5183_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5183_cast_fp16 = slice_by_index(begin = var_5183_begin_0, end = var_5183_end_0, end_mask = var_5183_end_mask_0, squeeze_mask = var_5183_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5183_cast_fp16")]; + bool var_5185_transpose_x_1 = const()[name = string("op_5185_transpose_x_1"), val = bool(true)]; + bool var_5185_transpose_y_1 = const()[name = string("op_5185_transpose_y_1"), val = bool(false)]; + tensor var_5185_cast_fp16 = matmul(transpose_x = var_5185_transpose_x_1, transpose_y = var_5185_transpose_y_1, x = var_5183_cast_fp16, y = v_new_273_cast_fp16)[name = string("op_5185_cast_fp16")]; + tensor state_275_cast_fp16 = add(x = var_5181_cast_fp16, y = var_5185_cast_fp16)[name = string("state_275_cast_fp16")]; + tensor var_5188_begin_0 = const()[name = string("op_5188_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_5188_end_0 = const()[name = string("op_5188_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_5188_end_mask_0 = const()[name = string("op_5188_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5188_squeeze_mask_0 = const()[name = string("op_5188_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5188_cast_fp16 = slice_by_index(begin = var_5188_begin_0, end = var_5188_end_0, end_mask = var_5188_end_mask_0, squeeze_mask = var_5188_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5188_cast_fp16")]; + tensor var_5190_begin_0 = const()[name = string("op_5190_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_5190_end_0 = const()[name = string("op_5190_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_5190_end_mask_0 = const()[name = string("op_5190_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5190_squeeze_mask_0 = const()[name = string("op_5190_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5190_cast_fp16 = slice_by_index(begin = var_5190_begin_0, end = var_5190_end_0, end_mask = var_5190_end_mask_0, squeeze_mask = var_5190_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5190_cast_fp16")]; + bool var_5191_transpose_x_0 = const()[name = string("op_5191_transpose_x_0"), val = bool(false)]; + bool var_5191_transpose_y_0 = const()[name = string("op_5191_transpose_y_0"), val = bool(false)]; + tensor var_5191_cast_fp16 = matmul(transpose_x = var_5191_transpose_x_0, transpose_y = var_5191_transpose_y_0, x = var_5190_cast_fp16, y = state_275_cast_fp16)[name = string("op_5191_cast_fp16")]; + tensor v_new_275_cast_fp16 = sub(x = var_5188_cast_fp16, y = var_5191_cast_fp16)[name = string("v_new_275_cast_fp16")]; + tensor var_5194_begin_0 = const()[name = string("op_5194_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_5194_end_0 = const()[name = string("op_5194_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_5194_end_mask_0 = const()[name = string("op_5194_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5194_squeeze_mask_0 = const()[name = string("op_5194_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5194_cast_fp16 = slice_by_index(begin = var_5194_begin_0, end = var_5194_end_0, end_mask = var_5194_end_mask_0, squeeze_mask = var_5194_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5194_cast_fp16")]; + bool var_5195_transpose_x_0 = const()[name = string("op_5195_transpose_x_0"), val = bool(false)]; + bool var_5195_transpose_y_0 = const()[name = string("op_5195_transpose_y_0"), val = bool(false)]; + tensor var_5195_cast_fp16 = matmul(transpose_x = var_5195_transpose_x_0, transpose_y = var_5195_transpose_y_0, x = var_5194_cast_fp16, y = state_275_cast_fp16)[name = string("op_5195_cast_fp16")]; + tensor var_5197_begin_0 = const()[name = string("op_5197_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_5197_end_0 = const()[name = string("op_5197_end_0"), val = tensor([32, 10, 64, 64])]; + tensor var_5197_end_mask_0 = const()[name = string("op_5197_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5197_squeeze_mask_0 = const()[name = string("op_5197_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5197_cast_fp16 = slice_by_index(begin = var_5197_begin_0, end = var_5197_end_0, end_mask = var_5197_end_mask_0, squeeze_mask = var_5197_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5197_cast_fp16")]; + bool var_5198_transpose_x_0 = const()[name = string("op_5198_transpose_x_0"), val = bool(false)]; + bool var_5198_transpose_y_0 = const()[name = string("op_5198_transpose_y_0"), val = bool(false)]; + tensor var_5198_cast_fp16 = matmul(transpose_x = var_5198_transpose_x_0, transpose_y = var_5198_transpose_y_0, x = var_5197_cast_fp16, y = v_new_275_cast_fp16)[name = string("op_5198_cast_fp16")]; + tensor var_5199_cast_fp16 = add(x = var_5195_cast_fp16, y = var_5198_cast_fp16)[name = string("op_5199_cast_fp16")]; + tensor var_5201_begin_0 = const()[name = string("op_5201_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_5201_end_0 = const()[name = string("op_5201_end_0"), val = tensor([32, 10, 1, 1])]; + tensor var_5201_end_mask_0 = const()[name = string("op_5201_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5201_squeeze_mask_0 = const()[name = string("op_5201_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5201_cast_fp16 = slice_by_index(begin = var_5201_begin_0, end = var_5201_end_0, end_mask = var_5201_end_mask_0, squeeze_mask = var_5201_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5201_cast_fp16")]; + tensor var_5202_cast_fp16 = mul(x = state_275_cast_fp16, y = var_5201_cast_fp16)[name = string("op_5202_cast_fp16")]; + tensor var_5204_begin_0 = const()[name = string("op_5204_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_5204_end_0 = const()[name = string("op_5204_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_5204_end_mask_0 = const()[name = string("op_5204_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5204_squeeze_mask_0 = const()[name = string("op_5204_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5204_cast_fp16 = slice_by_index(begin = var_5204_begin_0, end = var_5204_end_0, end_mask = var_5204_end_mask_0, squeeze_mask = var_5204_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5204_cast_fp16")]; + bool var_5206_transpose_x_1 = const()[name = string("op_5206_transpose_x_1"), val = bool(true)]; + bool var_5206_transpose_y_1 = const()[name = string("op_5206_transpose_y_1"), val = bool(false)]; + tensor var_5206_cast_fp16 = matmul(transpose_x = var_5206_transpose_x_1, transpose_y = var_5206_transpose_y_1, x = var_5204_cast_fp16, y = v_new_275_cast_fp16)[name = string("op_5206_cast_fp16")]; + tensor state_277_cast_fp16 = add(x = var_5202_cast_fp16, y = var_5206_cast_fp16)[name = string("state_277_cast_fp16")]; + tensor var_5209_begin_0 = const()[name = string("op_5209_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_5209_end_0 = const()[name = string("op_5209_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_5209_end_mask_0 = const()[name = string("op_5209_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5209_squeeze_mask_0 = const()[name = string("op_5209_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5209_cast_fp16 = slice_by_index(begin = var_5209_begin_0, end = var_5209_end_0, end_mask = var_5209_end_mask_0, squeeze_mask = var_5209_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5209_cast_fp16")]; + tensor var_5211_begin_0 = const()[name = string("op_5211_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_5211_end_0 = const()[name = string("op_5211_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_5211_end_mask_0 = const()[name = string("op_5211_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5211_squeeze_mask_0 = const()[name = string("op_5211_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5211_cast_fp16 = slice_by_index(begin = var_5211_begin_0, end = var_5211_end_0, end_mask = var_5211_end_mask_0, squeeze_mask = var_5211_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5211_cast_fp16")]; + bool var_5212_transpose_x_0 = const()[name = string("op_5212_transpose_x_0"), val = bool(false)]; + bool var_5212_transpose_y_0 = const()[name = string("op_5212_transpose_y_0"), val = bool(false)]; + tensor var_5212_cast_fp16 = matmul(transpose_x = var_5212_transpose_x_0, transpose_y = var_5212_transpose_y_0, x = var_5211_cast_fp16, y = state_277_cast_fp16)[name = string("op_5212_cast_fp16")]; + tensor v_new_277_cast_fp16 = sub(x = var_5209_cast_fp16, y = var_5212_cast_fp16)[name = string("v_new_277_cast_fp16")]; + tensor var_5215_begin_0 = const()[name = string("op_5215_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_5215_end_0 = const()[name = string("op_5215_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_5215_end_mask_0 = const()[name = string("op_5215_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5215_squeeze_mask_0 = const()[name = string("op_5215_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5215_cast_fp16 = slice_by_index(begin = var_5215_begin_0, end = var_5215_end_0, end_mask = var_5215_end_mask_0, squeeze_mask = var_5215_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5215_cast_fp16")]; + bool var_5216_transpose_x_0 = const()[name = string("op_5216_transpose_x_0"), val = bool(false)]; + bool var_5216_transpose_y_0 = const()[name = string("op_5216_transpose_y_0"), val = bool(false)]; + tensor var_5216_cast_fp16 = matmul(transpose_x = var_5216_transpose_x_0, transpose_y = var_5216_transpose_y_0, x = var_5215_cast_fp16, y = state_277_cast_fp16)[name = string("op_5216_cast_fp16")]; + tensor var_5218_begin_0 = const()[name = string("op_5218_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_5218_end_0 = const()[name = string("op_5218_end_0"), val = tensor([32, 11, 64, 64])]; + tensor var_5218_end_mask_0 = const()[name = string("op_5218_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5218_squeeze_mask_0 = const()[name = string("op_5218_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5218_cast_fp16 = slice_by_index(begin = var_5218_begin_0, end = var_5218_end_0, end_mask = var_5218_end_mask_0, squeeze_mask = var_5218_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5218_cast_fp16")]; + bool var_5219_transpose_x_0 = const()[name = string("op_5219_transpose_x_0"), val = bool(false)]; + bool var_5219_transpose_y_0 = const()[name = string("op_5219_transpose_y_0"), val = bool(false)]; + tensor var_5219_cast_fp16 = matmul(transpose_x = var_5219_transpose_x_0, transpose_y = var_5219_transpose_y_0, x = var_5218_cast_fp16, y = v_new_277_cast_fp16)[name = string("op_5219_cast_fp16")]; + tensor var_5220_cast_fp16 = add(x = var_5216_cast_fp16, y = var_5219_cast_fp16)[name = string("op_5220_cast_fp16")]; + tensor var_5222_begin_0 = const()[name = string("op_5222_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_5222_end_0 = const()[name = string("op_5222_end_0"), val = tensor([32, 11, 1, 1])]; + tensor var_5222_end_mask_0 = const()[name = string("op_5222_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5222_squeeze_mask_0 = const()[name = string("op_5222_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, squeeze_mask = var_5222_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5222_cast_fp16")]; + tensor var_5223_cast_fp16 = mul(x = state_277_cast_fp16, y = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")]; + tensor var_5225_begin_0 = const()[name = string("op_5225_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_5225_end_0 = const()[name = string("op_5225_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_5225_end_mask_0 = const()[name = string("op_5225_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5225_squeeze_mask_0 = const()[name = string("op_5225_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5225_cast_fp16 = slice_by_index(begin = var_5225_begin_0, end = var_5225_end_0, end_mask = var_5225_end_mask_0, squeeze_mask = var_5225_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5225_cast_fp16")]; + bool var_5227_transpose_x_1 = const()[name = string("op_5227_transpose_x_1"), val = bool(true)]; + bool var_5227_transpose_y_1 = const()[name = string("op_5227_transpose_y_1"), val = bool(false)]; + tensor var_5227_cast_fp16 = matmul(transpose_x = var_5227_transpose_x_1, transpose_y = var_5227_transpose_y_1, x = var_5225_cast_fp16, y = v_new_277_cast_fp16)[name = string("op_5227_cast_fp16")]; + tensor state_279_cast_fp16 = add(x = var_5223_cast_fp16, y = var_5227_cast_fp16)[name = string("state_279_cast_fp16")]; + tensor var_5230_begin_0 = const()[name = string("op_5230_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_5230_end_0 = const()[name = string("op_5230_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_5230_end_mask_0 = const()[name = string("op_5230_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5230_squeeze_mask_0 = const()[name = string("op_5230_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5230_cast_fp16 = slice_by_index(begin = var_5230_begin_0, end = var_5230_end_0, end_mask = var_5230_end_mask_0, squeeze_mask = var_5230_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5230_cast_fp16")]; + tensor var_5232_begin_0 = const()[name = string("op_5232_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_5232_end_0 = const()[name = string("op_5232_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_5232_end_mask_0 = const()[name = string("op_5232_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5232_squeeze_mask_0 = const()[name = string("op_5232_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5232_cast_fp16 = slice_by_index(begin = var_5232_begin_0, end = var_5232_end_0, end_mask = var_5232_end_mask_0, squeeze_mask = var_5232_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5232_cast_fp16")]; + bool var_5233_transpose_x_0 = const()[name = string("op_5233_transpose_x_0"), val = bool(false)]; + bool var_5233_transpose_y_0 = const()[name = string("op_5233_transpose_y_0"), val = bool(false)]; + tensor var_5233_cast_fp16 = matmul(transpose_x = var_5233_transpose_x_0, transpose_y = var_5233_transpose_y_0, x = var_5232_cast_fp16, y = state_279_cast_fp16)[name = string("op_5233_cast_fp16")]; + tensor v_new_279_cast_fp16 = sub(x = var_5230_cast_fp16, y = var_5233_cast_fp16)[name = string("v_new_279_cast_fp16")]; + tensor var_5236_begin_0 = const()[name = string("op_5236_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_5236_end_0 = const()[name = string("op_5236_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_5236_end_mask_0 = const()[name = string("op_5236_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5236_squeeze_mask_0 = const()[name = string("op_5236_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5236_cast_fp16 = slice_by_index(begin = var_5236_begin_0, end = var_5236_end_0, end_mask = var_5236_end_mask_0, squeeze_mask = var_5236_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5236_cast_fp16")]; + bool var_5237_transpose_x_0 = const()[name = string("op_5237_transpose_x_0"), val = bool(false)]; + bool var_5237_transpose_y_0 = const()[name = string("op_5237_transpose_y_0"), val = bool(false)]; + tensor var_5237_cast_fp16 = matmul(transpose_x = var_5237_transpose_x_0, transpose_y = var_5237_transpose_y_0, x = var_5236_cast_fp16, y = state_279_cast_fp16)[name = string("op_5237_cast_fp16")]; + tensor var_5239_begin_0 = const()[name = string("op_5239_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_5239_end_0 = const()[name = string("op_5239_end_0"), val = tensor([32, 12, 64, 64])]; + tensor var_5239_end_mask_0 = const()[name = string("op_5239_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5239_squeeze_mask_0 = const()[name = string("op_5239_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5239_cast_fp16 = slice_by_index(begin = var_5239_begin_0, end = var_5239_end_0, end_mask = var_5239_end_mask_0, squeeze_mask = var_5239_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5239_cast_fp16")]; + bool var_5240_transpose_x_0 = const()[name = string("op_5240_transpose_x_0"), val = bool(false)]; + bool var_5240_transpose_y_0 = const()[name = string("op_5240_transpose_y_0"), val = bool(false)]; + tensor var_5240_cast_fp16 = matmul(transpose_x = var_5240_transpose_x_0, transpose_y = var_5240_transpose_y_0, x = var_5239_cast_fp16, y = v_new_279_cast_fp16)[name = string("op_5240_cast_fp16")]; + tensor var_5241_cast_fp16 = add(x = var_5237_cast_fp16, y = var_5240_cast_fp16)[name = string("op_5241_cast_fp16")]; + tensor var_5243_begin_0 = const()[name = string("op_5243_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_5243_end_0 = const()[name = string("op_5243_end_0"), val = tensor([32, 12, 1, 1])]; + tensor var_5243_end_mask_0 = const()[name = string("op_5243_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5243_squeeze_mask_0 = const()[name = string("op_5243_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5243_cast_fp16 = slice_by_index(begin = var_5243_begin_0, end = var_5243_end_0, end_mask = var_5243_end_mask_0, squeeze_mask = var_5243_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5243_cast_fp16")]; + tensor var_5244_cast_fp16 = mul(x = state_279_cast_fp16, y = var_5243_cast_fp16)[name = string("op_5244_cast_fp16")]; + tensor var_5246_begin_0 = const()[name = string("op_5246_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_5246_end_0 = const()[name = string("op_5246_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_5246_end_mask_0 = const()[name = string("op_5246_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5246_squeeze_mask_0 = const()[name = string("op_5246_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5246_cast_fp16 = slice_by_index(begin = var_5246_begin_0, end = var_5246_end_0, end_mask = var_5246_end_mask_0, squeeze_mask = var_5246_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5246_cast_fp16")]; + bool var_5248_transpose_x_1 = const()[name = string("op_5248_transpose_x_1"), val = bool(true)]; + bool var_5248_transpose_y_1 = const()[name = string("op_5248_transpose_y_1"), val = bool(false)]; + tensor var_5248_cast_fp16 = matmul(transpose_x = var_5248_transpose_x_1, transpose_y = var_5248_transpose_y_1, x = var_5246_cast_fp16, y = v_new_279_cast_fp16)[name = string("op_5248_cast_fp16")]; + tensor state_281_cast_fp16 = add(x = var_5244_cast_fp16, y = var_5248_cast_fp16)[name = string("state_281_cast_fp16")]; + tensor var_5251_begin_0 = const()[name = string("op_5251_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5251_end_0 = const()[name = string("op_5251_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_5251_end_mask_0 = const()[name = string("op_5251_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5251_squeeze_mask_0 = const()[name = string("op_5251_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5251_cast_fp16 = slice_by_index(begin = var_5251_begin_0, end = var_5251_end_0, end_mask = var_5251_end_mask_0, squeeze_mask = var_5251_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5251_cast_fp16")]; + tensor var_5253_begin_0 = const()[name = string("op_5253_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5253_end_0 = const()[name = string("op_5253_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_5253_end_mask_0 = const()[name = string("op_5253_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5253_squeeze_mask_0 = const()[name = string("op_5253_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5253_cast_fp16 = slice_by_index(begin = var_5253_begin_0, end = var_5253_end_0, end_mask = var_5253_end_mask_0, squeeze_mask = var_5253_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5253_cast_fp16")]; + bool var_5254_transpose_x_0 = const()[name = string("op_5254_transpose_x_0"), val = bool(false)]; + bool var_5254_transpose_y_0 = const()[name = string("op_5254_transpose_y_0"), val = bool(false)]; + tensor var_5254_cast_fp16 = matmul(transpose_x = var_5254_transpose_x_0, transpose_y = var_5254_transpose_y_0, x = var_5253_cast_fp16, y = state_281_cast_fp16)[name = string("op_5254_cast_fp16")]; + tensor v_new_281_cast_fp16 = sub(x = var_5251_cast_fp16, y = var_5254_cast_fp16)[name = string("v_new_281_cast_fp16")]; + tensor var_5257_begin_0 = const()[name = string("op_5257_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5257_end_0 = const()[name = string("op_5257_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_5257_end_mask_0 = const()[name = string("op_5257_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5257_squeeze_mask_0 = const()[name = string("op_5257_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5257_cast_fp16 = slice_by_index(begin = var_5257_begin_0, end = var_5257_end_0, end_mask = var_5257_end_mask_0, squeeze_mask = var_5257_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5257_cast_fp16")]; + bool var_5258_transpose_x_0 = const()[name = string("op_5258_transpose_x_0"), val = bool(false)]; + bool var_5258_transpose_y_0 = const()[name = string("op_5258_transpose_y_0"), val = bool(false)]; + tensor var_5258_cast_fp16 = matmul(transpose_x = var_5258_transpose_x_0, transpose_y = var_5258_transpose_y_0, x = var_5257_cast_fp16, y = state_281_cast_fp16)[name = string("op_5258_cast_fp16")]; + tensor var_5260_begin_0 = const()[name = string("op_5260_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5260_end_0 = const()[name = string("op_5260_end_0"), val = tensor([32, 13, 64, 64])]; + tensor var_5260_end_mask_0 = const()[name = string("op_5260_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5260_squeeze_mask_0 = const()[name = string("op_5260_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5260_cast_fp16 = slice_by_index(begin = var_5260_begin_0, end = var_5260_end_0, end_mask = var_5260_end_mask_0, squeeze_mask = var_5260_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5260_cast_fp16")]; + bool var_5261_transpose_x_0 = const()[name = string("op_5261_transpose_x_0"), val = bool(false)]; + bool var_5261_transpose_y_0 = const()[name = string("op_5261_transpose_y_0"), val = bool(false)]; + tensor var_5261_cast_fp16 = matmul(transpose_x = var_5261_transpose_x_0, transpose_y = var_5261_transpose_y_0, x = var_5260_cast_fp16, y = v_new_281_cast_fp16)[name = string("op_5261_cast_fp16")]; + tensor var_5262_cast_fp16 = add(x = var_5258_cast_fp16, y = var_5261_cast_fp16)[name = string("op_5262_cast_fp16")]; + tensor var_5264_begin_0 = const()[name = string("op_5264_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5264_end_0 = const()[name = string("op_5264_end_0"), val = tensor([32, 13, 1, 1])]; + tensor var_5264_end_mask_0 = const()[name = string("op_5264_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5264_squeeze_mask_0 = const()[name = string("op_5264_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5264_cast_fp16 = slice_by_index(begin = var_5264_begin_0, end = var_5264_end_0, end_mask = var_5264_end_mask_0, squeeze_mask = var_5264_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5264_cast_fp16")]; + tensor var_5265_cast_fp16 = mul(x = state_281_cast_fp16, y = var_5264_cast_fp16)[name = string("op_5265_cast_fp16")]; + tensor var_5267_begin_0 = const()[name = string("op_5267_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5267_end_0 = const()[name = string("op_5267_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_5267_end_mask_0 = const()[name = string("op_5267_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5267_squeeze_mask_0 = const()[name = string("op_5267_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5267_cast_fp16 = slice_by_index(begin = var_5267_begin_0, end = var_5267_end_0, end_mask = var_5267_end_mask_0, squeeze_mask = var_5267_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5267_cast_fp16")]; + bool var_5269_transpose_x_1 = const()[name = string("op_5269_transpose_x_1"), val = bool(true)]; + bool var_5269_transpose_y_1 = const()[name = string("op_5269_transpose_y_1"), val = bool(false)]; + tensor var_5269_cast_fp16 = matmul(transpose_x = var_5269_transpose_x_1, transpose_y = var_5269_transpose_y_1, x = var_5267_cast_fp16, y = v_new_281_cast_fp16)[name = string("op_5269_cast_fp16")]; + tensor state_283_cast_fp16 = add(x = var_5265_cast_fp16, y = var_5269_cast_fp16)[name = string("state_283_cast_fp16")]; + tensor var_5272_begin_0 = const()[name = string("op_5272_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_5272_end_0 = const()[name = string("op_5272_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_5272_end_mask_0 = const()[name = string("op_5272_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5272_squeeze_mask_0 = const()[name = string("op_5272_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5272_cast_fp16 = slice_by_index(begin = var_5272_begin_0, end = var_5272_end_0, end_mask = var_5272_end_mask_0, squeeze_mask = var_5272_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5272_cast_fp16")]; + tensor var_5274_begin_0 = const()[name = string("op_5274_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_5274_end_0 = const()[name = string("op_5274_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_5274_end_mask_0 = const()[name = string("op_5274_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5274_squeeze_mask_0 = const()[name = string("op_5274_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, squeeze_mask = var_5274_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5274_cast_fp16")]; + bool var_5275_transpose_x_0 = const()[name = string("op_5275_transpose_x_0"), val = bool(false)]; + bool var_5275_transpose_y_0 = const()[name = string("op_5275_transpose_y_0"), val = bool(false)]; + tensor var_5275_cast_fp16 = matmul(transpose_x = var_5275_transpose_x_0, transpose_y = var_5275_transpose_y_0, x = var_5274_cast_fp16, y = state_283_cast_fp16)[name = string("op_5275_cast_fp16")]; + tensor v_new_283_cast_fp16 = sub(x = var_5272_cast_fp16, y = var_5275_cast_fp16)[name = string("v_new_283_cast_fp16")]; + tensor var_5278_begin_0 = const()[name = string("op_5278_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_5278_end_0 = const()[name = string("op_5278_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_5278_end_mask_0 = const()[name = string("op_5278_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5278_squeeze_mask_0 = const()[name = string("op_5278_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5278_cast_fp16 = slice_by_index(begin = var_5278_begin_0, end = var_5278_end_0, end_mask = var_5278_end_mask_0, squeeze_mask = var_5278_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5278_cast_fp16")]; + bool var_5279_transpose_x_0 = const()[name = string("op_5279_transpose_x_0"), val = bool(false)]; + bool var_5279_transpose_y_0 = const()[name = string("op_5279_transpose_y_0"), val = bool(false)]; + tensor var_5279_cast_fp16 = matmul(transpose_x = var_5279_transpose_x_0, transpose_y = var_5279_transpose_y_0, x = var_5278_cast_fp16, y = state_283_cast_fp16)[name = string("op_5279_cast_fp16")]; + tensor var_5281_begin_0 = const()[name = string("op_5281_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_5281_end_0 = const()[name = string("op_5281_end_0"), val = tensor([32, 14, 64, 64])]; + tensor var_5281_end_mask_0 = const()[name = string("op_5281_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5281_squeeze_mask_0 = const()[name = string("op_5281_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5281_cast_fp16 = slice_by_index(begin = var_5281_begin_0, end = var_5281_end_0, end_mask = var_5281_end_mask_0, squeeze_mask = var_5281_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5281_cast_fp16")]; + bool var_5282_transpose_x_0 = const()[name = string("op_5282_transpose_x_0"), val = bool(false)]; + bool var_5282_transpose_y_0 = const()[name = string("op_5282_transpose_y_0"), val = bool(false)]; + tensor var_5282_cast_fp16 = matmul(transpose_x = var_5282_transpose_x_0, transpose_y = var_5282_transpose_y_0, x = var_5281_cast_fp16, y = v_new_283_cast_fp16)[name = string("op_5282_cast_fp16")]; + tensor var_5283_cast_fp16 = add(x = var_5279_cast_fp16, y = var_5282_cast_fp16)[name = string("op_5283_cast_fp16")]; + tensor var_5285_begin_0 = const()[name = string("op_5285_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_5285_end_0 = const()[name = string("op_5285_end_0"), val = tensor([32, 14, 1, 1])]; + tensor var_5285_end_mask_0 = const()[name = string("op_5285_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5285_squeeze_mask_0 = const()[name = string("op_5285_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5285_cast_fp16 = slice_by_index(begin = var_5285_begin_0, end = var_5285_end_0, end_mask = var_5285_end_mask_0, squeeze_mask = var_5285_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5285_cast_fp16")]; + tensor var_5286_cast_fp16 = mul(x = state_283_cast_fp16, y = var_5285_cast_fp16)[name = string("op_5286_cast_fp16")]; + tensor var_5288_begin_0 = const()[name = string("op_5288_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_5288_end_0 = const()[name = string("op_5288_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_5288_end_mask_0 = const()[name = string("op_5288_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5288_squeeze_mask_0 = const()[name = string("op_5288_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5288_cast_fp16 = slice_by_index(begin = var_5288_begin_0, end = var_5288_end_0, end_mask = var_5288_end_mask_0, squeeze_mask = var_5288_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5288_cast_fp16")]; + bool var_5290_transpose_x_1 = const()[name = string("op_5290_transpose_x_1"), val = bool(true)]; + bool var_5290_transpose_y_1 = const()[name = string("op_5290_transpose_y_1"), val = bool(false)]; + tensor var_5290_cast_fp16 = matmul(transpose_x = var_5290_transpose_x_1, transpose_y = var_5290_transpose_y_1, x = var_5288_cast_fp16, y = v_new_283_cast_fp16)[name = string("op_5290_cast_fp16")]; + tensor state_285_cast_fp16 = add(x = var_5286_cast_fp16, y = var_5290_cast_fp16)[name = string("state_285_cast_fp16")]; + tensor var_5293_begin_0 = const()[name = string("op_5293_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5293_end_0 = const()[name = string("op_5293_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_5293_end_mask_0 = const()[name = string("op_5293_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5293_squeeze_mask_0 = const()[name = string("op_5293_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5293_cast_fp16 = slice_by_index(begin = var_5293_begin_0, end = var_5293_end_0, end_mask = var_5293_end_mask_0, squeeze_mask = var_5293_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5293_cast_fp16")]; + tensor var_5295_begin_0 = const()[name = string("op_5295_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5295_end_0 = const()[name = string("op_5295_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_5295_end_mask_0 = const()[name = string("op_5295_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5295_squeeze_mask_0 = const()[name = string("op_5295_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5295_cast_fp16 = slice_by_index(begin = var_5295_begin_0, end = var_5295_end_0, end_mask = var_5295_end_mask_0, squeeze_mask = var_5295_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5295_cast_fp16")]; + bool var_5296_transpose_x_0 = const()[name = string("op_5296_transpose_x_0"), val = bool(false)]; + bool var_5296_transpose_y_0 = const()[name = string("op_5296_transpose_y_0"), val = bool(false)]; + tensor var_5296_cast_fp16 = matmul(transpose_x = var_5296_transpose_x_0, transpose_y = var_5296_transpose_y_0, x = var_5295_cast_fp16, y = state_285_cast_fp16)[name = string("op_5296_cast_fp16")]; + tensor v_new_285_cast_fp16 = sub(x = var_5293_cast_fp16, y = var_5296_cast_fp16)[name = string("v_new_285_cast_fp16")]; + tensor var_5299_begin_0 = const()[name = string("op_5299_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5299_end_0 = const()[name = string("op_5299_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_5299_end_mask_0 = const()[name = string("op_5299_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5299_squeeze_mask_0 = const()[name = string("op_5299_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5299_cast_fp16 = slice_by_index(begin = var_5299_begin_0, end = var_5299_end_0, end_mask = var_5299_end_mask_0, squeeze_mask = var_5299_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5299_cast_fp16")]; + bool var_5300_transpose_x_0 = const()[name = string("op_5300_transpose_x_0"), val = bool(false)]; + bool var_5300_transpose_y_0 = const()[name = string("op_5300_transpose_y_0"), val = bool(false)]; + tensor var_5300_cast_fp16 = matmul(transpose_x = var_5300_transpose_x_0, transpose_y = var_5300_transpose_y_0, x = var_5299_cast_fp16, y = state_285_cast_fp16)[name = string("op_5300_cast_fp16")]; + tensor var_5302_begin_0 = const()[name = string("op_5302_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5302_end_0 = const()[name = string("op_5302_end_0"), val = tensor([32, 15, 64, 64])]; + tensor var_5302_end_mask_0 = const()[name = string("op_5302_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5302_squeeze_mask_0 = const()[name = string("op_5302_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5302_cast_fp16 = slice_by_index(begin = var_5302_begin_0, end = var_5302_end_0, end_mask = var_5302_end_mask_0, squeeze_mask = var_5302_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5302_cast_fp16")]; + bool var_5303_transpose_x_0 = const()[name = string("op_5303_transpose_x_0"), val = bool(false)]; + bool var_5303_transpose_y_0 = const()[name = string("op_5303_transpose_y_0"), val = bool(false)]; + tensor var_5303_cast_fp16 = matmul(transpose_x = var_5303_transpose_x_0, transpose_y = var_5303_transpose_y_0, x = var_5302_cast_fp16, y = v_new_285_cast_fp16)[name = string("op_5303_cast_fp16")]; + tensor var_5304_cast_fp16 = add(x = var_5300_cast_fp16, y = var_5303_cast_fp16)[name = string("op_5304_cast_fp16")]; + tensor var_5306_begin_0 = const()[name = string("op_5306_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5306_end_0 = const()[name = string("op_5306_end_0"), val = tensor([32, 15, 1, 1])]; + tensor var_5306_end_mask_0 = const()[name = string("op_5306_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5306_squeeze_mask_0 = const()[name = string("op_5306_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5306_cast_fp16 = slice_by_index(begin = var_5306_begin_0, end = var_5306_end_0, end_mask = var_5306_end_mask_0, squeeze_mask = var_5306_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5306_cast_fp16")]; + tensor var_5307_cast_fp16 = mul(x = state_285_cast_fp16, y = var_5306_cast_fp16)[name = string("op_5307_cast_fp16")]; + tensor var_5309_begin_0 = const()[name = string("op_5309_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5309_end_0 = const()[name = string("op_5309_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_5309_end_mask_0 = const()[name = string("op_5309_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5309_squeeze_mask_0 = const()[name = string("op_5309_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5309_cast_fp16 = slice_by_index(begin = var_5309_begin_0, end = var_5309_end_0, end_mask = var_5309_end_mask_0, squeeze_mask = var_5309_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5309_cast_fp16")]; + bool var_5311_transpose_x_1 = const()[name = string("op_5311_transpose_x_1"), val = bool(true)]; + bool var_5311_transpose_y_1 = const()[name = string("op_5311_transpose_y_1"), val = bool(false)]; + tensor var_5311_cast_fp16 = matmul(transpose_x = var_5311_transpose_x_1, transpose_y = var_5311_transpose_y_1, x = var_5309_cast_fp16, y = v_new_285_cast_fp16)[name = string("op_5311_cast_fp16")]; + tensor state_287_cast_fp16 = add(x = var_5307_cast_fp16, y = var_5311_cast_fp16)[name = string("state_287_cast_fp16")]; + tensor var_5314_begin_0 = const()[name = string("op_5314_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5314_end_0 = const()[name = string("op_5314_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_5314_end_mask_0 = const()[name = string("op_5314_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5314_squeeze_mask_0 = const()[name = string("op_5314_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5314_cast_fp16 = slice_by_index(begin = var_5314_begin_0, end = var_5314_end_0, end_mask = var_5314_end_mask_0, squeeze_mask = var_5314_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5314_cast_fp16")]; + tensor var_5316_begin_0 = const()[name = string("op_5316_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5316_end_0 = const()[name = string("op_5316_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_5316_end_mask_0 = const()[name = string("op_5316_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5316_squeeze_mask_0 = const()[name = string("op_5316_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5316_cast_fp16 = slice_by_index(begin = var_5316_begin_0, end = var_5316_end_0, end_mask = var_5316_end_mask_0, squeeze_mask = var_5316_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5316_cast_fp16")]; + bool var_5317_transpose_x_0 = const()[name = string("op_5317_transpose_x_0"), val = bool(false)]; + bool var_5317_transpose_y_0 = const()[name = string("op_5317_transpose_y_0"), val = bool(false)]; + tensor var_5317_cast_fp16 = matmul(transpose_x = var_5317_transpose_x_0, transpose_y = var_5317_transpose_y_0, x = var_5316_cast_fp16, y = state_287_cast_fp16)[name = string("op_5317_cast_fp16")]; + tensor v_new_287_cast_fp16 = sub(x = var_5314_cast_fp16, y = var_5317_cast_fp16)[name = string("v_new_287_cast_fp16")]; + tensor var_5320_begin_0 = const()[name = string("op_5320_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5320_end_0 = const()[name = string("op_5320_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_5320_end_mask_0 = const()[name = string("op_5320_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5320_squeeze_mask_0 = const()[name = string("op_5320_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5320_cast_fp16 = slice_by_index(begin = var_5320_begin_0, end = var_5320_end_0, end_mask = var_5320_end_mask_0, squeeze_mask = var_5320_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5320_cast_fp16")]; + bool var_5321_transpose_x_0 = const()[name = string("op_5321_transpose_x_0"), val = bool(false)]; + bool var_5321_transpose_y_0 = const()[name = string("op_5321_transpose_y_0"), val = bool(false)]; + tensor var_5321_cast_fp16 = matmul(transpose_x = var_5321_transpose_x_0, transpose_y = var_5321_transpose_y_0, x = var_5320_cast_fp16, y = state_287_cast_fp16)[name = string("op_5321_cast_fp16")]; + tensor var_5323_begin_0 = const()[name = string("op_5323_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5323_end_0 = const()[name = string("op_5323_end_0"), val = tensor([32, 16, 64, 64])]; + tensor var_5323_end_mask_0 = const()[name = string("op_5323_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5323_squeeze_mask_0 = const()[name = string("op_5323_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5323_cast_fp16 = slice_by_index(begin = var_5323_begin_0, end = var_5323_end_0, end_mask = var_5323_end_mask_0, squeeze_mask = var_5323_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5323_cast_fp16")]; + bool var_5324_transpose_x_0 = const()[name = string("op_5324_transpose_x_0"), val = bool(false)]; + bool var_5324_transpose_y_0 = const()[name = string("op_5324_transpose_y_0"), val = bool(false)]; + tensor var_5324_cast_fp16 = matmul(transpose_x = var_5324_transpose_x_0, transpose_y = var_5324_transpose_y_0, x = var_5323_cast_fp16, y = v_new_287_cast_fp16)[name = string("op_5324_cast_fp16")]; + tensor var_5325_cast_fp16 = add(x = var_5321_cast_fp16, y = var_5324_cast_fp16)[name = string("op_5325_cast_fp16")]; + tensor var_5327_begin_0 = const()[name = string("op_5327_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5327_end_0 = const()[name = string("op_5327_end_0"), val = tensor([32, 16, 1, 1])]; + tensor var_5327_end_mask_0 = const()[name = string("op_5327_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5327_squeeze_mask_0 = const()[name = string("op_5327_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5327_cast_fp16 = slice_by_index(begin = var_5327_begin_0, end = var_5327_end_0, end_mask = var_5327_end_mask_0, squeeze_mask = var_5327_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5327_cast_fp16")]; + tensor var_5328_cast_fp16 = mul(x = state_287_cast_fp16, y = var_5327_cast_fp16)[name = string("op_5328_cast_fp16")]; + tensor var_5330_begin_0 = const()[name = string("op_5330_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5330_end_0 = const()[name = string("op_5330_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_5330_end_mask_0 = const()[name = string("op_5330_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5330_squeeze_mask_0 = const()[name = string("op_5330_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5330_cast_fp16 = slice_by_index(begin = var_5330_begin_0, end = var_5330_end_0, end_mask = var_5330_end_mask_0, squeeze_mask = var_5330_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5330_cast_fp16")]; + bool var_5332_transpose_x_1 = const()[name = string("op_5332_transpose_x_1"), val = bool(true)]; + bool var_5332_transpose_y_1 = const()[name = string("op_5332_transpose_y_1"), val = bool(false)]; + tensor var_5332_cast_fp16 = matmul(transpose_x = var_5332_transpose_x_1, transpose_y = var_5332_transpose_y_1, x = var_5330_cast_fp16, y = v_new_287_cast_fp16)[name = string("op_5332_cast_fp16")]; + tensor state_289_cast_fp16 = add(x = var_5328_cast_fp16, y = var_5332_cast_fp16)[name = string("state_289_cast_fp16")]; + tensor var_5335_begin_0 = const()[name = string("op_5335_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_5335_end_0 = const()[name = string("op_5335_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_5335_end_mask_0 = const()[name = string("op_5335_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5335_squeeze_mask_0 = const()[name = string("op_5335_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5335_cast_fp16 = slice_by_index(begin = var_5335_begin_0, end = var_5335_end_0, end_mask = var_5335_end_mask_0, squeeze_mask = var_5335_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5335_cast_fp16")]; + tensor var_5337_begin_0 = const()[name = string("op_5337_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_5337_end_0 = const()[name = string("op_5337_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_5337_end_mask_0 = const()[name = string("op_5337_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5337_squeeze_mask_0 = const()[name = string("op_5337_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5337_cast_fp16 = slice_by_index(begin = var_5337_begin_0, end = var_5337_end_0, end_mask = var_5337_end_mask_0, squeeze_mask = var_5337_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5337_cast_fp16")]; + bool var_5338_transpose_x_0 = const()[name = string("op_5338_transpose_x_0"), val = bool(false)]; + bool var_5338_transpose_y_0 = const()[name = string("op_5338_transpose_y_0"), val = bool(false)]; + tensor var_5338_cast_fp16 = matmul(transpose_x = var_5338_transpose_x_0, transpose_y = var_5338_transpose_y_0, x = var_5337_cast_fp16, y = state_289_cast_fp16)[name = string("op_5338_cast_fp16")]; + tensor v_new_289_cast_fp16 = sub(x = var_5335_cast_fp16, y = var_5338_cast_fp16)[name = string("v_new_289_cast_fp16")]; + tensor var_5341_begin_0 = const()[name = string("op_5341_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_5341_end_0 = const()[name = string("op_5341_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_5341_end_mask_0 = const()[name = string("op_5341_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5341_squeeze_mask_0 = const()[name = string("op_5341_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5341_cast_fp16 = slice_by_index(begin = var_5341_begin_0, end = var_5341_end_0, end_mask = var_5341_end_mask_0, squeeze_mask = var_5341_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5341_cast_fp16")]; + bool var_5342_transpose_x_0 = const()[name = string("op_5342_transpose_x_0"), val = bool(false)]; + bool var_5342_transpose_y_0 = const()[name = string("op_5342_transpose_y_0"), val = bool(false)]; + tensor var_5342_cast_fp16 = matmul(transpose_x = var_5342_transpose_x_0, transpose_y = var_5342_transpose_y_0, x = var_5341_cast_fp16, y = state_289_cast_fp16)[name = string("op_5342_cast_fp16")]; + tensor var_5344_begin_0 = const()[name = string("op_5344_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_5344_end_0 = const()[name = string("op_5344_end_0"), val = tensor([32, 17, 64, 64])]; + tensor var_5344_end_mask_0 = const()[name = string("op_5344_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5344_squeeze_mask_0 = const()[name = string("op_5344_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5344_cast_fp16 = slice_by_index(begin = var_5344_begin_0, end = var_5344_end_0, end_mask = var_5344_end_mask_0, squeeze_mask = var_5344_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5344_cast_fp16")]; + bool var_5345_transpose_x_0 = const()[name = string("op_5345_transpose_x_0"), val = bool(false)]; + bool var_5345_transpose_y_0 = const()[name = string("op_5345_transpose_y_0"), val = bool(false)]; + tensor var_5345_cast_fp16 = matmul(transpose_x = var_5345_transpose_x_0, transpose_y = var_5345_transpose_y_0, x = var_5344_cast_fp16, y = v_new_289_cast_fp16)[name = string("op_5345_cast_fp16")]; + tensor var_5346_cast_fp16 = add(x = var_5342_cast_fp16, y = var_5345_cast_fp16)[name = string("op_5346_cast_fp16")]; + tensor var_5348_begin_0 = const()[name = string("op_5348_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_5348_end_0 = const()[name = string("op_5348_end_0"), val = tensor([32, 17, 1, 1])]; + tensor var_5348_end_mask_0 = const()[name = string("op_5348_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5348_squeeze_mask_0 = const()[name = string("op_5348_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5348_cast_fp16 = slice_by_index(begin = var_5348_begin_0, end = var_5348_end_0, end_mask = var_5348_end_mask_0, squeeze_mask = var_5348_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5348_cast_fp16")]; + tensor var_5349_cast_fp16 = mul(x = state_289_cast_fp16, y = var_5348_cast_fp16)[name = string("op_5349_cast_fp16")]; + tensor var_5351_begin_0 = const()[name = string("op_5351_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_5351_end_0 = const()[name = string("op_5351_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_5351_end_mask_0 = const()[name = string("op_5351_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5351_squeeze_mask_0 = const()[name = string("op_5351_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5351_cast_fp16 = slice_by_index(begin = var_5351_begin_0, end = var_5351_end_0, end_mask = var_5351_end_mask_0, squeeze_mask = var_5351_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5351_cast_fp16")]; + bool var_5353_transpose_x_1 = const()[name = string("op_5353_transpose_x_1"), val = bool(true)]; + bool var_5353_transpose_y_1 = const()[name = string("op_5353_transpose_y_1"), val = bool(false)]; + tensor var_5353_cast_fp16 = matmul(transpose_x = var_5353_transpose_x_1, transpose_y = var_5353_transpose_y_1, x = var_5351_cast_fp16, y = v_new_289_cast_fp16)[name = string("op_5353_cast_fp16")]; + tensor state_291_cast_fp16 = add(x = var_5349_cast_fp16, y = var_5353_cast_fp16)[name = string("state_291_cast_fp16")]; + tensor var_5356_begin_0 = const()[name = string("op_5356_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_5356_end_0 = const()[name = string("op_5356_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_5356_end_mask_0 = const()[name = string("op_5356_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5356_squeeze_mask_0 = const()[name = string("op_5356_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5356_cast_fp16 = slice_by_index(begin = var_5356_begin_0, end = var_5356_end_0, end_mask = var_5356_end_mask_0, squeeze_mask = var_5356_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5356_cast_fp16")]; + tensor var_5358_begin_0 = const()[name = string("op_5358_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_5358_end_0 = const()[name = string("op_5358_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_5358_end_mask_0 = const()[name = string("op_5358_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5358_squeeze_mask_0 = const()[name = string("op_5358_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5358_cast_fp16 = slice_by_index(begin = var_5358_begin_0, end = var_5358_end_0, end_mask = var_5358_end_mask_0, squeeze_mask = var_5358_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5358_cast_fp16")]; + bool var_5359_transpose_x_0 = const()[name = string("op_5359_transpose_x_0"), val = bool(false)]; + bool var_5359_transpose_y_0 = const()[name = string("op_5359_transpose_y_0"), val = bool(false)]; + tensor var_5359_cast_fp16 = matmul(transpose_x = var_5359_transpose_x_0, transpose_y = var_5359_transpose_y_0, x = var_5358_cast_fp16, y = state_291_cast_fp16)[name = string("op_5359_cast_fp16")]; + tensor v_new_291_cast_fp16 = sub(x = var_5356_cast_fp16, y = var_5359_cast_fp16)[name = string("v_new_291_cast_fp16")]; + tensor var_5362_begin_0 = const()[name = string("op_5362_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_5362_end_0 = const()[name = string("op_5362_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_5362_end_mask_0 = const()[name = string("op_5362_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5362_squeeze_mask_0 = const()[name = string("op_5362_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5362_cast_fp16 = slice_by_index(begin = var_5362_begin_0, end = var_5362_end_0, end_mask = var_5362_end_mask_0, squeeze_mask = var_5362_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5362_cast_fp16")]; + bool var_5363_transpose_x_0 = const()[name = string("op_5363_transpose_x_0"), val = bool(false)]; + bool var_5363_transpose_y_0 = const()[name = string("op_5363_transpose_y_0"), val = bool(false)]; + tensor var_5363_cast_fp16 = matmul(transpose_x = var_5363_transpose_x_0, transpose_y = var_5363_transpose_y_0, x = var_5362_cast_fp16, y = state_291_cast_fp16)[name = string("op_5363_cast_fp16")]; + tensor var_5365_begin_0 = const()[name = string("op_5365_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_5365_end_0 = const()[name = string("op_5365_end_0"), val = tensor([32, 18, 64, 64])]; + tensor var_5365_end_mask_0 = const()[name = string("op_5365_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5365_squeeze_mask_0 = const()[name = string("op_5365_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5365_cast_fp16 = slice_by_index(begin = var_5365_begin_0, end = var_5365_end_0, end_mask = var_5365_end_mask_0, squeeze_mask = var_5365_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5365_cast_fp16")]; + bool var_5366_transpose_x_0 = const()[name = string("op_5366_transpose_x_0"), val = bool(false)]; + bool var_5366_transpose_y_0 = const()[name = string("op_5366_transpose_y_0"), val = bool(false)]; + tensor var_5366_cast_fp16 = matmul(transpose_x = var_5366_transpose_x_0, transpose_y = var_5366_transpose_y_0, x = var_5365_cast_fp16, y = v_new_291_cast_fp16)[name = string("op_5366_cast_fp16")]; + tensor var_5367_cast_fp16 = add(x = var_5363_cast_fp16, y = var_5366_cast_fp16)[name = string("op_5367_cast_fp16")]; + tensor var_5369_begin_0 = const()[name = string("op_5369_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_5369_end_0 = const()[name = string("op_5369_end_0"), val = tensor([32, 18, 1, 1])]; + tensor var_5369_end_mask_0 = const()[name = string("op_5369_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5369_squeeze_mask_0 = const()[name = string("op_5369_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5369_cast_fp16 = slice_by_index(begin = var_5369_begin_0, end = var_5369_end_0, end_mask = var_5369_end_mask_0, squeeze_mask = var_5369_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5369_cast_fp16")]; + tensor var_5370_cast_fp16 = mul(x = state_291_cast_fp16, y = var_5369_cast_fp16)[name = string("op_5370_cast_fp16")]; + tensor var_5372_begin_0 = const()[name = string("op_5372_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_5372_end_0 = const()[name = string("op_5372_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_5372_end_mask_0 = const()[name = string("op_5372_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5372_squeeze_mask_0 = const()[name = string("op_5372_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5372_cast_fp16 = slice_by_index(begin = var_5372_begin_0, end = var_5372_end_0, end_mask = var_5372_end_mask_0, squeeze_mask = var_5372_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5372_cast_fp16")]; + bool var_5374_transpose_x_1 = const()[name = string("op_5374_transpose_x_1"), val = bool(true)]; + bool var_5374_transpose_y_1 = const()[name = string("op_5374_transpose_y_1"), val = bool(false)]; + tensor var_5374_cast_fp16 = matmul(transpose_x = var_5374_transpose_x_1, transpose_y = var_5374_transpose_y_1, x = var_5372_cast_fp16, y = v_new_291_cast_fp16)[name = string("op_5374_cast_fp16")]; + tensor state_293_cast_fp16 = add(x = var_5370_cast_fp16, y = var_5374_cast_fp16)[name = string("state_293_cast_fp16")]; + tensor var_5377_begin_0 = const()[name = string("op_5377_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_5377_end_0 = const()[name = string("op_5377_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_5377_end_mask_0 = const()[name = string("op_5377_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5377_squeeze_mask_0 = const()[name = string("op_5377_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5377_cast_fp16 = slice_by_index(begin = var_5377_begin_0, end = var_5377_end_0, end_mask = var_5377_end_mask_0, squeeze_mask = var_5377_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5377_cast_fp16")]; + tensor var_5379_begin_0 = const()[name = string("op_5379_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_5379_end_0 = const()[name = string("op_5379_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_5379_end_mask_0 = const()[name = string("op_5379_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5379_squeeze_mask_0 = const()[name = string("op_5379_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5379_cast_fp16 = slice_by_index(begin = var_5379_begin_0, end = var_5379_end_0, end_mask = var_5379_end_mask_0, squeeze_mask = var_5379_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5379_cast_fp16")]; + bool var_5380_transpose_x_0 = const()[name = string("op_5380_transpose_x_0"), val = bool(false)]; + bool var_5380_transpose_y_0 = const()[name = string("op_5380_transpose_y_0"), val = bool(false)]; + tensor var_5380_cast_fp16 = matmul(transpose_x = var_5380_transpose_x_0, transpose_y = var_5380_transpose_y_0, x = var_5379_cast_fp16, y = state_293_cast_fp16)[name = string("op_5380_cast_fp16")]; + tensor v_new_293_cast_fp16 = sub(x = var_5377_cast_fp16, y = var_5380_cast_fp16)[name = string("v_new_293_cast_fp16")]; + tensor var_5383_begin_0 = const()[name = string("op_5383_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_5383_end_0 = const()[name = string("op_5383_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_5383_end_mask_0 = const()[name = string("op_5383_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5383_squeeze_mask_0 = const()[name = string("op_5383_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5383_cast_fp16 = slice_by_index(begin = var_5383_begin_0, end = var_5383_end_0, end_mask = var_5383_end_mask_0, squeeze_mask = var_5383_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5383_cast_fp16")]; + bool var_5384_transpose_x_0 = const()[name = string("op_5384_transpose_x_0"), val = bool(false)]; + bool var_5384_transpose_y_0 = const()[name = string("op_5384_transpose_y_0"), val = bool(false)]; + tensor var_5384_cast_fp16 = matmul(transpose_x = var_5384_transpose_x_0, transpose_y = var_5384_transpose_y_0, x = var_5383_cast_fp16, y = state_293_cast_fp16)[name = string("op_5384_cast_fp16")]; + tensor var_5386_begin_0 = const()[name = string("op_5386_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_5386_end_0 = const()[name = string("op_5386_end_0"), val = tensor([32, 19, 64, 64])]; + tensor var_5386_end_mask_0 = const()[name = string("op_5386_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5386_squeeze_mask_0 = const()[name = string("op_5386_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5386_cast_fp16 = slice_by_index(begin = var_5386_begin_0, end = var_5386_end_0, end_mask = var_5386_end_mask_0, squeeze_mask = var_5386_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5386_cast_fp16")]; + bool var_5387_transpose_x_0 = const()[name = string("op_5387_transpose_x_0"), val = bool(false)]; + bool var_5387_transpose_y_0 = const()[name = string("op_5387_transpose_y_0"), val = bool(false)]; + tensor var_5387_cast_fp16 = matmul(transpose_x = var_5387_transpose_x_0, transpose_y = var_5387_transpose_y_0, x = var_5386_cast_fp16, y = v_new_293_cast_fp16)[name = string("op_5387_cast_fp16")]; + tensor var_5388_cast_fp16 = add(x = var_5384_cast_fp16, y = var_5387_cast_fp16)[name = string("op_5388_cast_fp16")]; + tensor var_5390_begin_0 = const()[name = string("op_5390_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_5390_end_0 = const()[name = string("op_5390_end_0"), val = tensor([32, 19, 1, 1])]; + tensor var_5390_end_mask_0 = const()[name = string("op_5390_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5390_squeeze_mask_0 = const()[name = string("op_5390_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5390_cast_fp16 = slice_by_index(begin = var_5390_begin_0, end = var_5390_end_0, end_mask = var_5390_end_mask_0, squeeze_mask = var_5390_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5390_cast_fp16")]; + tensor var_5391_cast_fp16 = mul(x = state_293_cast_fp16, y = var_5390_cast_fp16)[name = string("op_5391_cast_fp16")]; + tensor var_5393_begin_0 = const()[name = string("op_5393_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_5393_end_0 = const()[name = string("op_5393_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_5393_end_mask_0 = const()[name = string("op_5393_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5393_squeeze_mask_0 = const()[name = string("op_5393_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5393_cast_fp16 = slice_by_index(begin = var_5393_begin_0, end = var_5393_end_0, end_mask = var_5393_end_mask_0, squeeze_mask = var_5393_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5393_cast_fp16")]; + bool var_5395_transpose_x_1 = const()[name = string("op_5395_transpose_x_1"), val = bool(true)]; + bool var_5395_transpose_y_1 = const()[name = string("op_5395_transpose_y_1"), val = bool(false)]; + tensor var_5395_cast_fp16 = matmul(transpose_x = var_5395_transpose_x_1, transpose_y = var_5395_transpose_y_1, x = var_5393_cast_fp16, y = v_new_293_cast_fp16)[name = string("op_5395_cast_fp16")]; + tensor state_295_cast_fp16 = add(x = var_5391_cast_fp16, y = var_5395_cast_fp16)[name = string("state_295_cast_fp16")]; + tensor var_5398_begin_0 = const()[name = string("op_5398_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_5398_end_0 = const()[name = string("op_5398_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_5398_end_mask_0 = const()[name = string("op_5398_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5398_squeeze_mask_0 = const()[name = string("op_5398_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5398_cast_fp16 = slice_by_index(begin = var_5398_begin_0, end = var_5398_end_0, end_mask = var_5398_end_mask_0, squeeze_mask = var_5398_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5398_cast_fp16")]; + tensor var_5400_begin_0 = const()[name = string("op_5400_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_5400_end_0 = const()[name = string("op_5400_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_5400_end_mask_0 = const()[name = string("op_5400_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5400_squeeze_mask_0 = const()[name = string("op_5400_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5400_cast_fp16 = slice_by_index(begin = var_5400_begin_0, end = var_5400_end_0, end_mask = var_5400_end_mask_0, squeeze_mask = var_5400_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5400_cast_fp16")]; + bool var_5401_transpose_x_0 = const()[name = string("op_5401_transpose_x_0"), val = bool(false)]; + bool var_5401_transpose_y_0 = const()[name = string("op_5401_transpose_y_0"), val = bool(false)]; + tensor var_5401_cast_fp16 = matmul(transpose_x = var_5401_transpose_x_0, transpose_y = var_5401_transpose_y_0, x = var_5400_cast_fp16, y = state_295_cast_fp16)[name = string("op_5401_cast_fp16")]; + tensor v_new_295_cast_fp16 = sub(x = var_5398_cast_fp16, y = var_5401_cast_fp16)[name = string("v_new_295_cast_fp16")]; + tensor var_5404_begin_0 = const()[name = string("op_5404_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_5404_end_0 = const()[name = string("op_5404_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_5404_end_mask_0 = const()[name = string("op_5404_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5404_squeeze_mask_0 = const()[name = string("op_5404_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5404_cast_fp16 = slice_by_index(begin = var_5404_begin_0, end = var_5404_end_0, end_mask = var_5404_end_mask_0, squeeze_mask = var_5404_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5404_cast_fp16")]; + bool var_5405_transpose_x_0 = const()[name = string("op_5405_transpose_x_0"), val = bool(false)]; + bool var_5405_transpose_y_0 = const()[name = string("op_5405_transpose_y_0"), val = bool(false)]; + tensor var_5405_cast_fp16 = matmul(transpose_x = var_5405_transpose_x_0, transpose_y = var_5405_transpose_y_0, x = var_5404_cast_fp16, y = state_295_cast_fp16)[name = string("op_5405_cast_fp16")]; + tensor var_5407_begin_0 = const()[name = string("op_5407_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_5407_end_0 = const()[name = string("op_5407_end_0"), val = tensor([32, 20, 64, 64])]; + tensor var_5407_end_mask_0 = const()[name = string("op_5407_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5407_squeeze_mask_0 = const()[name = string("op_5407_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5407_cast_fp16 = slice_by_index(begin = var_5407_begin_0, end = var_5407_end_0, end_mask = var_5407_end_mask_0, squeeze_mask = var_5407_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5407_cast_fp16")]; + bool var_5408_transpose_x_0 = const()[name = string("op_5408_transpose_x_0"), val = bool(false)]; + bool var_5408_transpose_y_0 = const()[name = string("op_5408_transpose_y_0"), val = bool(false)]; + tensor var_5408_cast_fp16 = matmul(transpose_x = var_5408_transpose_x_0, transpose_y = var_5408_transpose_y_0, x = var_5407_cast_fp16, y = v_new_295_cast_fp16)[name = string("op_5408_cast_fp16")]; + tensor var_5409_cast_fp16 = add(x = var_5405_cast_fp16, y = var_5408_cast_fp16)[name = string("op_5409_cast_fp16")]; + tensor var_5411_begin_0 = const()[name = string("op_5411_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_5411_end_0 = const()[name = string("op_5411_end_0"), val = tensor([32, 20, 1, 1])]; + tensor var_5411_end_mask_0 = const()[name = string("op_5411_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5411_squeeze_mask_0 = const()[name = string("op_5411_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5411_cast_fp16 = slice_by_index(begin = var_5411_begin_0, end = var_5411_end_0, end_mask = var_5411_end_mask_0, squeeze_mask = var_5411_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5411_cast_fp16")]; + tensor var_5412_cast_fp16 = mul(x = state_295_cast_fp16, y = var_5411_cast_fp16)[name = string("op_5412_cast_fp16")]; + tensor var_5414_begin_0 = const()[name = string("op_5414_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_5414_end_0 = const()[name = string("op_5414_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_5414_end_mask_0 = const()[name = string("op_5414_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5414_squeeze_mask_0 = const()[name = string("op_5414_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5414_cast_fp16 = slice_by_index(begin = var_5414_begin_0, end = var_5414_end_0, end_mask = var_5414_end_mask_0, squeeze_mask = var_5414_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5414_cast_fp16")]; + bool var_5416_transpose_x_1 = const()[name = string("op_5416_transpose_x_1"), val = bool(true)]; + bool var_5416_transpose_y_1 = const()[name = string("op_5416_transpose_y_1"), val = bool(false)]; + tensor var_5416_cast_fp16 = matmul(transpose_x = var_5416_transpose_x_1, transpose_y = var_5416_transpose_y_1, x = var_5414_cast_fp16, y = v_new_295_cast_fp16)[name = string("op_5416_cast_fp16")]; + tensor state_297_cast_fp16 = add(x = var_5412_cast_fp16, y = var_5416_cast_fp16)[name = string("state_297_cast_fp16")]; + tensor var_5419_begin_0 = const()[name = string("op_5419_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_5419_end_0 = const()[name = string("op_5419_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_5419_end_mask_0 = const()[name = string("op_5419_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5419_squeeze_mask_0 = const()[name = string("op_5419_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5419_cast_fp16 = slice_by_index(begin = var_5419_begin_0, end = var_5419_end_0, end_mask = var_5419_end_mask_0, squeeze_mask = var_5419_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5419_cast_fp16")]; + tensor var_5421_begin_0 = const()[name = string("op_5421_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_5421_end_0 = const()[name = string("op_5421_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_5421_end_mask_0 = const()[name = string("op_5421_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5421_squeeze_mask_0 = const()[name = string("op_5421_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5421_cast_fp16 = slice_by_index(begin = var_5421_begin_0, end = var_5421_end_0, end_mask = var_5421_end_mask_0, squeeze_mask = var_5421_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5421_cast_fp16")]; + bool var_5422_transpose_x_0 = const()[name = string("op_5422_transpose_x_0"), val = bool(false)]; + bool var_5422_transpose_y_0 = const()[name = string("op_5422_transpose_y_0"), val = bool(false)]; + tensor var_5422_cast_fp16 = matmul(transpose_x = var_5422_transpose_x_0, transpose_y = var_5422_transpose_y_0, x = var_5421_cast_fp16, y = state_297_cast_fp16)[name = string("op_5422_cast_fp16")]; + tensor v_new_297_cast_fp16 = sub(x = var_5419_cast_fp16, y = var_5422_cast_fp16)[name = string("v_new_297_cast_fp16")]; + tensor var_5425_begin_0 = const()[name = string("op_5425_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_5425_end_0 = const()[name = string("op_5425_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_5425_end_mask_0 = const()[name = string("op_5425_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5425_squeeze_mask_0 = const()[name = string("op_5425_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5425_cast_fp16 = slice_by_index(begin = var_5425_begin_0, end = var_5425_end_0, end_mask = var_5425_end_mask_0, squeeze_mask = var_5425_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5425_cast_fp16")]; + bool var_5426_transpose_x_0 = const()[name = string("op_5426_transpose_x_0"), val = bool(false)]; + bool var_5426_transpose_y_0 = const()[name = string("op_5426_transpose_y_0"), val = bool(false)]; + tensor var_5426_cast_fp16 = matmul(transpose_x = var_5426_transpose_x_0, transpose_y = var_5426_transpose_y_0, x = var_5425_cast_fp16, y = state_297_cast_fp16)[name = string("op_5426_cast_fp16")]; + tensor var_5428_begin_0 = const()[name = string("op_5428_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_5428_end_0 = const()[name = string("op_5428_end_0"), val = tensor([32, 21, 64, 64])]; + tensor var_5428_end_mask_0 = const()[name = string("op_5428_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5428_squeeze_mask_0 = const()[name = string("op_5428_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5428_cast_fp16 = slice_by_index(begin = var_5428_begin_0, end = var_5428_end_0, end_mask = var_5428_end_mask_0, squeeze_mask = var_5428_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5428_cast_fp16")]; + bool var_5429_transpose_x_0 = const()[name = string("op_5429_transpose_x_0"), val = bool(false)]; + bool var_5429_transpose_y_0 = const()[name = string("op_5429_transpose_y_0"), val = bool(false)]; + tensor var_5429_cast_fp16 = matmul(transpose_x = var_5429_transpose_x_0, transpose_y = var_5429_transpose_y_0, x = var_5428_cast_fp16, y = v_new_297_cast_fp16)[name = string("op_5429_cast_fp16")]; + tensor var_5430_cast_fp16 = add(x = var_5426_cast_fp16, y = var_5429_cast_fp16)[name = string("op_5430_cast_fp16")]; + tensor var_5432_begin_0 = const()[name = string("op_5432_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_5432_end_0 = const()[name = string("op_5432_end_0"), val = tensor([32, 21, 1, 1])]; + tensor var_5432_end_mask_0 = const()[name = string("op_5432_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5432_squeeze_mask_0 = const()[name = string("op_5432_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5432_cast_fp16 = slice_by_index(begin = var_5432_begin_0, end = var_5432_end_0, end_mask = var_5432_end_mask_0, squeeze_mask = var_5432_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5432_cast_fp16")]; + tensor var_5433_cast_fp16 = mul(x = state_297_cast_fp16, y = var_5432_cast_fp16)[name = string("op_5433_cast_fp16")]; + tensor var_5435_begin_0 = const()[name = string("op_5435_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_5435_end_0 = const()[name = string("op_5435_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_5435_end_mask_0 = const()[name = string("op_5435_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5435_squeeze_mask_0 = const()[name = string("op_5435_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5435_cast_fp16 = slice_by_index(begin = var_5435_begin_0, end = var_5435_end_0, end_mask = var_5435_end_mask_0, squeeze_mask = var_5435_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5435_cast_fp16")]; + bool var_5437_transpose_x_1 = const()[name = string("op_5437_transpose_x_1"), val = bool(true)]; + bool var_5437_transpose_y_1 = const()[name = string("op_5437_transpose_y_1"), val = bool(false)]; + tensor var_5437_cast_fp16 = matmul(transpose_x = var_5437_transpose_x_1, transpose_y = var_5437_transpose_y_1, x = var_5435_cast_fp16, y = v_new_297_cast_fp16)[name = string("op_5437_cast_fp16")]; + tensor state_299_cast_fp16 = add(x = var_5433_cast_fp16, y = var_5437_cast_fp16)[name = string("state_299_cast_fp16")]; + tensor var_5440_begin_0 = const()[name = string("op_5440_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_5440_end_0 = const()[name = string("op_5440_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_5440_end_mask_0 = const()[name = string("op_5440_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5440_squeeze_mask_0 = const()[name = string("op_5440_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5440_cast_fp16 = slice_by_index(begin = var_5440_begin_0, end = var_5440_end_0, end_mask = var_5440_end_mask_0, squeeze_mask = var_5440_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5440_cast_fp16")]; + tensor var_5442_begin_0 = const()[name = string("op_5442_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_5442_end_0 = const()[name = string("op_5442_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_5442_end_mask_0 = const()[name = string("op_5442_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5442_squeeze_mask_0 = const()[name = string("op_5442_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5442_cast_fp16 = slice_by_index(begin = var_5442_begin_0, end = var_5442_end_0, end_mask = var_5442_end_mask_0, squeeze_mask = var_5442_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5442_cast_fp16")]; + bool var_5443_transpose_x_0 = const()[name = string("op_5443_transpose_x_0"), val = bool(false)]; + bool var_5443_transpose_y_0 = const()[name = string("op_5443_transpose_y_0"), val = bool(false)]; + tensor var_5443_cast_fp16 = matmul(transpose_x = var_5443_transpose_x_0, transpose_y = var_5443_transpose_y_0, x = var_5442_cast_fp16, y = state_299_cast_fp16)[name = string("op_5443_cast_fp16")]; + tensor v_new_299_cast_fp16 = sub(x = var_5440_cast_fp16, y = var_5443_cast_fp16)[name = string("v_new_299_cast_fp16")]; + tensor var_5446_begin_0 = const()[name = string("op_5446_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_5446_end_0 = const()[name = string("op_5446_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_5446_end_mask_0 = const()[name = string("op_5446_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5446_squeeze_mask_0 = const()[name = string("op_5446_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5446_cast_fp16 = slice_by_index(begin = var_5446_begin_0, end = var_5446_end_0, end_mask = var_5446_end_mask_0, squeeze_mask = var_5446_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5446_cast_fp16")]; + bool var_5447_transpose_x_0 = const()[name = string("op_5447_transpose_x_0"), val = bool(false)]; + bool var_5447_transpose_y_0 = const()[name = string("op_5447_transpose_y_0"), val = bool(false)]; + tensor var_5447_cast_fp16 = matmul(transpose_x = var_5447_transpose_x_0, transpose_y = var_5447_transpose_y_0, x = var_5446_cast_fp16, y = state_299_cast_fp16)[name = string("op_5447_cast_fp16")]; + tensor var_5449_begin_0 = const()[name = string("op_5449_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_5449_end_0 = const()[name = string("op_5449_end_0"), val = tensor([32, 22, 64, 64])]; + tensor var_5449_end_mask_0 = const()[name = string("op_5449_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5449_squeeze_mask_0 = const()[name = string("op_5449_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5449_cast_fp16 = slice_by_index(begin = var_5449_begin_0, end = var_5449_end_0, end_mask = var_5449_end_mask_0, squeeze_mask = var_5449_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5449_cast_fp16")]; + bool var_5450_transpose_x_0 = const()[name = string("op_5450_transpose_x_0"), val = bool(false)]; + bool var_5450_transpose_y_0 = const()[name = string("op_5450_transpose_y_0"), val = bool(false)]; + tensor var_5450_cast_fp16 = matmul(transpose_x = var_5450_transpose_x_0, transpose_y = var_5450_transpose_y_0, x = var_5449_cast_fp16, y = v_new_299_cast_fp16)[name = string("op_5450_cast_fp16")]; + tensor var_5451_cast_fp16 = add(x = var_5447_cast_fp16, y = var_5450_cast_fp16)[name = string("op_5451_cast_fp16")]; + tensor var_5453_begin_0 = const()[name = string("op_5453_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_5453_end_0 = const()[name = string("op_5453_end_0"), val = tensor([32, 22, 1, 1])]; + tensor var_5453_end_mask_0 = const()[name = string("op_5453_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5453_squeeze_mask_0 = const()[name = string("op_5453_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5453_cast_fp16 = slice_by_index(begin = var_5453_begin_0, end = var_5453_end_0, end_mask = var_5453_end_mask_0, squeeze_mask = var_5453_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5453_cast_fp16")]; + tensor var_5454_cast_fp16 = mul(x = state_299_cast_fp16, y = var_5453_cast_fp16)[name = string("op_5454_cast_fp16")]; + tensor var_5456_begin_0 = const()[name = string("op_5456_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_5456_end_0 = const()[name = string("op_5456_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_5456_end_mask_0 = const()[name = string("op_5456_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5456_squeeze_mask_0 = const()[name = string("op_5456_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5456_cast_fp16 = slice_by_index(begin = var_5456_begin_0, end = var_5456_end_0, end_mask = var_5456_end_mask_0, squeeze_mask = var_5456_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5456_cast_fp16")]; + bool var_5458_transpose_x_1 = const()[name = string("op_5458_transpose_x_1"), val = bool(true)]; + bool var_5458_transpose_y_1 = const()[name = string("op_5458_transpose_y_1"), val = bool(false)]; + tensor var_5458_cast_fp16 = matmul(transpose_x = var_5458_transpose_x_1, transpose_y = var_5458_transpose_y_1, x = var_5456_cast_fp16, y = v_new_299_cast_fp16)[name = string("op_5458_cast_fp16")]; + tensor state_301_cast_fp16 = add(x = var_5454_cast_fp16, y = var_5458_cast_fp16)[name = string("state_301_cast_fp16")]; + tensor var_5461_begin_0 = const()[name = string("op_5461_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_5461_end_0 = const()[name = string("op_5461_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_5461_end_mask_0 = const()[name = string("op_5461_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5461_squeeze_mask_0 = const()[name = string("op_5461_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5461_cast_fp16 = slice_by_index(begin = var_5461_begin_0, end = var_5461_end_0, end_mask = var_5461_end_mask_0, squeeze_mask = var_5461_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5461_cast_fp16")]; + tensor var_5463_begin_0 = const()[name = string("op_5463_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_5463_end_0 = const()[name = string("op_5463_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_5463_end_mask_0 = const()[name = string("op_5463_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5463_squeeze_mask_0 = const()[name = string("op_5463_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5463_cast_fp16 = slice_by_index(begin = var_5463_begin_0, end = var_5463_end_0, end_mask = var_5463_end_mask_0, squeeze_mask = var_5463_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5463_cast_fp16")]; + bool var_5464_transpose_x_0 = const()[name = string("op_5464_transpose_x_0"), val = bool(false)]; + bool var_5464_transpose_y_0 = const()[name = string("op_5464_transpose_y_0"), val = bool(false)]; + tensor var_5464_cast_fp16 = matmul(transpose_x = var_5464_transpose_x_0, transpose_y = var_5464_transpose_y_0, x = var_5463_cast_fp16, y = state_301_cast_fp16)[name = string("op_5464_cast_fp16")]; + tensor v_new_301_cast_fp16 = sub(x = var_5461_cast_fp16, y = var_5464_cast_fp16)[name = string("v_new_301_cast_fp16")]; + tensor var_5467_begin_0 = const()[name = string("op_5467_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_5467_end_0 = const()[name = string("op_5467_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_5467_end_mask_0 = const()[name = string("op_5467_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5467_squeeze_mask_0 = const()[name = string("op_5467_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5467_cast_fp16 = slice_by_index(begin = var_5467_begin_0, end = var_5467_end_0, end_mask = var_5467_end_mask_0, squeeze_mask = var_5467_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5467_cast_fp16")]; + bool var_5468_transpose_x_0 = const()[name = string("op_5468_transpose_x_0"), val = bool(false)]; + bool var_5468_transpose_y_0 = const()[name = string("op_5468_transpose_y_0"), val = bool(false)]; + tensor var_5468_cast_fp16 = matmul(transpose_x = var_5468_transpose_x_0, transpose_y = var_5468_transpose_y_0, x = var_5467_cast_fp16, y = state_301_cast_fp16)[name = string("op_5468_cast_fp16")]; + tensor var_5470_begin_0 = const()[name = string("op_5470_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_5470_end_0 = const()[name = string("op_5470_end_0"), val = tensor([32, 23, 64, 64])]; + tensor var_5470_end_mask_0 = const()[name = string("op_5470_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5470_squeeze_mask_0 = const()[name = string("op_5470_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5470_cast_fp16 = slice_by_index(begin = var_5470_begin_0, end = var_5470_end_0, end_mask = var_5470_end_mask_0, squeeze_mask = var_5470_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5470_cast_fp16")]; + bool var_5471_transpose_x_0 = const()[name = string("op_5471_transpose_x_0"), val = bool(false)]; + bool var_5471_transpose_y_0 = const()[name = string("op_5471_transpose_y_0"), val = bool(false)]; + tensor var_5471_cast_fp16 = matmul(transpose_x = var_5471_transpose_x_0, transpose_y = var_5471_transpose_y_0, x = var_5470_cast_fp16, y = v_new_301_cast_fp16)[name = string("op_5471_cast_fp16")]; + tensor var_5472_cast_fp16 = add(x = var_5468_cast_fp16, y = var_5471_cast_fp16)[name = string("op_5472_cast_fp16")]; + tensor var_5474_begin_0 = const()[name = string("op_5474_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_5474_end_0 = const()[name = string("op_5474_end_0"), val = tensor([32, 23, 1, 1])]; + tensor var_5474_end_mask_0 = const()[name = string("op_5474_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5474_squeeze_mask_0 = const()[name = string("op_5474_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5474_cast_fp16 = slice_by_index(begin = var_5474_begin_0, end = var_5474_end_0, end_mask = var_5474_end_mask_0, squeeze_mask = var_5474_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5474_cast_fp16")]; + tensor var_5475_cast_fp16 = mul(x = state_301_cast_fp16, y = var_5474_cast_fp16)[name = string("op_5475_cast_fp16")]; + tensor var_5477_begin_0 = const()[name = string("op_5477_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_5477_end_0 = const()[name = string("op_5477_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_5477_end_mask_0 = const()[name = string("op_5477_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5477_squeeze_mask_0 = const()[name = string("op_5477_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5477_cast_fp16 = slice_by_index(begin = var_5477_begin_0, end = var_5477_end_0, end_mask = var_5477_end_mask_0, squeeze_mask = var_5477_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5477_cast_fp16")]; + bool var_5479_transpose_x_1 = const()[name = string("op_5479_transpose_x_1"), val = bool(true)]; + bool var_5479_transpose_y_1 = const()[name = string("op_5479_transpose_y_1"), val = bool(false)]; + tensor var_5479_cast_fp16 = matmul(transpose_x = var_5479_transpose_x_1, transpose_y = var_5479_transpose_y_1, x = var_5477_cast_fp16, y = v_new_301_cast_fp16)[name = string("op_5479_cast_fp16")]; + tensor state_303_cast_fp16 = add(x = var_5475_cast_fp16, y = var_5479_cast_fp16)[name = string("state_303_cast_fp16")]; + tensor var_5482_begin_0 = const()[name = string("op_5482_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_5482_end_0 = const()[name = string("op_5482_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_5482_end_mask_0 = const()[name = string("op_5482_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5482_squeeze_mask_0 = const()[name = string("op_5482_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5482_cast_fp16 = slice_by_index(begin = var_5482_begin_0, end = var_5482_end_0, end_mask = var_5482_end_mask_0, squeeze_mask = var_5482_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5482_cast_fp16")]; + tensor var_5484_begin_0 = const()[name = string("op_5484_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_5484_end_0 = const()[name = string("op_5484_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_5484_end_mask_0 = const()[name = string("op_5484_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5484_squeeze_mask_0 = const()[name = string("op_5484_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5484_cast_fp16 = slice_by_index(begin = var_5484_begin_0, end = var_5484_end_0, end_mask = var_5484_end_mask_0, squeeze_mask = var_5484_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5484_cast_fp16")]; + bool var_5485_transpose_x_0 = const()[name = string("op_5485_transpose_x_0"), val = bool(false)]; + bool var_5485_transpose_y_0 = const()[name = string("op_5485_transpose_y_0"), val = bool(false)]; + tensor var_5485_cast_fp16 = matmul(transpose_x = var_5485_transpose_x_0, transpose_y = var_5485_transpose_y_0, x = var_5484_cast_fp16, y = state_303_cast_fp16)[name = string("op_5485_cast_fp16")]; + tensor v_new_303_cast_fp16 = sub(x = var_5482_cast_fp16, y = var_5485_cast_fp16)[name = string("v_new_303_cast_fp16")]; + tensor var_5488_begin_0 = const()[name = string("op_5488_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_5488_end_0 = const()[name = string("op_5488_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_5488_end_mask_0 = const()[name = string("op_5488_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5488_squeeze_mask_0 = const()[name = string("op_5488_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5488_cast_fp16 = slice_by_index(begin = var_5488_begin_0, end = var_5488_end_0, end_mask = var_5488_end_mask_0, squeeze_mask = var_5488_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5488_cast_fp16")]; + bool var_5489_transpose_x_0 = const()[name = string("op_5489_transpose_x_0"), val = bool(false)]; + bool var_5489_transpose_y_0 = const()[name = string("op_5489_transpose_y_0"), val = bool(false)]; + tensor var_5489_cast_fp16 = matmul(transpose_x = var_5489_transpose_x_0, transpose_y = var_5489_transpose_y_0, x = var_5488_cast_fp16, y = state_303_cast_fp16)[name = string("op_5489_cast_fp16")]; + tensor var_5491_begin_0 = const()[name = string("op_5491_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_5491_end_0 = const()[name = string("op_5491_end_0"), val = tensor([32, 24, 64, 64])]; + tensor var_5491_end_mask_0 = const()[name = string("op_5491_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5491_squeeze_mask_0 = const()[name = string("op_5491_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5491_cast_fp16 = slice_by_index(begin = var_5491_begin_0, end = var_5491_end_0, end_mask = var_5491_end_mask_0, squeeze_mask = var_5491_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5491_cast_fp16")]; + bool var_5492_transpose_x_0 = const()[name = string("op_5492_transpose_x_0"), val = bool(false)]; + bool var_5492_transpose_y_0 = const()[name = string("op_5492_transpose_y_0"), val = bool(false)]; + tensor var_5492_cast_fp16 = matmul(transpose_x = var_5492_transpose_x_0, transpose_y = var_5492_transpose_y_0, x = var_5491_cast_fp16, y = v_new_303_cast_fp16)[name = string("op_5492_cast_fp16")]; + tensor var_5493_cast_fp16 = add(x = var_5489_cast_fp16, y = var_5492_cast_fp16)[name = string("op_5493_cast_fp16")]; + tensor var_5495_begin_0 = const()[name = string("op_5495_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_5495_end_0 = const()[name = string("op_5495_end_0"), val = tensor([32, 24, 1, 1])]; + tensor var_5495_end_mask_0 = const()[name = string("op_5495_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5495_squeeze_mask_0 = const()[name = string("op_5495_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5495_cast_fp16 = slice_by_index(begin = var_5495_begin_0, end = var_5495_end_0, end_mask = var_5495_end_mask_0, squeeze_mask = var_5495_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5495_cast_fp16")]; + tensor var_5496_cast_fp16 = mul(x = state_303_cast_fp16, y = var_5495_cast_fp16)[name = string("op_5496_cast_fp16")]; + tensor var_5498_begin_0 = const()[name = string("op_5498_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_5498_end_0 = const()[name = string("op_5498_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_5498_end_mask_0 = const()[name = string("op_5498_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5498_squeeze_mask_0 = const()[name = string("op_5498_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5498_cast_fp16 = slice_by_index(begin = var_5498_begin_0, end = var_5498_end_0, end_mask = var_5498_end_mask_0, squeeze_mask = var_5498_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5498_cast_fp16")]; + bool var_5500_transpose_x_1 = const()[name = string("op_5500_transpose_x_1"), val = bool(true)]; + bool var_5500_transpose_y_1 = const()[name = string("op_5500_transpose_y_1"), val = bool(false)]; + tensor var_5500_cast_fp16 = matmul(transpose_x = var_5500_transpose_x_1, transpose_y = var_5500_transpose_y_1, x = var_5498_cast_fp16, y = v_new_303_cast_fp16)[name = string("op_5500_cast_fp16")]; + tensor state_305_cast_fp16 = add(x = var_5496_cast_fp16, y = var_5500_cast_fp16)[name = string("state_305_cast_fp16")]; + tensor var_5503_begin_0 = const()[name = string("op_5503_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_5503_end_0 = const()[name = string("op_5503_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_5503_end_mask_0 = const()[name = string("op_5503_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5503_squeeze_mask_0 = const()[name = string("op_5503_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5503_cast_fp16 = slice_by_index(begin = var_5503_begin_0, end = var_5503_end_0, end_mask = var_5503_end_mask_0, squeeze_mask = var_5503_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5503_cast_fp16")]; + tensor var_5505_begin_0 = const()[name = string("op_5505_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_5505_end_0 = const()[name = string("op_5505_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_5505_end_mask_0 = const()[name = string("op_5505_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5505_squeeze_mask_0 = const()[name = string("op_5505_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5505_cast_fp16 = slice_by_index(begin = var_5505_begin_0, end = var_5505_end_0, end_mask = var_5505_end_mask_0, squeeze_mask = var_5505_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5505_cast_fp16")]; + bool var_5506_transpose_x_0 = const()[name = string("op_5506_transpose_x_0"), val = bool(false)]; + bool var_5506_transpose_y_0 = const()[name = string("op_5506_transpose_y_0"), val = bool(false)]; + tensor var_5506_cast_fp16 = matmul(transpose_x = var_5506_transpose_x_0, transpose_y = var_5506_transpose_y_0, x = var_5505_cast_fp16, y = state_305_cast_fp16)[name = string("op_5506_cast_fp16")]; + tensor v_new_305_cast_fp16 = sub(x = var_5503_cast_fp16, y = var_5506_cast_fp16)[name = string("v_new_305_cast_fp16")]; + tensor var_5509_begin_0 = const()[name = string("op_5509_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_5509_end_0 = const()[name = string("op_5509_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_5509_end_mask_0 = const()[name = string("op_5509_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5509_squeeze_mask_0 = const()[name = string("op_5509_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5509_cast_fp16 = slice_by_index(begin = var_5509_begin_0, end = var_5509_end_0, end_mask = var_5509_end_mask_0, squeeze_mask = var_5509_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5509_cast_fp16")]; + bool var_5510_transpose_x_0 = const()[name = string("op_5510_transpose_x_0"), val = bool(false)]; + bool var_5510_transpose_y_0 = const()[name = string("op_5510_transpose_y_0"), val = bool(false)]; + tensor var_5510_cast_fp16 = matmul(transpose_x = var_5510_transpose_x_0, transpose_y = var_5510_transpose_y_0, x = var_5509_cast_fp16, y = state_305_cast_fp16)[name = string("op_5510_cast_fp16")]; + tensor var_5512_begin_0 = const()[name = string("op_5512_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_5512_end_0 = const()[name = string("op_5512_end_0"), val = tensor([32, 25, 64, 64])]; + tensor var_5512_end_mask_0 = const()[name = string("op_5512_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5512_squeeze_mask_0 = const()[name = string("op_5512_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5512_cast_fp16 = slice_by_index(begin = var_5512_begin_0, end = var_5512_end_0, end_mask = var_5512_end_mask_0, squeeze_mask = var_5512_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5512_cast_fp16")]; + bool var_5513_transpose_x_0 = const()[name = string("op_5513_transpose_x_0"), val = bool(false)]; + bool var_5513_transpose_y_0 = const()[name = string("op_5513_transpose_y_0"), val = bool(false)]; + tensor var_5513_cast_fp16 = matmul(transpose_x = var_5513_transpose_x_0, transpose_y = var_5513_transpose_y_0, x = var_5512_cast_fp16, y = v_new_305_cast_fp16)[name = string("op_5513_cast_fp16")]; + tensor var_5514_cast_fp16 = add(x = var_5510_cast_fp16, y = var_5513_cast_fp16)[name = string("op_5514_cast_fp16")]; + tensor var_5516_begin_0 = const()[name = string("op_5516_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_5516_end_0 = const()[name = string("op_5516_end_0"), val = tensor([32, 25, 1, 1])]; + tensor var_5516_end_mask_0 = const()[name = string("op_5516_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5516_squeeze_mask_0 = const()[name = string("op_5516_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5516_cast_fp16 = slice_by_index(begin = var_5516_begin_0, end = var_5516_end_0, end_mask = var_5516_end_mask_0, squeeze_mask = var_5516_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5516_cast_fp16")]; + tensor var_5517_cast_fp16 = mul(x = state_305_cast_fp16, y = var_5516_cast_fp16)[name = string("op_5517_cast_fp16")]; + tensor var_5519_begin_0 = const()[name = string("op_5519_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_5519_end_0 = const()[name = string("op_5519_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_5519_end_mask_0 = const()[name = string("op_5519_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5519_squeeze_mask_0 = const()[name = string("op_5519_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5519_cast_fp16 = slice_by_index(begin = var_5519_begin_0, end = var_5519_end_0, end_mask = var_5519_end_mask_0, squeeze_mask = var_5519_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5519_cast_fp16")]; + bool var_5521_transpose_x_1 = const()[name = string("op_5521_transpose_x_1"), val = bool(true)]; + bool var_5521_transpose_y_1 = const()[name = string("op_5521_transpose_y_1"), val = bool(false)]; + tensor var_5521_cast_fp16 = matmul(transpose_x = var_5521_transpose_x_1, transpose_y = var_5521_transpose_y_1, x = var_5519_cast_fp16, y = v_new_305_cast_fp16)[name = string("op_5521_cast_fp16")]; + tensor state_307_cast_fp16 = add(x = var_5517_cast_fp16, y = var_5521_cast_fp16)[name = string("state_307_cast_fp16")]; + tensor var_5524_begin_0 = const()[name = string("op_5524_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_5524_end_0 = const()[name = string("op_5524_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_5524_end_mask_0 = const()[name = string("op_5524_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5524_squeeze_mask_0 = const()[name = string("op_5524_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5524_cast_fp16 = slice_by_index(begin = var_5524_begin_0, end = var_5524_end_0, end_mask = var_5524_end_mask_0, squeeze_mask = var_5524_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5524_cast_fp16")]; + tensor var_5526_begin_0 = const()[name = string("op_5526_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_5526_end_0 = const()[name = string("op_5526_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_5526_end_mask_0 = const()[name = string("op_5526_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5526_squeeze_mask_0 = const()[name = string("op_5526_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5526_cast_fp16 = slice_by_index(begin = var_5526_begin_0, end = var_5526_end_0, end_mask = var_5526_end_mask_0, squeeze_mask = var_5526_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5526_cast_fp16")]; + bool var_5527_transpose_x_0 = const()[name = string("op_5527_transpose_x_0"), val = bool(false)]; + bool var_5527_transpose_y_0 = const()[name = string("op_5527_transpose_y_0"), val = bool(false)]; + tensor var_5527_cast_fp16 = matmul(transpose_x = var_5527_transpose_x_0, transpose_y = var_5527_transpose_y_0, x = var_5526_cast_fp16, y = state_307_cast_fp16)[name = string("op_5527_cast_fp16")]; + tensor v_new_307_cast_fp16 = sub(x = var_5524_cast_fp16, y = var_5527_cast_fp16)[name = string("v_new_307_cast_fp16")]; + tensor var_5530_begin_0 = const()[name = string("op_5530_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_5530_end_0 = const()[name = string("op_5530_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_5530_end_mask_0 = const()[name = string("op_5530_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5530_squeeze_mask_0 = const()[name = string("op_5530_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5530_cast_fp16 = slice_by_index(begin = var_5530_begin_0, end = var_5530_end_0, end_mask = var_5530_end_mask_0, squeeze_mask = var_5530_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5530_cast_fp16")]; + bool var_5531_transpose_x_0 = const()[name = string("op_5531_transpose_x_0"), val = bool(false)]; + bool var_5531_transpose_y_0 = const()[name = string("op_5531_transpose_y_0"), val = bool(false)]; + tensor var_5531_cast_fp16 = matmul(transpose_x = var_5531_transpose_x_0, transpose_y = var_5531_transpose_y_0, x = var_5530_cast_fp16, y = state_307_cast_fp16)[name = string("op_5531_cast_fp16")]; + tensor var_5533_begin_0 = const()[name = string("op_5533_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_5533_end_0 = const()[name = string("op_5533_end_0"), val = tensor([32, 26, 64, 64])]; + tensor var_5533_end_mask_0 = const()[name = string("op_5533_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5533_squeeze_mask_0 = const()[name = string("op_5533_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5533_cast_fp16 = slice_by_index(begin = var_5533_begin_0, end = var_5533_end_0, end_mask = var_5533_end_mask_0, squeeze_mask = var_5533_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5533_cast_fp16")]; + bool var_5534_transpose_x_0 = const()[name = string("op_5534_transpose_x_0"), val = bool(false)]; + bool var_5534_transpose_y_0 = const()[name = string("op_5534_transpose_y_0"), val = bool(false)]; + tensor var_5534_cast_fp16 = matmul(transpose_x = var_5534_transpose_x_0, transpose_y = var_5534_transpose_y_0, x = var_5533_cast_fp16, y = v_new_307_cast_fp16)[name = string("op_5534_cast_fp16")]; + tensor var_5535_cast_fp16 = add(x = var_5531_cast_fp16, y = var_5534_cast_fp16)[name = string("op_5535_cast_fp16")]; + tensor var_5537_begin_0 = const()[name = string("op_5537_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_5537_end_0 = const()[name = string("op_5537_end_0"), val = tensor([32, 26, 1, 1])]; + tensor var_5537_end_mask_0 = const()[name = string("op_5537_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5537_squeeze_mask_0 = const()[name = string("op_5537_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5537_cast_fp16 = slice_by_index(begin = var_5537_begin_0, end = var_5537_end_0, end_mask = var_5537_end_mask_0, squeeze_mask = var_5537_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5537_cast_fp16")]; + tensor var_5538_cast_fp16 = mul(x = state_307_cast_fp16, y = var_5537_cast_fp16)[name = string("op_5538_cast_fp16")]; + tensor var_5540_begin_0 = const()[name = string("op_5540_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_5540_end_0 = const()[name = string("op_5540_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_5540_end_mask_0 = const()[name = string("op_5540_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5540_squeeze_mask_0 = const()[name = string("op_5540_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5540_cast_fp16 = slice_by_index(begin = var_5540_begin_0, end = var_5540_end_0, end_mask = var_5540_end_mask_0, squeeze_mask = var_5540_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5540_cast_fp16")]; + bool var_5542_transpose_x_1 = const()[name = string("op_5542_transpose_x_1"), val = bool(true)]; + bool var_5542_transpose_y_1 = const()[name = string("op_5542_transpose_y_1"), val = bool(false)]; + tensor var_5542_cast_fp16 = matmul(transpose_x = var_5542_transpose_x_1, transpose_y = var_5542_transpose_y_1, x = var_5540_cast_fp16, y = v_new_307_cast_fp16)[name = string("op_5542_cast_fp16")]; + tensor state_309_cast_fp16 = add(x = var_5538_cast_fp16, y = var_5542_cast_fp16)[name = string("state_309_cast_fp16")]; + tensor var_5545_begin_0 = const()[name = string("op_5545_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_5545_end_0 = const()[name = string("op_5545_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_5545_end_mask_0 = const()[name = string("op_5545_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5545_squeeze_mask_0 = const()[name = string("op_5545_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5545_cast_fp16 = slice_by_index(begin = var_5545_begin_0, end = var_5545_end_0, end_mask = var_5545_end_mask_0, squeeze_mask = var_5545_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5545_cast_fp16")]; + tensor var_5547_begin_0 = const()[name = string("op_5547_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_5547_end_0 = const()[name = string("op_5547_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_5547_end_mask_0 = const()[name = string("op_5547_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5547_squeeze_mask_0 = const()[name = string("op_5547_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5547_cast_fp16 = slice_by_index(begin = var_5547_begin_0, end = var_5547_end_0, end_mask = var_5547_end_mask_0, squeeze_mask = var_5547_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5547_cast_fp16")]; + bool var_5548_transpose_x_0 = const()[name = string("op_5548_transpose_x_0"), val = bool(false)]; + bool var_5548_transpose_y_0 = const()[name = string("op_5548_transpose_y_0"), val = bool(false)]; + tensor var_5548_cast_fp16 = matmul(transpose_x = var_5548_transpose_x_0, transpose_y = var_5548_transpose_y_0, x = var_5547_cast_fp16, y = state_309_cast_fp16)[name = string("op_5548_cast_fp16")]; + tensor v_new_309_cast_fp16 = sub(x = var_5545_cast_fp16, y = var_5548_cast_fp16)[name = string("v_new_309_cast_fp16")]; + tensor var_5551_begin_0 = const()[name = string("op_5551_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_5551_end_0 = const()[name = string("op_5551_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_5551_end_mask_0 = const()[name = string("op_5551_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5551_squeeze_mask_0 = const()[name = string("op_5551_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5551_cast_fp16 = slice_by_index(begin = var_5551_begin_0, end = var_5551_end_0, end_mask = var_5551_end_mask_0, squeeze_mask = var_5551_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5551_cast_fp16")]; + bool var_5552_transpose_x_0 = const()[name = string("op_5552_transpose_x_0"), val = bool(false)]; + bool var_5552_transpose_y_0 = const()[name = string("op_5552_transpose_y_0"), val = bool(false)]; + tensor var_5552_cast_fp16 = matmul(transpose_x = var_5552_transpose_x_0, transpose_y = var_5552_transpose_y_0, x = var_5551_cast_fp16, y = state_309_cast_fp16)[name = string("op_5552_cast_fp16")]; + tensor var_5554_begin_0 = const()[name = string("op_5554_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_5554_end_0 = const()[name = string("op_5554_end_0"), val = tensor([32, 27, 64, 64])]; + tensor var_5554_end_mask_0 = const()[name = string("op_5554_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5554_squeeze_mask_0 = const()[name = string("op_5554_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5554_cast_fp16 = slice_by_index(begin = var_5554_begin_0, end = var_5554_end_0, end_mask = var_5554_end_mask_0, squeeze_mask = var_5554_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5554_cast_fp16")]; + bool var_5555_transpose_x_0 = const()[name = string("op_5555_transpose_x_0"), val = bool(false)]; + bool var_5555_transpose_y_0 = const()[name = string("op_5555_transpose_y_0"), val = bool(false)]; + tensor var_5555_cast_fp16 = matmul(transpose_x = var_5555_transpose_x_0, transpose_y = var_5555_transpose_y_0, x = var_5554_cast_fp16, y = v_new_309_cast_fp16)[name = string("op_5555_cast_fp16")]; + tensor var_5556_cast_fp16 = add(x = var_5552_cast_fp16, y = var_5555_cast_fp16)[name = string("op_5556_cast_fp16")]; + tensor var_5558_begin_0 = const()[name = string("op_5558_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_5558_end_0 = const()[name = string("op_5558_end_0"), val = tensor([32, 27, 1, 1])]; + tensor var_5558_end_mask_0 = const()[name = string("op_5558_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5558_squeeze_mask_0 = const()[name = string("op_5558_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5558_cast_fp16 = slice_by_index(begin = var_5558_begin_0, end = var_5558_end_0, end_mask = var_5558_end_mask_0, squeeze_mask = var_5558_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5558_cast_fp16")]; + tensor var_5559_cast_fp16 = mul(x = state_309_cast_fp16, y = var_5558_cast_fp16)[name = string("op_5559_cast_fp16")]; + tensor var_5561_begin_0 = const()[name = string("op_5561_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_5561_end_0 = const()[name = string("op_5561_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_5561_end_mask_0 = const()[name = string("op_5561_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5561_squeeze_mask_0 = const()[name = string("op_5561_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5561_cast_fp16 = slice_by_index(begin = var_5561_begin_0, end = var_5561_end_0, end_mask = var_5561_end_mask_0, squeeze_mask = var_5561_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5561_cast_fp16")]; + bool var_5563_transpose_x_1 = const()[name = string("op_5563_transpose_x_1"), val = bool(true)]; + bool var_5563_transpose_y_1 = const()[name = string("op_5563_transpose_y_1"), val = bool(false)]; + tensor var_5563_cast_fp16 = matmul(transpose_x = var_5563_transpose_x_1, transpose_y = var_5563_transpose_y_1, x = var_5561_cast_fp16, y = v_new_309_cast_fp16)[name = string("op_5563_cast_fp16")]; + tensor state_311_cast_fp16 = add(x = var_5559_cast_fp16, y = var_5563_cast_fp16)[name = string("state_311_cast_fp16")]; + tensor var_5566_begin_0 = const()[name = string("op_5566_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_5566_end_0 = const()[name = string("op_5566_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_5566_end_mask_0 = const()[name = string("op_5566_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5566_squeeze_mask_0 = const()[name = string("op_5566_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5566_cast_fp16 = slice_by_index(begin = var_5566_begin_0, end = var_5566_end_0, end_mask = var_5566_end_mask_0, squeeze_mask = var_5566_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5566_cast_fp16")]; + tensor var_5568_begin_0 = const()[name = string("op_5568_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_5568_end_0 = const()[name = string("op_5568_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_5568_end_mask_0 = const()[name = string("op_5568_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5568_squeeze_mask_0 = const()[name = string("op_5568_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5568_cast_fp16 = slice_by_index(begin = var_5568_begin_0, end = var_5568_end_0, end_mask = var_5568_end_mask_0, squeeze_mask = var_5568_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5568_cast_fp16")]; + bool var_5569_transpose_x_0 = const()[name = string("op_5569_transpose_x_0"), val = bool(false)]; + bool var_5569_transpose_y_0 = const()[name = string("op_5569_transpose_y_0"), val = bool(false)]; + tensor var_5569_cast_fp16 = matmul(transpose_x = var_5569_transpose_x_0, transpose_y = var_5569_transpose_y_0, x = var_5568_cast_fp16, y = state_311_cast_fp16)[name = string("op_5569_cast_fp16")]; + tensor v_new_311_cast_fp16 = sub(x = var_5566_cast_fp16, y = var_5569_cast_fp16)[name = string("v_new_311_cast_fp16")]; + tensor var_5572_begin_0 = const()[name = string("op_5572_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_5572_end_0 = const()[name = string("op_5572_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_5572_end_mask_0 = const()[name = string("op_5572_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5572_squeeze_mask_0 = const()[name = string("op_5572_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5572_cast_fp16 = slice_by_index(begin = var_5572_begin_0, end = var_5572_end_0, end_mask = var_5572_end_mask_0, squeeze_mask = var_5572_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5572_cast_fp16")]; + bool var_5573_transpose_x_0 = const()[name = string("op_5573_transpose_x_0"), val = bool(false)]; + bool var_5573_transpose_y_0 = const()[name = string("op_5573_transpose_y_0"), val = bool(false)]; + tensor var_5573_cast_fp16 = matmul(transpose_x = var_5573_transpose_x_0, transpose_y = var_5573_transpose_y_0, x = var_5572_cast_fp16, y = state_311_cast_fp16)[name = string("op_5573_cast_fp16")]; + tensor var_5575_begin_0 = const()[name = string("op_5575_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_5575_end_0 = const()[name = string("op_5575_end_0"), val = tensor([32, 28, 64, 64])]; + tensor var_5575_end_mask_0 = const()[name = string("op_5575_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5575_squeeze_mask_0 = const()[name = string("op_5575_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5575_cast_fp16 = slice_by_index(begin = var_5575_begin_0, end = var_5575_end_0, end_mask = var_5575_end_mask_0, squeeze_mask = var_5575_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5575_cast_fp16")]; + bool var_5576_transpose_x_0 = const()[name = string("op_5576_transpose_x_0"), val = bool(false)]; + bool var_5576_transpose_y_0 = const()[name = string("op_5576_transpose_y_0"), val = bool(false)]; + tensor var_5576_cast_fp16 = matmul(transpose_x = var_5576_transpose_x_0, transpose_y = var_5576_transpose_y_0, x = var_5575_cast_fp16, y = v_new_311_cast_fp16)[name = string("op_5576_cast_fp16")]; + tensor var_5577_cast_fp16 = add(x = var_5573_cast_fp16, y = var_5576_cast_fp16)[name = string("op_5577_cast_fp16")]; + tensor var_5579_begin_0 = const()[name = string("op_5579_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_5579_end_0 = const()[name = string("op_5579_end_0"), val = tensor([32, 28, 1, 1])]; + tensor var_5579_end_mask_0 = const()[name = string("op_5579_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5579_squeeze_mask_0 = const()[name = string("op_5579_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5579_cast_fp16 = slice_by_index(begin = var_5579_begin_0, end = var_5579_end_0, end_mask = var_5579_end_mask_0, squeeze_mask = var_5579_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5579_cast_fp16")]; + tensor var_5580_cast_fp16 = mul(x = state_311_cast_fp16, y = var_5579_cast_fp16)[name = string("op_5580_cast_fp16")]; + tensor var_5582_begin_0 = const()[name = string("op_5582_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_5582_end_0 = const()[name = string("op_5582_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_5582_end_mask_0 = const()[name = string("op_5582_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5582_squeeze_mask_0 = const()[name = string("op_5582_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5582_cast_fp16 = slice_by_index(begin = var_5582_begin_0, end = var_5582_end_0, end_mask = var_5582_end_mask_0, squeeze_mask = var_5582_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5582_cast_fp16")]; + bool var_5584_transpose_x_1 = const()[name = string("op_5584_transpose_x_1"), val = bool(true)]; + bool var_5584_transpose_y_1 = const()[name = string("op_5584_transpose_y_1"), val = bool(false)]; + tensor var_5584_cast_fp16 = matmul(transpose_x = var_5584_transpose_x_1, transpose_y = var_5584_transpose_y_1, x = var_5582_cast_fp16, y = v_new_311_cast_fp16)[name = string("op_5584_cast_fp16")]; + tensor state_313_cast_fp16 = add(x = var_5580_cast_fp16, y = var_5584_cast_fp16)[name = string("state_313_cast_fp16")]; + tensor var_5587_begin_0 = const()[name = string("op_5587_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_5587_end_0 = const()[name = string("op_5587_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_5587_end_mask_0 = const()[name = string("op_5587_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5587_squeeze_mask_0 = const()[name = string("op_5587_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5587_cast_fp16 = slice_by_index(begin = var_5587_begin_0, end = var_5587_end_0, end_mask = var_5587_end_mask_0, squeeze_mask = var_5587_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5587_cast_fp16")]; + tensor var_5589_begin_0 = const()[name = string("op_5589_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_5589_end_0 = const()[name = string("op_5589_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_5589_end_mask_0 = const()[name = string("op_5589_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5589_squeeze_mask_0 = const()[name = string("op_5589_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5589_cast_fp16 = slice_by_index(begin = var_5589_begin_0, end = var_5589_end_0, end_mask = var_5589_end_mask_0, squeeze_mask = var_5589_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5589_cast_fp16")]; + bool var_5590_transpose_x_0 = const()[name = string("op_5590_transpose_x_0"), val = bool(false)]; + bool var_5590_transpose_y_0 = const()[name = string("op_5590_transpose_y_0"), val = bool(false)]; + tensor var_5590_cast_fp16 = matmul(transpose_x = var_5590_transpose_x_0, transpose_y = var_5590_transpose_y_0, x = var_5589_cast_fp16, y = state_313_cast_fp16)[name = string("op_5590_cast_fp16")]; + tensor v_new_313_cast_fp16 = sub(x = var_5587_cast_fp16, y = var_5590_cast_fp16)[name = string("v_new_313_cast_fp16")]; + tensor var_5593_begin_0 = const()[name = string("op_5593_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_5593_end_0 = const()[name = string("op_5593_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_5593_end_mask_0 = const()[name = string("op_5593_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5593_squeeze_mask_0 = const()[name = string("op_5593_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5593_cast_fp16 = slice_by_index(begin = var_5593_begin_0, end = var_5593_end_0, end_mask = var_5593_end_mask_0, squeeze_mask = var_5593_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5593_cast_fp16")]; + bool var_5594_transpose_x_0 = const()[name = string("op_5594_transpose_x_0"), val = bool(false)]; + bool var_5594_transpose_y_0 = const()[name = string("op_5594_transpose_y_0"), val = bool(false)]; + tensor var_5594_cast_fp16 = matmul(transpose_x = var_5594_transpose_x_0, transpose_y = var_5594_transpose_y_0, x = var_5593_cast_fp16, y = state_313_cast_fp16)[name = string("op_5594_cast_fp16")]; + tensor var_5596_begin_0 = const()[name = string("op_5596_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_5596_end_0 = const()[name = string("op_5596_end_0"), val = tensor([32, 29, 64, 64])]; + tensor var_5596_end_mask_0 = const()[name = string("op_5596_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5596_squeeze_mask_0 = const()[name = string("op_5596_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5596_cast_fp16 = slice_by_index(begin = var_5596_begin_0, end = var_5596_end_0, end_mask = var_5596_end_mask_0, squeeze_mask = var_5596_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5596_cast_fp16")]; + bool var_5597_transpose_x_0 = const()[name = string("op_5597_transpose_x_0"), val = bool(false)]; + bool var_5597_transpose_y_0 = const()[name = string("op_5597_transpose_y_0"), val = bool(false)]; + tensor var_5597_cast_fp16 = matmul(transpose_x = var_5597_transpose_x_0, transpose_y = var_5597_transpose_y_0, x = var_5596_cast_fp16, y = v_new_313_cast_fp16)[name = string("op_5597_cast_fp16")]; + tensor var_5598_cast_fp16 = add(x = var_5594_cast_fp16, y = var_5597_cast_fp16)[name = string("op_5598_cast_fp16")]; + tensor var_5600_begin_0 = const()[name = string("op_5600_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_5600_end_0 = const()[name = string("op_5600_end_0"), val = tensor([32, 29, 1, 1])]; + tensor var_5600_end_mask_0 = const()[name = string("op_5600_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5600_squeeze_mask_0 = const()[name = string("op_5600_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5600_cast_fp16 = slice_by_index(begin = var_5600_begin_0, end = var_5600_end_0, end_mask = var_5600_end_mask_0, squeeze_mask = var_5600_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5600_cast_fp16")]; + tensor var_5601_cast_fp16 = mul(x = state_313_cast_fp16, y = var_5600_cast_fp16)[name = string("op_5601_cast_fp16")]; + tensor var_5603_begin_0 = const()[name = string("op_5603_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_5603_end_0 = const()[name = string("op_5603_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_5603_end_mask_0 = const()[name = string("op_5603_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5603_squeeze_mask_0 = const()[name = string("op_5603_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5603_cast_fp16 = slice_by_index(begin = var_5603_begin_0, end = var_5603_end_0, end_mask = var_5603_end_mask_0, squeeze_mask = var_5603_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5603_cast_fp16")]; + bool var_5605_transpose_x_1 = const()[name = string("op_5605_transpose_x_1"), val = bool(true)]; + bool var_5605_transpose_y_1 = const()[name = string("op_5605_transpose_y_1"), val = bool(false)]; + tensor var_5605_cast_fp16 = matmul(transpose_x = var_5605_transpose_x_1, transpose_y = var_5605_transpose_y_1, x = var_5603_cast_fp16, y = v_new_313_cast_fp16)[name = string("op_5605_cast_fp16")]; + tensor state_315_cast_fp16 = add(x = var_5601_cast_fp16, y = var_5605_cast_fp16)[name = string("state_315_cast_fp16")]; + tensor var_5608_begin_0 = const()[name = string("op_5608_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_5608_end_0 = const()[name = string("op_5608_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_5608_end_mask_0 = const()[name = string("op_5608_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5608_squeeze_mask_0 = const()[name = string("op_5608_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5608_cast_fp16 = slice_by_index(begin = var_5608_begin_0, end = var_5608_end_0, end_mask = var_5608_end_mask_0, squeeze_mask = var_5608_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5608_cast_fp16")]; + tensor var_5610_begin_0 = const()[name = string("op_5610_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_5610_end_0 = const()[name = string("op_5610_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_5610_end_mask_0 = const()[name = string("op_5610_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5610_squeeze_mask_0 = const()[name = string("op_5610_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5610_cast_fp16 = slice_by_index(begin = var_5610_begin_0, end = var_5610_end_0, end_mask = var_5610_end_mask_0, squeeze_mask = var_5610_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5610_cast_fp16")]; + bool var_5611_transpose_x_0 = const()[name = string("op_5611_transpose_x_0"), val = bool(false)]; + bool var_5611_transpose_y_0 = const()[name = string("op_5611_transpose_y_0"), val = bool(false)]; + tensor var_5611_cast_fp16 = matmul(transpose_x = var_5611_transpose_x_0, transpose_y = var_5611_transpose_y_0, x = var_5610_cast_fp16, y = state_315_cast_fp16)[name = string("op_5611_cast_fp16")]; + tensor v_new_315_cast_fp16 = sub(x = var_5608_cast_fp16, y = var_5611_cast_fp16)[name = string("v_new_315_cast_fp16")]; + tensor var_5614_begin_0 = const()[name = string("op_5614_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_5614_end_0 = const()[name = string("op_5614_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_5614_end_mask_0 = const()[name = string("op_5614_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5614_squeeze_mask_0 = const()[name = string("op_5614_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5614_cast_fp16 = slice_by_index(begin = var_5614_begin_0, end = var_5614_end_0, end_mask = var_5614_end_mask_0, squeeze_mask = var_5614_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5614_cast_fp16")]; + bool var_5615_transpose_x_0 = const()[name = string("op_5615_transpose_x_0"), val = bool(false)]; + bool var_5615_transpose_y_0 = const()[name = string("op_5615_transpose_y_0"), val = bool(false)]; + tensor var_5615_cast_fp16 = matmul(transpose_x = var_5615_transpose_x_0, transpose_y = var_5615_transpose_y_0, x = var_5614_cast_fp16, y = state_315_cast_fp16)[name = string("op_5615_cast_fp16")]; + tensor var_5617_begin_0 = const()[name = string("op_5617_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_5617_end_0 = const()[name = string("op_5617_end_0"), val = tensor([32, 30, 64, 64])]; + tensor var_5617_end_mask_0 = const()[name = string("op_5617_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5617_squeeze_mask_0 = const()[name = string("op_5617_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5617_cast_fp16 = slice_by_index(begin = var_5617_begin_0, end = var_5617_end_0, end_mask = var_5617_end_mask_0, squeeze_mask = var_5617_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5617_cast_fp16")]; + bool var_5618_transpose_x_0 = const()[name = string("op_5618_transpose_x_0"), val = bool(false)]; + bool var_5618_transpose_y_0 = const()[name = string("op_5618_transpose_y_0"), val = bool(false)]; + tensor var_5618_cast_fp16 = matmul(transpose_x = var_5618_transpose_x_0, transpose_y = var_5618_transpose_y_0, x = var_5617_cast_fp16, y = v_new_315_cast_fp16)[name = string("op_5618_cast_fp16")]; + tensor var_5619_cast_fp16 = add(x = var_5615_cast_fp16, y = var_5618_cast_fp16)[name = string("op_5619_cast_fp16")]; + tensor var_5621_begin_0 = const()[name = string("op_5621_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_5621_end_0 = const()[name = string("op_5621_end_0"), val = tensor([32, 30, 1, 1])]; + tensor var_5621_end_mask_0 = const()[name = string("op_5621_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5621_squeeze_mask_0 = const()[name = string("op_5621_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5621_cast_fp16 = slice_by_index(begin = var_5621_begin_0, end = var_5621_end_0, end_mask = var_5621_end_mask_0, squeeze_mask = var_5621_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5621_cast_fp16")]; + tensor var_5622_cast_fp16 = mul(x = state_315_cast_fp16, y = var_5621_cast_fp16)[name = string("op_5622_cast_fp16")]; + tensor var_5624_begin_0 = const()[name = string("op_5624_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_5624_end_0 = const()[name = string("op_5624_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_5624_end_mask_0 = const()[name = string("op_5624_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5624_squeeze_mask_0 = const()[name = string("op_5624_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5624_cast_fp16 = slice_by_index(begin = var_5624_begin_0, end = var_5624_end_0, end_mask = var_5624_end_mask_0, squeeze_mask = var_5624_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5624_cast_fp16")]; + bool var_5626_transpose_x_1 = const()[name = string("op_5626_transpose_x_1"), val = bool(true)]; + bool var_5626_transpose_y_1 = const()[name = string("op_5626_transpose_y_1"), val = bool(false)]; + tensor var_5626_cast_fp16 = matmul(transpose_x = var_5626_transpose_x_1, transpose_y = var_5626_transpose_y_1, x = var_5624_cast_fp16, y = v_new_315_cast_fp16)[name = string("op_5626_cast_fp16")]; + tensor state_317_cast_fp16 = add(x = var_5622_cast_fp16, y = var_5626_cast_fp16)[name = string("state_317_cast_fp16")]; + tensor var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5629_squeeze_mask_0 = const()[name = string("op_5629_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, squeeze_mask = var_5629_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5629_cast_fp16")]; + tensor var_5631_begin_0 = const()[name = string("op_5631_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_5631_end_0 = const()[name = string("op_5631_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_5631_end_mask_0 = const()[name = string("op_5631_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5631_squeeze_mask_0 = const()[name = string("op_5631_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5631_cast_fp16 = slice_by_index(begin = var_5631_begin_0, end = var_5631_end_0, end_mask = var_5631_end_mask_0, squeeze_mask = var_5631_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5631_cast_fp16")]; + bool var_5632_transpose_x_0 = const()[name = string("op_5632_transpose_x_0"), val = bool(false)]; + bool var_5632_transpose_y_0 = const()[name = string("op_5632_transpose_y_0"), val = bool(false)]; + tensor var_5632_cast_fp16 = matmul(transpose_x = var_5632_transpose_x_0, transpose_y = var_5632_transpose_y_0, x = var_5631_cast_fp16, y = state_317_cast_fp16)[name = string("op_5632_cast_fp16")]; + tensor v_new_317_cast_fp16 = sub(x = var_5629_cast_fp16, y = var_5632_cast_fp16)[name = string("v_new_317_cast_fp16")]; + tensor var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5635_squeeze_mask_0 = const()[name = string("op_5635_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, squeeze_mask = var_5635_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5635_cast_fp16")]; + bool var_5636_transpose_x_0 = const()[name = string("op_5636_transpose_x_0"), val = bool(false)]; + bool var_5636_transpose_y_0 = const()[name = string("op_5636_transpose_y_0"), val = bool(false)]; + tensor var_5636_cast_fp16 = matmul(transpose_x = var_5636_transpose_x_0, transpose_y = var_5636_transpose_y_0, x = var_5635_cast_fp16, y = state_317_cast_fp16)[name = string("op_5636_cast_fp16")]; + tensor var_5638_begin_0 = const()[name = string("op_5638_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_5638_end_0 = const()[name = string("op_5638_end_0"), val = tensor([32, 31, 64, 64])]; + tensor var_5638_end_mask_0 = const()[name = string("op_5638_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5638_squeeze_mask_0 = const()[name = string("op_5638_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5638_cast_fp16 = slice_by_index(begin = var_5638_begin_0, end = var_5638_end_0, end_mask = var_5638_end_mask_0, squeeze_mask = var_5638_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5638_cast_fp16")]; + bool var_5639_transpose_x_0 = const()[name = string("op_5639_transpose_x_0"), val = bool(false)]; + bool var_5639_transpose_y_0 = const()[name = string("op_5639_transpose_y_0"), val = bool(false)]; + tensor var_5639_cast_fp16 = matmul(transpose_x = var_5639_transpose_x_0, transpose_y = var_5639_transpose_y_0, x = var_5638_cast_fp16, y = v_new_317_cast_fp16)[name = string("op_5639_cast_fp16")]; + tensor var_5640_cast_fp16 = add(x = var_5636_cast_fp16, y = var_5639_cast_fp16)[name = string("op_5640_cast_fp16")]; + tensor var_5642_begin_0 = const()[name = string("op_5642_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_5642_end_0 = const()[name = string("op_5642_end_0"), val = tensor([32, 31, 1, 1])]; + tensor var_5642_end_mask_0 = const()[name = string("op_5642_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5642_squeeze_mask_0 = const()[name = string("op_5642_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5642_cast_fp16 = slice_by_index(begin = var_5642_begin_0, end = var_5642_end_0, end_mask = var_5642_end_mask_0, squeeze_mask = var_5642_squeeze_mask_0, x = chunk_decay_9_cast_fp16)[name = string("op_5642_cast_fp16")]; + tensor var_5643_cast_fp16 = mul(x = state_317_cast_fp16, y = var_5642_cast_fp16)[name = string("op_5643_cast_fp16")]; + tensor var_5645_begin_0 = const()[name = string("op_5645_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_5645_end_0 = const()[name = string("op_5645_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_5645_end_mask_0 = const()[name = string("op_5645_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5645_squeeze_mask_0 = const()[name = string("op_5645_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5645_cast_fp16 = slice_by_index(begin = var_5645_begin_0, end = var_5645_end_0, end_mask = var_5645_end_mask_0, squeeze_mask = var_5645_squeeze_mask_0, x = k_dec_9_cast_fp16)[name = string("op_5645_cast_fp16")]; + bool var_5647_transpose_x_1 = const()[name = string("op_5647_transpose_x_1"), val = bool(true)]; + bool var_5647_transpose_y_1 = const()[name = string("op_5647_transpose_y_1"), val = bool(false)]; + tensor var_5647_cast_fp16 = matmul(transpose_x = var_5647_transpose_x_1, transpose_y = var_5647_transpose_y_1, x = var_5645_cast_fp16, y = v_new_317_cast_fp16)[name = string("op_5647_cast_fp16")]; + tensor state_319_cast_fp16 = add(x = var_5643_cast_fp16, y = var_5647_cast_fp16)[name = string("state_319_cast_fp16")]; + tensor var_5650_begin_0 = const()[name = string("op_5650_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_5650_end_0 = const()[name = string("op_5650_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_5650_end_mask_0 = const()[name = string("op_5650_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5650_squeeze_mask_0 = const()[name = string("op_5650_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5650_cast_fp16 = slice_by_index(begin = var_5650_begin_0, end = var_5650_end_0, end_mask = var_5650_end_mask_0, squeeze_mask = var_5650_squeeze_mask_0, x = new_v_9_cast_fp16)[name = string("op_5650_cast_fp16")]; + tensor var_5652_begin_0 = const()[name = string("op_5652_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_5652_end_0 = const()[name = string("op_5652_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_5652_end_mask_0 = const()[name = string("op_5652_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5652_squeeze_mask_0 = const()[name = string("op_5652_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5652_cast_fp16 = slice_by_index(begin = var_5652_begin_0, end = var_5652_end_0, end_mask = var_5652_end_mask_0, squeeze_mask = var_5652_squeeze_mask_0, x = k_cumdecay_9_cast_fp16)[name = string("op_5652_cast_fp16")]; + bool var_5653_transpose_x_0 = const()[name = string("op_5653_transpose_x_0"), val = bool(false)]; + bool var_5653_transpose_y_0 = const()[name = string("op_5653_transpose_y_0"), val = bool(false)]; + tensor var_5653_cast_fp16 = matmul(transpose_x = var_5653_transpose_x_0, transpose_y = var_5653_transpose_y_0, x = var_5652_cast_fp16, y = state_319_cast_fp16)[name = string("op_5653_cast_fp16")]; + tensor v_new_319_cast_fp16 = sub(x = var_5650_cast_fp16, y = var_5653_cast_fp16)[name = string("v_new_319_cast_fp16")]; + tensor var_5656_begin_0 = const()[name = string("op_5656_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_5656_end_0 = const()[name = string("op_5656_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_5656_end_mask_0 = const()[name = string("op_5656_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5656_squeeze_mask_0 = const()[name = string("op_5656_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5656_cast_fp16 = slice_by_index(begin = var_5656_begin_0, end = var_5656_end_0, end_mask = var_5656_end_mask_0, squeeze_mask = var_5656_squeeze_mask_0, x = q_dec_9_cast_fp16)[name = string("op_5656_cast_fp16")]; + bool var_5657_transpose_x_0 = const()[name = string("op_5657_transpose_x_0"), val = bool(false)]; + bool var_5657_transpose_y_0 = const()[name = string("op_5657_transpose_y_0"), val = bool(false)]; + tensor var_5657_cast_fp16 = matmul(transpose_x = var_5657_transpose_x_0, transpose_y = var_5657_transpose_y_0, x = var_5656_cast_fp16, y = state_319_cast_fp16)[name = string("op_5657_cast_fp16")]; + tensor var_5659_begin_0 = const()[name = string("op_5659_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_5659_end_0 = const()[name = string("op_5659_end_0"), val = tensor([32, 32, 64, 64])]; + tensor var_5659_end_mask_0 = const()[name = string("op_5659_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5659_squeeze_mask_0 = const()[name = string("op_5659_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_5659_cast_fp16 = slice_by_index(begin = var_5659_begin_0, end = var_5659_end_0, end_mask = var_5659_end_mask_0, squeeze_mask = var_5659_squeeze_mask_0, x = attn_intra_9_cast_fp16)[name = string("op_5659_cast_fp16")]; + bool var_5660_transpose_x_0 = const()[name = string("op_5660_transpose_x_0"), val = bool(false)]; + bool var_5660_transpose_y_0 = const()[name = string("op_5660_transpose_y_0"), val = bool(false)]; + tensor var_5660_cast_fp16 = matmul(transpose_x = var_5660_transpose_x_0, transpose_y = var_5660_transpose_y_0, x = var_5659_cast_fp16, y = v_new_319_cast_fp16)[name = string("op_5660_cast_fp16")]; + tensor var_5661_cast_fp16 = add(x = var_5657_cast_fp16, y = var_5660_cast_fp16)[name = string("op_5661_cast_fp16")]; + int32 var_5663_axis_0 = const()[name = string("op_5663_axis_0"), val = int32(1)]; + tensor var_5663_cast_fp16 = stack(axis = var_5663_axis_0, values = (var_5010_cast_fp16, var_5031_cast_fp16, var_5052_cast_fp16, var_5073_cast_fp16, var_5094_cast_fp16, var_5115_cast_fp16, var_5136_cast_fp16, var_5157_cast_fp16, var_5178_cast_fp16, var_5199_cast_fp16, var_5220_cast_fp16, var_5241_cast_fp16, var_5262_cast_fp16, var_5283_cast_fp16, var_5304_cast_fp16, var_5325_cast_fp16, var_5346_cast_fp16, var_5367_cast_fp16, var_5388_cast_fp16, var_5409_cast_fp16, var_5430_cast_fp16, var_5451_cast_fp16, var_5472_cast_fp16, var_5493_cast_fp16, var_5514_cast_fp16, var_5535_cast_fp16, var_5556_cast_fp16, var_5577_cast_fp16, var_5598_cast_fp16, var_5619_cast_fp16, var_5640_cast_fp16, var_5661_cast_fp16))[name = string("op_5663_cast_fp16")]; + tensor var_5664 = const()[name = string("op_5664"), val = tensor([32, 2048, 128])]; + tensor var_5665_cast_fp16 = reshape(shape = var_5664, x = var_5663_cast_fp16)[name = string("op_5665_cast_fp16")]; + tensor var_5666 = const()[name = string("op_5666"), val = tensor([1, 0, 2])]; + tensor input_79_begin_0 = const()[name = string("input_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_end_0 = const()[name = string("input_79_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor input_79_end_mask_0 = const()[name = string("input_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor input_79_squeeze_mask_0 = const()[name = string("input_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor input_79_cast_fp16 = slice_by_index(begin = input_79_begin_0, end = input_79_end_0, end_mask = input_79_end_mask_0, squeeze_mask = input_79_squeeze_mask_0, x = z_9_cast_fp16)[name = string("input_79_cast_fp16")]; + fp16 var_4646_promoted_1_to_fp16 = const()[name = string("op_4646_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor x_143_cast_fp16 = transpose(perm = var_5666, x = var_5665_cast_fp16)[name = string("transpose_12")]; + tensor var_5670_cast_fp16 = pow(x = x_143_cast_fp16, y = var_4646_promoted_1_to_fp16)[name = string("op_5670_cast_fp16")]; + tensor var_5672_axes_0 = const()[name = string("op_5672_axes_0"), val = tensor([-1])]; + bool var_5672_keep_dims_0 = const()[name = string("op_5672_keep_dims_0"), val = bool(true)]; + tensor var_5672_cast_fp16 = reduce_mean(axes = var_5672_axes_0, keep_dims = var_5672_keep_dims_0, x = var_5670_cast_fp16)[name = string("op_5672_cast_fp16")]; + fp16 var_5673_to_fp16 = const()[name = string("op_5673_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_5674_cast_fp16 = add(x = var_5672_cast_fp16, y = var_5673_to_fp16)[name = string("op_5674_cast_fp16")]; + fp32 var_5675_epsilon_0 = const()[name = string("op_5675_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5675_cast_fp16 = rsqrt(epsilon = var_5675_epsilon_0, x = var_5674_cast_fp16)[name = string("op_5675_cast_fp16")]; + tensor x_145_cast_fp16 = mul(x = x_143_cast_fp16, y = var_5675_cast_fp16)[name = string("x_145_cast_fp16")]; + tensor layers_5_linear_attn_norm_weight_to_fp16 = const()[name = string("layers_5_linear_attn_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606136384)))]; + tensor var_5677_cast_fp16 = mul(x = layers_5_linear_attn_norm_weight_to_fp16, y = x_145_cast_fp16)[name = string("op_5677_cast_fp16")]; + tensor var_5678_cast_fp16 = silu(x = input_79_cast_fp16)[name = string("op_5678_cast_fp16")]; + tensor core_9_cast_fp16 = mul(x = var_5677_cast_fp16, y = var_5678_cast_fp16)[name = string("core_9_cast_fp16")]; + tensor var_5680 = const()[name = string("op_5680"), val = tensor([1, 2048, 4096])]; + tensor input_81_cast_fp16 = reshape(shape = var_5680, x = core_9_cast_fp16)[name = string("input_81_cast_fp16")]; + tensor layers_5_linear_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606136704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616622528))))[name = string("layers_5_linear_attn_out_proj_weight_to_fp16_quantized")]; + tensor linear_43_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_linear_attn_out_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = x_125_cast_fp16, y = linear_43_cast_fp16)[name = string("x_147_cast_fp16")]; + fp16 var_4646_promoted_2_to_fp16 = const()[name = string("op_4646_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor var_5686_cast_fp16 = pow(x = x_147_cast_fp16, y = var_4646_promoted_2_to_fp16)[name = string("op_5686_cast_fp16")]; + tensor var_5688_axes_0 = const()[name = string("op_5688_axes_0"), val = tensor([-1])]; + bool var_5688_keep_dims_0 = const()[name = string("op_5688_keep_dims_0"), val = bool(true)]; + tensor var_5688_cast_fp16 = reduce_mean(axes = var_5688_axes_0, keep_dims = var_5688_keep_dims_0, x = var_5686_cast_fp16)[name = string("op_5688_cast_fp16")]; + fp16 var_5689_to_fp16 = const()[name = string("op_5689_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_5690_cast_fp16 = add(x = var_5688_cast_fp16, y = var_5689_to_fp16)[name = string("op_5690_cast_fp16")]; + fp32 var_5691_epsilon_0 = const()[name = string("op_5691_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5691_cast_fp16 = rsqrt(epsilon = var_5691_epsilon_0, x = var_5690_cast_fp16)[name = string("op_5691_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_147_cast_fp16, y = var_5691_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor var_5694_to_fp16 = const()[name = string("op_5694_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616627712)))]; + tensor input_83_cast_fp16 = mul(x = x_149_cast_fp16, y = var_5694_to_fp16)[name = string("input_83_cast_fp16")]; + tensor layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616632896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640225920))))[name = string("layers_5_mlp_gate_proj_weight_to_fp16_quantized")]; + tensor linear_44_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor var_5701_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("op_5701_cast_fp16")]; + tensor layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640244416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663837440))))[name = string("layers_5_mlp_up_proj_weight_to_fp16_quantized")]; + tensor linear_45_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_45_cast_fp16")]; + tensor input_87_cast_fp16 = mul(x = var_5701_cast_fp16, y = linear_45_cast_fp16)[name = string("input_87_cast_fp16")]; + tensor layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663855936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687448960))))[name = string("layers_5_mlp_down_proj_weight_to_fp16_quantized")]; + tensor linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor x_151_cast_fp16 = add(x = x_147_cast_fp16, y = linear_46_cast_fp16)[name = string("x_151_cast_fp16")]; + int32 var_5721 = const()[name = string("op_5721"), val = int32(-2)]; + int32 var_5750 = const()[name = string("op_5750"), val = int32(-1)]; + fp16 var_5749_promoted_to_fp16 = const()[name = string("op_5749_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_5759_cast_fp16 = pow(x = x_151_cast_fp16, y = var_5749_promoted_to_fp16)[name = string("op_5759_cast_fp16")]; + tensor var_5761_axes_0 = const()[name = string("op_5761_axes_0"), val = tensor([-1])]; + bool var_5761_keep_dims_0 = const()[name = string("op_5761_keep_dims_0"), val = bool(true)]; + tensor var_5761_cast_fp16 = reduce_mean(axes = var_5761_axes_0, keep_dims = var_5761_keep_dims_0, x = var_5759_cast_fp16)[name = string("op_5761_cast_fp16")]; + fp16 var_5762_to_fp16 = const()[name = string("op_5762_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_5763_cast_fp16 = add(x = var_5761_cast_fp16, y = var_5762_to_fp16)[name = string("op_5763_cast_fp16")]; + fp32 var_5764_epsilon_0 = const()[name = string("op_5764_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5764_cast_fp16 = rsqrt(epsilon = var_5764_epsilon_0, x = var_5763_cast_fp16)[name = string("op_5764_cast_fp16")]; + tensor x_153_cast_fp16 = mul(x = x_151_cast_fp16, y = var_5764_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor var_5767_to_fp16 = const()[name = string("op_5767_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687454144)))]; + tensor x_155_cast_fp16 = mul(x = x_153_cast_fp16, y = var_5767_to_fp16)[name = string("x_155_cast_fp16")]; + tensor layers_6_linear_attn_in_proj_qkv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687459328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708430912))))[name = string("layers_6_linear_attn_in_proj_qkv_weight_to_fp16_quantized")]; + tensor linear_47_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_linear_attn_in_proj_qkv_weight_to_fp16_quantized, x = x_155_cast_fp16)[name = string("linear_47_cast_fp16")]; + tensor input_89_perm_0 = const()[name = string("input_89_perm_0"), val = tensor([0, 2, 1])]; + tensor input_91_pad_0 = const()[name = string("input_91_pad_0"), val = tensor([0, 0, 0, 0, 3, 0])]; + string input_91_mode_0 = const()[name = string("input_91_mode_0"), val = string("constant")]; + fp16 const_66_to_fp16 = const()[name = string("const_66_to_fp16"), val = fp16(0x0p+0)]; + tensor input_89_cast_fp16 = transpose(perm = input_89_perm_0, x = linear_47_cast_fp16)[name = string("transpose_11")]; + tensor input_91_cast_fp16 = pad(constant_val = const_66_to_fp16, mode = input_91_mode_0, pad = input_91_pad_0, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")]; + string input_93_pad_type_0 = const()[name = string("input_93_pad_type_0"), val = string("valid")]; + int32 input_93_groups_0 = const()[name = string("input_93_groups_0"), val = int32(8192)]; + tensor input_93_strides_0 = const()[name = string("input_93_strides_0"), val = tensor([1])]; + tensor input_93_pad_0 = const()[name = string("input_93_pad_0"), val = tensor([0, 0])]; + tensor input_93_dilations_0 = const()[name = string("input_93_dilations_0"), val = tensor([1])]; + tensor layers_6_linear_attn_conv1d_weight_to_fp16 = const()[name = string("layers_6_linear_attn_conv1d_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708447360)))]; + tensor input_93_cast_fp16 = conv(dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = layers_6_linear_attn_conv1d_weight_to_fp16, x = input_91_cast_fp16)[name = string("input_93_cast_fp16")]; + tensor var_5802_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("op_5802_cast_fp16")]; + tensor qkv_perm_0 = const()[name = string("qkv_perm_0"), val = tensor([0, 2, 1])]; + tensor var_5804_begin_0 = const()[name = string("op_5804_begin_0"), val = tensor([0, 0, 0])]; + tensor var_5804_end_0 = const()[name = string("op_5804_end_0"), val = tensor([1, 2048, 2048])]; + tensor var_5804_end_mask_0 = const()[name = string("op_5804_end_mask_0"), val = tensor([true, true, false])]; + tensor qkv_cast_fp16 = transpose(perm = qkv_perm_0, x = var_5802_cast_fp16)[name = string("transpose_10")]; + tensor var_5804_cast_fp16 = slice_by_index(begin = var_5804_begin_0, end = var_5804_end_0, end_mask = var_5804_end_mask_0, x = qkv_cast_fp16)[name = string("op_5804_cast_fp16")]; + tensor var_5805 = const()[name = string("op_5805"), val = tensor([1, 2048, 16, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_5805, x = var_5804_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor var_5807_begin_0 = const()[name = string("op_5807_begin_0"), val = tensor([0, 0, 2048])]; + tensor var_5807_end_0 = const()[name = string("op_5807_end_0"), val = tensor([1, 2048, 4096])]; + tensor var_5807_end_mask_0 = const()[name = string("op_5807_end_mask_0"), val = tensor([true, true, false])]; + tensor var_5807_cast_fp16 = slice_by_index(begin = var_5807_begin_0, end = var_5807_end_0, end_mask = var_5807_end_mask_0, x = qkv_cast_fp16)[name = string("op_5807_cast_fp16")]; + tensor var_5808 = const()[name = string("op_5808"), val = tensor([1, 2048, 16, 128])]; + tensor k_47_cast_fp16 = reshape(shape = var_5808, x = var_5807_cast_fp16)[name = string("k_47_cast_fp16")]; + tensor var_5810_begin_0 = const()[name = string("op_5810_begin_0"), val = tensor([0, 0, 4096])]; + tensor var_5810_end_0 = const()[name = string("op_5810_end_0"), val = tensor([1, 2048, 8192])]; + tensor var_5810_end_mask_0 = const()[name = string("op_5810_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5810_cast_fp16 = slice_by_index(begin = var_5810_begin_0, end = var_5810_end_0, end_mask = var_5810_end_mask_0, x = qkv_cast_fp16)[name = string("op_5810_cast_fp16")]; + tensor var_5811 = const()[name = string("op_5811"), val = tensor([1, 2048, 32, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_5811, x = var_5810_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor layers_6_linear_attn_in_proj_z_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708512960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(718998784))))[name = string("layers_6_linear_attn_in_proj_z_weight_to_fp16_quantized")]; + tensor linear_48_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_linear_attn_in_proj_z_weight_to_fp16_quantized, x = x_155_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_5815 = const()[name = string("op_5815"), val = tensor([1, 2048, 32, 128])]; + tensor z_cast_fp16 = reshape(shape = var_5815, x = linear_48_cast_fp16)[name = string("z_cast_fp16")]; + tensor layers_6_linear_attn_in_proj_b_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719007040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719089024))))[name = string("layers_6_linear_attn_in_proj_b_weight_to_fp16_quantized")]; + tensor linear_49_bias_0_to_fp16 = const()[name = string("linear_49_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719089152)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_49_bias_0_to_fp16, weight = layers_6_linear_attn_in_proj_b_weight_to_fp16_quantized, x = x_155_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor beta_21_cast_fp16 = sigmoid(x = linear_49_cast_fp16)[name = string("beta_21_cast_fp16")]; + tensor var_5824_weight_0_to_fp16 = const()[name = string("op_5824_weight_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719089280)))]; + tensor var_5824_bias_0_to_fp16 = const()[name = string("op_5824_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719253184)))]; + tensor var_5824_cast_fp16 = linear(bias = var_5824_bias_0_to_fp16, weight = var_5824_weight_0_to_fp16, x = x_155_cast_fp16)[name = string("op_5824_cast_fp16")]; + tensor var_5825_cast_fp16 = softplus(x = var_5824_cast_fp16)[name = string("op_5825_cast_fp16")]; + tensor var_5821_to_fp16 = const()[name = string("op_5821_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719253312)))]; + tensor g_21_cast_fp16 = mul(x = var_5821_to_fp16, y = var_5825_cast_fp16)[name = string("g_21_cast_fp16")]; + tensor var_5827_cast_fp16 = mul(x = q_55_cast_fp16, y = q_55_cast_fp16)[name = string("op_5827_cast_fp16")]; + tensor var_5829_axes_0 = const()[name = string("op_5829_axes_0"), val = tensor([-1])]; + bool var_5829_keep_dims_0 = const()[name = string("op_5829_keep_dims_0"), val = bool(true)]; + tensor var_5829_cast_fp16 = reduce_sum(axes = var_5829_axes_0, keep_dims = var_5829_keep_dims_0, x = var_5827_cast_fp16)[name = string("op_5829_cast_fp16")]; + fp16 var_5830_to_fp16 = const()[name = string("op_5830_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_5831_cast_fp16 = add(x = var_5829_cast_fp16, y = var_5830_to_fp16)[name = string("op_5831_cast_fp16")]; + fp32 var_5832_epsilon_0 = const()[name = string("op_5832_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5832_cast_fp16 = rsqrt(epsilon = var_5832_epsilon_0, x = var_5831_cast_fp16)[name = string("op_5832_cast_fp16")]; + tensor q_57_cast_fp16 = mul(x = q_55_cast_fp16, y = var_5832_cast_fp16)[name = string("q_57_cast_fp16")]; + tensor var_5834_cast_fp16 = mul(x = k_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_5834_cast_fp16")]; + tensor var_5836_axes_0 = const()[name = string("op_5836_axes_0"), val = tensor([-1])]; + bool var_5836_keep_dims_0 = const()[name = string("op_5836_keep_dims_0"), val = bool(true)]; + tensor var_5836_cast_fp16 = reduce_sum(axes = var_5836_axes_0, keep_dims = var_5836_keep_dims_0, x = var_5834_cast_fp16)[name = string("op_5836_cast_fp16")]; + fp16 var_5837_to_fp16 = const()[name = string("op_5837_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_5838_cast_fp16 = add(x = var_5836_cast_fp16, y = var_5837_to_fp16)[name = string("op_5838_cast_fp16")]; + fp32 var_5839_epsilon_0 = const()[name = string("op_5839_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5839_cast_fp16 = rsqrt(epsilon = var_5839_epsilon_0, x = var_5838_cast_fp16)[name = string("op_5839_cast_fp16")]; + tensor k_49_cast_fp16 = mul(x = k_47_cast_fp16, y = var_5839_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor var_5844_axes_0 = const()[name = string("op_5844_axes_0"), val = tensor([3])]; + tensor var_5844_cast_fp16 = expand_dims(axes = var_5844_axes_0, x = q_57_cast_fp16)[name = string("op_5844_cast_fp16")]; + tensor var_5846_reps_0 = const()[name = string("op_5846_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_5846_cast_fp16 = tile(reps = var_5846_reps_0, x = var_5844_cast_fp16)[name = string("op_5846_cast_fp16")]; + tensor var_5847 = const()[name = string("op_5847"), val = tensor([1, 2048, 32, 128])]; + tensor q_59_cast_fp16 = reshape(shape = var_5847, x = var_5846_cast_fp16)[name = string("q_59_cast_fp16")]; + tensor var_5852_axes_0 = const()[name = string("op_5852_axes_0"), val = tensor([3])]; + tensor var_5852_cast_fp16 = expand_dims(axes = var_5852_axes_0, x = k_49_cast_fp16)[name = string("op_5852_cast_fp16")]; + tensor var_5854_reps_0 = const()[name = string("op_5854_reps_0"), val = tensor([1, 1, 1, 2, 1])]; + tensor var_5854_cast_fp16 = tile(reps = var_5854_reps_0, x = var_5852_cast_fp16)[name = string("op_5854_cast_fp16")]; + tensor var_5855 = const()[name = string("op_5855"), val = tensor([1, 2048, 32, 128])]; + tensor k_51_cast_fp16 = reshape(shape = var_5855, x = var_5854_cast_fp16)[name = string("k_51_cast_fp16")]; + fp16 var_5857_to_fp16 = const()[name = string("op_5857_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor q_61_cast_fp16 = mul(x = q_59_cast_fp16, y = var_5857_to_fp16)[name = string("q_61_cast_fp16")]; + tensor var_5859_begin_0 = const()[name = string("op_5859_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5859_end_0 = const()[name = string("op_5859_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_5859_end_mask_0 = const()[name = string("op_5859_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_5859_squeeze_mask_0 = const()[name = string("op_5859_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_5859_cast_fp16 = slice_by_index(begin = var_5859_begin_0, end = var_5859_end_0, end_mask = var_5859_end_mask_0, squeeze_mask = var_5859_squeeze_mask_0, x = q_61_cast_fp16)[name = string("op_5859_cast_fp16")]; + tensor var_5860 = const()[name = string("op_5860"), val = tensor([1, 0, 2])]; + tensor var_5862 = const()[name = string("op_5862"), val = tensor([32, 32, 64, 128])]; + tensor var_5861_cast_fp16 = transpose(perm = var_5860, x = var_5859_cast_fp16)[name = string("transpose_9")]; + tensor q_63_cast_fp16 = reshape(shape = var_5862, x = var_5861_cast_fp16)[name = string("q_63_cast_fp16")]; + tensor var_5864_begin_0 = const()[name = string("op_5864_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5864_end_0 = const()[name = string("op_5864_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_5864_end_mask_0 = const()[name = string("op_5864_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_5864_squeeze_mask_0 = const()[name = string("op_5864_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_5864_cast_fp16 = slice_by_index(begin = var_5864_begin_0, end = var_5864_end_0, end_mask = var_5864_end_mask_0, squeeze_mask = var_5864_squeeze_mask_0, x = k_51_cast_fp16)[name = string("op_5864_cast_fp16")]; + tensor var_5865 = const()[name = string("op_5865"), val = tensor([1, 0, 2])]; + tensor var_5867 = const()[name = string("op_5867"), val = tensor([32, 32, 64, 128])]; + tensor var_5866_cast_fp16 = transpose(perm = var_5865, x = var_5864_cast_fp16)[name = string("transpose_8")]; + tensor k_53_cast_fp16 = reshape(shape = var_5867, x = var_5866_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor var_5869_begin_0 = const()[name = string("op_5869_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5869_end_0 = const()[name = string("op_5869_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor var_5869_end_mask_0 = const()[name = string("op_5869_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_5869_squeeze_mask_0 = const()[name = string("op_5869_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor var_5869_cast_fp16 = slice_by_index(begin = var_5869_begin_0, end = var_5869_end_0, end_mask = var_5869_end_mask_0, squeeze_mask = var_5869_squeeze_mask_0, x = v_25_cast_fp16)[name = string("op_5869_cast_fp16")]; + tensor var_5870 = const()[name = string("op_5870"), val = tensor([1, 0, 2])]; + tensor var_5872 = const()[name = string("op_5872"), val = tensor([32, 32, 64, 128])]; + tensor var_5871_cast_fp16 = transpose(perm = var_5870, x = var_5869_cast_fp16)[name = string("transpose_7")]; + tensor v_27_cast_fp16 = reshape(shape = var_5872, x = var_5871_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor var_5874_begin_0 = const()[name = string("op_5874_begin_0"), val = tensor([0, 0, 0])]; + tensor var_5874_end_0 = const()[name = string("op_5874_end_0"), val = tensor([1, 2048, 32])]; + tensor var_5874_end_mask_0 = const()[name = string("op_5874_end_mask_0"), val = tensor([false, true, true])]; + tensor var_5874_squeeze_mask_0 = const()[name = string("op_5874_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_5874_cast_fp16 = slice_by_index(begin = var_5874_begin_0, end = var_5874_end_0, end_mask = var_5874_end_mask_0, squeeze_mask = var_5874_squeeze_mask_0, x = beta_21_cast_fp16)[name = string("op_5874_cast_fp16")]; + tensor var_5875_perm_0 = const()[name = string("op_5875_perm_0"), val = tensor([1, 0])]; + tensor var_5876 = const()[name = string("op_5876"), val = tensor([32, 32, 64, 1])]; + tensor var_5875_cast_fp16 = transpose(perm = var_5875_perm_0, x = var_5874_cast_fp16)[name = string("transpose_6")]; + tensor beta_cast_fp16 = reshape(shape = var_5876, x = var_5875_cast_fp16)[name = string("beta_cast_fp16")]; + tensor var_5878_begin_0 = const()[name = string("op_5878_begin_0"), val = tensor([0, 0, 0])]; + tensor var_5878_end_0 = const()[name = string("op_5878_end_0"), val = tensor([1, 2048, 32])]; + tensor var_5878_end_mask_0 = const()[name = string("op_5878_end_mask_0"), val = tensor([false, true, true])]; + tensor var_5878_squeeze_mask_0 = const()[name = string("op_5878_squeeze_mask_0"), val = tensor([true, false, false])]; + tensor var_5878_cast_fp16 = slice_by_index(begin = var_5878_begin_0, end = var_5878_end_0, end_mask = var_5878_end_mask_0, squeeze_mask = var_5878_squeeze_mask_0, x = g_21_cast_fp16)[name = string("op_5878_cast_fp16")]; + tensor var_5879_perm_0 = const()[name = string("op_5879_perm_0"), val = tensor([1, 0])]; + tensor var_5880 = const()[name = string("op_5880"), val = tensor([32, 32, 64])]; + tensor var_5879_cast_fp16 = transpose(perm = var_5879_perm_0, x = var_5878_cast_fp16)[name = string("transpose_5")]; + tensor g_cast_fp16 = reshape(shape = var_5880, x = var_5879_cast_fp16)[name = string("g_cast_fp16")]; + tensor g_col_axes_0 = const()[name = string("g_col_axes_0"), val = tensor([-1])]; + tensor g_col_cast_fp16 = expand_dims(axes = g_col_axes_0, x = g_cast_fp16)[name = string("g_col_cast_fp16")]; + bool var_5883_transpose_x_0 = const()[name = string("op_5883_transpose_x_0"), val = bool(false)]; + bool var_5883_transpose_y_0 = const()[name = string("op_5883_transpose_y_0"), val = bool(false)]; + tensor var_5883_cast_fp16 = matmul(transpose_x = var_5883_transpose_x_0, transpose_y = var_5883_transpose_y_0, x = layers_0_linear_attn_lower_incl_to_fp16, y = g_col_cast_fp16)[name = string("op_5883_cast_fp16")]; + tensor cum_axes_0 = const()[name = string("cum_axes_0"), val = tensor([-1])]; + tensor cum_cast_fp16 = squeeze(axes = cum_axes_0, x = var_5883_cast_fp16)[name = string("cum_cast_fp16")]; + bool var_5885_transpose_x_0 = const()[name = string("op_5885_transpose_x_0"), val = bool(false)]; + bool var_5885_transpose_y_0 = const()[name = string("op_5885_transpose_y_0"), val = bool(false)]; + tensor var_5885_cast_fp16 = matmul(transpose_x = var_5885_transpose_x_0, transpose_y = var_5885_transpose_y_0, x = layers_0_linear_attn_suffix_sum_t_to_fp16, y = g_col_cast_fp16)[name = string("op_5885_cast_fp16")]; + tensor to_end_axes_0 = const()[name = string("to_end_axes_0"), val = tensor([-1])]; + tensor to_end_cast_fp16 = squeeze(axes = to_end_axes_0, x = var_5885_cast_fp16)[name = string("to_end_cast_fp16")]; + bool var_5887_transpose_x_0 = const()[name = string("op_5887_transpose_x_0"), val = bool(false)]; + bool var_5887_transpose_y_0 = const()[name = string("op_5887_transpose_y_0"), val = bool(false)]; + tensor var_5887_cast_fp16 = matmul(transpose_x = var_5887_transpose_x_0, transpose_y = var_5887_transpose_y_0, x = layers_0_linear_attn_pair_sum_t_to_fp16, y = g_col_cast_fp16)[name = string("op_5887_cast_fp16")]; + tensor var_5888 = const()[name = string("op_5888"), val = tensor([32, 32, 64, 64])]; + tensor pair_71_cast_fp16 = reshape(shape = var_5888, x = var_5887_cast_fp16)[name = string("pair_71_cast_fp16")]; + tensor var_5890_cast_fp16 = exp(x = pair_71_cast_fp16)[name = string("op_5890_cast_fp16")]; + tensor pair_decay_cast_fp16 = mul(x = var_5890_cast_fp16, y = layers_0_linear_attn_lower_incl_to_fp16)[name = string("pair_decay_cast_fp16")]; + tensor k_beta_cast_fp16 = mul(x = k_53_cast_fp16, y = beta_cast_fp16)[name = string("k_beta_cast_fp16")]; + tensor v_beta_cast_fp16 = mul(x = v_27_cast_fp16, y = beta_cast_fp16)[name = string("v_beta_cast_fp16")]; + bool var_5895_transpose_x_1 = const()[name = string("op_5895_transpose_x_1"), val = bool(false)]; + bool var_5895_transpose_y_1 = const()[name = string("op_5895_transpose_y_1"), val = bool(true)]; + tensor var_5895_cast_fp16 = matmul(transpose_x = var_5895_transpose_x_1, transpose_y = var_5895_transpose_y_1, x = k_beta_cast_fp16, y = k_53_cast_fp16)[name = string("op_5895_cast_fp16")]; + tensor kkt_cast_fp16 = mul(x = var_5895_cast_fp16, y = pair_decay_cast_fp16)[name = string("kkt_cast_fp16")]; + bool var_5898_transpose_x_1 = const()[name = string("op_5898_transpose_x_1"), val = bool(false)]; + bool var_5898_transpose_y_1 = const()[name = string("op_5898_transpose_y_1"), val = bool(true)]; + tensor var_5898_cast_fp16 = matmul(transpose_x = var_5898_transpose_x_1, transpose_y = var_5898_transpose_y_1, x = q_63_cast_fp16, y = k_53_cast_fp16)[name = string("op_5898_cast_fp16")]; + tensor attn_intra_cast_fp16 = mul(x = var_5898_cast_fp16, y = pair_decay_cast_fp16)[name = string("attn_intra_cast_fp16")]; + tensor var_5900_cast_fp16 = mul(x = kkt_cast_fp16, y = layers_0_linear_attn_strict_lower_to_fp16)[name = string("op_5900_cast_fp16")]; + tensor t_21_cast_fp16 = add(x = layers_0_linear_attn_eye_to_fp16, y = var_5900_cast_fp16)[name = string("t_21_cast_fp16")]; + tensor var_5904 = const()[name = string("op_5904"), val = tensor([-1, 64, 64])]; + tensor t_cast_fp16 = reshape(shape = var_5904, x = t_21_cast_fp16)[name = string("t_cast_fp16")]; + tensor var_5909 = const()[name = string("op_5909"), val = tensor([1024, 32, 2, 32, 2])]; + tensor var_5910_cast_fp16 = reshape(shape = var_5909, x = t_cast_fp16)[name = string("op_5910_cast_fp16")]; + tensor var_5915_cast_fp16 = mul(x = var_5910_cast_fp16, y = var_227_to_fp16)[name = string("op_5915_cast_fp16")]; + tensor blocks_61_axes_0 = const()[name = string("blocks_61_axes_0"), val = tensor([-2])]; + bool blocks_61_keep_dims_0 = const()[name = string("blocks_61_keep_dims_0"), val = bool(false)]; + tensor blocks_61_cast_fp16 = reduce_sum(axes = blocks_61_axes_0, keep_dims = blocks_61_keep_dims_0, x = var_5915_cast_fp16)[name = string("blocks_61_cast_fp16")]; + tensor var_5918_begin_0 = const()[name = string("op_5918_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5918_end_0 = const()[name = string("op_5918_end_0"), val = tensor([1024, 32, 2, 2])]; + tensor var_5918_end_mask_0 = const()[name = string("op_5918_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5918_cast_fp16 = slice_by_index(begin = var_5918_begin_0, end = var_5918_end_0, end_mask = var_5918_end_mask_0, x = blocks_61_cast_fp16)[name = string("op_5918_cast_fp16")]; + tensor x_157_begin_0 = const()[name = string("x_157_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_157_end_0 = const()[name = string("x_157_end_0"), val = tensor([1024, 32, 1, 1])]; + tensor x_157_end_mask_0 = const()[name = string("x_157_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_157_cast_fp16 = slice_by_index(begin = x_157_begin_0, end = x_157_end_0, end_mask = x_157_end_mask_0, x = var_5918_cast_fp16)[name = string("x_157_cast_fp16")]; + bool var_5928_transpose_x_0 = const()[name = string("op_5928_transpose_x_0"), val = bool(false)]; + bool var_5928_transpose_y_0 = const()[name = string("op_5928_transpose_y_0"), val = bool(false)]; + tensor var_5928_cast_fp16 = matmul(transpose_x = var_5928_transpose_x_0, transpose_y = var_5928_transpose_y_0, x = x_157_cast_fp16, y = var_235_to_fp16)[name = string("op_5928_cast_fp16")]; + bool var_5929_transpose_x_0 = const()[name = string("op_5929_transpose_x_0"), val = bool(false)]; + bool var_5929_transpose_y_0 = const()[name = string("op_5929_transpose_y_0"), val = bool(false)]; + tensor var_5929_cast_fp16 = matmul(transpose_x = var_5929_transpose_x_0, transpose_y = var_5929_transpose_y_0, x = var_235_to_fp16, y = var_5928_cast_fp16)[name = string("op_5929_cast_fp16")]; + fp16 const_71_promoted_to_fp16 = const()[name = string("const_71_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_61_cast_fp16 = mul(x = var_5929_cast_fp16, y = const_71_promoted_to_fp16)[name = string("lower_61_cast_fp16")]; + bool bottom_61_interleave_0 = const()[name = string("bottom_61_interleave_0"), val = bool(false)]; + tensor bottom_61_cast_fp16 = concat(axis = var_5750, interleave = bottom_61_interleave_0, values = (lower_61_cast_fp16, var_235_to_fp16))[name = string("bottom_61_cast_fp16")]; + bool inv_83_interleave_0 = const()[name = string("inv_83_interleave_0"), val = bool(false)]; + tensor inv_83_cast_fp16 = concat(axis = var_5721, interleave = inv_83_interleave_0, values = (top_1_to_fp16, bottom_61_cast_fp16))[name = string("inv_83_cast_fp16")]; + tensor var_5938 = const()[name = string("op_5938"), val = tensor([1024, 16, 4, 16, 4])]; + tensor var_5939_cast_fp16 = reshape(shape = var_5938, x = t_cast_fp16)[name = string("op_5939_cast_fp16")]; + tensor var_5944_cast_fp16 = mul(x = var_5939_cast_fp16, y = var_256_to_fp16)[name = string("op_5944_cast_fp16")]; + tensor blocks_63_axes_0 = const()[name = string("blocks_63_axes_0"), val = tensor([-2])]; + bool blocks_63_keep_dims_0 = const()[name = string("blocks_63_keep_dims_0"), val = bool(false)]; + tensor blocks_63_cast_fp16 = reduce_sum(axes = blocks_63_axes_0, keep_dims = blocks_63_keep_dims_0, x = var_5944_cast_fp16)[name = string("blocks_63_cast_fp16")]; + tensor var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor([0, 0, 2, 0])]; + tensor var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor([1024, 16, 4, 4])]; + tensor var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = blocks_63_cast_fp16)[name = string("op_5947_cast_fp16")]; + tensor x_159_begin_0 = const()[name = string("x_159_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_159_end_0 = const()[name = string("x_159_end_0"), val = tensor([1024, 16, 2, 2])]; + tensor x_159_end_mask_0 = const()[name = string("x_159_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_159_cast_fp16 = slice_by_index(begin = x_159_begin_0, end = x_159_end_0, end_mask = x_159_end_mask_0, x = var_5947_cast_fp16)[name = string("x_159_cast_fp16")]; + tensor var_5949 = const()[name = string("op_5949"), val = tensor([1024, 16, 2, 2, 2])]; + tensor pair_75_cast_fp16 = reshape(shape = var_5949, x = inv_83_cast_fp16)[name = string("pair_75_cast_fp16")]; + tensor var_5951_begin_0 = const()[name = string("op_5951_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5951_end_0 = const()[name = string("op_5951_end_0"), val = tensor([1024, 16, 1, 2, 2])]; + tensor var_5951_end_mask_0 = const()[name = string("op_5951_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_5951_squeeze_mask_0 = const()[name = string("op_5951_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_5951_cast_fp16 = slice_by_index(begin = var_5951_begin_0, end = var_5951_end_0, end_mask = var_5951_end_mask_0, squeeze_mask = var_5951_squeeze_mask_0, x = pair_75_cast_fp16)[name = string("op_5951_cast_fp16")]; + tensor var_5954_begin_0 = const()[name = string("op_5954_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_5954_end_0 = const()[name = string("op_5954_end_0"), val = tensor([1024, 16, 2, 2, 2])]; + tensor var_5954_end_mask_0 = const()[name = string("op_5954_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_5954_squeeze_mask_0 = const()[name = string("op_5954_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_5954_cast_fp16 = slice_by_index(begin = var_5954_begin_0, end = var_5954_end_0, end_mask = var_5954_end_mask_0, squeeze_mask = var_5954_squeeze_mask_0, x = pair_75_cast_fp16)[name = string("op_5954_cast_fp16")]; + bool var_5957_transpose_x_0 = const()[name = string("op_5957_transpose_x_0"), val = bool(false)]; + bool var_5957_transpose_y_0 = const()[name = string("op_5957_transpose_y_0"), val = bool(false)]; + tensor var_5957_cast_fp16 = matmul(transpose_x = var_5957_transpose_x_0, transpose_y = var_5957_transpose_y_0, x = x_159_cast_fp16, y = var_5951_cast_fp16)[name = string("op_5957_cast_fp16")]; + bool var_5958_transpose_x_0 = const()[name = string("op_5958_transpose_x_0"), val = bool(false)]; + bool var_5958_transpose_y_0 = const()[name = string("op_5958_transpose_y_0"), val = bool(false)]; + tensor var_5958_cast_fp16 = matmul(transpose_x = var_5958_transpose_x_0, transpose_y = var_5958_transpose_y_0, x = var_5954_cast_fp16, y = var_5957_cast_fp16)[name = string("op_5958_cast_fp16")]; + fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_63_cast_fp16 = mul(x = var_5958_cast_fp16, y = const_72_promoted_to_fp16)[name = string("lower_63_cast_fp16")]; + bool top_63_interleave_0 = const()[name = string("top_63_interleave_0"), val = bool(false)]; + tensor top_63_cast_fp16 = concat(axis = var_5750, interleave = top_63_interleave_0, values = (var_5951_cast_fp16, var_273_to_fp16))[name = string("top_63_cast_fp16")]; + bool bottom_63_interleave_0 = const()[name = string("bottom_63_interleave_0"), val = bool(false)]; + tensor bottom_63_cast_fp16 = concat(axis = var_5750, interleave = bottom_63_interleave_0, values = (lower_63_cast_fp16, var_5954_cast_fp16))[name = string("bottom_63_cast_fp16")]; + bool inv_85_interleave_0 = const()[name = string("inv_85_interleave_0"), val = bool(false)]; + tensor inv_85_cast_fp16 = concat(axis = var_5721, interleave = inv_85_interleave_0, values = (top_63_cast_fp16, bottom_63_cast_fp16))[name = string("inv_85_cast_fp16")]; + tensor var_5967 = const()[name = string("op_5967"), val = tensor([1024, 8, 8, 8, 8])]; + tensor var_5968_cast_fp16 = reshape(shape = var_5967, x = t_cast_fp16)[name = string("op_5968_cast_fp16")]; + tensor var_5972_to_fp16 = const()[name = string("op_5972_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719253440)))]; + tensor var_5973_cast_fp16 = mul(x = var_5968_cast_fp16, y = var_5972_to_fp16)[name = string("op_5973_cast_fp16")]; + tensor blocks_65_axes_0 = const()[name = string("blocks_65_axes_0"), val = tensor([-2])]; + bool blocks_65_keep_dims_0 = const()[name = string("blocks_65_keep_dims_0"), val = bool(false)]; + tensor blocks_65_cast_fp16 = reduce_sum(axes = blocks_65_axes_0, keep_dims = blocks_65_keep_dims_0, x = var_5973_cast_fp16)[name = string("blocks_65_cast_fp16")]; + tensor var_5976_begin_0 = const()[name = string("op_5976_begin_0"), val = tensor([0, 0, 4, 0])]; + tensor var_5976_end_0 = const()[name = string("op_5976_end_0"), val = tensor([1024, 8, 8, 8])]; + tensor var_5976_end_mask_0 = const()[name = string("op_5976_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5976_cast_fp16 = slice_by_index(begin = var_5976_begin_0, end = var_5976_end_0, end_mask = var_5976_end_mask_0, x = blocks_65_cast_fp16)[name = string("op_5976_cast_fp16")]; + tensor x_161_begin_0 = const()[name = string("x_161_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_161_end_0 = const()[name = string("x_161_end_0"), val = tensor([1024, 8, 4, 4])]; + tensor x_161_end_mask_0 = const()[name = string("x_161_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_161_cast_fp16 = slice_by_index(begin = x_161_begin_0, end = x_161_end_0, end_mask = x_161_end_mask_0, x = var_5976_cast_fp16)[name = string("x_161_cast_fp16")]; + tensor var_5978 = const()[name = string("op_5978"), val = tensor([1024, 8, 2, 4, 4])]; + tensor pair_77_cast_fp16 = reshape(shape = var_5978, x = inv_85_cast_fp16)[name = string("pair_77_cast_fp16")]; + tensor var_5980_begin_0 = const()[name = string("op_5980_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5980_end_0 = const()[name = string("op_5980_end_0"), val = tensor([1024, 8, 1, 4, 4])]; + tensor var_5980_end_mask_0 = const()[name = string("op_5980_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_5980_squeeze_mask_0 = const()[name = string("op_5980_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_5980_cast_fp16 = slice_by_index(begin = var_5980_begin_0, end = var_5980_end_0, end_mask = var_5980_end_mask_0, squeeze_mask = var_5980_squeeze_mask_0, x = pair_77_cast_fp16)[name = string("op_5980_cast_fp16")]; + tensor var_5983_begin_0 = const()[name = string("op_5983_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_5983_end_0 = const()[name = string("op_5983_end_0"), val = tensor([1024, 8, 2, 4, 4])]; + tensor var_5983_end_mask_0 = const()[name = string("op_5983_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_5983_squeeze_mask_0 = const()[name = string("op_5983_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_5983_cast_fp16 = slice_by_index(begin = var_5983_begin_0, end = var_5983_end_0, end_mask = var_5983_end_mask_0, squeeze_mask = var_5983_squeeze_mask_0, x = pair_77_cast_fp16)[name = string("op_5983_cast_fp16")]; + bool var_5986_transpose_x_0 = const()[name = string("op_5986_transpose_x_0"), val = bool(false)]; + bool var_5986_transpose_y_0 = const()[name = string("op_5986_transpose_y_0"), val = bool(false)]; + tensor var_5986_cast_fp16 = matmul(transpose_x = var_5986_transpose_x_0, transpose_y = var_5986_transpose_y_0, x = x_161_cast_fp16, y = var_5980_cast_fp16)[name = string("op_5986_cast_fp16")]; + bool var_5987_transpose_x_0 = const()[name = string("op_5987_transpose_x_0"), val = bool(false)]; + bool var_5987_transpose_y_0 = const()[name = string("op_5987_transpose_y_0"), val = bool(false)]; + tensor var_5987_cast_fp16 = matmul(transpose_x = var_5987_transpose_x_0, transpose_y = var_5987_transpose_y_0, x = var_5983_cast_fp16, y = var_5986_cast_fp16)[name = string("op_5987_cast_fp16")]; + fp16 const_73_promoted_to_fp16 = const()[name = string("const_73_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_65_cast_fp16 = mul(x = var_5987_cast_fp16, y = const_73_promoted_to_fp16)[name = string("lower_65_cast_fp16")]; + bool top_65_interleave_0 = const()[name = string("top_65_interleave_0"), val = bool(false)]; + tensor top_65_cast_fp16 = concat(axis = var_5750, interleave = top_65_interleave_0, values = (var_5980_cast_fp16, var_302_to_fp16))[name = string("top_65_cast_fp16")]; + bool bottom_65_interleave_0 = const()[name = string("bottom_65_interleave_0"), val = bool(false)]; + tensor bottom_65_cast_fp16 = concat(axis = var_5750, interleave = bottom_65_interleave_0, values = (lower_65_cast_fp16, var_5983_cast_fp16))[name = string("bottom_65_cast_fp16")]; + bool inv_87_interleave_0 = const()[name = string("inv_87_interleave_0"), val = bool(false)]; + tensor inv_87_cast_fp16 = concat(axis = var_5721, interleave = inv_87_interleave_0, values = (top_65_cast_fp16, bottom_65_cast_fp16))[name = string("inv_87_cast_fp16")]; + tensor var_5996 = const()[name = string("op_5996"), val = tensor([1024, 4, 16, 4, 16])]; + tensor var_5997_cast_fp16 = reshape(shape = var_5996, x = t_cast_fp16)[name = string("op_5997_cast_fp16")]; + tensor var_6001_to_fp16 = const()[name = string("op_6001_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719253632)))]; + tensor var_6002_cast_fp16 = mul(x = var_5997_cast_fp16, y = var_6001_to_fp16)[name = string("op_6002_cast_fp16")]; + tensor blocks_67_axes_0 = const()[name = string("blocks_67_axes_0"), val = tensor([-2])]; + bool blocks_67_keep_dims_0 = const()[name = string("blocks_67_keep_dims_0"), val = bool(false)]; + tensor blocks_67_cast_fp16 = reduce_sum(axes = blocks_67_axes_0, keep_dims = blocks_67_keep_dims_0, x = var_6002_cast_fp16)[name = string("blocks_67_cast_fp16")]; + tensor var_6005_begin_0 = const()[name = string("op_6005_begin_0"), val = tensor([0, 0, 8, 0])]; + tensor var_6005_end_0 = const()[name = string("op_6005_end_0"), val = tensor([1024, 4, 16, 16])]; + tensor var_6005_end_mask_0 = const()[name = string("op_6005_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6005_cast_fp16 = slice_by_index(begin = var_6005_begin_0, end = var_6005_end_0, end_mask = var_6005_end_mask_0, x = blocks_67_cast_fp16)[name = string("op_6005_cast_fp16")]; + tensor x_163_begin_0 = const()[name = string("x_163_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_163_end_0 = const()[name = string("x_163_end_0"), val = tensor([1024, 4, 8, 8])]; + tensor x_163_end_mask_0 = const()[name = string("x_163_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_163_cast_fp16 = slice_by_index(begin = x_163_begin_0, end = x_163_end_0, end_mask = x_163_end_mask_0, x = var_6005_cast_fp16)[name = string("x_163_cast_fp16")]; + tensor var_6007 = const()[name = string("op_6007"), val = tensor([1024, 4, 2, 8, 8])]; + tensor pair_79_cast_fp16 = reshape(shape = var_6007, x = inv_87_cast_fp16)[name = string("pair_79_cast_fp16")]; + tensor var_6009_begin_0 = const()[name = string("op_6009_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6009_end_0 = const()[name = string("op_6009_end_0"), val = tensor([1024, 4, 1, 8, 8])]; + tensor var_6009_end_mask_0 = const()[name = string("op_6009_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_6009_squeeze_mask_0 = const()[name = string("op_6009_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_6009_cast_fp16 = slice_by_index(begin = var_6009_begin_0, end = var_6009_end_0, end_mask = var_6009_end_mask_0, squeeze_mask = var_6009_squeeze_mask_0, x = pair_79_cast_fp16)[name = string("op_6009_cast_fp16")]; + tensor var_6012_begin_0 = const()[name = string("op_6012_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_6012_end_0 = const()[name = string("op_6012_end_0"), val = tensor([1024, 4, 2, 8, 8])]; + tensor var_6012_end_mask_0 = const()[name = string("op_6012_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_6012_squeeze_mask_0 = const()[name = string("op_6012_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_6012_cast_fp16 = slice_by_index(begin = var_6012_begin_0, end = var_6012_end_0, end_mask = var_6012_end_mask_0, squeeze_mask = var_6012_squeeze_mask_0, x = pair_79_cast_fp16)[name = string("op_6012_cast_fp16")]; + bool var_6015_transpose_x_0 = const()[name = string("op_6015_transpose_x_0"), val = bool(false)]; + bool var_6015_transpose_y_0 = const()[name = string("op_6015_transpose_y_0"), val = bool(false)]; + tensor var_6015_cast_fp16 = matmul(transpose_x = var_6015_transpose_x_0, transpose_y = var_6015_transpose_y_0, x = x_163_cast_fp16, y = var_6009_cast_fp16)[name = string("op_6015_cast_fp16")]; + bool var_6016_transpose_x_0 = const()[name = string("op_6016_transpose_x_0"), val = bool(false)]; + bool var_6016_transpose_y_0 = const()[name = string("op_6016_transpose_y_0"), val = bool(false)]; + tensor var_6016_cast_fp16 = matmul(transpose_x = var_6016_transpose_x_0, transpose_y = var_6016_transpose_y_0, x = var_6012_cast_fp16, y = var_6015_cast_fp16)[name = string("op_6016_cast_fp16")]; + fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_67_cast_fp16 = mul(x = var_6016_cast_fp16, y = const_74_promoted_to_fp16)[name = string("lower_67_cast_fp16")]; + bool top_67_interleave_0 = const()[name = string("top_67_interleave_0"), val = bool(false)]; + tensor top_67_cast_fp16 = concat(axis = var_5750, interleave = top_67_interleave_0, values = (var_6009_cast_fp16, var_331_to_fp16))[name = string("top_67_cast_fp16")]; + bool bottom_67_interleave_0 = const()[name = string("bottom_67_interleave_0"), val = bool(false)]; + tensor bottom_67_cast_fp16 = concat(axis = var_5750, interleave = bottom_67_interleave_0, values = (lower_67_cast_fp16, var_6012_cast_fp16))[name = string("bottom_67_cast_fp16")]; + bool inv_89_interleave_0 = const()[name = string("inv_89_interleave_0"), val = bool(false)]; + tensor inv_89_cast_fp16 = concat(axis = var_5721, interleave = inv_89_interleave_0, values = (top_67_cast_fp16, bottom_67_cast_fp16))[name = string("inv_89_cast_fp16")]; + tensor var_6025 = const()[name = string("op_6025"), val = tensor([1024, 2, 32, 2, 32])]; + tensor var_6026_cast_fp16 = reshape(shape = var_6025, x = t_cast_fp16)[name = string("op_6026_cast_fp16")]; + tensor var_6030_to_fp16 = const()[name = string("op_6030_to_fp16"), val = tensor([[[[0x1p+0], [0x0p+0]]], [[[0x0p+0], [0x1p+0]]]])]; + tensor var_6031_cast_fp16 = mul(x = var_6026_cast_fp16, y = var_6030_to_fp16)[name = string("op_6031_cast_fp16")]; + tensor blocks_69_axes_0 = const()[name = string("blocks_69_axes_0"), val = tensor([-2])]; + bool blocks_69_keep_dims_0 = const()[name = string("blocks_69_keep_dims_0"), val = bool(false)]; + tensor blocks_69_cast_fp16 = reduce_sum(axes = blocks_69_axes_0, keep_dims = blocks_69_keep_dims_0, x = var_6031_cast_fp16)[name = string("blocks_69_cast_fp16")]; + tensor var_6034_begin_0 = const()[name = string("op_6034_begin_0"), val = tensor([0, 0, 16, 0])]; + tensor var_6034_end_0 = const()[name = string("op_6034_end_0"), val = tensor([1024, 2, 32, 32])]; + tensor var_6034_end_mask_0 = const()[name = string("op_6034_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6034_cast_fp16 = slice_by_index(begin = var_6034_begin_0, end = var_6034_end_0, end_mask = var_6034_end_mask_0, x = blocks_69_cast_fp16)[name = string("op_6034_cast_fp16")]; + tensor x_165_begin_0 = const()[name = string("x_165_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_165_end_0 = const()[name = string("x_165_end_0"), val = tensor([1024, 2, 16, 16])]; + tensor x_165_end_mask_0 = const()[name = string("x_165_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_165_cast_fp16 = slice_by_index(begin = x_165_begin_0, end = x_165_end_0, end_mask = x_165_end_mask_0, x = var_6034_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_6036 = const()[name = string("op_6036"), val = tensor([1024, 2, 2, 16, 16])]; + tensor pair_81_cast_fp16 = reshape(shape = var_6036, x = inv_89_cast_fp16)[name = string("pair_81_cast_fp16")]; + tensor var_6038_begin_0 = const()[name = string("op_6038_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6038_end_0 = const()[name = string("op_6038_end_0"), val = tensor([1024, 2, 1, 16, 16])]; + tensor var_6038_end_mask_0 = const()[name = string("op_6038_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_6038_squeeze_mask_0 = const()[name = string("op_6038_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_6038_cast_fp16 = slice_by_index(begin = var_6038_begin_0, end = var_6038_end_0, end_mask = var_6038_end_mask_0, squeeze_mask = var_6038_squeeze_mask_0, x = pair_81_cast_fp16)[name = string("op_6038_cast_fp16")]; + tensor var_6041_begin_0 = const()[name = string("op_6041_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_6041_end_0 = const()[name = string("op_6041_end_0"), val = tensor([1024, 2, 2, 16, 16])]; + tensor var_6041_end_mask_0 = const()[name = string("op_6041_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_6041_squeeze_mask_0 = const()[name = string("op_6041_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_6041_cast_fp16 = slice_by_index(begin = var_6041_begin_0, end = var_6041_end_0, end_mask = var_6041_end_mask_0, squeeze_mask = var_6041_squeeze_mask_0, x = pair_81_cast_fp16)[name = string("op_6041_cast_fp16")]; + bool var_6044_transpose_x_0 = const()[name = string("op_6044_transpose_x_0"), val = bool(false)]; + bool var_6044_transpose_y_0 = const()[name = string("op_6044_transpose_y_0"), val = bool(false)]; + tensor var_6044_cast_fp16 = matmul(transpose_x = var_6044_transpose_x_0, transpose_y = var_6044_transpose_y_0, x = x_165_cast_fp16, y = var_6038_cast_fp16)[name = string("op_6044_cast_fp16")]; + bool var_6045_transpose_x_0 = const()[name = string("op_6045_transpose_x_0"), val = bool(false)]; + bool var_6045_transpose_y_0 = const()[name = string("op_6045_transpose_y_0"), val = bool(false)]; + tensor var_6045_cast_fp16 = matmul(transpose_x = var_6045_transpose_x_0, transpose_y = var_6045_transpose_y_0, x = var_6041_cast_fp16, y = var_6044_cast_fp16)[name = string("op_6045_cast_fp16")]; + fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_69_cast_fp16 = mul(x = var_6045_cast_fp16, y = const_75_promoted_to_fp16)[name = string("lower_69_cast_fp16")]; + bool top_69_interleave_0 = const()[name = string("top_69_interleave_0"), val = bool(false)]; + tensor top_69_cast_fp16 = concat(axis = var_5750, interleave = top_69_interleave_0, values = (var_6038_cast_fp16, var_360_to_fp16))[name = string("top_69_cast_fp16")]; + bool bottom_69_interleave_0 = const()[name = string("bottom_69_interleave_0"), val = bool(false)]; + tensor bottom_69_cast_fp16 = concat(axis = var_5750, interleave = bottom_69_interleave_0, values = (lower_69_cast_fp16, var_6041_cast_fp16))[name = string("bottom_69_cast_fp16")]; + bool inv_91_interleave_0 = const()[name = string("inv_91_interleave_0"), val = bool(false)]; + tensor inv_91_cast_fp16 = concat(axis = var_5721, interleave = inv_91_interleave_0, values = (top_69_cast_fp16, bottom_69_cast_fp16))[name = string("inv_91_cast_fp16")]; + tensor var_6054 = const()[name = string("op_6054"), val = tensor([1024, 1, 64, 1, 64])]; + tensor var_6055_cast_fp16 = reshape(shape = var_6054, x = t_cast_fp16)[name = string("op_6055_cast_fp16")]; + tensor blocks_axes_0 = const()[name = string("blocks_axes_0"), val = tensor([-2])]; + bool blocks_keep_dims_0 = const()[name = string("blocks_keep_dims_0"), val = bool(false)]; + tensor blocks_cast_fp16 = reduce_sum(axes = blocks_axes_0, keep_dims = blocks_keep_dims_0, x = var_6055_cast_fp16)[name = string("blocks_cast_fp16")]; + tensor var_6063_begin_0 = const()[name = string("op_6063_begin_0"), val = tensor([0, 0, 32, 0])]; + tensor var_6063_end_0 = const()[name = string("op_6063_end_0"), val = tensor([1024, 1, 64, 64])]; + tensor var_6063_end_mask_0 = const()[name = string("op_6063_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6063_cast_fp16 = slice_by_index(begin = var_6063_begin_0, end = var_6063_end_0, end_mask = var_6063_end_mask_0, x = blocks_cast_fp16)[name = string("op_6063_cast_fp16")]; + tensor x_167_begin_0 = const()[name = string("x_167_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_167_end_0 = const()[name = string("x_167_end_0"), val = tensor([1024, 1, 32, 32])]; + tensor x_167_end_mask_0 = const()[name = string("x_167_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_167_cast_fp16 = slice_by_index(begin = x_167_begin_0, end = x_167_end_0, end_mask = x_167_end_mask_0, x = var_6063_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_6065 = const()[name = string("op_6065"), val = tensor([1024, 1, 2, 32, 32])]; + tensor pair_cast_fp16 = reshape(shape = var_6065, x = inv_91_cast_fp16)[name = string("pair_cast_fp16")]; + tensor var_6067_begin_0 = const()[name = string("op_6067_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6067_end_0 = const()[name = string("op_6067_end_0"), val = tensor([1024, 1, 1, 32, 32])]; + tensor var_6067_end_mask_0 = const()[name = string("op_6067_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_6067_squeeze_mask_0 = const()[name = string("op_6067_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_6067_cast_fp16 = slice_by_index(begin = var_6067_begin_0, end = var_6067_end_0, end_mask = var_6067_end_mask_0, squeeze_mask = var_6067_squeeze_mask_0, x = pair_cast_fp16)[name = string("op_6067_cast_fp16")]; + tensor var_6070_begin_0 = const()[name = string("op_6070_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor var_6070_end_0 = const()[name = string("op_6070_end_0"), val = tensor([1024, 1, 2, 32, 32])]; + tensor var_6070_end_mask_0 = const()[name = string("op_6070_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor var_6070_squeeze_mask_0 = const()[name = string("op_6070_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor var_6070_cast_fp16 = slice_by_index(begin = var_6070_begin_0, end = var_6070_end_0, end_mask = var_6070_end_mask_0, squeeze_mask = var_6070_squeeze_mask_0, x = pair_cast_fp16)[name = string("op_6070_cast_fp16")]; + bool var_6073_transpose_x_0 = const()[name = string("op_6073_transpose_x_0"), val = bool(false)]; + bool var_6073_transpose_y_0 = const()[name = string("op_6073_transpose_y_0"), val = bool(false)]; + tensor var_6073_cast_fp16 = matmul(transpose_x = var_6073_transpose_x_0, transpose_y = var_6073_transpose_y_0, x = x_167_cast_fp16, y = var_6067_cast_fp16)[name = string("op_6073_cast_fp16")]; + bool var_6074_transpose_x_0 = const()[name = string("op_6074_transpose_x_0"), val = bool(false)]; + bool var_6074_transpose_y_0 = const()[name = string("op_6074_transpose_y_0"), val = bool(false)]; + tensor var_6074_cast_fp16 = matmul(transpose_x = var_6074_transpose_x_0, transpose_y = var_6074_transpose_y_0, x = var_6070_cast_fp16, y = var_6073_cast_fp16)[name = string("op_6074_cast_fp16")]; + fp16 const_76_promoted_to_fp16 = const()[name = string("const_76_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor lower_cast_fp16 = mul(x = var_6074_cast_fp16, y = const_76_promoted_to_fp16)[name = string("lower_cast_fp16")]; + bool top_interleave_0 = const()[name = string("top_interleave_0"), val = bool(false)]; + tensor top_cast_fp16 = concat(axis = var_5750, interleave = top_interleave_0, values = (var_6067_cast_fp16, var_389_to_fp16))[name = string("top_cast_fp16")]; + bool bottom_interleave_0 = const()[name = string("bottom_interleave_0"), val = bool(false)]; + tensor bottom_cast_fp16 = concat(axis = var_5750, interleave = bottom_interleave_0, values = (lower_cast_fp16, var_6070_cast_fp16))[name = string("bottom_cast_fp16")]; + bool inv_93_interleave_0 = const()[name = string("inv_93_interleave_0"), val = bool(false)]; + tensor inv_93_cast_fp16 = concat(axis = var_5721, interleave = inv_93_interleave_0, values = (top_cast_fp16, bottom_cast_fp16))[name = string("inv_93_cast_fp16")]; + tensor var_6083 = const()[name = string("op_6083"), val = tensor([32, 32, 64, 64])]; + tensor inv_cast_fp16 = reshape(shape = var_6083, x = inv_93_cast_fp16)[name = string("inv_cast_fp16")]; + bool new_v_transpose_x_0 = const()[name = string("new_v_transpose_x_0"), val = bool(false)]; + bool new_v_transpose_y_0 = const()[name = string("new_v_transpose_y_0"), val = bool(false)]; + tensor new_v_cast_fp16 = matmul(transpose_x = new_v_transpose_x_0, transpose_y = new_v_transpose_y_0, x = inv_cast_fp16, y = v_beta_cast_fp16)[name = string("new_v_cast_fp16")]; + tensor var_6086_cast_fp16 = exp(x = cum_cast_fp16)[name = string("op_6086_cast_fp16")]; + tensor var_6087_axes_0 = const()[name = string("op_6087_axes_0"), val = tensor([3])]; + tensor var_6087_cast_fp16 = expand_dims(axes = var_6087_axes_0, x = var_6086_cast_fp16)[name = string("op_6087_cast_fp16")]; + tensor var_6088_cast_fp16 = mul(x = k_beta_cast_fp16, y = var_6087_cast_fp16)[name = string("op_6088_cast_fp16")]; + bool k_cumdecay_transpose_x_0 = const()[name = string("k_cumdecay_transpose_x_0"), val = bool(false)]; + bool k_cumdecay_transpose_y_0 = const()[name = string("k_cumdecay_transpose_y_0"), val = bool(false)]; + tensor k_cumdecay_cast_fp16 = matmul(transpose_x = k_cumdecay_transpose_x_0, transpose_y = k_cumdecay_transpose_y_0, x = inv_cast_fp16, y = var_6088_cast_fp16)[name = string("k_cumdecay_cast_fp16")]; + tensor q_dec_cast_fp16 = mul(x = q_63_cast_fp16, y = var_6087_cast_fp16)[name = string("q_dec_cast_fp16")]; + tensor var_6093_cast_fp16 = exp(x = to_end_cast_fp16)[name = string("op_6093_cast_fp16")]; + tensor var_6094_axes_0 = const()[name = string("op_6094_axes_0"), val = tensor([3])]; + tensor var_6094_cast_fp16 = expand_dims(axes = var_6094_axes_0, x = var_6093_cast_fp16)[name = string("op_6094_cast_fp16")]; + tensor k_dec_cast_fp16 = mul(x = k_53_cast_fp16, y = var_6094_cast_fp16)[name = string("k_dec_cast_fp16")]; + tensor var_6096_begin_0 = const()[name = string("op_6096_begin_0"), val = tensor([0, 0, -1])]; + tensor var_6096_end_0 = const()[name = string("op_6096_end_0"), val = tensor([32, 32, 64])]; + tensor var_6096_end_mask_0 = const()[name = string("op_6096_end_mask_0"), val = tensor([true, true, true])]; + tensor var_6096_cast_fp16 = slice_by_index(begin = var_6096_begin_0, end = var_6096_end_0, end_mask = var_6096_end_mask_0, x = cum_cast_fp16)[name = string("op_6096_cast_fp16")]; + tensor var_6097_cast_fp16 = exp(x = var_6096_cast_fp16)[name = string("op_6097_cast_fp16")]; + tensor chunk_decay_axes_0 = const()[name = string("chunk_decay_axes_0"), val = tensor([3])]; + tensor chunk_decay_cast_fp16 = expand_dims(axes = chunk_decay_axes_0, x = var_6097_cast_fp16)[name = string("chunk_decay_cast_fp16")]; + tensor var_6102_begin_0 = const()[name = string("op_6102_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6102_end_0 = const()[name = string("op_6102_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_6102_end_mask_0 = const()[name = string("op_6102_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6102_squeeze_mask_0 = const()[name = string("op_6102_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6102_cast_fp16 = slice_by_index(begin = var_6102_begin_0, end = var_6102_end_0, end_mask = var_6102_end_mask_0, squeeze_mask = var_6102_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6102_cast_fp16")]; + tensor var_6104_begin_0 = const()[name = string("op_6104_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6104_end_0 = const()[name = string("op_6104_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_6104_end_mask_0 = const()[name = string("op_6104_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6104_squeeze_mask_0 = const()[name = string("op_6104_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6104_cast_fp16 = slice_by_index(begin = var_6104_begin_0, end = var_6104_end_0, end_mask = var_6104_end_mask_0, squeeze_mask = var_6104_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6104_cast_fp16")]; + bool var_6105_transpose_x_0 = const()[name = string("op_6105_transpose_x_0"), val = bool(false)]; + bool var_6105_transpose_y_0 = const()[name = string("op_6105_transpose_y_0"), val = bool(false)]; + tensor var_6105_cast_fp16 = matmul(transpose_x = var_6105_transpose_x_0, transpose_y = var_6105_transpose_y_0, x = var_6104_cast_fp16, y = state_1_to_fp16)[name = string("op_6105_cast_fp16")]; + tensor v_new_321_cast_fp16 = sub(x = var_6102_cast_fp16, y = var_6105_cast_fp16)[name = string("v_new_321_cast_fp16")]; + tensor var_6108_begin_0 = const()[name = string("op_6108_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6108_end_0 = const()[name = string("op_6108_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_6108_end_mask_0 = const()[name = string("op_6108_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6108_squeeze_mask_0 = const()[name = string("op_6108_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6108_cast_fp16 = slice_by_index(begin = var_6108_begin_0, end = var_6108_end_0, end_mask = var_6108_end_mask_0, squeeze_mask = var_6108_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6108_cast_fp16")]; + bool var_6109_transpose_x_0 = const()[name = string("op_6109_transpose_x_0"), val = bool(false)]; + bool var_6109_transpose_y_0 = const()[name = string("op_6109_transpose_y_0"), val = bool(false)]; + tensor var_6109_cast_fp16 = matmul(transpose_x = var_6109_transpose_x_0, transpose_y = var_6109_transpose_y_0, x = var_6108_cast_fp16, y = state_1_to_fp16)[name = string("op_6109_cast_fp16")]; + tensor var_6111_begin_0 = const()[name = string("op_6111_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6111_end_0 = const()[name = string("op_6111_end_0"), val = tensor([32, 1, 64, 64])]; + tensor var_6111_end_mask_0 = const()[name = string("op_6111_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6111_squeeze_mask_0 = const()[name = string("op_6111_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6111_cast_fp16 = slice_by_index(begin = var_6111_begin_0, end = var_6111_end_0, end_mask = var_6111_end_mask_0, squeeze_mask = var_6111_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6111_cast_fp16")]; + bool var_6112_transpose_x_0 = const()[name = string("op_6112_transpose_x_0"), val = bool(false)]; + bool var_6112_transpose_y_0 = const()[name = string("op_6112_transpose_y_0"), val = bool(false)]; + tensor var_6112_cast_fp16 = matmul(transpose_x = var_6112_transpose_x_0, transpose_y = var_6112_transpose_y_0, x = var_6111_cast_fp16, y = v_new_321_cast_fp16)[name = string("op_6112_cast_fp16")]; + tensor var_6113_cast_fp16 = add(x = var_6109_cast_fp16, y = var_6112_cast_fp16)[name = string("op_6113_cast_fp16")]; + tensor var_6115_begin_0 = const()[name = string("op_6115_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6115_end_0 = const()[name = string("op_6115_end_0"), val = tensor([32, 1, 1, 1])]; + tensor var_6115_end_mask_0 = const()[name = string("op_6115_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6115_squeeze_mask_0 = const()[name = string("op_6115_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6115_cast_fp16 = slice_by_index(begin = var_6115_begin_0, end = var_6115_end_0, end_mask = var_6115_end_mask_0, squeeze_mask = var_6115_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6115_cast_fp16")]; + tensor var_6116_cast_fp16 = mul(x = state_1_to_fp16, y = var_6115_cast_fp16)[name = string("op_6116_cast_fp16")]; + tensor var_6118_begin_0 = const()[name = string("op_6118_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6118_end_0 = const()[name = string("op_6118_end_0"), val = tensor([32, 1, 64, 128])]; + tensor var_6118_end_mask_0 = const()[name = string("op_6118_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6118_squeeze_mask_0 = const()[name = string("op_6118_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6118_cast_fp16 = slice_by_index(begin = var_6118_begin_0, end = var_6118_end_0, end_mask = var_6118_end_mask_0, squeeze_mask = var_6118_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6118_cast_fp16")]; + bool var_6120_transpose_x_1 = const()[name = string("op_6120_transpose_x_1"), val = bool(true)]; + bool var_6120_transpose_y_1 = const()[name = string("op_6120_transpose_y_1"), val = bool(false)]; + tensor var_6120_cast_fp16 = matmul(transpose_x = var_6120_transpose_x_1, transpose_y = var_6120_transpose_y_1, x = var_6118_cast_fp16, y = v_new_321_cast_fp16)[name = string("op_6120_cast_fp16")]; + tensor state_323_cast_fp16 = add(x = var_6116_cast_fp16, y = var_6120_cast_fp16)[name = string("state_323_cast_fp16")]; + tensor var_6123_begin_0 = const()[name = string("op_6123_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_6123_end_0 = const()[name = string("op_6123_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_6123_end_mask_0 = const()[name = string("op_6123_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6123_squeeze_mask_0 = const()[name = string("op_6123_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6123_cast_fp16 = slice_by_index(begin = var_6123_begin_0, end = var_6123_end_0, end_mask = var_6123_end_mask_0, squeeze_mask = var_6123_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6123_cast_fp16")]; + tensor var_6125_begin_0 = const()[name = string("op_6125_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_6125_end_0 = const()[name = string("op_6125_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_6125_end_mask_0 = const()[name = string("op_6125_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6125_squeeze_mask_0 = const()[name = string("op_6125_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6125_cast_fp16 = slice_by_index(begin = var_6125_begin_0, end = var_6125_end_0, end_mask = var_6125_end_mask_0, squeeze_mask = var_6125_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6125_cast_fp16")]; + bool var_6126_transpose_x_0 = const()[name = string("op_6126_transpose_x_0"), val = bool(false)]; + bool var_6126_transpose_y_0 = const()[name = string("op_6126_transpose_y_0"), val = bool(false)]; + tensor var_6126_cast_fp16 = matmul(transpose_x = var_6126_transpose_x_0, transpose_y = var_6126_transpose_y_0, x = var_6125_cast_fp16, y = state_323_cast_fp16)[name = string("op_6126_cast_fp16")]; + tensor v_new_323_cast_fp16 = sub(x = var_6123_cast_fp16, y = var_6126_cast_fp16)[name = string("v_new_323_cast_fp16")]; + tensor var_6129_begin_0 = const()[name = string("op_6129_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_6129_end_0 = const()[name = string("op_6129_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_6129_end_mask_0 = const()[name = string("op_6129_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6129_squeeze_mask_0 = const()[name = string("op_6129_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6129_cast_fp16 = slice_by_index(begin = var_6129_begin_0, end = var_6129_end_0, end_mask = var_6129_end_mask_0, squeeze_mask = var_6129_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6129_cast_fp16")]; + bool var_6130_transpose_x_0 = const()[name = string("op_6130_transpose_x_0"), val = bool(false)]; + bool var_6130_transpose_y_0 = const()[name = string("op_6130_transpose_y_0"), val = bool(false)]; + tensor var_6130_cast_fp16 = matmul(transpose_x = var_6130_transpose_x_0, transpose_y = var_6130_transpose_y_0, x = var_6129_cast_fp16, y = state_323_cast_fp16)[name = string("op_6130_cast_fp16")]; + tensor var_6132_begin_0 = const()[name = string("op_6132_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_6132_end_0 = const()[name = string("op_6132_end_0"), val = tensor([32, 2, 64, 64])]; + tensor var_6132_end_mask_0 = const()[name = string("op_6132_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6132_squeeze_mask_0 = const()[name = string("op_6132_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6132_cast_fp16 = slice_by_index(begin = var_6132_begin_0, end = var_6132_end_0, end_mask = var_6132_end_mask_0, squeeze_mask = var_6132_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6132_cast_fp16")]; + bool var_6133_transpose_x_0 = const()[name = string("op_6133_transpose_x_0"), val = bool(false)]; + bool var_6133_transpose_y_0 = const()[name = string("op_6133_transpose_y_0"), val = bool(false)]; + tensor var_6133_cast_fp16 = matmul(transpose_x = var_6133_transpose_x_0, transpose_y = var_6133_transpose_y_0, x = var_6132_cast_fp16, y = v_new_323_cast_fp16)[name = string("op_6133_cast_fp16")]; + tensor var_6134_cast_fp16 = add(x = var_6130_cast_fp16, y = var_6133_cast_fp16)[name = string("op_6134_cast_fp16")]; + tensor var_6136_begin_0 = const()[name = string("op_6136_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_6136_end_0 = const()[name = string("op_6136_end_0"), val = tensor([32, 2, 1, 1])]; + tensor var_6136_end_mask_0 = const()[name = string("op_6136_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6136_squeeze_mask_0 = const()[name = string("op_6136_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6136_cast_fp16 = slice_by_index(begin = var_6136_begin_0, end = var_6136_end_0, end_mask = var_6136_end_mask_0, squeeze_mask = var_6136_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6136_cast_fp16")]; + tensor var_6137_cast_fp16 = mul(x = state_323_cast_fp16, y = var_6136_cast_fp16)[name = string("op_6137_cast_fp16")]; + tensor var_6139_begin_0 = const()[name = string("op_6139_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_6139_end_0 = const()[name = string("op_6139_end_0"), val = tensor([32, 2, 64, 128])]; + tensor var_6139_end_mask_0 = const()[name = string("op_6139_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6139_squeeze_mask_0 = const()[name = string("op_6139_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6139_cast_fp16 = slice_by_index(begin = var_6139_begin_0, end = var_6139_end_0, end_mask = var_6139_end_mask_0, squeeze_mask = var_6139_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6139_cast_fp16")]; + bool var_6141_transpose_x_1 = const()[name = string("op_6141_transpose_x_1"), val = bool(true)]; + bool var_6141_transpose_y_1 = const()[name = string("op_6141_transpose_y_1"), val = bool(false)]; + tensor var_6141_cast_fp16 = matmul(transpose_x = var_6141_transpose_x_1, transpose_y = var_6141_transpose_y_1, x = var_6139_cast_fp16, y = v_new_323_cast_fp16)[name = string("op_6141_cast_fp16")]; + tensor state_325_cast_fp16 = add(x = var_6137_cast_fp16, y = var_6141_cast_fp16)[name = string("state_325_cast_fp16")]; + tensor var_6144_begin_0 = const()[name = string("op_6144_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_6144_end_0 = const()[name = string("op_6144_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_6144_end_mask_0 = const()[name = string("op_6144_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6144_squeeze_mask_0 = const()[name = string("op_6144_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6144_cast_fp16 = slice_by_index(begin = var_6144_begin_0, end = var_6144_end_0, end_mask = var_6144_end_mask_0, squeeze_mask = var_6144_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6144_cast_fp16")]; + tensor var_6146_begin_0 = const()[name = string("op_6146_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_6146_end_0 = const()[name = string("op_6146_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_6146_end_mask_0 = const()[name = string("op_6146_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6146_squeeze_mask_0 = const()[name = string("op_6146_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6146_cast_fp16 = slice_by_index(begin = var_6146_begin_0, end = var_6146_end_0, end_mask = var_6146_end_mask_0, squeeze_mask = var_6146_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6146_cast_fp16")]; + bool var_6147_transpose_x_0 = const()[name = string("op_6147_transpose_x_0"), val = bool(false)]; + bool var_6147_transpose_y_0 = const()[name = string("op_6147_transpose_y_0"), val = bool(false)]; + tensor var_6147_cast_fp16 = matmul(transpose_x = var_6147_transpose_x_0, transpose_y = var_6147_transpose_y_0, x = var_6146_cast_fp16, y = state_325_cast_fp16)[name = string("op_6147_cast_fp16")]; + tensor v_new_325_cast_fp16 = sub(x = var_6144_cast_fp16, y = var_6147_cast_fp16)[name = string("v_new_325_cast_fp16")]; + tensor var_6150_begin_0 = const()[name = string("op_6150_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_6150_end_0 = const()[name = string("op_6150_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_6150_end_mask_0 = const()[name = string("op_6150_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6150_squeeze_mask_0 = const()[name = string("op_6150_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6150_cast_fp16 = slice_by_index(begin = var_6150_begin_0, end = var_6150_end_0, end_mask = var_6150_end_mask_0, squeeze_mask = var_6150_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6150_cast_fp16")]; + bool var_6151_transpose_x_0 = const()[name = string("op_6151_transpose_x_0"), val = bool(false)]; + bool var_6151_transpose_y_0 = const()[name = string("op_6151_transpose_y_0"), val = bool(false)]; + tensor var_6151_cast_fp16 = matmul(transpose_x = var_6151_transpose_x_0, transpose_y = var_6151_transpose_y_0, x = var_6150_cast_fp16, y = state_325_cast_fp16)[name = string("op_6151_cast_fp16")]; + tensor var_6153_begin_0 = const()[name = string("op_6153_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_6153_end_0 = const()[name = string("op_6153_end_0"), val = tensor([32, 3, 64, 64])]; + tensor var_6153_end_mask_0 = const()[name = string("op_6153_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6153_squeeze_mask_0 = const()[name = string("op_6153_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6153_cast_fp16 = slice_by_index(begin = var_6153_begin_0, end = var_6153_end_0, end_mask = var_6153_end_mask_0, squeeze_mask = var_6153_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6153_cast_fp16")]; + bool var_6154_transpose_x_0 = const()[name = string("op_6154_transpose_x_0"), val = bool(false)]; + bool var_6154_transpose_y_0 = const()[name = string("op_6154_transpose_y_0"), val = bool(false)]; + tensor var_6154_cast_fp16 = matmul(transpose_x = var_6154_transpose_x_0, transpose_y = var_6154_transpose_y_0, x = var_6153_cast_fp16, y = v_new_325_cast_fp16)[name = string("op_6154_cast_fp16")]; + tensor var_6155_cast_fp16 = add(x = var_6151_cast_fp16, y = var_6154_cast_fp16)[name = string("op_6155_cast_fp16")]; + tensor var_6157_begin_0 = const()[name = string("op_6157_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_6157_end_0 = const()[name = string("op_6157_end_0"), val = tensor([32, 3, 1, 1])]; + tensor var_6157_end_mask_0 = const()[name = string("op_6157_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6157_squeeze_mask_0 = const()[name = string("op_6157_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6157_cast_fp16 = slice_by_index(begin = var_6157_begin_0, end = var_6157_end_0, end_mask = var_6157_end_mask_0, squeeze_mask = var_6157_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6157_cast_fp16")]; + tensor var_6158_cast_fp16 = mul(x = state_325_cast_fp16, y = var_6157_cast_fp16)[name = string("op_6158_cast_fp16")]; + tensor var_6160_begin_0 = const()[name = string("op_6160_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_6160_end_0 = const()[name = string("op_6160_end_0"), val = tensor([32, 3, 64, 128])]; + tensor var_6160_end_mask_0 = const()[name = string("op_6160_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6160_squeeze_mask_0 = const()[name = string("op_6160_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6160_cast_fp16 = slice_by_index(begin = var_6160_begin_0, end = var_6160_end_0, end_mask = var_6160_end_mask_0, squeeze_mask = var_6160_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6160_cast_fp16")]; + bool var_6162_transpose_x_1 = const()[name = string("op_6162_transpose_x_1"), val = bool(true)]; + bool var_6162_transpose_y_1 = const()[name = string("op_6162_transpose_y_1"), val = bool(false)]; + tensor var_6162_cast_fp16 = matmul(transpose_x = var_6162_transpose_x_1, transpose_y = var_6162_transpose_y_1, x = var_6160_cast_fp16, y = v_new_325_cast_fp16)[name = string("op_6162_cast_fp16")]; + tensor state_327_cast_fp16 = add(x = var_6158_cast_fp16, y = var_6162_cast_fp16)[name = string("state_327_cast_fp16")]; + tensor var_6165_begin_0 = const()[name = string("op_6165_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_6165_end_0 = const()[name = string("op_6165_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_6165_end_mask_0 = const()[name = string("op_6165_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6165_squeeze_mask_0 = const()[name = string("op_6165_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6165_cast_fp16 = slice_by_index(begin = var_6165_begin_0, end = var_6165_end_0, end_mask = var_6165_end_mask_0, squeeze_mask = var_6165_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6165_cast_fp16")]; + tensor var_6167_begin_0 = const()[name = string("op_6167_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_6167_end_0 = const()[name = string("op_6167_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_6167_end_mask_0 = const()[name = string("op_6167_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6167_squeeze_mask_0 = const()[name = string("op_6167_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6167_cast_fp16 = slice_by_index(begin = var_6167_begin_0, end = var_6167_end_0, end_mask = var_6167_end_mask_0, squeeze_mask = var_6167_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6167_cast_fp16")]; + bool var_6168_transpose_x_0 = const()[name = string("op_6168_transpose_x_0"), val = bool(false)]; + bool var_6168_transpose_y_0 = const()[name = string("op_6168_transpose_y_0"), val = bool(false)]; + tensor var_6168_cast_fp16 = matmul(transpose_x = var_6168_transpose_x_0, transpose_y = var_6168_transpose_y_0, x = var_6167_cast_fp16, y = state_327_cast_fp16)[name = string("op_6168_cast_fp16")]; + tensor v_new_327_cast_fp16 = sub(x = var_6165_cast_fp16, y = var_6168_cast_fp16)[name = string("v_new_327_cast_fp16")]; + tensor var_6171_begin_0 = const()[name = string("op_6171_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_6171_end_0 = const()[name = string("op_6171_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_6171_end_mask_0 = const()[name = string("op_6171_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6171_squeeze_mask_0 = const()[name = string("op_6171_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6171_cast_fp16 = slice_by_index(begin = var_6171_begin_0, end = var_6171_end_0, end_mask = var_6171_end_mask_0, squeeze_mask = var_6171_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6171_cast_fp16")]; + bool var_6172_transpose_x_0 = const()[name = string("op_6172_transpose_x_0"), val = bool(false)]; + bool var_6172_transpose_y_0 = const()[name = string("op_6172_transpose_y_0"), val = bool(false)]; + tensor var_6172_cast_fp16 = matmul(transpose_x = var_6172_transpose_x_0, transpose_y = var_6172_transpose_y_0, x = var_6171_cast_fp16, y = state_327_cast_fp16)[name = string("op_6172_cast_fp16")]; + tensor var_6174_begin_0 = const()[name = string("op_6174_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_6174_end_0 = const()[name = string("op_6174_end_0"), val = tensor([32, 4, 64, 64])]; + tensor var_6174_end_mask_0 = const()[name = string("op_6174_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6174_squeeze_mask_0 = const()[name = string("op_6174_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6174_cast_fp16 = slice_by_index(begin = var_6174_begin_0, end = var_6174_end_0, end_mask = var_6174_end_mask_0, squeeze_mask = var_6174_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6174_cast_fp16")]; + bool var_6175_transpose_x_0 = const()[name = string("op_6175_transpose_x_0"), val = bool(false)]; + bool var_6175_transpose_y_0 = const()[name = string("op_6175_transpose_y_0"), val = bool(false)]; + tensor var_6175_cast_fp16 = matmul(transpose_x = var_6175_transpose_x_0, transpose_y = var_6175_transpose_y_0, x = var_6174_cast_fp16, y = v_new_327_cast_fp16)[name = string("op_6175_cast_fp16")]; + tensor var_6176_cast_fp16 = add(x = var_6172_cast_fp16, y = var_6175_cast_fp16)[name = string("op_6176_cast_fp16")]; + tensor var_6178_begin_0 = const()[name = string("op_6178_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_6178_end_0 = const()[name = string("op_6178_end_0"), val = tensor([32, 4, 1, 1])]; + tensor var_6178_end_mask_0 = const()[name = string("op_6178_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6178_squeeze_mask_0 = const()[name = string("op_6178_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6178_cast_fp16 = slice_by_index(begin = var_6178_begin_0, end = var_6178_end_0, end_mask = var_6178_end_mask_0, squeeze_mask = var_6178_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6178_cast_fp16")]; + tensor var_6179_cast_fp16 = mul(x = state_327_cast_fp16, y = var_6178_cast_fp16)[name = string("op_6179_cast_fp16")]; + tensor var_6181_begin_0 = const()[name = string("op_6181_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_6181_end_0 = const()[name = string("op_6181_end_0"), val = tensor([32, 4, 64, 128])]; + tensor var_6181_end_mask_0 = const()[name = string("op_6181_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6181_squeeze_mask_0 = const()[name = string("op_6181_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6181_cast_fp16 = slice_by_index(begin = var_6181_begin_0, end = var_6181_end_0, end_mask = var_6181_end_mask_0, squeeze_mask = var_6181_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6181_cast_fp16")]; + bool var_6183_transpose_x_1 = const()[name = string("op_6183_transpose_x_1"), val = bool(true)]; + bool var_6183_transpose_y_1 = const()[name = string("op_6183_transpose_y_1"), val = bool(false)]; + tensor var_6183_cast_fp16 = matmul(transpose_x = var_6183_transpose_x_1, transpose_y = var_6183_transpose_y_1, x = var_6181_cast_fp16, y = v_new_327_cast_fp16)[name = string("op_6183_cast_fp16")]; + tensor state_329_cast_fp16 = add(x = var_6179_cast_fp16, y = var_6183_cast_fp16)[name = string("state_329_cast_fp16")]; + tensor var_6186_begin_0 = const()[name = string("op_6186_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_6186_end_0 = const()[name = string("op_6186_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_6186_end_mask_0 = const()[name = string("op_6186_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6186_squeeze_mask_0 = const()[name = string("op_6186_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6186_cast_fp16 = slice_by_index(begin = var_6186_begin_0, end = var_6186_end_0, end_mask = var_6186_end_mask_0, squeeze_mask = var_6186_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6186_cast_fp16")]; + tensor var_6188_begin_0 = const()[name = string("op_6188_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_6188_end_0 = const()[name = string("op_6188_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_6188_end_mask_0 = const()[name = string("op_6188_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6188_squeeze_mask_0 = const()[name = string("op_6188_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6188_cast_fp16 = slice_by_index(begin = var_6188_begin_0, end = var_6188_end_0, end_mask = var_6188_end_mask_0, squeeze_mask = var_6188_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6188_cast_fp16")]; + bool var_6189_transpose_x_0 = const()[name = string("op_6189_transpose_x_0"), val = bool(false)]; + bool var_6189_transpose_y_0 = const()[name = string("op_6189_transpose_y_0"), val = bool(false)]; + tensor var_6189_cast_fp16 = matmul(transpose_x = var_6189_transpose_x_0, transpose_y = var_6189_transpose_y_0, x = var_6188_cast_fp16, y = state_329_cast_fp16)[name = string("op_6189_cast_fp16")]; + tensor v_new_329_cast_fp16 = sub(x = var_6186_cast_fp16, y = var_6189_cast_fp16)[name = string("v_new_329_cast_fp16")]; + tensor var_6192_begin_0 = const()[name = string("op_6192_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_6192_end_0 = const()[name = string("op_6192_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_6192_end_mask_0 = const()[name = string("op_6192_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6192_squeeze_mask_0 = const()[name = string("op_6192_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6192_cast_fp16 = slice_by_index(begin = var_6192_begin_0, end = var_6192_end_0, end_mask = var_6192_end_mask_0, squeeze_mask = var_6192_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6192_cast_fp16")]; + bool var_6193_transpose_x_0 = const()[name = string("op_6193_transpose_x_0"), val = bool(false)]; + bool var_6193_transpose_y_0 = const()[name = string("op_6193_transpose_y_0"), val = bool(false)]; + tensor var_6193_cast_fp16 = matmul(transpose_x = var_6193_transpose_x_0, transpose_y = var_6193_transpose_y_0, x = var_6192_cast_fp16, y = state_329_cast_fp16)[name = string("op_6193_cast_fp16")]; + tensor var_6195_begin_0 = const()[name = string("op_6195_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_6195_end_0 = const()[name = string("op_6195_end_0"), val = tensor([32, 5, 64, 64])]; + tensor var_6195_end_mask_0 = const()[name = string("op_6195_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6195_squeeze_mask_0 = const()[name = string("op_6195_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6195_cast_fp16 = slice_by_index(begin = var_6195_begin_0, end = var_6195_end_0, end_mask = var_6195_end_mask_0, squeeze_mask = var_6195_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6195_cast_fp16")]; + bool var_6196_transpose_x_0 = const()[name = string("op_6196_transpose_x_0"), val = bool(false)]; + bool var_6196_transpose_y_0 = const()[name = string("op_6196_transpose_y_0"), val = bool(false)]; + tensor var_6196_cast_fp16 = matmul(transpose_x = var_6196_transpose_x_0, transpose_y = var_6196_transpose_y_0, x = var_6195_cast_fp16, y = v_new_329_cast_fp16)[name = string("op_6196_cast_fp16")]; + tensor var_6197_cast_fp16 = add(x = var_6193_cast_fp16, y = var_6196_cast_fp16)[name = string("op_6197_cast_fp16")]; + tensor var_6199_begin_0 = const()[name = string("op_6199_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_6199_end_0 = const()[name = string("op_6199_end_0"), val = tensor([32, 5, 1, 1])]; + tensor var_6199_end_mask_0 = const()[name = string("op_6199_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6199_squeeze_mask_0 = const()[name = string("op_6199_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6199_cast_fp16 = slice_by_index(begin = var_6199_begin_0, end = var_6199_end_0, end_mask = var_6199_end_mask_0, squeeze_mask = var_6199_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6199_cast_fp16")]; + tensor var_6200_cast_fp16 = mul(x = state_329_cast_fp16, y = var_6199_cast_fp16)[name = string("op_6200_cast_fp16")]; + tensor var_6202_begin_0 = const()[name = string("op_6202_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_6202_end_0 = const()[name = string("op_6202_end_0"), val = tensor([32, 5, 64, 128])]; + tensor var_6202_end_mask_0 = const()[name = string("op_6202_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6202_squeeze_mask_0 = const()[name = string("op_6202_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6202_cast_fp16 = slice_by_index(begin = var_6202_begin_0, end = var_6202_end_0, end_mask = var_6202_end_mask_0, squeeze_mask = var_6202_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6202_cast_fp16")]; + bool var_6204_transpose_x_1 = const()[name = string("op_6204_transpose_x_1"), val = bool(true)]; + bool var_6204_transpose_y_1 = const()[name = string("op_6204_transpose_y_1"), val = bool(false)]; + tensor var_6204_cast_fp16 = matmul(transpose_x = var_6204_transpose_x_1, transpose_y = var_6204_transpose_y_1, x = var_6202_cast_fp16, y = v_new_329_cast_fp16)[name = string("op_6204_cast_fp16")]; + tensor state_331_cast_fp16 = add(x = var_6200_cast_fp16, y = var_6204_cast_fp16)[name = string("state_331_cast_fp16")]; + tensor var_6207_begin_0 = const()[name = string("op_6207_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_6207_end_0 = const()[name = string("op_6207_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_6207_end_mask_0 = const()[name = string("op_6207_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6207_squeeze_mask_0 = const()[name = string("op_6207_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6207_cast_fp16 = slice_by_index(begin = var_6207_begin_0, end = var_6207_end_0, end_mask = var_6207_end_mask_0, squeeze_mask = var_6207_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6207_cast_fp16")]; + tensor var_6209_begin_0 = const()[name = string("op_6209_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_6209_end_0 = const()[name = string("op_6209_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_6209_end_mask_0 = const()[name = string("op_6209_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6209_squeeze_mask_0 = const()[name = string("op_6209_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6209_cast_fp16 = slice_by_index(begin = var_6209_begin_0, end = var_6209_end_0, end_mask = var_6209_end_mask_0, squeeze_mask = var_6209_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6209_cast_fp16")]; + bool var_6210_transpose_x_0 = const()[name = string("op_6210_transpose_x_0"), val = bool(false)]; + bool var_6210_transpose_y_0 = const()[name = string("op_6210_transpose_y_0"), val = bool(false)]; + tensor var_6210_cast_fp16 = matmul(transpose_x = var_6210_transpose_x_0, transpose_y = var_6210_transpose_y_0, x = var_6209_cast_fp16, y = state_331_cast_fp16)[name = string("op_6210_cast_fp16")]; + tensor v_new_331_cast_fp16 = sub(x = var_6207_cast_fp16, y = var_6210_cast_fp16)[name = string("v_new_331_cast_fp16")]; + tensor var_6213_begin_0 = const()[name = string("op_6213_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_6213_end_0 = const()[name = string("op_6213_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_6213_end_mask_0 = const()[name = string("op_6213_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6213_squeeze_mask_0 = const()[name = string("op_6213_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6213_cast_fp16 = slice_by_index(begin = var_6213_begin_0, end = var_6213_end_0, end_mask = var_6213_end_mask_0, squeeze_mask = var_6213_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6213_cast_fp16")]; + bool var_6214_transpose_x_0 = const()[name = string("op_6214_transpose_x_0"), val = bool(false)]; + bool var_6214_transpose_y_0 = const()[name = string("op_6214_transpose_y_0"), val = bool(false)]; + tensor var_6214_cast_fp16 = matmul(transpose_x = var_6214_transpose_x_0, transpose_y = var_6214_transpose_y_0, x = var_6213_cast_fp16, y = state_331_cast_fp16)[name = string("op_6214_cast_fp16")]; + tensor var_6216_begin_0 = const()[name = string("op_6216_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_6216_end_0 = const()[name = string("op_6216_end_0"), val = tensor([32, 6, 64, 64])]; + tensor var_6216_end_mask_0 = const()[name = string("op_6216_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6216_squeeze_mask_0 = const()[name = string("op_6216_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6216_cast_fp16 = slice_by_index(begin = var_6216_begin_0, end = var_6216_end_0, end_mask = var_6216_end_mask_0, squeeze_mask = var_6216_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6216_cast_fp16")]; + bool var_6217_transpose_x_0 = const()[name = string("op_6217_transpose_x_0"), val = bool(false)]; + bool var_6217_transpose_y_0 = const()[name = string("op_6217_transpose_y_0"), val = bool(false)]; + tensor var_6217_cast_fp16 = matmul(transpose_x = var_6217_transpose_x_0, transpose_y = var_6217_transpose_y_0, x = var_6216_cast_fp16, y = v_new_331_cast_fp16)[name = string("op_6217_cast_fp16")]; + tensor var_6218_cast_fp16 = add(x = var_6214_cast_fp16, y = var_6217_cast_fp16)[name = string("op_6218_cast_fp16")]; + tensor var_6220_begin_0 = const()[name = string("op_6220_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_6220_end_0 = const()[name = string("op_6220_end_0"), val = tensor([32, 6, 1, 1])]; + tensor var_6220_end_mask_0 = const()[name = string("op_6220_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6220_squeeze_mask_0 = const()[name = string("op_6220_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6220_cast_fp16 = slice_by_index(begin = var_6220_begin_0, end = var_6220_end_0, end_mask = var_6220_end_mask_0, squeeze_mask = var_6220_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6220_cast_fp16")]; + tensor var_6221_cast_fp16 = mul(x = state_331_cast_fp16, y = var_6220_cast_fp16)[name = string("op_6221_cast_fp16")]; + tensor var_6223_begin_0 = const()[name = string("op_6223_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_6223_end_0 = const()[name = string("op_6223_end_0"), val = tensor([32, 6, 64, 128])]; + tensor var_6223_end_mask_0 = const()[name = string("op_6223_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6223_squeeze_mask_0 = const()[name = string("op_6223_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6223_cast_fp16 = slice_by_index(begin = var_6223_begin_0, end = var_6223_end_0, end_mask = var_6223_end_mask_0, squeeze_mask = var_6223_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6223_cast_fp16")]; + bool var_6225_transpose_x_1 = const()[name = string("op_6225_transpose_x_1"), val = bool(true)]; + bool var_6225_transpose_y_1 = const()[name = string("op_6225_transpose_y_1"), val = bool(false)]; + tensor var_6225_cast_fp16 = matmul(transpose_x = var_6225_transpose_x_1, transpose_y = var_6225_transpose_y_1, x = var_6223_cast_fp16, y = v_new_331_cast_fp16)[name = string("op_6225_cast_fp16")]; + tensor state_333_cast_fp16 = add(x = var_6221_cast_fp16, y = var_6225_cast_fp16)[name = string("state_333_cast_fp16")]; + tensor var_6228_begin_0 = const()[name = string("op_6228_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_6228_end_0 = const()[name = string("op_6228_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_6228_end_mask_0 = const()[name = string("op_6228_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6228_squeeze_mask_0 = const()[name = string("op_6228_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6228_cast_fp16 = slice_by_index(begin = var_6228_begin_0, end = var_6228_end_0, end_mask = var_6228_end_mask_0, squeeze_mask = var_6228_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6228_cast_fp16")]; + tensor var_6230_begin_0 = const()[name = string("op_6230_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_6230_end_0 = const()[name = string("op_6230_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_6230_end_mask_0 = const()[name = string("op_6230_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6230_squeeze_mask_0 = const()[name = string("op_6230_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6230_cast_fp16 = slice_by_index(begin = var_6230_begin_0, end = var_6230_end_0, end_mask = var_6230_end_mask_0, squeeze_mask = var_6230_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6230_cast_fp16")]; + bool var_6231_transpose_x_0 = const()[name = string("op_6231_transpose_x_0"), val = bool(false)]; + bool var_6231_transpose_y_0 = const()[name = string("op_6231_transpose_y_0"), val = bool(false)]; + tensor var_6231_cast_fp16 = matmul(transpose_x = var_6231_transpose_x_0, transpose_y = var_6231_transpose_y_0, x = var_6230_cast_fp16, y = state_333_cast_fp16)[name = string("op_6231_cast_fp16")]; + tensor v_new_333_cast_fp16 = sub(x = var_6228_cast_fp16, y = var_6231_cast_fp16)[name = string("v_new_333_cast_fp16")]; + tensor var_6234_begin_0 = const()[name = string("op_6234_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_6234_end_0 = const()[name = string("op_6234_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_6234_end_mask_0 = const()[name = string("op_6234_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6234_squeeze_mask_0 = const()[name = string("op_6234_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6234_cast_fp16 = slice_by_index(begin = var_6234_begin_0, end = var_6234_end_0, end_mask = var_6234_end_mask_0, squeeze_mask = var_6234_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6234_cast_fp16")]; + bool var_6235_transpose_x_0 = const()[name = string("op_6235_transpose_x_0"), val = bool(false)]; + bool var_6235_transpose_y_0 = const()[name = string("op_6235_transpose_y_0"), val = bool(false)]; + tensor var_6235_cast_fp16 = matmul(transpose_x = var_6235_transpose_x_0, transpose_y = var_6235_transpose_y_0, x = var_6234_cast_fp16, y = state_333_cast_fp16)[name = string("op_6235_cast_fp16")]; + tensor var_6237_begin_0 = const()[name = string("op_6237_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_6237_end_0 = const()[name = string("op_6237_end_0"), val = tensor([32, 7, 64, 64])]; + tensor var_6237_end_mask_0 = const()[name = string("op_6237_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6237_squeeze_mask_0 = const()[name = string("op_6237_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6237_cast_fp16 = slice_by_index(begin = var_6237_begin_0, end = var_6237_end_0, end_mask = var_6237_end_mask_0, squeeze_mask = var_6237_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6237_cast_fp16")]; + bool var_6238_transpose_x_0 = const()[name = string("op_6238_transpose_x_0"), val = bool(false)]; + bool var_6238_transpose_y_0 = const()[name = string("op_6238_transpose_y_0"), val = bool(false)]; + tensor var_6238_cast_fp16 = matmul(transpose_x = var_6238_transpose_x_0, transpose_y = var_6238_transpose_y_0, x = var_6237_cast_fp16, y = v_new_333_cast_fp16)[name = string("op_6238_cast_fp16")]; + tensor var_6239_cast_fp16 = add(x = var_6235_cast_fp16, y = var_6238_cast_fp16)[name = string("op_6239_cast_fp16")]; + tensor var_6241_begin_0 = const()[name = string("op_6241_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_6241_end_0 = const()[name = string("op_6241_end_0"), val = tensor([32, 7, 1, 1])]; + tensor var_6241_end_mask_0 = const()[name = string("op_6241_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6241_squeeze_mask_0 = const()[name = string("op_6241_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6241_cast_fp16 = slice_by_index(begin = var_6241_begin_0, end = var_6241_end_0, end_mask = var_6241_end_mask_0, squeeze_mask = var_6241_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6241_cast_fp16")]; + tensor var_6242_cast_fp16 = mul(x = state_333_cast_fp16, y = var_6241_cast_fp16)[name = string("op_6242_cast_fp16")]; + tensor var_6244_begin_0 = const()[name = string("op_6244_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_6244_end_0 = const()[name = string("op_6244_end_0"), val = tensor([32, 7, 64, 128])]; + tensor var_6244_end_mask_0 = const()[name = string("op_6244_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6244_squeeze_mask_0 = const()[name = string("op_6244_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6244_cast_fp16 = slice_by_index(begin = var_6244_begin_0, end = var_6244_end_0, end_mask = var_6244_end_mask_0, squeeze_mask = var_6244_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6244_cast_fp16")]; + bool var_6246_transpose_x_1 = const()[name = string("op_6246_transpose_x_1"), val = bool(true)]; + bool var_6246_transpose_y_1 = const()[name = string("op_6246_transpose_y_1"), val = bool(false)]; + tensor var_6246_cast_fp16 = matmul(transpose_x = var_6246_transpose_x_1, transpose_y = var_6246_transpose_y_1, x = var_6244_cast_fp16, y = v_new_333_cast_fp16)[name = string("op_6246_cast_fp16")]; + tensor state_335_cast_fp16 = add(x = var_6242_cast_fp16, y = var_6246_cast_fp16)[name = string("state_335_cast_fp16")]; + tensor var_6249_begin_0 = const()[name = string("op_6249_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_6249_end_0 = const()[name = string("op_6249_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_6249_end_mask_0 = const()[name = string("op_6249_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6249_squeeze_mask_0 = const()[name = string("op_6249_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6249_cast_fp16 = slice_by_index(begin = var_6249_begin_0, end = var_6249_end_0, end_mask = var_6249_end_mask_0, squeeze_mask = var_6249_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6249_cast_fp16")]; + tensor var_6251_begin_0 = const()[name = string("op_6251_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_6251_end_0 = const()[name = string("op_6251_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_6251_end_mask_0 = const()[name = string("op_6251_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6251_squeeze_mask_0 = const()[name = string("op_6251_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6251_cast_fp16 = slice_by_index(begin = var_6251_begin_0, end = var_6251_end_0, end_mask = var_6251_end_mask_0, squeeze_mask = var_6251_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6251_cast_fp16")]; + bool var_6252_transpose_x_0 = const()[name = string("op_6252_transpose_x_0"), val = bool(false)]; + bool var_6252_transpose_y_0 = const()[name = string("op_6252_transpose_y_0"), val = bool(false)]; + tensor var_6252_cast_fp16 = matmul(transpose_x = var_6252_transpose_x_0, transpose_y = var_6252_transpose_y_0, x = var_6251_cast_fp16, y = state_335_cast_fp16)[name = string("op_6252_cast_fp16")]; + tensor v_new_335_cast_fp16 = sub(x = var_6249_cast_fp16, y = var_6252_cast_fp16)[name = string("v_new_335_cast_fp16")]; + tensor var_6255_begin_0 = const()[name = string("op_6255_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_6255_end_0 = const()[name = string("op_6255_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_6255_end_mask_0 = const()[name = string("op_6255_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6255_squeeze_mask_0 = const()[name = string("op_6255_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6255_cast_fp16 = slice_by_index(begin = var_6255_begin_0, end = var_6255_end_0, end_mask = var_6255_end_mask_0, squeeze_mask = var_6255_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6255_cast_fp16")]; + bool var_6256_transpose_x_0 = const()[name = string("op_6256_transpose_x_0"), val = bool(false)]; + bool var_6256_transpose_y_0 = const()[name = string("op_6256_transpose_y_0"), val = bool(false)]; + tensor var_6256_cast_fp16 = matmul(transpose_x = var_6256_transpose_x_0, transpose_y = var_6256_transpose_y_0, x = var_6255_cast_fp16, y = state_335_cast_fp16)[name = string("op_6256_cast_fp16")]; + tensor var_6258_begin_0 = const()[name = string("op_6258_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_6258_end_0 = const()[name = string("op_6258_end_0"), val = tensor([32, 8, 64, 64])]; + tensor var_6258_end_mask_0 = const()[name = string("op_6258_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6258_squeeze_mask_0 = const()[name = string("op_6258_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6258_cast_fp16 = slice_by_index(begin = var_6258_begin_0, end = var_6258_end_0, end_mask = var_6258_end_mask_0, squeeze_mask = var_6258_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6258_cast_fp16")]; + bool var_6259_transpose_x_0 = const()[name = string("op_6259_transpose_x_0"), val = bool(false)]; + bool var_6259_transpose_y_0 = const()[name = string("op_6259_transpose_y_0"), val = bool(false)]; + tensor var_6259_cast_fp16 = matmul(transpose_x = var_6259_transpose_x_0, transpose_y = var_6259_transpose_y_0, x = var_6258_cast_fp16, y = v_new_335_cast_fp16)[name = string("op_6259_cast_fp16")]; + tensor var_6260_cast_fp16 = add(x = var_6256_cast_fp16, y = var_6259_cast_fp16)[name = string("op_6260_cast_fp16")]; + tensor var_6262_begin_0 = const()[name = string("op_6262_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_6262_end_0 = const()[name = string("op_6262_end_0"), val = tensor([32, 8, 1, 1])]; + tensor var_6262_end_mask_0 = const()[name = string("op_6262_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6262_squeeze_mask_0 = const()[name = string("op_6262_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6262_cast_fp16 = slice_by_index(begin = var_6262_begin_0, end = var_6262_end_0, end_mask = var_6262_end_mask_0, squeeze_mask = var_6262_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6262_cast_fp16")]; + tensor var_6263_cast_fp16 = mul(x = state_335_cast_fp16, y = var_6262_cast_fp16)[name = string("op_6263_cast_fp16")]; + tensor var_6265_begin_0 = const()[name = string("op_6265_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_6265_end_0 = const()[name = string("op_6265_end_0"), val = tensor([32, 8, 64, 128])]; + tensor var_6265_end_mask_0 = const()[name = string("op_6265_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6265_squeeze_mask_0 = const()[name = string("op_6265_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6265_cast_fp16 = slice_by_index(begin = var_6265_begin_0, end = var_6265_end_0, end_mask = var_6265_end_mask_0, squeeze_mask = var_6265_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6265_cast_fp16")]; + bool var_6267_transpose_x_1 = const()[name = string("op_6267_transpose_x_1"), val = bool(true)]; + bool var_6267_transpose_y_1 = const()[name = string("op_6267_transpose_y_1"), val = bool(false)]; + tensor var_6267_cast_fp16 = matmul(transpose_x = var_6267_transpose_x_1, transpose_y = var_6267_transpose_y_1, x = var_6265_cast_fp16, y = v_new_335_cast_fp16)[name = string("op_6267_cast_fp16")]; + tensor state_337_cast_fp16 = add(x = var_6263_cast_fp16, y = var_6267_cast_fp16)[name = string("state_337_cast_fp16")]; + tensor var_6270_begin_0 = const()[name = string("op_6270_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_6270_end_0 = const()[name = string("op_6270_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_6270_end_mask_0 = const()[name = string("op_6270_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6270_squeeze_mask_0 = const()[name = string("op_6270_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6270_cast_fp16 = slice_by_index(begin = var_6270_begin_0, end = var_6270_end_0, end_mask = var_6270_end_mask_0, squeeze_mask = var_6270_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6270_cast_fp16")]; + tensor var_6272_begin_0 = const()[name = string("op_6272_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_6272_end_0 = const()[name = string("op_6272_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_6272_end_mask_0 = const()[name = string("op_6272_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6272_squeeze_mask_0 = const()[name = string("op_6272_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6272_cast_fp16 = slice_by_index(begin = var_6272_begin_0, end = var_6272_end_0, end_mask = var_6272_end_mask_0, squeeze_mask = var_6272_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6272_cast_fp16")]; + bool var_6273_transpose_x_0 = const()[name = string("op_6273_transpose_x_0"), val = bool(false)]; + bool var_6273_transpose_y_0 = const()[name = string("op_6273_transpose_y_0"), val = bool(false)]; + tensor var_6273_cast_fp16 = matmul(transpose_x = var_6273_transpose_x_0, transpose_y = var_6273_transpose_y_0, x = var_6272_cast_fp16, y = state_337_cast_fp16)[name = string("op_6273_cast_fp16")]; + tensor v_new_337_cast_fp16 = sub(x = var_6270_cast_fp16, y = var_6273_cast_fp16)[name = string("v_new_337_cast_fp16")]; + tensor var_6276_begin_0 = const()[name = string("op_6276_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_6276_end_0 = const()[name = string("op_6276_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_6276_end_mask_0 = const()[name = string("op_6276_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6276_squeeze_mask_0 = const()[name = string("op_6276_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6276_cast_fp16 = slice_by_index(begin = var_6276_begin_0, end = var_6276_end_0, end_mask = var_6276_end_mask_0, squeeze_mask = var_6276_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6276_cast_fp16")]; + bool var_6277_transpose_x_0 = const()[name = string("op_6277_transpose_x_0"), val = bool(false)]; + bool var_6277_transpose_y_0 = const()[name = string("op_6277_transpose_y_0"), val = bool(false)]; + tensor var_6277_cast_fp16 = matmul(transpose_x = var_6277_transpose_x_0, transpose_y = var_6277_transpose_y_0, x = var_6276_cast_fp16, y = state_337_cast_fp16)[name = string("op_6277_cast_fp16")]; + tensor var_6279_begin_0 = const()[name = string("op_6279_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_6279_end_0 = const()[name = string("op_6279_end_0"), val = tensor([32, 9, 64, 64])]; + tensor var_6279_end_mask_0 = const()[name = string("op_6279_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6279_squeeze_mask_0 = const()[name = string("op_6279_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6279_cast_fp16 = slice_by_index(begin = var_6279_begin_0, end = var_6279_end_0, end_mask = var_6279_end_mask_0, squeeze_mask = var_6279_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6279_cast_fp16")]; + bool var_6280_transpose_x_0 = const()[name = string("op_6280_transpose_x_0"), val = bool(false)]; + bool var_6280_transpose_y_0 = const()[name = string("op_6280_transpose_y_0"), val = bool(false)]; + tensor var_6280_cast_fp16 = matmul(transpose_x = var_6280_transpose_x_0, transpose_y = var_6280_transpose_y_0, x = var_6279_cast_fp16, y = v_new_337_cast_fp16)[name = string("op_6280_cast_fp16")]; + tensor var_6281_cast_fp16 = add(x = var_6277_cast_fp16, y = var_6280_cast_fp16)[name = string("op_6281_cast_fp16")]; + tensor var_6283_begin_0 = const()[name = string("op_6283_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_6283_end_0 = const()[name = string("op_6283_end_0"), val = tensor([32, 9, 1, 1])]; + tensor var_6283_end_mask_0 = const()[name = string("op_6283_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6283_squeeze_mask_0 = const()[name = string("op_6283_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6283_cast_fp16 = slice_by_index(begin = var_6283_begin_0, end = var_6283_end_0, end_mask = var_6283_end_mask_0, squeeze_mask = var_6283_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6283_cast_fp16")]; + tensor var_6284_cast_fp16 = mul(x = state_337_cast_fp16, y = var_6283_cast_fp16)[name = string("op_6284_cast_fp16")]; + tensor var_6286_begin_0 = const()[name = string("op_6286_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_6286_end_0 = const()[name = string("op_6286_end_0"), val = tensor([32, 9, 64, 128])]; + tensor var_6286_end_mask_0 = const()[name = string("op_6286_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6286_squeeze_mask_0 = const()[name = string("op_6286_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6286_cast_fp16 = slice_by_index(begin = var_6286_begin_0, end = var_6286_end_0, end_mask = var_6286_end_mask_0, squeeze_mask = var_6286_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6286_cast_fp16")]; + bool var_6288_transpose_x_1 = const()[name = string("op_6288_transpose_x_1"), val = bool(true)]; + bool var_6288_transpose_y_1 = const()[name = string("op_6288_transpose_y_1"), val = bool(false)]; + tensor var_6288_cast_fp16 = matmul(transpose_x = var_6288_transpose_x_1, transpose_y = var_6288_transpose_y_1, x = var_6286_cast_fp16, y = v_new_337_cast_fp16)[name = string("op_6288_cast_fp16")]; + tensor state_339_cast_fp16 = add(x = var_6284_cast_fp16, y = var_6288_cast_fp16)[name = string("state_339_cast_fp16")]; + tensor var_6291_begin_0 = const()[name = string("op_6291_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_6291_end_0 = const()[name = string("op_6291_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_6291_end_mask_0 = const()[name = string("op_6291_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6291_squeeze_mask_0 = const()[name = string("op_6291_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6291_cast_fp16 = slice_by_index(begin = var_6291_begin_0, end = var_6291_end_0, end_mask = var_6291_end_mask_0, squeeze_mask = var_6291_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6291_cast_fp16")]; + tensor var_6293_begin_0 = const()[name = string("op_6293_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_6293_end_0 = const()[name = string("op_6293_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_6293_end_mask_0 = const()[name = string("op_6293_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6293_squeeze_mask_0 = const()[name = string("op_6293_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6293_cast_fp16 = slice_by_index(begin = var_6293_begin_0, end = var_6293_end_0, end_mask = var_6293_end_mask_0, squeeze_mask = var_6293_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6293_cast_fp16")]; + bool var_6294_transpose_x_0 = const()[name = string("op_6294_transpose_x_0"), val = bool(false)]; + bool var_6294_transpose_y_0 = const()[name = string("op_6294_transpose_y_0"), val = bool(false)]; + tensor var_6294_cast_fp16 = matmul(transpose_x = var_6294_transpose_x_0, transpose_y = var_6294_transpose_y_0, x = var_6293_cast_fp16, y = state_339_cast_fp16)[name = string("op_6294_cast_fp16")]; + tensor v_new_339_cast_fp16 = sub(x = var_6291_cast_fp16, y = var_6294_cast_fp16)[name = string("v_new_339_cast_fp16")]; + tensor var_6297_begin_0 = const()[name = string("op_6297_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_6297_end_0 = const()[name = string("op_6297_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_6297_end_mask_0 = const()[name = string("op_6297_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6297_squeeze_mask_0 = const()[name = string("op_6297_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6297_cast_fp16 = slice_by_index(begin = var_6297_begin_0, end = var_6297_end_0, end_mask = var_6297_end_mask_0, squeeze_mask = var_6297_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6297_cast_fp16")]; + bool var_6298_transpose_x_0 = const()[name = string("op_6298_transpose_x_0"), val = bool(false)]; + bool var_6298_transpose_y_0 = const()[name = string("op_6298_transpose_y_0"), val = bool(false)]; + tensor var_6298_cast_fp16 = matmul(transpose_x = var_6298_transpose_x_0, transpose_y = var_6298_transpose_y_0, x = var_6297_cast_fp16, y = state_339_cast_fp16)[name = string("op_6298_cast_fp16")]; + tensor var_6300_begin_0 = const()[name = string("op_6300_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_6300_end_0 = const()[name = string("op_6300_end_0"), val = tensor([32, 10, 64, 64])]; + tensor var_6300_end_mask_0 = const()[name = string("op_6300_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6300_squeeze_mask_0 = const()[name = string("op_6300_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6300_cast_fp16 = slice_by_index(begin = var_6300_begin_0, end = var_6300_end_0, end_mask = var_6300_end_mask_0, squeeze_mask = var_6300_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6300_cast_fp16")]; + bool var_6301_transpose_x_0 = const()[name = string("op_6301_transpose_x_0"), val = bool(false)]; + bool var_6301_transpose_y_0 = const()[name = string("op_6301_transpose_y_0"), val = bool(false)]; + tensor var_6301_cast_fp16 = matmul(transpose_x = var_6301_transpose_x_0, transpose_y = var_6301_transpose_y_0, x = var_6300_cast_fp16, y = v_new_339_cast_fp16)[name = string("op_6301_cast_fp16")]; + tensor var_6302_cast_fp16 = add(x = var_6298_cast_fp16, y = var_6301_cast_fp16)[name = string("op_6302_cast_fp16")]; + tensor var_6304_begin_0 = const()[name = string("op_6304_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_6304_end_0 = const()[name = string("op_6304_end_0"), val = tensor([32, 10, 1, 1])]; + tensor var_6304_end_mask_0 = const()[name = string("op_6304_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6304_squeeze_mask_0 = const()[name = string("op_6304_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6304_cast_fp16 = slice_by_index(begin = var_6304_begin_0, end = var_6304_end_0, end_mask = var_6304_end_mask_0, squeeze_mask = var_6304_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6304_cast_fp16")]; + tensor var_6305_cast_fp16 = mul(x = state_339_cast_fp16, y = var_6304_cast_fp16)[name = string("op_6305_cast_fp16")]; + tensor var_6307_begin_0 = const()[name = string("op_6307_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_6307_end_0 = const()[name = string("op_6307_end_0"), val = tensor([32, 10, 64, 128])]; + tensor var_6307_end_mask_0 = const()[name = string("op_6307_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6307_squeeze_mask_0 = const()[name = string("op_6307_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6307_cast_fp16 = slice_by_index(begin = var_6307_begin_0, end = var_6307_end_0, end_mask = var_6307_end_mask_0, squeeze_mask = var_6307_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6307_cast_fp16")]; + bool var_6309_transpose_x_1 = const()[name = string("op_6309_transpose_x_1"), val = bool(true)]; + bool var_6309_transpose_y_1 = const()[name = string("op_6309_transpose_y_1"), val = bool(false)]; + tensor var_6309_cast_fp16 = matmul(transpose_x = var_6309_transpose_x_1, transpose_y = var_6309_transpose_y_1, x = var_6307_cast_fp16, y = v_new_339_cast_fp16)[name = string("op_6309_cast_fp16")]; + tensor state_341_cast_fp16 = add(x = var_6305_cast_fp16, y = var_6309_cast_fp16)[name = string("state_341_cast_fp16")]; + tensor var_6312_begin_0 = const()[name = string("op_6312_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_6312_end_0 = const()[name = string("op_6312_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_6312_end_mask_0 = const()[name = string("op_6312_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6312_squeeze_mask_0 = const()[name = string("op_6312_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6312_cast_fp16 = slice_by_index(begin = var_6312_begin_0, end = var_6312_end_0, end_mask = var_6312_end_mask_0, squeeze_mask = var_6312_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6312_cast_fp16")]; + tensor var_6314_begin_0 = const()[name = string("op_6314_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_6314_end_0 = const()[name = string("op_6314_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_6314_end_mask_0 = const()[name = string("op_6314_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6314_squeeze_mask_0 = const()[name = string("op_6314_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6314_cast_fp16 = slice_by_index(begin = var_6314_begin_0, end = var_6314_end_0, end_mask = var_6314_end_mask_0, squeeze_mask = var_6314_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6314_cast_fp16")]; + bool var_6315_transpose_x_0 = const()[name = string("op_6315_transpose_x_0"), val = bool(false)]; + bool var_6315_transpose_y_0 = const()[name = string("op_6315_transpose_y_0"), val = bool(false)]; + tensor var_6315_cast_fp16 = matmul(transpose_x = var_6315_transpose_x_0, transpose_y = var_6315_transpose_y_0, x = var_6314_cast_fp16, y = state_341_cast_fp16)[name = string("op_6315_cast_fp16")]; + tensor v_new_341_cast_fp16 = sub(x = var_6312_cast_fp16, y = var_6315_cast_fp16)[name = string("v_new_341_cast_fp16")]; + tensor var_6318_begin_0 = const()[name = string("op_6318_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_6318_end_0 = const()[name = string("op_6318_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_6318_end_mask_0 = const()[name = string("op_6318_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6318_squeeze_mask_0 = const()[name = string("op_6318_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6318_cast_fp16 = slice_by_index(begin = var_6318_begin_0, end = var_6318_end_0, end_mask = var_6318_end_mask_0, squeeze_mask = var_6318_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6318_cast_fp16")]; + bool var_6319_transpose_x_0 = const()[name = string("op_6319_transpose_x_0"), val = bool(false)]; + bool var_6319_transpose_y_0 = const()[name = string("op_6319_transpose_y_0"), val = bool(false)]; + tensor var_6319_cast_fp16 = matmul(transpose_x = var_6319_transpose_x_0, transpose_y = var_6319_transpose_y_0, x = var_6318_cast_fp16, y = state_341_cast_fp16)[name = string("op_6319_cast_fp16")]; + tensor var_6321_begin_0 = const()[name = string("op_6321_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_6321_end_0 = const()[name = string("op_6321_end_0"), val = tensor([32, 11, 64, 64])]; + tensor var_6321_end_mask_0 = const()[name = string("op_6321_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6321_squeeze_mask_0 = const()[name = string("op_6321_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6321_cast_fp16 = slice_by_index(begin = var_6321_begin_0, end = var_6321_end_0, end_mask = var_6321_end_mask_0, squeeze_mask = var_6321_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6321_cast_fp16")]; + bool var_6322_transpose_x_0 = const()[name = string("op_6322_transpose_x_0"), val = bool(false)]; + bool var_6322_transpose_y_0 = const()[name = string("op_6322_transpose_y_0"), val = bool(false)]; + tensor var_6322_cast_fp16 = matmul(transpose_x = var_6322_transpose_x_0, transpose_y = var_6322_transpose_y_0, x = var_6321_cast_fp16, y = v_new_341_cast_fp16)[name = string("op_6322_cast_fp16")]; + tensor var_6323_cast_fp16 = add(x = var_6319_cast_fp16, y = var_6322_cast_fp16)[name = string("op_6323_cast_fp16")]; + tensor var_6325_begin_0 = const()[name = string("op_6325_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_6325_end_0 = const()[name = string("op_6325_end_0"), val = tensor([32, 11, 1, 1])]; + tensor var_6325_end_mask_0 = const()[name = string("op_6325_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6325_squeeze_mask_0 = const()[name = string("op_6325_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6325_cast_fp16 = slice_by_index(begin = var_6325_begin_0, end = var_6325_end_0, end_mask = var_6325_end_mask_0, squeeze_mask = var_6325_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6325_cast_fp16")]; + tensor var_6326_cast_fp16 = mul(x = state_341_cast_fp16, y = var_6325_cast_fp16)[name = string("op_6326_cast_fp16")]; + tensor var_6328_begin_0 = const()[name = string("op_6328_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_6328_end_0 = const()[name = string("op_6328_end_0"), val = tensor([32, 11, 64, 128])]; + tensor var_6328_end_mask_0 = const()[name = string("op_6328_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6328_squeeze_mask_0 = const()[name = string("op_6328_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6328_cast_fp16 = slice_by_index(begin = var_6328_begin_0, end = var_6328_end_0, end_mask = var_6328_end_mask_0, squeeze_mask = var_6328_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6328_cast_fp16")]; + bool var_6330_transpose_x_1 = const()[name = string("op_6330_transpose_x_1"), val = bool(true)]; + bool var_6330_transpose_y_1 = const()[name = string("op_6330_transpose_y_1"), val = bool(false)]; + tensor var_6330_cast_fp16 = matmul(transpose_x = var_6330_transpose_x_1, transpose_y = var_6330_transpose_y_1, x = var_6328_cast_fp16, y = v_new_341_cast_fp16)[name = string("op_6330_cast_fp16")]; + tensor state_343_cast_fp16 = add(x = var_6326_cast_fp16, y = var_6330_cast_fp16)[name = string("state_343_cast_fp16")]; + tensor var_6333_begin_0 = const()[name = string("op_6333_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_6333_end_0 = const()[name = string("op_6333_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_6333_end_mask_0 = const()[name = string("op_6333_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6333_squeeze_mask_0 = const()[name = string("op_6333_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6333_cast_fp16 = slice_by_index(begin = var_6333_begin_0, end = var_6333_end_0, end_mask = var_6333_end_mask_0, squeeze_mask = var_6333_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6333_cast_fp16")]; + tensor var_6335_begin_0 = const()[name = string("op_6335_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_6335_end_0 = const()[name = string("op_6335_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_6335_end_mask_0 = const()[name = string("op_6335_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6335_squeeze_mask_0 = const()[name = string("op_6335_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6335_cast_fp16 = slice_by_index(begin = var_6335_begin_0, end = var_6335_end_0, end_mask = var_6335_end_mask_0, squeeze_mask = var_6335_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6335_cast_fp16")]; + bool var_6336_transpose_x_0 = const()[name = string("op_6336_transpose_x_0"), val = bool(false)]; + bool var_6336_transpose_y_0 = const()[name = string("op_6336_transpose_y_0"), val = bool(false)]; + tensor var_6336_cast_fp16 = matmul(transpose_x = var_6336_transpose_x_0, transpose_y = var_6336_transpose_y_0, x = var_6335_cast_fp16, y = state_343_cast_fp16)[name = string("op_6336_cast_fp16")]; + tensor v_new_343_cast_fp16 = sub(x = var_6333_cast_fp16, y = var_6336_cast_fp16)[name = string("v_new_343_cast_fp16")]; + tensor var_6339_begin_0 = const()[name = string("op_6339_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_6339_end_0 = const()[name = string("op_6339_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_6339_end_mask_0 = const()[name = string("op_6339_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6339_squeeze_mask_0 = const()[name = string("op_6339_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6339_cast_fp16 = slice_by_index(begin = var_6339_begin_0, end = var_6339_end_0, end_mask = var_6339_end_mask_0, squeeze_mask = var_6339_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6339_cast_fp16")]; + bool var_6340_transpose_x_0 = const()[name = string("op_6340_transpose_x_0"), val = bool(false)]; + bool var_6340_transpose_y_0 = const()[name = string("op_6340_transpose_y_0"), val = bool(false)]; + tensor var_6340_cast_fp16 = matmul(transpose_x = var_6340_transpose_x_0, transpose_y = var_6340_transpose_y_0, x = var_6339_cast_fp16, y = state_343_cast_fp16)[name = string("op_6340_cast_fp16")]; + tensor var_6342_begin_0 = const()[name = string("op_6342_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_6342_end_0 = const()[name = string("op_6342_end_0"), val = tensor([32, 12, 64, 64])]; + tensor var_6342_end_mask_0 = const()[name = string("op_6342_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6342_squeeze_mask_0 = const()[name = string("op_6342_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6342_cast_fp16 = slice_by_index(begin = var_6342_begin_0, end = var_6342_end_0, end_mask = var_6342_end_mask_0, squeeze_mask = var_6342_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6342_cast_fp16")]; + bool var_6343_transpose_x_0 = const()[name = string("op_6343_transpose_x_0"), val = bool(false)]; + bool var_6343_transpose_y_0 = const()[name = string("op_6343_transpose_y_0"), val = bool(false)]; + tensor var_6343_cast_fp16 = matmul(transpose_x = var_6343_transpose_x_0, transpose_y = var_6343_transpose_y_0, x = var_6342_cast_fp16, y = v_new_343_cast_fp16)[name = string("op_6343_cast_fp16")]; + tensor var_6344_cast_fp16 = add(x = var_6340_cast_fp16, y = var_6343_cast_fp16)[name = string("op_6344_cast_fp16")]; + tensor var_6346_begin_0 = const()[name = string("op_6346_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_6346_end_0 = const()[name = string("op_6346_end_0"), val = tensor([32, 12, 1, 1])]; + tensor var_6346_end_mask_0 = const()[name = string("op_6346_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6346_squeeze_mask_0 = const()[name = string("op_6346_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6346_cast_fp16 = slice_by_index(begin = var_6346_begin_0, end = var_6346_end_0, end_mask = var_6346_end_mask_0, squeeze_mask = var_6346_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6346_cast_fp16")]; + tensor var_6347_cast_fp16 = mul(x = state_343_cast_fp16, y = var_6346_cast_fp16)[name = string("op_6347_cast_fp16")]; + tensor var_6349_begin_0 = const()[name = string("op_6349_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_6349_end_0 = const()[name = string("op_6349_end_0"), val = tensor([32, 12, 64, 128])]; + tensor var_6349_end_mask_0 = const()[name = string("op_6349_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6349_squeeze_mask_0 = const()[name = string("op_6349_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6349_cast_fp16 = slice_by_index(begin = var_6349_begin_0, end = var_6349_end_0, end_mask = var_6349_end_mask_0, squeeze_mask = var_6349_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6349_cast_fp16")]; + bool var_6351_transpose_x_1 = const()[name = string("op_6351_transpose_x_1"), val = bool(true)]; + bool var_6351_transpose_y_1 = const()[name = string("op_6351_transpose_y_1"), val = bool(false)]; + tensor var_6351_cast_fp16 = matmul(transpose_x = var_6351_transpose_x_1, transpose_y = var_6351_transpose_y_1, x = var_6349_cast_fp16, y = v_new_343_cast_fp16)[name = string("op_6351_cast_fp16")]; + tensor state_345_cast_fp16 = add(x = var_6347_cast_fp16, y = var_6351_cast_fp16)[name = string("state_345_cast_fp16")]; + tensor var_6354_begin_0 = const()[name = string("op_6354_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_6354_end_0 = const()[name = string("op_6354_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_6354_end_mask_0 = const()[name = string("op_6354_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6354_squeeze_mask_0 = const()[name = string("op_6354_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6354_cast_fp16 = slice_by_index(begin = var_6354_begin_0, end = var_6354_end_0, end_mask = var_6354_end_mask_0, squeeze_mask = var_6354_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6354_cast_fp16")]; + tensor var_6356_begin_0 = const()[name = string("op_6356_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_6356_end_0 = const()[name = string("op_6356_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_6356_end_mask_0 = const()[name = string("op_6356_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6356_squeeze_mask_0 = const()[name = string("op_6356_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6356_cast_fp16 = slice_by_index(begin = var_6356_begin_0, end = var_6356_end_0, end_mask = var_6356_end_mask_0, squeeze_mask = var_6356_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6356_cast_fp16")]; + bool var_6357_transpose_x_0 = const()[name = string("op_6357_transpose_x_0"), val = bool(false)]; + bool var_6357_transpose_y_0 = const()[name = string("op_6357_transpose_y_0"), val = bool(false)]; + tensor var_6357_cast_fp16 = matmul(transpose_x = var_6357_transpose_x_0, transpose_y = var_6357_transpose_y_0, x = var_6356_cast_fp16, y = state_345_cast_fp16)[name = string("op_6357_cast_fp16")]; + tensor v_new_345_cast_fp16 = sub(x = var_6354_cast_fp16, y = var_6357_cast_fp16)[name = string("v_new_345_cast_fp16")]; + tensor var_6360_begin_0 = const()[name = string("op_6360_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_6360_end_0 = const()[name = string("op_6360_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_6360_end_mask_0 = const()[name = string("op_6360_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6360_squeeze_mask_0 = const()[name = string("op_6360_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6360_cast_fp16 = slice_by_index(begin = var_6360_begin_0, end = var_6360_end_0, end_mask = var_6360_end_mask_0, squeeze_mask = var_6360_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6360_cast_fp16")]; + bool var_6361_transpose_x_0 = const()[name = string("op_6361_transpose_x_0"), val = bool(false)]; + bool var_6361_transpose_y_0 = const()[name = string("op_6361_transpose_y_0"), val = bool(false)]; + tensor var_6361_cast_fp16 = matmul(transpose_x = var_6361_transpose_x_0, transpose_y = var_6361_transpose_y_0, x = var_6360_cast_fp16, y = state_345_cast_fp16)[name = string("op_6361_cast_fp16")]; + tensor var_6363_begin_0 = const()[name = string("op_6363_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_6363_end_0 = const()[name = string("op_6363_end_0"), val = tensor([32, 13, 64, 64])]; + tensor var_6363_end_mask_0 = const()[name = string("op_6363_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6363_squeeze_mask_0 = const()[name = string("op_6363_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6363_cast_fp16 = slice_by_index(begin = var_6363_begin_0, end = var_6363_end_0, end_mask = var_6363_end_mask_0, squeeze_mask = var_6363_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6363_cast_fp16")]; + bool var_6364_transpose_x_0 = const()[name = string("op_6364_transpose_x_0"), val = bool(false)]; + bool var_6364_transpose_y_0 = const()[name = string("op_6364_transpose_y_0"), val = bool(false)]; + tensor var_6364_cast_fp16 = matmul(transpose_x = var_6364_transpose_x_0, transpose_y = var_6364_transpose_y_0, x = var_6363_cast_fp16, y = v_new_345_cast_fp16)[name = string("op_6364_cast_fp16")]; + tensor var_6365_cast_fp16 = add(x = var_6361_cast_fp16, y = var_6364_cast_fp16)[name = string("op_6365_cast_fp16")]; + tensor var_6367_begin_0 = const()[name = string("op_6367_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_6367_end_0 = const()[name = string("op_6367_end_0"), val = tensor([32, 13, 1, 1])]; + tensor var_6367_end_mask_0 = const()[name = string("op_6367_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6367_squeeze_mask_0 = const()[name = string("op_6367_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6367_cast_fp16 = slice_by_index(begin = var_6367_begin_0, end = var_6367_end_0, end_mask = var_6367_end_mask_0, squeeze_mask = var_6367_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6367_cast_fp16")]; + tensor var_6368_cast_fp16 = mul(x = state_345_cast_fp16, y = var_6367_cast_fp16)[name = string("op_6368_cast_fp16")]; + tensor var_6370_begin_0 = const()[name = string("op_6370_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_6370_end_0 = const()[name = string("op_6370_end_0"), val = tensor([32, 13, 64, 128])]; + tensor var_6370_end_mask_0 = const()[name = string("op_6370_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6370_squeeze_mask_0 = const()[name = string("op_6370_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6370_cast_fp16 = slice_by_index(begin = var_6370_begin_0, end = var_6370_end_0, end_mask = var_6370_end_mask_0, squeeze_mask = var_6370_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6370_cast_fp16")]; + bool var_6372_transpose_x_1 = const()[name = string("op_6372_transpose_x_1"), val = bool(true)]; + bool var_6372_transpose_y_1 = const()[name = string("op_6372_transpose_y_1"), val = bool(false)]; + tensor var_6372_cast_fp16 = matmul(transpose_x = var_6372_transpose_x_1, transpose_y = var_6372_transpose_y_1, x = var_6370_cast_fp16, y = v_new_345_cast_fp16)[name = string("op_6372_cast_fp16")]; + tensor state_347_cast_fp16 = add(x = var_6368_cast_fp16, y = var_6372_cast_fp16)[name = string("state_347_cast_fp16")]; + tensor var_6375_begin_0 = const()[name = string("op_6375_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_6375_end_0 = const()[name = string("op_6375_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_6375_end_mask_0 = const()[name = string("op_6375_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6375_squeeze_mask_0 = const()[name = string("op_6375_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6375_cast_fp16 = slice_by_index(begin = var_6375_begin_0, end = var_6375_end_0, end_mask = var_6375_end_mask_0, squeeze_mask = var_6375_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6375_cast_fp16")]; + tensor var_6377_begin_0 = const()[name = string("op_6377_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_6377_end_0 = const()[name = string("op_6377_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_6377_end_mask_0 = const()[name = string("op_6377_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6377_squeeze_mask_0 = const()[name = string("op_6377_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6377_cast_fp16 = slice_by_index(begin = var_6377_begin_0, end = var_6377_end_0, end_mask = var_6377_end_mask_0, squeeze_mask = var_6377_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6377_cast_fp16")]; + bool var_6378_transpose_x_0 = const()[name = string("op_6378_transpose_x_0"), val = bool(false)]; + bool var_6378_transpose_y_0 = const()[name = string("op_6378_transpose_y_0"), val = bool(false)]; + tensor var_6378_cast_fp16 = matmul(transpose_x = var_6378_transpose_x_0, transpose_y = var_6378_transpose_y_0, x = var_6377_cast_fp16, y = state_347_cast_fp16)[name = string("op_6378_cast_fp16")]; + tensor v_new_347_cast_fp16 = sub(x = var_6375_cast_fp16, y = var_6378_cast_fp16)[name = string("v_new_347_cast_fp16")]; + tensor var_6381_begin_0 = const()[name = string("op_6381_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_6381_end_0 = const()[name = string("op_6381_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_6381_end_mask_0 = const()[name = string("op_6381_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6381_squeeze_mask_0 = const()[name = string("op_6381_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6381_cast_fp16 = slice_by_index(begin = var_6381_begin_0, end = var_6381_end_0, end_mask = var_6381_end_mask_0, squeeze_mask = var_6381_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6381_cast_fp16")]; + bool var_6382_transpose_x_0 = const()[name = string("op_6382_transpose_x_0"), val = bool(false)]; + bool var_6382_transpose_y_0 = const()[name = string("op_6382_transpose_y_0"), val = bool(false)]; + tensor var_6382_cast_fp16 = matmul(transpose_x = var_6382_transpose_x_0, transpose_y = var_6382_transpose_y_0, x = var_6381_cast_fp16, y = state_347_cast_fp16)[name = string("op_6382_cast_fp16")]; + tensor var_6384_begin_0 = const()[name = string("op_6384_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_6384_end_0 = const()[name = string("op_6384_end_0"), val = tensor([32, 14, 64, 64])]; + tensor var_6384_end_mask_0 = const()[name = string("op_6384_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6384_squeeze_mask_0 = const()[name = string("op_6384_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6384_cast_fp16 = slice_by_index(begin = var_6384_begin_0, end = var_6384_end_0, end_mask = var_6384_end_mask_0, squeeze_mask = var_6384_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6384_cast_fp16")]; + bool var_6385_transpose_x_0 = const()[name = string("op_6385_transpose_x_0"), val = bool(false)]; + bool var_6385_transpose_y_0 = const()[name = string("op_6385_transpose_y_0"), val = bool(false)]; + tensor var_6385_cast_fp16 = matmul(transpose_x = var_6385_transpose_x_0, transpose_y = var_6385_transpose_y_0, x = var_6384_cast_fp16, y = v_new_347_cast_fp16)[name = string("op_6385_cast_fp16")]; + tensor var_6386_cast_fp16 = add(x = var_6382_cast_fp16, y = var_6385_cast_fp16)[name = string("op_6386_cast_fp16")]; + tensor var_6388_begin_0 = const()[name = string("op_6388_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_6388_end_0 = const()[name = string("op_6388_end_0"), val = tensor([32, 14, 1, 1])]; + tensor var_6388_end_mask_0 = const()[name = string("op_6388_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6388_squeeze_mask_0 = const()[name = string("op_6388_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6388_cast_fp16 = slice_by_index(begin = var_6388_begin_0, end = var_6388_end_0, end_mask = var_6388_end_mask_0, squeeze_mask = var_6388_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6388_cast_fp16")]; + tensor var_6389_cast_fp16 = mul(x = state_347_cast_fp16, y = var_6388_cast_fp16)[name = string("op_6389_cast_fp16")]; + tensor var_6391_begin_0 = const()[name = string("op_6391_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_6391_end_0 = const()[name = string("op_6391_end_0"), val = tensor([32, 14, 64, 128])]; + tensor var_6391_end_mask_0 = const()[name = string("op_6391_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6391_squeeze_mask_0 = const()[name = string("op_6391_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6391_cast_fp16 = slice_by_index(begin = var_6391_begin_0, end = var_6391_end_0, end_mask = var_6391_end_mask_0, squeeze_mask = var_6391_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6391_cast_fp16")]; + bool var_6393_transpose_x_1 = const()[name = string("op_6393_transpose_x_1"), val = bool(true)]; + bool var_6393_transpose_y_1 = const()[name = string("op_6393_transpose_y_1"), val = bool(false)]; + tensor var_6393_cast_fp16 = matmul(transpose_x = var_6393_transpose_x_1, transpose_y = var_6393_transpose_y_1, x = var_6391_cast_fp16, y = v_new_347_cast_fp16)[name = string("op_6393_cast_fp16")]; + tensor state_349_cast_fp16 = add(x = var_6389_cast_fp16, y = var_6393_cast_fp16)[name = string("state_349_cast_fp16")]; + tensor var_6396_begin_0 = const()[name = string("op_6396_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_6396_end_0 = const()[name = string("op_6396_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_6396_end_mask_0 = const()[name = string("op_6396_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6396_squeeze_mask_0 = const()[name = string("op_6396_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6396_cast_fp16 = slice_by_index(begin = var_6396_begin_0, end = var_6396_end_0, end_mask = var_6396_end_mask_0, squeeze_mask = var_6396_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6396_cast_fp16")]; + tensor var_6398_begin_0 = const()[name = string("op_6398_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_6398_end_0 = const()[name = string("op_6398_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_6398_end_mask_0 = const()[name = string("op_6398_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6398_squeeze_mask_0 = const()[name = string("op_6398_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6398_cast_fp16 = slice_by_index(begin = var_6398_begin_0, end = var_6398_end_0, end_mask = var_6398_end_mask_0, squeeze_mask = var_6398_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6398_cast_fp16")]; + bool var_6399_transpose_x_0 = const()[name = string("op_6399_transpose_x_0"), val = bool(false)]; + bool var_6399_transpose_y_0 = const()[name = string("op_6399_transpose_y_0"), val = bool(false)]; + tensor var_6399_cast_fp16 = matmul(transpose_x = var_6399_transpose_x_0, transpose_y = var_6399_transpose_y_0, x = var_6398_cast_fp16, y = state_349_cast_fp16)[name = string("op_6399_cast_fp16")]; + tensor v_new_349_cast_fp16 = sub(x = var_6396_cast_fp16, y = var_6399_cast_fp16)[name = string("v_new_349_cast_fp16")]; + tensor var_6402_begin_0 = const()[name = string("op_6402_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_6402_end_0 = const()[name = string("op_6402_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_6402_end_mask_0 = const()[name = string("op_6402_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6402_squeeze_mask_0 = const()[name = string("op_6402_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6402_cast_fp16 = slice_by_index(begin = var_6402_begin_0, end = var_6402_end_0, end_mask = var_6402_end_mask_0, squeeze_mask = var_6402_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6402_cast_fp16")]; + bool var_6403_transpose_x_0 = const()[name = string("op_6403_transpose_x_0"), val = bool(false)]; + bool var_6403_transpose_y_0 = const()[name = string("op_6403_transpose_y_0"), val = bool(false)]; + tensor var_6403_cast_fp16 = matmul(transpose_x = var_6403_transpose_x_0, transpose_y = var_6403_transpose_y_0, x = var_6402_cast_fp16, y = state_349_cast_fp16)[name = string("op_6403_cast_fp16")]; + tensor var_6405_begin_0 = const()[name = string("op_6405_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_6405_end_0 = const()[name = string("op_6405_end_0"), val = tensor([32, 15, 64, 64])]; + tensor var_6405_end_mask_0 = const()[name = string("op_6405_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6405_squeeze_mask_0 = const()[name = string("op_6405_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6405_cast_fp16 = slice_by_index(begin = var_6405_begin_0, end = var_6405_end_0, end_mask = var_6405_end_mask_0, squeeze_mask = var_6405_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6405_cast_fp16")]; + bool var_6406_transpose_x_0 = const()[name = string("op_6406_transpose_x_0"), val = bool(false)]; + bool var_6406_transpose_y_0 = const()[name = string("op_6406_transpose_y_0"), val = bool(false)]; + tensor var_6406_cast_fp16 = matmul(transpose_x = var_6406_transpose_x_0, transpose_y = var_6406_transpose_y_0, x = var_6405_cast_fp16, y = v_new_349_cast_fp16)[name = string("op_6406_cast_fp16")]; + tensor var_6407_cast_fp16 = add(x = var_6403_cast_fp16, y = var_6406_cast_fp16)[name = string("op_6407_cast_fp16")]; + tensor var_6409_begin_0 = const()[name = string("op_6409_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_6409_end_0 = const()[name = string("op_6409_end_0"), val = tensor([32, 15, 1, 1])]; + tensor var_6409_end_mask_0 = const()[name = string("op_6409_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6409_squeeze_mask_0 = const()[name = string("op_6409_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6409_cast_fp16 = slice_by_index(begin = var_6409_begin_0, end = var_6409_end_0, end_mask = var_6409_end_mask_0, squeeze_mask = var_6409_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6409_cast_fp16")]; + tensor var_6410_cast_fp16 = mul(x = state_349_cast_fp16, y = var_6409_cast_fp16)[name = string("op_6410_cast_fp16")]; + tensor var_6412_begin_0 = const()[name = string("op_6412_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_6412_end_0 = const()[name = string("op_6412_end_0"), val = tensor([32, 15, 64, 128])]; + tensor var_6412_end_mask_0 = const()[name = string("op_6412_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6412_squeeze_mask_0 = const()[name = string("op_6412_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6412_cast_fp16 = slice_by_index(begin = var_6412_begin_0, end = var_6412_end_0, end_mask = var_6412_end_mask_0, squeeze_mask = var_6412_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6412_cast_fp16")]; + bool var_6414_transpose_x_1 = const()[name = string("op_6414_transpose_x_1"), val = bool(true)]; + bool var_6414_transpose_y_1 = const()[name = string("op_6414_transpose_y_1"), val = bool(false)]; + tensor var_6414_cast_fp16 = matmul(transpose_x = var_6414_transpose_x_1, transpose_y = var_6414_transpose_y_1, x = var_6412_cast_fp16, y = v_new_349_cast_fp16)[name = string("op_6414_cast_fp16")]; + tensor state_351_cast_fp16 = add(x = var_6410_cast_fp16, y = var_6414_cast_fp16)[name = string("state_351_cast_fp16")]; + tensor var_6417_begin_0 = const()[name = string("op_6417_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_6417_end_0 = const()[name = string("op_6417_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_6417_end_mask_0 = const()[name = string("op_6417_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6417_squeeze_mask_0 = const()[name = string("op_6417_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6417_cast_fp16 = slice_by_index(begin = var_6417_begin_0, end = var_6417_end_0, end_mask = var_6417_end_mask_0, squeeze_mask = var_6417_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6417_cast_fp16")]; + tensor var_6419_begin_0 = const()[name = string("op_6419_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_6419_end_0 = const()[name = string("op_6419_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_6419_end_mask_0 = const()[name = string("op_6419_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6419_squeeze_mask_0 = const()[name = string("op_6419_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6419_cast_fp16 = slice_by_index(begin = var_6419_begin_0, end = var_6419_end_0, end_mask = var_6419_end_mask_0, squeeze_mask = var_6419_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6419_cast_fp16")]; + bool var_6420_transpose_x_0 = const()[name = string("op_6420_transpose_x_0"), val = bool(false)]; + bool var_6420_transpose_y_0 = const()[name = string("op_6420_transpose_y_0"), val = bool(false)]; + tensor var_6420_cast_fp16 = matmul(transpose_x = var_6420_transpose_x_0, transpose_y = var_6420_transpose_y_0, x = var_6419_cast_fp16, y = state_351_cast_fp16)[name = string("op_6420_cast_fp16")]; + tensor v_new_351_cast_fp16 = sub(x = var_6417_cast_fp16, y = var_6420_cast_fp16)[name = string("v_new_351_cast_fp16")]; + tensor var_6423_begin_0 = const()[name = string("op_6423_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_6423_end_0 = const()[name = string("op_6423_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_6423_end_mask_0 = const()[name = string("op_6423_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6423_squeeze_mask_0 = const()[name = string("op_6423_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6423_cast_fp16 = slice_by_index(begin = var_6423_begin_0, end = var_6423_end_0, end_mask = var_6423_end_mask_0, squeeze_mask = var_6423_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6423_cast_fp16")]; + bool var_6424_transpose_x_0 = const()[name = string("op_6424_transpose_x_0"), val = bool(false)]; + bool var_6424_transpose_y_0 = const()[name = string("op_6424_transpose_y_0"), val = bool(false)]; + tensor var_6424_cast_fp16 = matmul(transpose_x = var_6424_transpose_x_0, transpose_y = var_6424_transpose_y_0, x = var_6423_cast_fp16, y = state_351_cast_fp16)[name = string("op_6424_cast_fp16")]; + tensor var_6426_begin_0 = const()[name = string("op_6426_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_6426_end_0 = const()[name = string("op_6426_end_0"), val = tensor([32, 16, 64, 64])]; + tensor var_6426_end_mask_0 = const()[name = string("op_6426_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6426_squeeze_mask_0 = const()[name = string("op_6426_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6426_cast_fp16 = slice_by_index(begin = var_6426_begin_0, end = var_6426_end_0, end_mask = var_6426_end_mask_0, squeeze_mask = var_6426_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6426_cast_fp16")]; + bool var_6427_transpose_x_0 = const()[name = string("op_6427_transpose_x_0"), val = bool(false)]; + bool var_6427_transpose_y_0 = const()[name = string("op_6427_transpose_y_0"), val = bool(false)]; + tensor var_6427_cast_fp16 = matmul(transpose_x = var_6427_transpose_x_0, transpose_y = var_6427_transpose_y_0, x = var_6426_cast_fp16, y = v_new_351_cast_fp16)[name = string("op_6427_cast_fp16")]; + tensor var_6428_cast_fp16 = add(x = var_6424_cast_fp16, y = var_6427_cast_fp16)[name = string("op_6428_cast_fp16")]; + tensor var_6430_begin_0 = const()[name = string("op_6430_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_6430_end_0 = const()[name = string("op_6430_end_0"), val = tensor([32, 16, 1, 1])]; + tensor var_6430_end_mask_0 = const()[name = string("op_6430_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6430_squeeze_mask_0 = const()[name = string("op_6430_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6430_cast_fp16 = slice_by_index(begin = var_6430_begin_0, end = var_6430_end_0, end_mask = var_6430_end_mask_0, squeeze_mask = var_6430_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6430_cast_fp16")]; + tensor var_6431_cast_fp16 = mul(x = state_351_cast_fp16, y = var_6430_cast_fp16)[name = string("op_6431_cast_fp16")]; + tensor var_6433_begin_0 = const()[name = string("op_6433_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_6433_end_0 = const()[name = string("op_6433_end_0"), val = tensor([32, 16, 64, 128])]; + tensor var_6433_end_mask_0 = const()[name = string("op_6433_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6433_squeeze_mask_0 = const()[name = string("op_6433_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6433_cast_fp16 = slice_by_index(begin = var_6433_begin_0, end = var_6433_end_0, end_mask = var_6433_end_mask_0, squeeze_mask = var_6433_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6433_cast_fp16")]; + bool var_6435_transpose_x_1 = const()[name = string("op_6435_transpose_x_1"), val = bool(true)]; + bool var_6435_transpose_y_1 = const()[name = string("op_6435_transpose_y_1"), val = bool(false)]; + tensor var_6435_cast_fp16 = matmul(transpose_x = var_6435_transpose_x_1, transpose_y = var_6435_transpose_y_1, x = var_6433_cast_fp16, y = v_new_351_cast_fp16)[name = string("op_6435_cast_fp16")]; + tensor state_353_cast_fp16 = add(x = var_6431_cast_fp16, y = var_6435_cast_fp16)[name = string("state_353_cast_fp16")]; + tensor var_6438_begin_0 = const()[name = string("op_6438_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_6438_end_0 = const()[name = string("op_6438_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_6438_end_mask_0 = const()[name = string("op_6438_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6438_squeeze_mask_0 = const()[name = string("op_6438_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6438_cast_fp16 = slice_by_index(begin = var_6438_begin_0, end = var_6438_end_0, end_mask = var_6438_end_mask_0, squeeze_mask = var_6438_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6438_cast_fp16")]; + tensor var_6440_begin_0 = const()[name = string("op_6440_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_6440_end_0 = const()[name = string("op_6440_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_6440_end_mask_0 = const()[name = string("op_6440_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6440_squeeze_mask_0 = const()[name = string("op_6440_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6440_cast_fp16 = slice_by_index(begin = var_6440_begin_0, end = var_6440_end_0, end_mask = var_6440_end_mask_0, squeeze_mask = var_6440_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6440_cast_fp16")]; + bool var_6441_transpose_x_0 = const()[name = string("op_6441_transpose_x_0"), val = bool(false)]; + bool var_6441_transpose_y_0 = const()[name = string("op_6441_transpose_y_0"), val = bool(false)]; + tensor var_6441_cast_fp16 = matmul(transpose_x = var_6441_transpose_x_0, transpose_y = var_6441_transpose_y_0, x = var_6440_cast_fp16, y = state_353_cast_fp16)[name = string("op_6441_cast_fp16")]; + tensor v_new_353_cast_fp16 = sub(x = var_6438_cast_fp16, y = var_6441_cast_fp16)[name = string("v_new_353_cast_fp16")]; + tensor var_6444_begin_0 = const()[name = string("op_6444_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_6444_end_0 = const()[name = string("op_6444_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_6444_end_mask_0 = const()[name = string("op_6444_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6444_squeeze_mask_0 = const()[name = string("op_6444_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6444_cast_fp16 = slice_by_index(begin = var_6444_begin_0, end = var_6444_end_0, end_mask = var_6444_end_mask_0, squeeze_mask = var_6444_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6444_cast_fp16")]; + bool var_6445_transpose_x_0 = const()[name = string("op_6445_transpose_x_0"), val = bool(false)]; + bool var_6445_transpose_y_0 = const()[name = string("op_6445_transpose_y_0"), val = bool(false)]; + tensor var_6445_cast_fp16 = matmul(transpose_x = var_6445_transpose_x_0, transpose_y = var_6445_transpose_y_0, x = var_6444_cast_fp16, y = state_353_cast_fp16)[name = string("op_6445_cast_fp16")]; + tensor var_6447_begin_0 = const()[name = string("op_6447_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_6447_end_0 = const()[name = string("op_6447_end_0"), val = tensor([32, 17, 64, 64])]; + tensor var_6447_end_mask_0 = const()[name = string("op_6447_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6447_squeeze_mask_0 = const()[name = string("op_6447_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6447_cast_fp16 = slice_by_index(begin = var_6447_begin_0, end = var_6447_end_0, end_mask = var_6447_end_mask_0, squeeze_mask = var_6447_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6447_cast_fp16")]; + bool var_6448_transpose_x_0 = const()[name = string("op_6448_transpose_x_0"), val = bool(false)]; + bool var_6448_transpose_y_0 = const()[name = string("op_6448_transpose_y_0"), val = bool(false)]; + tensor var_6448_cast_fp16 = matmul(transpose_x = var_6448_transpose_x_0, transpose_y = var_6448_transpose_y_0, x = var_6447_cast_fp16, y = v_new_353_cast_fp16)[name = string("op_6448_cast_fp16")]; + tensor var_6449_cast_fp16 = add(x = var_6445_cast_fp16, y = var_6448_cast_fp16)[name = string("op_6449_cast_fp16")]; + tensor var_6451_begin_0 = const()[name = string("op_6451_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_6451_end_0 = const()[name = string("op_6451_end_0"), val = tensor([32, 17, 1, 1])]; + tensor var_6451_end_mask_0 = const()[name = string("op_6451_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6451_squeeze_mask_0 = const()[name = string("op_6451_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6451_cast_fp16 = slice_by_index(begin = var_6451_begin_0, end = var_6451_end_0, end_mask = var_6451_end_mask_0, squeeze_mask = var_6451_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6451_cast_fp16")]; + tensor var_6452_cast_fp16 = mul(x = state_353_cast_fp16, y = var_6451_cast_fp16)[name = string("op_6452_cast_fp16")]; + tensor var_6454_begin_0 = const()[name = string("op_6454_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_6454_end_0 = const()[name = string("op_6454_end_0"), val = tensor([32, 17, 64, 128])]; + tensor var_6454_end_mask_0 = const()[name = string("op_6454_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6454_squeeze_mask_0 = const()[name = string("op_6454_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6454_cast_fp16 = slice_by_index(begin = var_6454_begin_0, end = var_6454_end_0, end_mask = var_6454_end_mask_0, squeeze_mask = var_6454_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6454_cast_fp16")]; + bool var_6456_transpose_x_1 = const()[name = string("op_6456_transpose_x_1"), val = bool(true)]; + bool var_6456_transpose_y_1 = const()[name = string("op_6456_transpose_y_1"), val = bool(false)]; + tensor var_6456_cast_fp16 = matmul(transpose_x = var_6456_transpose_x_1, transpose_y = var_6456_transpose_y_1, x = var_6454_cast_fp16, y = v_new_353_cast_fp16)[name = string("op_6456_cast_fp16")]; + tensor state_355_cast_fp16 = add(x = var_6452_cast_fp16, y = var_6456_cast_fp16)[name = string("state_355_cast_fp16")]; + tensor var_6459_begin_0 = const()[name = string("op_6459_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_6459_end_0 = const()[name = string("op_6459_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_6459_end_mask_0 = const()[name = string("op_6459_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6459_squeeze_mask_0 = const()[name = string("op_6459_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6459_cast_fp16 = slice_by_index(begin = var_6459_begin_0, end = var_6459_end_0, end_mask = var_6459_end_mask_0, squeeze_mask = var_6459_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6459_cast_fp16")]; + tensor var_6461_begin_0 = const()[name = string("op_6461_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_6461_end_0 = const()[name = string("op_6461_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_6461_end_mask_0 = const()[name = string("op_6461_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6461_squeeze_mask_0 = const()[name = string("op_6461_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6461_cast_fp16 = slice_by_index(begin = var_6461_begin_0, end = var_6461_end_0, end_mask = var_6461_end_mask_0, squeeze_mask = var_6461_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6461_cast_fp16")]; + bool var_6462_transpose_x_0 = const()[name = string("op_6462_transpose_x_0"), val = bool(false)]; + bool var_6462_transpose_y_0 = const()[name = string("op_6462_transpose_y_0"), val = bool(false)]; + tensor var_6462_cast_fp16 = matmul(transpose_x = var_6462_transpose_x_0, transpose_y = var_6462_transpose_y_0, x = var_6461_cast_fp16, y = state_355_cast_fp16)[name = string("op_6462_cast_fp16")]; + tensor v_new_355_cast_fp16 = sub(x = var_6459_cast_fp16, y = var_6462_cast_fp16)[name = string("v_new_355_cast_fp16")]; + tensor var_6465_begin_0 = const()[name = string("op_6465_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_6465_end_0 = const()[name = string("op_6465_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_6465_end_mask_0 = const()[name = string("op_6465_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6465_squeeze_mask_0 = const()[name = string("op_6465_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6465_cast_fp16 = slice_by_index(begin = var_6465_begin_0, end = var_6465_end_0, end_mask = var_6465_end_mask_0, squeeze_mask = var_6465_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6465_cast_fp16")]; + bool var_6466_transpose_x_0 = const()[name = string("op_6466_transpose_x_0"), val = bool(false)]; + bool var_6466_transpose_y_0 = const()[name = string("op_6466_transpose_y_0"), val = bool(false)]; + tensor var_6466_cast_fp16 = matmul(transpose_x = var_6466_transpose_x_0, transpose_y = var_6466_transpose_y_0, x = var_6465_cast_fp16, y = state_355_cast_fp16)[name = string("op_6466_cast_fp16")]; + tensor var_6468_begin_0 = const()[name = string("op_6468_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_6468_end_0 = const()[name = string("op_6468_end_0"), val = tensor([32, 18, 64, 64])]; + tensor var_6468_end_mask_0 = const()[name = string("op_6468_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6468_squeeze_mask_0 = const()[name = string("op_6468_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6468_cast_fp16 = slice_by_index(begin = var_6468_begin_0, end = var_6468_end_0, end_mask = var_6468_end_mask_0, squeeze_mask = var_6468_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6468_cast_fp16")]; + bool var_6469_transpose_x_0 = const()[name = string("op_6469_transpose_x_0"), val = bool(false)]; + bool var_6469_transpose_y_0 = const()[name = string("op_6469_transpose_y_0"), val = bool(false)]; + tensor var_6469_cast_fp16 = matmul(transpose_x = var_6469_transpose_x_0, transpose_y = var_6469_transpose_y_0, x = var_6468_cast_fp16, y = v_new_355_cast_fp16)[name = string("op_6469_cast_fp16")]; + tensor var_6470_cast_fp16 = add(x = var_6466_cast_fp16, y = var_6469_cast_fp16)[name = string("op_6470_cast_fp16")]; + tensor var_6472_begin_0 = const()[name = string("op_6472_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_6472_end_0 = const()[name = string("op_6472_end_0"), val = tensor([32, 18, 1, 1])]; + tensor var_6472_end_mask_0 = const()[name = string("op_6472_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6472_squeeze_mask_0 = const()[name = string("op_6472_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6472_cast_fp16 = slice_by_index(begin = var_6472_begin_0, end = var_6472_end_0, end_mask = var_6472_end_mask_0, squeeze_mask = var_6472_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6472_cast_fp16")]; + tensor var_6473_cast_fp16 = mul(x = state_355_cast_fp16, y = var_6472_cast_fp16)[name = string("op_6473_cast_fp16")]; + tensor var_6475_begin_0 = const()[name = string("op_6475_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_6475_end_0 = const()[name = string("op_6475_end_0"), val = tensor([32, 18, 64, 128])]; + tensor var_6475_end_mask_0 = const()[name = string("op_6475_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6475_squeeze_mask_0 = const()[name = string("op_6475_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6475_cast_fp16 = slice_by_index(begin = var_6475_begin_0, end = var_6475_end_0, end_mask = var_6475_end_mask_0, squeeze_mask = var_6475_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6475_cast_fp16")]; + bool var_6477_transpose_x_1 = const()[name = string("op_6477_transpose_x_1"), val = bool(true)]; + bool var_6477_transpose_y_1 = const()[name = string("op_6477_transpose_y_1"), val = bool(false)]; + tensor var_6477_cast_fp16 = matmul(transpose_x = var_6477_transpose_x_1, transpose_y = var_6477_transpose_y_1, x = var_6475_cast_fp16, y = v_new_355_cast_fp16)[name = string("op_6477_cast_fp16")]; + tensor state_357_cast_fp16 = add(x = var_6473_cast_fp16, y = var_6477_cast_fp16)[name = string("state_357_cast_fp16")]; + tensor var_6480_begin_0 = const()[name = string("op_6480_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_6480_end_0 = const()[name = string("op_6480_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_6480_end_mask_0 = const()[name = string("op_6480_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6480_squeeze_mask_0 = const()[name = string("op_6480_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6480_cast_fp16 = slice_by_index(begin = var_6480_begin_0, end = var_6480_end_0, end_mask = var_6480_end_mask_0, squeeze_mask = var_6480_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6480_cast_fp16")]; + tensor var_6482_begin_0 = const()[name = string("op_6482_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_6482_end_0 = const()[name = string("op_6482_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_6482_end_mask_0 = const()[name = string("op_6482_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6482_squeeze_mask_0 = const()[name = string("op_6482_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6482_cast_fp16 = slice_by_index(begin = var_6482_begin_0, end = var_6482_end_0, end_mask = var_6482_end_mask_0, squeeze_mask = var_6482_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6482_cast_fp16")]; + bool var_6483_transpose_x_0 = const()[name = string("op_6483_transpose_x_0"), val = bool(false)]; + bool var_6483_transpose_y_0 = const()[name = string("op_6483_transpose_y_0"), val = bool(false)]; + tensor var_6483_cast_fp16 = matmul(transpose_x = var_6483_transpose_x_0, transpose_y = var_6483_transpose_y_0, x = var_6482_cast_fp16, y = state_357_cast_fp16)[name = string("op_6483_cast_fp16")]; + tensor v_new_357_cast_fp16 = sub(x = var_6480_cast_fp16, y = var_6483_cast_fp16)[name = string("v_new_357_cast_fp16")]; + tensor var_6486_begin_0 = const()[name = string("op_6486_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_6486_end_0 = const()[name = string("op_6486_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_6486_end_mask_0 = const()[name = string("op_6486_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6486_squeeze_mask_0 = const()[name = string("op_6486_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6486_cast_fp16 = slice_by_index(begin = var_6486_begin_0, end = var_6486_end_0, end_mask = var_6486_end_mask_0, squeeze_mask = var_6486_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6486_cast_fp16")]; + bool var_6487_transpose_x_0 = const()[name = string("op_6487_transpose_x_0"), val = bool(false)]; + bool var_6487_transpose_y_0 = const()[name = string("op_6487_transpose_y_0"), val = bool(false)]; + tensor var_6487_cast_fp16 = matmul(transpose_x = var_6487_transpose_x_0, transpose_y = var_6487_transpose_y_0, x = var_6486_cast_fp16, y = state_357_cast_fp16)[name = string("op_6487_cast_fp16")]; + tensor var_6489_begin_0 = const()[name = string("op_6489_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_6489_end_0 = const()[name = string("op_6489_end_0"), val = tensor([32, 19, 64, 64])]; + tensor var_6489_end_mask_0 = const()[name = string("op_6489_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6489_squeeze_mask_0 = const()[name = string("op_6489_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6489_cast_fp16 = slice_by_index(begin = var_6489_begin_0, end = var_6489_end_0, end_mask = var_6489_end_mask_0, squeeze_mask = var_6489_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6489_cast_fp16")]; + bool var_6490_transpose_x_0 = const()[name = string("op_6490_transpose_x_0"), val = bool(false)]; + bool var_6490_transpose_y_0 = const()[name = string("op_6490_transpose_y_0"), val = bool(false)]; + tensor var_6490_cast_fp16 = matmul(transpose_x = var_6490_transpose_x_0, transpose_y = var_6490_transpose_y_0, x = var_6489_cast_fp16, y = v_new_357_cast_fp16)[name = string("op_6490_cast_fp16")]; + tensor var_6491_cast_fp16 = add(x = var_6487_cast_fp16, y = var_6490_cast_fp16)[name = string("op_6491_cast_fp16")]; + tensor var_6493_begin_0 = const()[name = string("op_6493_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_6493_end_0 = const()[name = string("op_6493_end_0"), val = tensor([32, 19, 1, 1])]; + tensor var_6493_end_mask_0 = const()[name = string("op_6493_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6493_squeeze_mask_0 = const()[name = string("op_6493_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6493_cast_fp16 = slice_by_index(begin = var_6493_begin_0, end = var_6493_end_0, end_mask = var_6493_end_mask_0, squeeze_mask = var_6493_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6493_cast_fp16")]; + tensor var_6494_cast_fp16 = mul(x = state_357_cast_fp16, y = var_6493_cast_fp16)[name = string("op_6494_cast_fp16")]; + tensor var_6496_begin_0 = const()[name = string("op_6496_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_6496_end_0 = const()[name = string("op_6496_end_0"), val = tensor([32, 19, 64, 128])]; + tensor var_6496_end_mask_0 = const()[name = string("op_6496_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6496_squeeze_mask_0 = const()[name = string("op_6496_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6496_cast_fp16 = slice_by_index(begin = var_6496_begin_0, end = var_6496_end_0, end_mask = var_6496_end_mask_0, squeeze_mask = var_6496_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6496_cast_fp16")]; + bool var_6498_transpose_x_1 = const()[name = string("op_6498_transpose_x_1"), val = bool(true)]; + bool var_6498_transpose_y_1 = const()[name = string("op_6498_transpose_y_1"), val = bool(false)]; + tensor var_6498_cast_fp16 = matmul(transpose_x = var_6498_transpose_x_1, transpose_y = var_6498_transpose_y_1, x = var_6496_cast_fp16, y = v_new_357_cast_fp16)[name = string("op_6498_cast_fp16")]; + tensor state_359_cast_fp16 = add(x = var_6494_cast_fp16, y = var_6498_cast_fp16)[name = string("state_359_cast_fp16")]; + tensor var_6501_begin_0 = const()[name = string("op_6501_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_6501_end_0 = const()[name = string("op_6501_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_6501_end_mask_0 = const()[name = string("op_6501_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6501_squeeze_mask_0 = const()[name = string("op_6501_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6501_cast_fp16 = slice_by_index(begin = var_6501_begin_0, end = var_6501_end_0, end_mask = var_6501_end_mask_0, squeeze_mask = var_6501_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6501_cast_fp16")]; + tensor var_6503_begin_0 = const()[name = string("op_6503_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_6503_end_0 = const()[name = string("op_6503_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_6503_end_mask_0 = const()[name = string("op_6503_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6503_squeeze_mask_0 = const()[name = string("op_6503_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6503_cast_fp16 = slice_by_index(begin = var_6503_begin_0, end = var_6503_end_0, end_mask = var_6503_end_mask_0, squeeze_mask = var_6503_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6503_cast_fp16")]; + bool var_6504_transpose_x_0 = const()[name = string("op_6504_transpose_x_0"), val = bool(false)]; + bool var_6504_transpose_y_0 = const()[name = string("op_6504_transpose_y_0"), val = bool(false)]; + tensor var_6504_cast_fp16 = matmul(transpose_x = var_6504_transpose_x_0, transpose_y = var_6504_transpose_y_0, x = var_6503_cast_fp16, y = state_359_cast_fp16)[name = string("op_6504_cast_fp16")]; + tensor v_new_359_cast_fp16 = sub(x = var_6501_cast_fp16, y = var_6504_cast_fp16)[name = string("v_new_359_cast_fp16")]; + tensor var_6507_begin_0 = const()[name = string("op_6507_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_6507_end_0 = const()[name = string("op_6507_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_6507_end_mask_0 = const()[name = string("op_6507_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6507_squeeze_mask_0 = const()[name = string("op_6507_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6507_cast_fp16 = slice_by_index(begin = var_6507_begin_0, end = var_6507_end_0, end_mask = var_6507_end_mask_0, squeeze_mask = var_6507_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6507_cast_fp16")]; + bool var_6508_transpose_x_0 = const()[name = string("op_6508_transpose_x_0"), val = bool(false)]; + bool var_6508_transpose_y_0 = const()[name = string("op_6508_transpose_y_0"), val = bool(false)]; + tensor var_6508_cast_fp16 = matmul(transpose_x = var_6508_transpose_x_0, transpose_y = var_6508_transpose_y_0, x = var_6507_cast_fp16, y = state_359_cast_fp16)[name = string("op_6508_cast_fp16")]; + tensor var_6510_begin_0 = const()[name = string("op_6510_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_6510_end_0 = const()[name = string("op_6510_end_0"), val = tensor([32, 20, 64, 64])]; + tensor var_6510_end_mask_0 = const()[name = string("op_6510_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6510_squeeze_mask_0 = const()[name = string("op_6510_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6510_cast_fp16 = slice_by_index(begin = var_6510_begin_0, end = var_6510_end_0, end_mask = var_6510_end_mask_0, squeeze_mask = var_6510_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6510_cast_fp16")]; + bool var_6511_transpose_x_0 = const()[name = string("op_6511_transpose_x_0"), val = bool(false)]; + bool var_6511_transpose_y_0 = const()[name = string("op_6511_transpose_y_0"), val = bool(false)]; + tensor var_6511_cast_fp16 = matmul(transpose_x = var_6511_transpose_x_0, transpose_y = var_6511_transpose_y_0, x = var_6510_cast_fp16, y = v_new_359_cast_fp16)[name = string("op_6511_cast_fp16")]; + tensor var_6512_cast_fp16 = add(x = var_6508_cast_fp16, y = var_6511_cast_fp16)[name = string("op_6512_cast_fp16")]; + tensor var_6514_begin_0 = const()[name = string("op_6514_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_6514_end_0 = const()[name = string("op_6514_end_0"), val = tensor([32, 20, 1, 1])]; + tensor var_6514_end_mask_0 = const()[name = string("op_6514_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6514_squeeze_mask_0 = const()[name = string("op_6514_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6514_cast_fp16 = slice_by_index(begin = var_6514_begin_0, end = var_6514_end_0, end_mask = var_6514_end_mask_0, squeeze_mask = var_6514_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6514_cast_fp16")]; + tensor var_6515_cast_fp16 = mul(x = state_359_cast_fp16, y = var_6514_cast_fp16)[name = string("op_6515_cast_fp16")]; + tensor var_6517_begin_0 = const()[name = string("op_6517_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_6517_end_0 = const()[name = string("op_6517_end_0"), val = tensor([32, 20, 64, 128])]; + tensor var_6517_end_mask_0 = const()[name = string("op_6517_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6517_squeeze_mask_0 = const()[name = string("op_6517_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6517_cast_fp16 = slice_by_index(begin = var_6517_begin_0, end = var_6517_end_0, end_mask = var_6517_end_mask_0, squeeze_mask = var_6517_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6517_cast_fp16")]; + bool var_6519_transpose_x_1 = const()[name = string("op_6519_transpose_x_1"), val = bool(true)]; + bool var_6519_transpose_y_1 = const()[name = string("op_6519_transpose_y_1"), val = bool(false)]; + tensor var_6519_cast_fp16 = matmul(transpose_x = var_6519_transpose_x_1, transpose_y = var_6519_transpose_y_1, x = var_6517_cast_fp16, y = v_new_359_cast_fp16)[name = string("op_6519_cast_fp16")]; + tensor state_361_cast_fp16 = add(x = var_6515_cast_fp16, y = var_6519_cast_fp16)[name = string("state_361_cast_fp16")]; + tensor var_6522_begin_0 = const()[name = string("op_6522_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_6522_end_0 = const()[name = string("op_6522_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_6522_end_mask_0 = const()[name = string("op_6522_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6522_squeeze_mask_0 = const()[name = string("op_6522_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6522_cast_fp16 = slice_by_index(begin = var_6522_begin_0, end = var_6522_end_0, end_mask = var_6522_end_mask_0, squeeze_mask = var_6522_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6522_cast_fp16")]; + tensor var_6524_begin_0 = const()[name = string("op_6524_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_6524_end_0 = const()[name = string("op_6524_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_6524_end_mask_0 = const()[name = string("op_6524_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6524_squeeze_mask_0 = const()[name = string("op_6524_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6524_cast_fp16 = slice_by_index(begin = var_6524_begin_0, end = var_6524_end_0, end_mask = var_6524_end_mask_0, squeeze_mask = var_6524_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6524_cast_fp16")]; + bool var_6525_transpose_x_0 = const()[name = string("op_6525_transpose_x_0"), val = bool(false)]; + bool var_6525_transpose_y_0 = const()[name = string("op_6525_transpose_y_0"), val = bool(false)]; + tensor var_6525_cast_fp16 = matmul(transpose_x = var_6525_transpose_x_0, transpose_y = var_6525_transpose_y_0, x = var_6524_cast_fp16, y = state_361_cast_fp16)[name = string("op_6525_cast_fp16")]; + tensor v_new_361_cast_fp16 = sub(x = var_6522_cast_fp16, y = var_6525_cast_fp16)[name = string("v_new_361_cast_fp16")]; + tensor var_6528_begin_0 = const()[name = string("op_6528_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_6528_end_0 = const()[name = string("op_6528_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_6528_end_mask_0 = const()[name = string("op_6528_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6528_squeeze_mask_0 = const()[name = string("op_6528_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6528_cast_fp16 = slice_by_index(begin = var_6528_begin_0, end = var_6528_end_0, end_mask = var_6528_end_mask_0, squeeze_mask = var_6528_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6528_cast_fp16")]; + bool var_6529_transpose_x_0 = const()[name = string("op_6529_transpose_x_0"), val = bool(false)]; + bool var_6529_transpose_y_0 = const()[name = string("op_6529_transpose_y_0"), val = bool(false)]; + tensor var_6529_cast_fp16 = matmul(transpose_x = var_6529_transpose_x_0, transpose_y = var_6529_transpose_y_0, x = var_6528_cast_fp16, y = state_361_cast_fp16)[name = string("op_6529_cast_fp16")]; + tensor var_6531_begin_0 = const()[name = string("op_6531_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_6531_end_0 = const()[name = string("op_6531_end_0"), val = tensor([32, 21, 64, 64])]; + tensor var_6531_end_mask_0 = const()[name = string("op_6531_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6531_squeeze_mask_0 = const()[name = string("op_6531_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6531_cast_fp16 = slice_by_index(begin = var_6531_begin_0, end = var_6531_end_0, end_mask = var_6531_end_mask_0, squeeze_mask = var_6531_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6531_cast_fp16")]; + bool var_6532_transpose_x_0 = const()[name = string("op_6532_transpose_x_0"), val = bool(false)]; + bool var_6532_transpose_y_0 = const()[name = string("op_6532_transpose_y_0"), val = bool(false)]; + tensor var_6532_cast_fp16 = matmul(transpose_x = var_6532_transpose_x_0, transpose_y = var_6532_transpose_y_0, x = var_6531_cast_fp16, y = v_new_361_cast_fp16)[name = string("op_6532_cast_fp16")]; + tensor var_6533_cast_fp16 = add(x = var_6529_cast_fp16, y = var_6532_cast_fp16)[name = string("op_6533_cast_fp16")]; + tensor var_6535_begin_0 = const()[name = string("op_6535_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_6535_end_0 = const()[name = string("op_6535_end_0"), val = tensor([32, 21, 1, 1])]; + tensor var_6535_end_mask_0 = const()[name = string("op_6535_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6535_squeeze_mask_0 = const()[name = string("op_6535_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6535_cast_fp16 = slice_by_index(begin = var_6535_begin_0, end = var_6535_end_0, end_mask = var_6535_end_mask_0, squeeze_mask = var_6535_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6535_cast_fp16")]; + tensor var_6536_cast_fp16 = mul(x = state_361_cast_fp16, y = var_6535_cast_fp16)[name = string("op_6536_cast_fp16")]; + tensor var_6538_begin_0 = const()[name = string("op_6538_begin_0"), val = tensor([0, 20, 0, 0])]; + tensor var_6538_end_0 = const()[name = string("op_6538_end_0"), val = tensor([32, 21, 64, 128])]; + tensor var_6538_end_mask_0 = const()[name = string("op_6538_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6538_squeeze_mask_0 = const()[name = string("op_6538_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6538_cast_fp16 = slice_by_index(begin = var_6538_begin_0, end = var_6538_end_0, end_mask = var_6538_end_mask_0, squeeze_mask = var_6538_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6538_cast_fp16")]; + bool var_6540_transpose_x_1 = const()[name = string("op_6540_transpose_x_1"), val = bool(true)]; + bool var_6540_transpose_y_1 = const()[name = string("op_6540_transpose_y_1"), val = bool(false)]; + tensor var_6540_cast_fp16 = matmul(transpose_x = var_6540_transpose_x_1, transpose_y = var_6540_transpose_y_1, x = var_6538_cast_fp16, y = v_new_361_cast_fp16)[name = string("op_6540_cast_fp16")]; + tensor state_363_cast_fp16 = add(x = var_6536_cast_fp16, y = var_6540_cast_fp16)[name = string("state_363_cast_fp16")]; + tensor var_6543_begin_0 = const()[name = string("op_6543_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_6543_end_0 = const()[name = string("op_6543_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_6543_end_mask_0 = const()[name = string("op_6543_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6543_squeeze_mask_0 = const()[name = string("op_6543_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6543_cast_fp16 = slice_by_index(begin = var_6543_begin_0, end = var_6543_end_0, end_mask = var_6543_end_mask_0, squeeze_mask = var_6543_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6543_cast_fp16")]; + tensor var_6545_begin_0 = const()[name = string("op_6545_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_6545_end_0 = const()[name = string("op_6545_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_6545_end_mask_0 = const()[name = string("op_6545_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6545_squeeze_mask_0 = const()[name = string("op_6545_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6545_cast_fp16 = slice_by_index(begin = var_6545_begin_0, end = var_6545_end_0, end_mask = var_6545_end_mask_0, squeeze_mask = var_6545_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6545_cast_fp16")]; + bool var_6546_transpose_x_0 = const()[name = string("op_6546_transpose_x_0"), val = bool(false)]; + bool var_6546_transpose_y_0 = const()[name = string("op_6546_transpose_y_0"), val = bool(false)]; + tensor var_6546_cast_fp16 = matmul(transpose_x = var_6546_transpose_x_0, transpose_y = var_6546_transpose_y_0, x = var_6545_cast_fp16, y = state_363_cast_fp16)[name = string("op_6546_cast_fp16")]; + tensor v_new_363_cast_fp16 = sub(x = var_6543_cast_fp16, y = var_6546_cast_fp16)[name = string("v_new_363_cast_fp16")]; + tensor var_6549_begin_0 = const()[name = string("op_6549_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_6549_end_0 = const()[name = string("op_6549_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_6549_end_mask_0 = const()[name = string("op_6549_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6549_squeeze_mask_0 = const()[name = string("op_6549_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6549_cast_fp16 = slice_by_index(begin = var_6549_begin_0, end = var_6549_end_0, end_mask = var_6549_end_mask_0, squeeze_mask = var_6549_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6549_cast_fp16")]; + bool var_6550_transpose_x_0 = const()[name = string("op_6550_transpose_x_0"), val = bool(false)]; + bool var_6550_transpose_y_0 = const()[name = string("op_6550_transpose_y_0"), val = bool(false)]; + tensor var_6550_cast_fp16 = matmul(transpose_x = var_6550_transpose_x_0, transpose_y = var_6550_transpose_y_0, x = var_6549_cast_fp16, y = state_363_cast_fp16)[name = string("op_6550_cast_fp16")]; + tensor var_6552_begin_0 = const()[name = string("op_6552_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_6552_end_0 = const()[name = string("op_6552_end_0"), val = tensor([32, 22, 64, 64])]; + tensor var_6552_end_mask_0 = const()[name = string("op_6552_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6552_squeeze_mask_0 = const()[name = string("op_6552_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6552_cast_fp16 = slice_by_index(begin = var_6552_begin_0, end = var_6552_end_0, end_mask = var_6552_end_mask_0, squeeze_mask = var_6552_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6552_cast_fp16")]; + bool var_6553_transpose_x_0 = const()[name = string("op_6553_transpose_x_0"), val = bool(false)]; + bool var_6553_transpose_y_0 = const()[name = string("op_6553_transpose_y_0"), val = bool(false)]; + tensor var_6553_cast_fp16 = matmul(transpose_x = var_6553_transpose_x_0, transpose_y = var_6553_transpose_y_0, x = var_6552_cast_fp16, y = v_new_363_cast_fp16)[name = string("op_6553_cast_fp16")]; + tensor var_6554_cast_fp16 = add(x = var_6550_cast_fp16, y = var_6553_cast_fp16)[name = string("op_6554_cast_fp16")]; + tensor var_6556_begin_0 = const()[name = string("op_6556_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_6556_end_0 = const()[name = string("op_6556_end_0"), val = tensor([32, 22, 1, 1])]; + tensor var_6556_end_mask_0 = const()[name = string("op_6556_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6556_squeeze_mask_0 = const()[name = string("op_6556_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6556_cast_fp16 = slice_by_index(begin = var_6556_begin_0, end = var_6556_end_0, end_mask = var_6556_end_mask_0, squeeze_mask = var_6556_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6556_cast_fp16")]; + tensor var_6557_cast_fp16 = mul(x = state_363_cast_fp16, y = var_6556_cast_fp16)[name = string("op_6557_cast_fp16")]; + tensor var_6559_begin_0 = const()[name = string("op_6559_begin_0"), val = tensor([0, 21, 0, 0])]; + tensor var_6559_end_0 = const()[name = string("op_6559_end_0"), val = tensor([32, 22, 64, 128])]; + tensor var_6559_end_mask_0 = const()[name = string("op_6559_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6559_squeeze_mask_0 = const()[name = string("op_6559_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6559_cast_fp16 = slice_by_index(begin = var_6559_begin_0, end = var_6559_end_0, end_mask = var_6559_end_mask_0, squeeze_mask = var_6559_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6559_cast_fp16")]; + bool var_6561_transpose_x_1 = const()[name = string("op_6561_transpose_x_1"), val = bool(true)]; + bool var_6561_transpose_y_1 = const()[name = string("op_6561_transpose_y_1"), val = bool(false)]; + tensor var_6561_cast_fp16 = matmul(transpose_x = var_6561_transpose_x_1, transpose_y = var_6561_transpose_y_1, x = var_6559_cast_fp16, y = v_new_363_cast_fp16)[name = string("op_6561_cast_fp16")]; + tensor state_365_cast_fp16 = add(x = var_6557_cast_fp16, y = var_6561_cast_fp16)[name = string("state_365_cast_fp16")]; + tensor var_6564_begin_0 = const()[name = string("op_6564_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_6564_end_0 = const()[name = string("op_6564_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_6564_end_mask_0 = const()[name = string("op_6564_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6564_squeeze_mask_0 = const()[name = string("op_6564_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6564_cast_fp16 = slice_by_index(begin = var_6564_begin_0, end = var_6564_end_0, end_mask = var_6564_end_mask_0, squeeze_mask = var_6564_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6564_cast_fp16")]; + tensor var_6566_begin_0 = const()[name = string("op_6566_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_6566_end_0 = const()[name = string("op_6566_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_6566_end_mask_0 = const()[name = string("op_6566_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6566_squeeze_mask_0 = const()[name = string("op_6566_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6566_cast_fp16 = slice_by_index(begin = var_6566_begin_0, end = var_6566_end_0, end_mask = var_6566_end_mask_0, squeeze_mask = var_6566_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6566_cast_fp16")]; + bool var_6567_transpose_x_0 = const()[name = string("op_6567_transpose_x_0"), val = bool(false)]; + bool var_6567_transpose_y_0 = const()[name = string("op_6567_transpose_y_0"), val = bool(false)]; + tensor var_6567_cast_fp16 = matmul(transpose_x = var_6567_transpose_x_0, transpose_y = var_6567_transpose_y_0, x = var_6566_cast_fp16, y = state_365_cast_fp16)[name = string("op_6567_cast_fp16")]; + tensor v_new_365_cast_fp16 = sub(x = var_6564_cast_fp16, y = var_6567_cast_fp16)[name = string("v_new_365_cast_fp16")]; + tensor var_6570_begin_0 = const()[name = string("op_6570_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_6570_end_0 = const()[name = string("op_6570_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_6570_end_mask_0 = const()[name = string("op_6570_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6570_squeeze_mask_0 = const()[name = string("op_6570_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6570_cast_fp16 = slice_by_index(begin = var_6570_begin_0, end = var_6570_end_0, end_mask = var_6570_end_mask_0, squeeze_mask = var_6570_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6570_cast_fp16")]; + bool var_6571_transpose_x_0 = const()[name = string("op_6571_transpose_x_0"), val = bool(false)]; + bool var_6571_transpose_y_0 = const()[name = string("op_6571_transpose_y_0"), val = bool(false)]; + tensor var_6571_cast_fp16 = matmul(transpose_x = var_6571_transpose_x_0, transpose_y = var_6571_transpose_y_0, x = var_6570_cast_fp16, y = state_365_cast_fp16)[name = string("op_6571_cast_fp16")]; + tensor var_6573_begin_0 = const()[name = string("op_6573_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_6573_end_0 = const()[name = string("op_6573_end_0"), val = tensor([32, 23, 64, 64])]; + tensor var_6573_end_mask_0 = const()[name = string("op_6573_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6573_squeeze_mask_0 = const()[name = string("op_6573_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6573_cast_fp16 = slice_by_index(begin = var_6573_begin_0, end = var_6573_end_0, end_mask = var_6573_end_mask_0, squeeze_mask = var_6573_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6573_cast_fp16")]; + bool var_6574_transpose_x_0 = const()[name = string("op_6574_transpose_x_0"), val = bool(false)]; + bool var_6574_transpose_y_0 = const()[name = string("op_6574_transpose_y_0"), val = bool(false)]; + tensor var_6574_cast_fp16 = matmul(transpose_x = var_6574_transpose_x_0, transpose_y = var_6574_transpose_y_0, x = var_6573_cast_fp16, y = v_new_365_cast_fp16)[name = string("op_6574_cast_fp16")]; + tensor var_6575_cast_fp16 = add(x = var_6571_cast_fp16, y = var_6574_cast_fp16)[name = string("op_6575_cast_fp16")]; + tensor var_6577_begin_0 = const()[name = string("op_6577_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_6577_end_0 = const()[name = string("op_6577_end_0"), val = tensor([32, 23, 1, 1])]; + tensor var_6577_end_mask_0 = const()[name = string("op_6577_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6577_squeeze_mask_0 = const()[name = string("op_6577_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6577_cast_fp16 = slice_by_index(begin = var_6577_begin_0, end = var_6577_end_0, end_mask = var_6577_end_mask_0, squeeze_mask = var_6577_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6577_cast_fp16")]; + tensor var_6578_cast_fp16 = mul(x = state_365_cast_fp16, y = var_6577_cast_fp16)[name = string("op_6578_cast_fp16")]; + tensor var_6580_begin_0 = const()[name = string("op_6580_begin_0"), val = tensor([0, 22, 0, 0])]; + tensor var_6580_end_0 = const()[name = string("op_6580_end_0"), val = tensor([32, 23, 64, 128])]; + tensor var_6580_end_mask_0 = const()[name = string("op_6580_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6580_squeeze_mask_0 = const()[name = string("op_6580_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6580_cast_fp16 = slice_by_index(begin = var_6580_begin_0, end = var_6580_end_0, end_mask = var_6580_end_mask_0, squeeze_mask = var_6580_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6580_cast_fp16")]; + bool var_6582_transpose_x_1 = const()[name = string("op_6582_transpose_x_1"), val = bool(true)]; + bool var_6582_transpose_y_1 = const()[name = string("op_6582_transpose_y_1"), val = bool(false)]; + tensor var_6582_cast_fp16 = matmul(transpose_x = var_6582_transpose_x_1, transpose_y = var_6582_transpose_y_1, x = var_6580_cast_fp16, y = v_new_365_cast_fp16)[name = string("op_6582_cast_fp16")]; + tensor state_367_cast_fp16 = add(x = var_6578_cast_fp16, y = var_6582_cast_fp16)[name = string("state_367_cast_fp16")]; + tensor var_6585_begin_0 = const()[name = string("op_6585_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_6585_end_0 = const()[name = string("op_6585_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_6585_end_mask_0 = const()[name = string("op_6585_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6585_squeeze_mask_0 = const()[name = string("op_6585_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6585_cast_fp16 = slice_by_index(begin = var_6585_begin_0, end = var_6585_end_0, end_mask = var_6585_end_mask_0, squeeze_mask = var_6585_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6585_cast_fp16")]; + tensor var_6587_begin_0 = const()[name = string("op_6587_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_6587_end_0 = const()[name = string("op_6587_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_6587_end_mask_0 = const()[name = string("op_6587_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6587_squeeze_mask_0 = const()[name = string("op_6587_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6587_cast_fp16 = slice_by_index(begin = var_6587_begin_0, end = var_6587_end_0, end_mask = var_6587_end_mask_0, squeeze_mask = var_6587_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6587_cast_fp16")]; + bool var_6588_transpose_x_0 = const()[name = string("op_6588_transpose_x_0"), val = bool(false)]; + bool var_6588_transpose_y_0 = const()[name = string("op_6588_transpose_y_0"), val = bool(false)]; + tensor var_6588_cast_fp16 = matmul(transpose_x = var_6588_transpose_x_0, transpose_y = var_6588_transpose_y_0, x = var_6587_cast_fp16, y = state_367_cast_fp16)[name = string("op_6588_cast_fp16")]; + tensor v_new_367_cast_fp16 = sub(x = var_6585_cast_fp16, y = var_6588_cast_fp16)[name = string("v_new_367_cast_fp16")]; + tensor var_6591_begin_0 = const()[name = string("op_6591_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_6591_end_0 = const()[name = string("op_6591_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_6591_end_mask_0 = const()[name = string("op_6591_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6591_squeeze_mask_0 = const()[name = string("op_6591_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6591_cast_fp16 = slice_by_index(begin = var_6591_begin_0, end = var_6591_end_0, end_mask = var_6591_end_mask_0, squeeze_mask = var_6591_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6591_cast_fp16")]; + bool var_6592_transpose_x_0 = const()[name = string("op_6592_transpose_x_0"), val = bool(false)]; + bool var_6592_transpose_y_0 = const()[name = string("op_6592_transpose_y_0"), val = bool(false)]; + tensor var_6592_cast_fp16 = matmul(transpose_x = var_6592_transpose_x_0, transpose_y = var_6592_transpose_y_0, x = var_6591_cast_fp16, y = state_367_cast_fp16)[name = string("op_6592_cast_fp16")]; + tensor var_6594_begin_0 = const()[name = string("op_6594_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_6594_end_0 = const()[name = string("op_6594_end_0"), val = tensor([32, 24, 64, 64])]; + tensor var_6594_end_mask_0 = const()[name = string("op_6594_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6594_squeeze_mask_0 = const()[name = string("op_6594_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6594_cast_fp16 = slice_by_index(begin = var_6594_begin_0, end = var_6594_end_0, end_mask = var_6594_end_mask_0, squeeze_mask = var_6594_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6594_cast_fp16")]; + bool var_6595_transpose_x_0 = const()[name = string("op_6595_transpose_x_0"), val = bool(false)]; + bool var_6595_transpose_y_0 = const()[name = string("op_6595_transpose_y_0"), val = bool(false)]; + tensor var_6595_cast_fp16 = matmul(transpose_x = var_6595_transpose_x_0, transpose_y = var_6595_transpose_y_0, x = var_6594_cast_fp16, y = v_new_367_cast_fp16)[name = string("op_6595_cast_fp16")]; + tensor var_6596_cast_fp16 = add(x = var_6592_cast_fp16, y = var_6595_cast_fp16)[name = string("op_6596_cast_fp16")]; + tensor var_6598_begin_0 = const()[name = string("op_6598_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_6598_end_0 = const()[name = string("op_6598_end_0"), val = tensor([32, 24, 1, 1])]; + tensor var_6598_end_mask_0 = const()[name = string("op_6598_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6598_squeeze_mask_0 = const()[name = string("op_6598_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6598_cast_fp16 = slice_by_index(begin = var_6598_begin_0, end = var_6598_end_0, end_mask = var_6598_end_mask_0, squeeze_mask = var_6598_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6598_cast_fp16")]; + tensor var_6599_cast_fp16 = mul(x = state_367_cast_fp16, y = var_6598_cast_fp16)[name = string("op_6599_cast_fp16")]; + tensor var_6601_begin_0 = const()[name = string("op_6601_begin_0"), val = tensor([0, 23, 0, 0])]; + tensor var_6601_end_0 = const()[name = string("op_6601_end_0"), val = tensor([32, 24, 64, 128])]; + tensor var_6601_end_mask_0 = const()[name = string("op_6601_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6601_squeeze_mask_0 = const()[name = string("op_6601_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6601_cast_fp16 = slice_by_index(begin = var_6601_begin_0, end = var_6601_end_0, end_mask = var_6601_end_mask_0, squeeze_mask = var_6601_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6601_cast_fp16")]; + bool var_6603_transpose_x_1 = const()[name = string("op_6603_transpose_x_1"), val = bool(true)]; + bool var_6603_transpose_y_1 = const()[name = string("op_6603_transpose_y_1"), val = bool(false)]; + tensor var_6603_cast_fp16 = matmul(transpose_x = var_6603_transpose_x_1, transpose_y = var_6603_transpose_y_1, x = var_6601_cast_fp16, y = v_new_367_cast_fp16)[name = string("op_6603_cast_fp16")]; + tensor state_369_cast_fp16 = add(x = var_6599_cast_fp16, y = var_6603_cast_fp16)[name = string("state_369_cast_fp16")]; + tensor var_6606_begin_0 = const()[name = string("op_6606_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_6606_end_0 = const()[name = string("op_6606_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_6606_end_mask_0 = const()[name = string("op_6606_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6606_squeeze_mask_0 = const()[name = string("op_6606_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6606_cast_fp16 = slice_by_index(begin = var_6606_begin_0, end = var_6606_end_0, end_mask = var_6606_end_mask_0, squeeze_mask = var_6606_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6606_cast_fp16")]; + tensor var_6608_begin_0 = const()[name = string("op_6608_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_6608_end_0 = const()[name = string("op_6608_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_6608_end_mask_0 = const()[name = string("op_6608_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6608_squeeze_mask_0 = const()[name = string("op_6608_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6608_cast_fp16 = slice_by_index(begin = var_6608_begin_0, end = var_6608_end_0, end_mask = var_6608_end_mask_0, squeeze_mask = var_6608_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6608_cast_fp16")]; + bool var_6609_transpose_x_0 = const()[name = string("op_6609_transpose_x_0"), val = bool(false)]; + bool var_6609_transpose_y_0 = const()[name = string("op_6609_transpose_y_0"), val = bool(false)]; + tensor var_6609_cast_fp16 = matmul(transpose_x = var_6609_transpose_x_0, transpose_y = var_6609_transpose_y_0, x = var_6608_cast_fp16, y = state_369_cast_fp16)[name = string("op_6609_cast_fp16")]; + tensor v_new_369_cast_fp16 = sub(x = var_6606_cast_fp16, y = var_6609_cast_fp16)[name = string("v_new_369_cast_fp16")]; + tensor var_6612_begin_0 = const()[name = string("op_6612_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_6612_end_0 = const()[name = string("op_6612_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_6612_end_mask_0 = const()[name = string("op_6612_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6612_squeeze_mask_0 = const()[name = string("op_6612_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6612_cast_fp16 = slice_by_index(begin = var_6612_begin_0, end = var_6612_end_0, end_mask = var_6612_end_mask_0, squeeze_mask = var_6612_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6612_cast_fp16")]; + bool var_6613_transpose_x_0 = const()[name = string("op_6613_transpose_x_0"), val = bool(false)]; + bool var_6613_transpose_y_0 = const()[name = string("op_6613_transpose_y_0"), val = bool(false)]; + tensor var_6613_cast_fp16 = matmul(transpose_x = var_6613_transpose_x_0, transpose_y = var_6613_transpose_y_0, x = var_6612_cast_fp16, y = state_369_cast_fp16)[name = string("op_6613_cast_fp16")]; + tensor var_6615_begin_0 = const()[name = string("op_6615_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_6615_end_0 = const()[name = string("op_6615_end_0"), val = tensor([32, 25, 64, 64])]; + tensor var_6615_end_mask_0 = const()[name = string("op_6615_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6615_squeeze_mask_0 = const()[name = string("op_6615_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6615_cast_fp16 = slice_by_index(begin = var_6615_begin_0, end = var_6615_end_0, end_mask = var_6615_end_mask_0, squeeze_mask = var_6615_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6615_cast_fp16")]; + bool var_6616_transpose_x_0 = const()[name = string("op_6616_transpose_x_0"), val = bool(false)]; + bool var_6616_transpose_y_0 = const()[name = string("op_6616_transpose_y_0"), val = bool(false)]; + tensor var_6616_cast_fp16 = matmul(transpose_x = var_6616_transpose_x_0, transpose_y = var_6616_transpose_y_0, x = var_6615_cast_fp16, y = v_new_369_cast_fp16)[name = string("op_6616_cast_fp16")]; + tensor var_6617_cast_fp16 = add(x = var_6613_cast_fp16, y = var_6616_cast_fp16)[name = string("op_6617_cast_fp16")]; + tensor var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor([32, 25, 1, 1])]; + tensor var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6619_squeeze_mask_0 = const()[name = string("op_6619_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, squeeze_mask = var_6619_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6619_cast_fp16")]; + tensor var_6620_cast_fp16 = mul(x = state_369_cast_fp16, y = var_6619_cast_fp16)[name = string("op_6620_cast_fp16")]; + tensor var_6622_begin_0 = const()[name = string("op_6622_begin_0"), val = tensor([0, 24, 0, 0])]; + tensor var_6622_end_0 = const()[name = string("op_6622_end_0"), val = tensor([32, 25, 64, 128])]; + tensor var_6622_end_mask_0 = const()[name = string("op_6622_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6622_squeeze_mask_0 = const()[name = string("op_6622_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6622_cast_fp16 = slice_by_index(begin = var_6622_begin_0, end = var_6622_end_0, end_mask = var_6622_end_mask_0, squeeze_mask = var_6622_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6622_cast_fp16")]; + bool var_6624_transpose_x_1 = const()[name = string("op_6624_transpose_x_1"), val = bool(true)]; + bool var_6624_transpose_y_1 = const()[name = string("op_6624_transpose_y_1"), val = bool(false)]; + tensor var_6624_cast_fp16 = matmul(transpose_x = var_6624_transpose_x_1, transpose_y = var_6624_transpose_y_1, x = var_6622_cast_fp16, y = v_new_369_cast_fp16)[name = string("op_6624_cast_fp16")]; + tensor state_371_cast_fp16 = add(x = var_6620_cast_fp16, y = var_6624_cast_fp16)[name = string("state_371_cast_fp16")]; + tensor var_6627_begin_0 = const()[name = string("op_6627_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_6627_end_0 = const()[name = string("op_6627_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_6627_end_mask_0 = const()[name = string("op_6627_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6627_squeeze_mask_0 = const()[name = string("op_6627_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6627_cast_fp16 = slice_by_index(begin = var_6627_begin_0, end = var_6627_end_0, end_mask = var_6627_end_mask_0, squeeze_mask = var_6627_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6627_cast_fp16")]; + tensor var_6629_begin_0 = const()[name = string("op_6629_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_6629_end_0 = const()[name = string("op_6629_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_6629_end_mask_0 = const()[name = string("op_6629_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6629_squeeze_mask_0 = const()[name = string("op_6629_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6629_cast_fp16 = slice_by_index(begin = var_6629_begin_0, end = var_6629_end_0, end_mask = var_6629_end_mask_0, squeeze_mask = var_6629_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6629_cast_fp16")]; + bool var_6630_transpose_x_0 = const()[name = string("op_6630_transpose_x_0"), val = bool(false)]; + bool var_6630_transpose_y_0 = const()[name = string("op_6630_transpose_y_0"), val = bool(false)]; + tensor var_6630_cast_fp16 = matmul(transpose_x = var_6630_transpose_x_0, transpose_y = var_6630_transpose_y_0, x = var_6629_cast_fp16, y = state_371_cast_fp16)[name = string("op_6630_cast_fp16")]; + tensor v_new_371_cast_fp16 = sub(x = var_6627_cast_fp16, y = var_6630_cast_fp16)[name = string("v_new_371_cast_fp16")]; + tensor var_6633_begin_0 = const()[name = string("op_6633_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_6633_end_0 = const()[name = string("op_6633_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_6633_end_mask_0 = const()[name = string("op_6633_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6633_squeeze_mask_0 = const()[name = string("op_6633_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6633_cast_fp16 = slice_by_index(begin = var_6633_begin_0, end = var_6633_end_0, end_mask = var_6633_end_mask_0, squeeze_mask = var_6633_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6633_cast_fp16")]; + bool var_6634_transpose_x_0 = const()[name = string("op_6634_transpose_x_0"), val = bool(false)]; + bool var_6634_transpose_y_0 = const()[name = string("op_6634_transpose_y_0"), val = bool(false)]; + tensor var_6634_cast_fp16 = matmul(transpose_x = var_6634_transpose_x_0, transpose_y = var_6634_transpose_y_0, x = var_6633_cast_fp16, y = state_371_cast_fp16)[name = string("op_6634_cast_fp16")]; + tensor var_6636_begin_0 = const()[name = string("op_6636_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_6636_end_0 = const()[name = string("op_6636_end_0"), val = tensor([32, 26, 64, 64])]; + tensor var_6636_end_mask_0 = const()[name = string("op_6636_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6636_squeeze_mask_0 = const()[name = string("op_6636_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6636_cast_fp16 = slice_by_index(begin = var_6636_begin_0, end = var_6636_end_0, end_mask = var_6636_end_mask_0, squeeze_mask = var_6636_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6636_cast_fp16")]; + bool var_6637_transpose_x_0 = const()[name = string("op_6637_transpose_x_0"), val = bool(false)]; + bool var_6637_transpose_y_0 = const()[name = string("op_6637_transpose_y_0"), val = bool(false)]; + tensor var_6637_cast_fp16 = matmul(transpose_x = var_6637_transpose_x_0, transpose_y = var_6637_transpose_y_0, x = var_6636_cast_fp16, y = v_new_371_cast_fp16)[name = string("op_6637_cast_fp16")]; + tensor var_6638_cast_fp16 = add(x = var_6634_cast_fp16, y = var_6637_cast_fp16)[name = string("op_6638_cast_fp16")]; + tensor var_6640_begin_0 = const()[name = string("op_6640_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_6640_end_0 = const()[name = string("op_6640_end_0"), val = tensor([32, 26, 1, 1])]; + tensor var_6640_end_mask_0 = const()[name = string("op_6640_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6640_squeeze_mask_0 = const()[name = string("op_6640_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6640_cast_fp16 = slice_by_index(begin = var_6640_begin_0, end = var_6640_end_0, end_mask = var_6640_end_mask_0, squeeze_mask = var_6640_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6640_cast_fp16")]; + tensor var_6641_cast_fp16 = mul(x = state_371_cast_fp16, y = var_6640_cast_fp16)[name = string("op_6641_cast_fp16")]; + tensor var_6643_begin_0 = const()[name = string("op_6643_begin_0"), val = tensor([0, 25, 0, 0])]; + tensor var_6643_end_0 = const()[name = string("op_6643_end_0"), val = tensor([32, 26, 64, 128])]; + tensor var_6643_end_mask_0 = const()[name = string("op_6643_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6643_squeeze_mask_0 = const()[name = string("op_6643_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6643_cast_fp16 = slice_by_index(begin = var_6643_begin_0, end = var_6643_end_0, end_mask = var_6643_end_mask_0, squeeze_mask = var_6643_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6643_cast_fp16")]; + bool var_6645_transpose_x_1 = const()[name = string("op_6645_transpose_x_1"), val = bool(true)]; + bool var_6645_transpose_y_1 = const()[name = string("op_6645_transpose_y_1"), val = bool(false)]; + tensor var_6645_cast_fp16 = matmul(transpose_x = var_6645_transpose_x_1, transpose_y = var_6645_transpose_y_1, x = var_6643_cast_fp16, y = v_new_371_cast_fp16)[name = string("op_6645_cast_fp16")]; + tensor state_373_cast_fp16 = add(x = var_6641_cast_fp16, y = var_6645_cast_fp16)[name = string("state_373_cast_fp16")]; + tensor var_6648_begin_0 = const()[name = string("op_6648_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_6648_end_0 = const()[name = string("op_6648_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_6648_end_mask_0 = const()[name = string("op_6648_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6648_squeeze_mask_0 = const()[name = string("op_6648_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6648_cast_fp16 = slice_by_index(begin = var_6648_begin_0, end = var_6648_end_0, end_mask = var_6648_end_mask_0, squeeze_mask = var_6648_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6648_cast_fp16")]; + tensor var_6650_begin_0 = const()[name = string("op_6650_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_6650_end_0 = const()[name = string("op_6650_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_6650_end_mask_0 = const()[name = string("op_6650_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6650_squeeze_mask_0 = const()[name = string("op_6650_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6650_cast_fp16 = slice_by_index(begin = var_6650_begin_0, end = var_6650_end_0, end_mask = var_6650_end_mask_0, squeeze_mask = var_6650_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6650_cast_fp16")]; + bool var_6651_transpose_x_0 = const()[name = string("op_6651_transpose_x_0"), val = bool(false)]; + bool var_6651_transpose_y_0 = const()[name = string("op_6651_transpose_y_0"), val = bool(false)]; + tensor var_6651_cast_fp16 = matmul(transpose_x = var_6651_transpose_x_0, transpose_y = var_6651_transpose_y_0, x = var_6650_cast_fp16, y = state_373_cast_fp16)[name = string("op_6651_cast_fp16")]; + tensor v_new_373_cast_fp16 = sub(x = var_6648_cast_fp16, y = var_6651_cast_fp16)[name = string("v_new_373_cast_fp16")]; + tensor var_6654_begin_0 = const()[name = string("op_6654_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_6654_end_0 = const()[name = string("op_6654_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_6654_end_mask_0 = const()[name = string("op_6654_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6654_squeeze_mask_0 = const()[name = string("op_6654_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6654_cast_fp16 = slice_by_index(begin = var_6654_begin_0, end = var_6654_end_0, end_mask = var_6654_end_mask_0, squeeze_mask = var_6654_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6654_cast_fp16")]; + bool var_6655_transpose_x_0 = const()[name = string("op_6655_transpose_x_0"), val = bool(false)]; + bool var_6655_transpose_y_0 = const()[name = string("op_6655_transpose_y_0"), val = bool(false)]; + tensor var_6655_cast_fp16 = matmul(transpose_x = var_6655_transpose_x_0, transpose_y = var_6655_transpose_y_0, x = var_6654_cast_fp16, y = state_373_cast_fp16)[name = string("op_6655_cast_fp16")]; + tensor var_6657_begin_0 = const()[name = string("op_6657_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_6657_end_0 = const()[name = string("op_6657_end_0"), val = tensor([32, 27, 64, 64])]; + tensor var_6657_end_mask_0 = const()[name = string("op_6657_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6657_squeeze_mask_0 = const()[name = string("op_6657_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6657_cast_fp16 = slice_by_index(begin = var_6657_begin_0, end = var_6657_end_0, end_mask = var_6657_end_mask_0, squeeze_mask = var_6657_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6657_cast_fp16")]; + bool var_6658_transpose_x_0 = const()[name = string("op_6658_transpose_x_0"), val = bool(false)]; + bool var_6658_transpose_y_0 = const()[name = string("op_6658_transpose_y_0"), val = bool(false)]; + tensor var_6658_cast_fp16 = matmul(transpose_x = var_6658_transpose_x_0, transpose_y = var_6658_transpose_y_0, x = var_6657_cast_fp16, y = v_new_373_cast_fp16)[name = string("op_6658_cast_fp16")]; + tensor var_6659_cast_fp16 = add(x = var_6655_cast_fp16, y = var_6658_cast_fp16)[name = string("op_6659_cast_fp16")]; + tensor var_6661_begin_0 = const()[name = string("op_6661_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_6661_end_0 = const()[name = string("op_6661_end_0"), val = tensor([32, 27, 1, 1])]; + tensor var_6661_end_mask_0 = const()[name = string("op_6661_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6661_squeeze_mask_0 = const()[name = string("op_6661_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6661_cast_fp16 = slice_by_index(begin = var_6661_begin_0, end = var_6661_end_0, end_mask = var_6661_end_mask_0, squeeze_mask = var_6661_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6661_cast_fp16")]; + tensor var_6662_cast_fp16 = mul(x = state_373_cast_fp16, y = var_6661_cast_fp16)[name = string("op_6662_cast_fp16")]; + tensor var_6664_begin_0 = const()[name = string("op_6664_begin_0"), val = tensor([0, 26, 0, 0])]; + tensor var_6664_end_0 = const()[name = string("op_6664_end_0"), val = tensor([32, 27, 64, 128])]; + tensor var_6664_end_mask_0 = const()[name = string("op_6664_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6664_squeeze_mask_0 = const()[name = string("op_6664_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6664_cast_fp16 = slice_by_index(begin = var_6664_begin_0, end = var_6664_end_0, end_mask = var_6664_end_mask_0, squeeze_mask = var_6664_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6664_cast_fp16")]; + bool var_6666_transpose_x_1 = const()[name = string("op_6666_transpose_x_1"), val = bool(true)]; + bool var_6666_transpose_y_1 = const()[name = string("op_6666_transpose_y_1"), val = bool(false)]; + tensor var_6666_cast_fp16 = matmul(transpose_x = var_6666_transpose_x_1, transpose_y = var_6666_transpose_y_1, x = var_6664_cast_fp16, y = v_new_373_cast_fp16)[name = string("op_6666_cast_fp16")]; + tensor state_375_cast_fp16 = add(x = var_6662_cast_fp16, y = var_6666_cast_fp16)[name = string("state_375_cast_fp16")]; + tensor var_6669_begin_0 = const()[name = string("op_6669_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_6669_end_0 = const()[name = string("op_6669_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_6669_end_mask_0 = const()[name = string("op_6669_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6669_squeeze_mask_0 = const()[name = string("op_6669_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6669_cast_fp16 = slice_by_index(begin = var_6669_begin_0, end = var_6669_end_0, end_mask = var_6669_end_mask_0, squeeze_mask = var_6669_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6669_cast_fp16")]; + tensor var_6671_begin_0 = const()[name = string("op_6671_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_6671_end_0 = const()[name = string("op_6671_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_6671_end_mask_0 = const()[name = string("op_6671_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6671_squeeze_mask_0 = const()[name = string("op_6671_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6671_cast_fp16 = slice_by_index(begin = var_6671_begin_0, end = var_6671_end_0, end_mask = var_6671_end_mask_0, squeeze_mask = var_6671_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6671_cast_fp16")]; + bool var_6672_transpose_x_0 = const()[name = string("op_6672_transpose_x_0"), val = bool(false)]; + bool var_6672_transpose_y_0 = const()[name = string("op_6672_transpose_y_0"), val = bool(false)]; + tensor var_6672_cast_fp16 = matmul(transpose_x = var_6672_transpose_x_0, transpose_y = var_6672_transpose_y_0, x = var_6671_cast_fp16, y = state_375_cast_fp16)[name = string("op_6672_cast_fp16")]; + tensor v_new_375_cast_fp16 = sub(x = var_6669_cast_fp16, y = var_6672_cast_fp16)[name = string("v_new_375_cast_fp16")]; + tensor var_6675_begin_0 = const()[name = string("op_6675_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_6675_end_0 = const()[name = string("op_6675_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_6675_end_mask_0 = const()[name = string("op_6675_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6675_squeeze_mask_0 = const()[name = string("op_6675_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6675_cast_fp16 = slice_by_index(begin = var_6675_begin_0, end = var_6675_end_0, end_mask = var_6675_end_mask_0, squeeze_mask = var_6675_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6675_cast_fp16")]; + bool var_6676_transpose_x_0 = const()[name = string("op_6676_transpose_x_0"), val = bool(false)]; + bool var_6676_transpose_y_0 = const()[name = string("op_6676_transpose_y_0"), val = bool(false)]; + tensor var_6676_cast_fp16 = matmul(transpose_x = var_6676_transpose_x_0, transpose_y = var_6676_transpose_y_0, x = var_6675_cast_fp16, y = state_375_cast_fp16)[name = string("op_6676_cast_fp16")]; + tensor var_6678_begin_0 = const()[name = string("op_6678_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_6678_end_0 = const()[name = string("op_6678_end_0"), val = tensor([32, 28, 64, 64])]; + tensor var_6678_end_mask_0 = const()[name = string("op_6678_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6678_squeeze_mask_0 = const()[name = string("op_6678_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6678_cast_fp16 = slice_by_index(begin = var_6678_begin_0, end = var_6678_end_0, end_mask = var_6678_end_mask_0, squeeze_mask = var_6678_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6678_cast_fp16")]; + bool var_6679_transpose_x_0 = const()[name = string("op_6679_transpose_x_0"), val = bool(false)]; + bool var_6679_transpose_y_0 = const()[name = string("op_6679_transpose_y_0"), val = bool(false)]; + tensor var_6679_cast_fp16 = matmul(transpose_x = var_6679_transpose_x_0, transpose_y = var_6679_transpose_y_0, x = var_6678_cast_fp16, y = v_new_375_cast_fp16)[name = string("op_6679_cast_fp16")]; + tensor var_6680_cast_fp16 = add(x = var_6676_cast_fp16, y = var_6679_cast_fp16)[name = string("op_6680_cast_fp16")]; + tensor var_6682_begin_0 = const()[name = string("op_6682_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_6682_end_0 = const()[name = string("op_6682_end_0"), val = tensor([32, 28, 1, 1])]; + tensor var_6682_end_mask_0 = const()[name = string("op_6682_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6682_squeeze_mask_0 = const()[name = string("op_6682_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6682_cast_fp16 = slice_by_index(begin = var_6682_begin_0, end = var_6682_end_0, end_mask = var_6682_end_mask_0, squeeze_mask = var_6682_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6682_cast_fp16")]; + tensor var_6683_cast_fp16 = mul(x = state_375_cast_fp16, y = var_6682_cast_fp16)[name = string("op_6683_cast_fp16")]; + tensor var_6685_begin_0 = const()[name = string("op_6685_begin_0"), val = tensor([0, 27, 0, 0])]; + tensor var_6685_end_0 = const()[name = string("op_6685_end_0"), val = tensor([32, 28, 64, 128])]; + tensor var_6685_end_mask_0 = const()[name = string("op_6685_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6685_squeeze_mask_0 = const()[name = string("op_6685_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6685_cast_fp16 = slice_by_index(begin = var_6685_begin_0, end = var_6685_end_0, end_mask = var_6685_end_mask_0, squeeze_mask = var_6685_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6685_cast_fp16")]; + bool var_6687_transpose_x_1 = const()[name = string("op_6687_transpose_x_1"), val = bool(true)]; + bool var_6687_transpose_y_1 = const()[name = string("op_6687_transpose_y_1"), val = bool(false)]; + tensor var_6687_cast_fp16 = matmul(transpose_x = var_6687_transpose_x_1, transpose_y = var_6687_transpose_y_1, x = var_6685_cast_fp16, y = v_new_375_cast_fp16)[name = string("op_6687_cast_fp16")]; + tensor state_377_cast_fp16 = add(x = var_6683_cast_fp16, y = var_6687_cast_fp16)[name = string("state_377_cast_fp16")]; + tensor var_6690_begin_0 = const()[name = string("op_6690_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_6690_end_0 = const()[name = string("op_6690_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_6690_end_mask_0 = const()[name = string("op_6690_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6690_squeeze_mask_0 = const()[name = string("op_6690_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6690_cast_fp16 = slice_by_index(begin = var_6690_begin_0, end = var_6690_end_0, end_mask = var_6690_end_mask_0, squeeze_mask = var_6690_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6690_cast_fp16")]; + tensor var_6692_begin_0 = const()[name = string("op_6692_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_6692_end_0 = const()[name = string("op_6692_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_6692_end_mask_0 = const()[name = string("op_6692_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6692_squeeze_mask_0 = const()[name = string("op_6692_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6692_cast_fp16 = slice_by_index(begin = var_6692_begin_0, end = var_6692_end_0, end_mask = var_6692_end_mask_0, squeeze_mask = var_6692_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6692_cast_fp16")]; + bool var_6693_transpose_x_0 = const()[name = string("op_6693_transpose_x_0"), val = bool(false)]; + bool var_6693_transpose_y_0 = const()[name = string("op_6693_transpose_y_0"), val = bool(false)]; + tensor var_6693_cast_fp16 = matmul(transpose_x = var_6693_transpose_x_0, transpose_y = var_6693_transpose_y_0, x = var_6692_cast_fp16, y = state_377_cast_fp16)[name = string("op_6693_cast_fp16")]; + tensor v_new_377_cast_fp16 = sub(x = var_6690_cast_fp16, y = var_6693_cast_fp16)[name = string("v_new_377_cast_fp16")]; + tensor var_6696_begin_0 = const()[name = string("op_6696_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_6696_end_0 = const()[name = string("op_6696_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_6696_end_mask_0 = const()[name = string("op_6696_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6696_squeeze_mask_0 = const()[name = string("op_6696_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6696_cast_fp16 = slice_by_index(begin = var_6696_begin_0, end = var_6696_end_0, end_mask = var_6696_end_mask_0, squeeze_mask = var_6696_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6696_cast_fp16")]; + bool var_6697_transpose_x_0 = const()[name = string("op_6697_transpose_x_0"), val = bool(false)]; + bool var_6697_transpose_y_0 = const()[name = string("op_6697_transpose_y_0"), val = bool(false)]; + tensor var_6697_cast_fp16 = matmul(transpose_x = var_6697_transpose_x_0, transpose_y = var_6697_transpose_y_0, x = var_6696_cast_fp16, y = state_377_cast_fp16)[name = string("op_6697_cast_fp16")]; + tensor var_6699_begin_0 = const()[name = string("op_6699_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_6699_end_0 = const()[name = string("op_6699_end_0"), val = tensor([32, 29, 64, 64])]; + tensor var_6699_end_mask_0 = const()[name = string("op_6699_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6699_squeeze_mask_0 = const()[name = string("op_6699_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6699_cast_fp16 = slice_by_index(begin = var_6699_begin_0, end = var_6699_end_0, end_mask = var_6699_end_mask_0, squeeze_mask = var_6699_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6699_cast_fp16")]; + bool var_6700_transpose_x_0 = const()[name = string("op_6700_transpose_x_0"), val = bool(false)]; + bool var_6700_transpose_y_0 = const()[name = string("op_6700_transpose_y_0"), val = bool(false)]; + tensor var_6700_cast_fp16 = matmul(transpose_x = var_6700_transpose_x_0, transpose_y = var_6700_transpose_y_0, x = var_6699_cast_fp16, y = v_new_377_cast_fp16)[name = string("op_6700_cast_fp16")]; + tensor var_6701_cast_fp16 = add(x = var_6697_cast_fp16, y = var_6700_cast_fp16)[name = string("op_6701_cast_fp16")]; + tensor var_6703_begin_0 = const()[name = string("op_6703_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_6703_end_0 = const()[name = string("op_6703_end_0"), val = tensor([32, 29, 1, 1])]; + tensor var_6703_end_mask_0 = const()[name = string("op_6703_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6703_squeeze_mask_0 = const()[name = string("op_6703_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6703_cast_fp16 = slice_by_index(begin = var_6703_begin_0, end = var_6703_end_0, end_mask = var_6703_end_mask_0, squeeze_mask = var_6703_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6703_cast_fp16")]; + tensor var_6704_cast_fp16 = mul(x = state_377_cast_fp16, y = var_6703_cast_fp16)[name = string("op_6704_cast_fp16")]; + tensor var_6706_begin_0 = const()[name = string("op_6706_begin_0"), val = tensor([0, 28, 0, 0])]; + tensor var_6706_end_0 = const()[name = string("op_6706_end_0"), val = tensor([32, 29, 64, 128])]; + tensor var_6706_end_mask_0 = const()[name = string("op_6706_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6706_squeeze_mask_0 = const()[name = string("op_6706_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6706_cast_fp16 = slice_by_index(begin = var_6706_begin_0, end = var_6706_end_0, end_mask = var_6706_end_mask_0, squeeze_mask = var_6706_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6706_cast_fp16")]; + bool var_6708_transpose_x_1 = const()[name = string("op_6708_transpose_x_1"), val = bool(true)]; + bool var_6708_transpose_y_1 = const()[name = string("op_6708_transpose_y_1"), val = bool(false)]; + tensor var_6708_cast_fp16 = matmul(transpose_x = var_6708_transpose_x_1, transpose_y = var_6708_transpose_y_1, x = var_6706_cast_fp16, y = v_new_377_cast_fp16)[name = string("op_6708_cast_fp16")]; + tensor state_379_cast_fp16 = add(x = var_6704_cast_fp16, y = var_6708_cast_fp16)[name = string("state_379_cast_fp16")]; + tensor var_6711_begin_0 = const()[name = string("op_6711_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_6711_end_0 = const()[name = string("op_6711_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_6711_end_mask_0 = const()[name = string("op_6711_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6711_squeeze_mask_0 = const()[name = string("op_6711_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6711_cast_fp16 = slice_by_index(begin = var_6711_begin_0, end = var_6711_end_0, end_mask = var_6711_end_mask_0, squeeze_mask = var_6711_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6711_cast_fp16")]; + tensor var_6713_begin_0 = const()[name = string("op_6713_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_6713_end_0 = const()[name = string("op_6713_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_6713_end_mask_0 = const()[name = string("op_6713_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6713_squeeze_mask_0 = const()[name = string("op_6713_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6713_cast_fp16 = slice_by_index(begin = var_6713_begin_0, end = var_6713_end_0, end_mask = var_6713_end_mask_0, squeeze_mask = var_6713_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6713_cast_fp16")]; + bool var_6714_transpose_x_0 = const()[name = string("op_6714_transpose_x_0"), val = bool(false)]; + bool var_6714_transpose_y_0 = const()[name = string("op_6714_transpose_y_0"), val = bool(false)]; + tensor var_6714_cast_fp16 = matmul(transpose_x = var_6714_transpose_x_0, transpose_y = var_6714_transpose_y_0, x = var_6713_cast_fp16, y = state_379_cast_fp16)[name = string("op_6714_cast_fp16")]; + tensor v_new_379_cast_fp16 = sub(x = var_6711_cast_fp16, y = var_6714_cast_fp16)[name = string("v_new_379_cast_fp16")]; + tensor var_6717_begin_0 = const()[name = string("op_6717_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_6717_end_0 = const()[name = string("op_6717_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_6717_end_mask_0 = const()[name = string("op_6717_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6717_squeeze_mask_0 = const()[name = string("op_6717_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6717_cast_fp16 = slice_by_index(begin = var_6717_begin_0, end = var_6717_end_0, end_mask = var_6717_end_mask_0, squeeze_mask = var_6717_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6717_cast_fp16")]; + bool var_6718_transpose_x_0 = const()[name = string("op_6718_transpose_x_0"), val = bool(false)]; + bool var_6718_transpose_y_0 = const()[name = string("op_6718_transpose_y_0"), val = bool(false)]; + tensor var_6718_cast_fp16 = matmul(transpose_x = var_6718_transpose_x_0, transpose_y = var_6718_transpose_y_0, x = var_6717_cast_fp16, y = state_379_cast_fp16)[name = string("op_6718_cast_fp16")]; + tensor var_6720_begin_0 = const()[name = string("op_6720_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_6720_end_0 = const()[name = string("op_6720_end_0"), val = tensor([32, 30, 64, 64])]; + tensor var_6720_end_mask_0 = const()[name = string("op_6720_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6720_squeeze_mask_0 = const()[name = string("op_6720_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6720_cast_fp16 = slice_by_index(begin = var_6720_begin_0, end = var_6720_end_0, end_mask = var_6720_end_mask_0, squeeze_mask = var_6720_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6720_cast_fp16")]; + bool var_6721_transpose_x_0 = const()[name = string("op_6721_transpose_x_0"), val = bool(false)]; + bool var_6721_transpose_y_0 = const()[name = string("op_6721_transpose_y_0"), val = bool(false)]; + tensor var_6721_cast_fp16 = matmul(transpose_x = var_6721_transpose_x_0, transpose_y = var_6721_transpose_y_0, x = var_6720_cast_fp16, y = v_new_379_cast_fp16)[name = string("op_6721_cast_fp16")]; + tensor var_6722_cast_fp16 = add(x = var_6718_cast_fp16, y = var_6721_cast_fp16)[name = string("op_6722_cast_fp16")]; + tensor var_6724_begin_0 = const()[name = string("op_6724_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_6724_end_0 = const()[name = string("op_6724_end_0"), val = tensor([32, 30, 1, 1])]; + tensor var_6724_end_mask_0 = const()[name = string("op_6724_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6724_squeeze_mask_0 = const()[name = string("op_6724_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6724_cast_fp16 = slice_by_index(begin = var_6724_begin_0, end = var_6724_end_0, end_mask = var_6724_end_mask_0, squeeze_mask = var_6724_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6724_cast_fp16")]; + tensor var_6725_cast_fp16 = mul(x = state_379_cast_fp16, y = var_6724_cast_fp16)[name = string("op_6725_cast_fp16")]; + tensor var_6727_begin_0 = const()[name = string("op_6727_begin_0"), val = tensor([0, 29, 0, 0])]; + tensor var_6727_end_0 = const()[name = string("op_6727_end_0"), val = tensor([32, 30, 64, 128])]; + tensor var_6727_end_mask_0 = const()[name = string("op_6727_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6727_squeeze_mask_0 = const()[name = string("op_6727_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6727_cast_fp16 = slice_by_index(begin = var_6727_begin_0, end = var_6727_end_0, end_mask = var_6727_end_mask_0, squeeze_mask = var_6727_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6727_cast_fp16")]; + bool var_6729_transpose_x_1 = const()[name = string("op_6729_transpose_x_1"), val = bool(true)]; + bool var_6729_transpose_y_1 = const()[name = string("op_6729_transpose_y_1"), val = bool(false)]; + tensor var_6729_cast_fp16 = matmul(transpose_x = var_6729_transpose_x_1, transpose_y = var_6729_transpose_y_1, x = var_6727_cast_fp16, y = v_new_379_cast_fp16)[name = string("op_6729_cast_fp16")]; + tensor state_381_cast_fp16 = add(x = var_6725_cast_fp16, y = var_6729_cast_fp16)[name = string("state_381_cast_fp16")]; + tensor var_6732_begin_0 = const()[name = string("op_6732_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_6732_end_0 = const()[name = string("op_6732_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_6732_end_mask_0 = const()[name = string("op_6732_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6732_squeeze_mask_0 = const()[name = string("op_6732_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6732_cast_fp16 = slice_by_index(begin = var_6732_begin_0, end = var_6732_end_0, end_mask = var_6732_end_mask_0, squeeze_mask = var_6732_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6732_cast_fp16")]; + tensor var_6734_begin_0 = const()[name = string("op_6734_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_6734_end_0 = const()[name = string("op_6734_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_6734_end_mask_0 = const()[name = string("op_6734_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6734_squeeze_mask_0 = const()[name = string("op_6734_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6734_cast_fp16 = slice_by_index(begin = var_6734_begin_0, end = var_6734_end_0, end_mask = var_6734_end_mask_0, squeeze_mask = var_6734_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6734_cast_fp16")]; + bool var_6735_transpose_x_0 = const()[name = string("op_6735_transpose_x_0"), val = bool(false)]; + bool var_6735_transpose_y_0 = const()[name = string("op_6735_transpose_y_0"), val = bool(false)]; + tensor var_6735_cast_fp16 = matmul(transpose_x = var_6735_transpose_x_0, transpose_y = var_6735_transpose_y_0, x = var_6734_cast_fp16, y = state_381_cast_fp16)[name = string("op_6735_cast_fp16")]; + tensor v_new_381_cast_fp16 = sub(x = var_6732_cast_fp16, y = var_6735_cast_fp16)[name = string("v_new_381_cast_fp16")]; + tensor var_6738_begin_0 = const()[name = string("op_6738_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_6738_end_0 = const()[name = string("op_6738_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_6738_end_mask_0 = const()[name = string("op_6738_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6738_squeeze_mask_0 = const()[name = string("op_6738_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6738_cast_fp16 = slice_by_index(begin = var_6738_begin_0, end = var_6738_end_0, end_mask = var_6738_end_mask_0, squeeze_mask = var_6738_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6738_cast_fp16")]; + bool var_6739_transpose_x_0 = const()[name = string("op_6739_transpose_x_0"), val = bool(false)]; + bool var_6739_transpose_y_0 = const()[name = string("op_6739_transpose_y_0"), val = bool(false)]; + tensor var_6739_cast_fp16 = matmul(transpose_x = var_6739_transpose_x_0, transpose_y = var_6739_transpose_y_0, x = var_6738_cast_fp16, y = state_381_cast_fp16)[name = string("op_6739_cast_fp16")]; + tensor var_6741_begin_0 = const()[name = string("op_6741_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_6741_end_0 = const()[name = string("op_6741_end_0"), val = tensor([32, 31, 64, 64])]; + tensor var_6741_end_mask_0 = const()[name = string("op_6741_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6741_squeeze_mask_0 = const()[name = string("op_6741_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6741_cast_fp16 = slice_by_index(begin = var_6741_begin_0, end = var_6741_end_0, end_mask = var_6741_end_mask_0, squeeze_mask = var_6741_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6741_cast_fp16")]; + bool var_6742_transpose_x_0 = const()[name = string("op_6742_transpose_x_0"), val = bool(false)]; + bool var_6742_transpose_y_0 = const()[name = string("op_6742_transpose_y_0"), val = bool(false)]; + tensor var_6742_cast_fp16 = matmul(transpose_x = var_6742_transpose_x_0, transpose_y = var_6742_transpose_y_0, x = var_6741_cast_fp16, y = v_new_381_cast_fp16)[name = string("op_6742_cast_fp16")]; + tensor var_6743_cast_fp16 = add(x = var_6739_cast_fp16, y = var_6742_cast_fp16)[name = string("op_6743_cast_fp16")]; + tensor var_6745_begin_0 = const()[name = string("op_6745_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_6745_end_0 = const()[name = string("op_6745_end_0"), val = tensor([32, 31, 1, 1])]; + tensor var_6745_end_mask_0 = const()[name = string("op_6745_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6745_squeeze_mask_0 = const()[name = string("op_6745_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6745_cast_fp16 = slice_by_index(begin = var_6745_begin_0, end = var_6745_end_0, end_mask = var_6745_end_mask_0, squeeze_mask = var_6745_squeeze_mask_0, x = chunk_decay_cast_fp16)[name = string("op_6745_cast_fp16")]; + tensor var_6746_cast_fp16 = mul(x = state_381_cast_fp16, y = var_6745_cast_fp16)[name = string("op_6746_cast_fp16")]; + tensor var_6748_begin_0 = const()[name = string("op_6748_begin_0"), val = tensor([0, 30, 0, 0])]; + tensor var_6748_end_0 = const()[name = string("op_6748_end_0"), val = tensor([32, 31, 64, 128])]; + tensor var_6748_end_mask_0 = const()[name = string("op_6748_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6748_squeeze_mask_0 = const()[name = string("op_6748_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6748_cast_fp16 = slice_by_index(begin = var_6748_begin_0, end = var_6748_end_0, end_mask = var_6748_end_mask_0, squeeze_mask = var_6748_squeeze_mask_0, x = k_dec_cast_fp16)[name = string("op_6748_cast_fp16")]; + bool var_6750_transpose_x_1 = const()[name = string("op_6750_transpose_x_1"), val = bool(true)]; + bool var_6750_transpose_y_1 = const()[name = string("op_6750_transpose_y_1"), val = bool(false)]; + tensor var_6750_cast_fp16 = matmul(transpose_x = var_6750_transpose_x_1, transpose_y = var_6750_transpose_y_1, x = var_6748_cast_fp16, y = v_new_381_cast_fp16)[name = string("op_6750_cast_fp16")]; + tensor state_cast_fp16 = add(x = var_6746_cast_fp16, y = var_6750_cast_fp16)[name = string("state_cast_fp16")]; + tensor var_6753_begin_0 = const()[name = string("op_6753_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_6753_end_0 = const()[name = string("op_6753_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_6753_end_mask_0 = const()[name = string("op_6753_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6753_squeeze_mask_0 = const()[name = string("op_6753_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6753_cast_fp16 = slice_by_index(begin = var_6753_begin_0, end = var_6753_end_0, end_mask = var_6753_end_mask_0, squeeze_mask = var_6753_squeeze_mask_0, x = new_v_cast_fp16)[name = string("op_6753_cast_fp16")]; + tensor var_6755_begin_0 = const()[name = string("op_6755_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_6755_end_0 = const()[name = string("op_6755_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_6755_end_mask_0 = const()[name = string("op_6755_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6755_squeeze_mask_0 = const()[name = string("op_6755_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6755_cast_fp16 = slice_by_index(begin = var_6755_begin_0, end = var_6755_end_0, end_mask = var_6755_end_mask_0, squeeze_mask = var_6755_squeeze_mask_0, x = k_cumdecay_cast_fp16)[name = string("op_6755_cast_fp16")]; + bool var_6756_transpose_x_0 = const()[name = string("op_6756_transpose_x_0"), val = bool(false)]; + bool var_6756_transpose_y_0 = const()[name = string("op_6756_transpose_y_0"), val = bool(false)]; + tensor var_6756_cast_fp16 = matmul(transpose_x = var_6756_transpose_x_0, transpose_y = var_6756_transpose_y_0, x = var_6755_cast_fp16, y = state_cast_fp16)[name = string("op_6756_cast_fp16")]; + tensor v_new_cast_fp16 = sub(x = var_6753_cast_fp16, y = var_6756_cast_fp16)[name = string("v_new_cast_fp16")]; + tensor var_6759_begin_0 = const()[name = string("op_6759_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_6759_end_0 = const()[name = string("op_6759_end_0"), val = tensor([32, 32, 64, 128])]; + tensor var_6759_end_mask_0 = const()[name = string("op_6759_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6759_squeeze_mask_0 = const()[name = string("op_6759_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6759_cast_fp16 = slice_by_index(begin = var_6759_begin_0, end = var_6759_end_0, end_mask = var_6759_end_mask_0, squeeze_mask = var_6759_squeeze_mask_0, x = q_dec_cast_fp16)[name = string("op_6759_cast_fp16")]; + bool var_6760_transpose_x_0 = const()[name = string("op_6760_transpose_x_0"), val = bool(false)]; + bool var_6760_transpose_y_0 = const()[name = string("op_6760_transpose_y_0"), val = bool(false)]; + tensor var_6760_cast_fp16 = matmul(transpose_x = var_6760_transpose_x_0, transpose_y = var_6760_transpose_y_0, x = var_6759_cast_fp16, y = state_cast_fp16)[name = string("op_6760_cast_fp16")]; + tensor var_6762_begin_0 = const()[name = string("op_6762_begin_0"), val = tensor([0, 31, 0, 0])]; + tensor var_6762_end_0 = const()[name = string("op_6762_end_0"), val = tensor([32, 32, 64, 64])]; + tensor var_6762_end_mask_0 = const()[name = string("op_6762_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6762_squeeze_mask_0 = const()[name = string("op_6762_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor var_6762_cast_fp16 = slice_by_index(begin = var_6762_begin_0, end = var_6762_end_0, end_mask = var_6762_end_mask_0, squeeze_mask = var_6762_squeeze_mask_0, x = attn_intra_cast_fp16)[name = string("op_6762_cast_fp16")]; + bool var_6763_transpose_x_0 = const()[name = string("op_6763_transpose_x_0"), val = bool(false)]; + bool var_6763_transpose_y_0 = const()[name = string("op_6763_transpose_y_0"), val = bool(false)]; + tensor var_6763_cast_fp16 = matmul(transpose_x = var_6763_transpose_x_0, transpose_y = var_6763_transpose_y_0, x = var_6762_cast_fp16, y = v_new_cast_fp16)[name = string("op_6763_cast_fp16")]; + tensor var_6764_cast_fp16 = add(x = var_6760_cast_fp16, y = var_6763_cast_fp16)[name = string("op_6764_cast_fp16")]; + int32 var_6766_axis_0 = const()[name = string("op_6766_axis_0"), val = int32(1)]; + tensor var_6766_cast_fp16 = stack(axis = var_6766_axis_0, values = (var_6113_cast_fp16, var_6134_cast_fp16, var_6155_cast_fp16, var_6176_cast_fp16, var_6197_cast_fp16, var_6218_cast_fp16, var_6239_cast_fp16, var_6260_cast_fp16, var_6281_cast_fp16, var_6302_cast_fp16, var_6323_cast_fp16, var_6344_cast_fp16, var_6365_cast_fp16, var_6386_cast_fp16, var_6407_cast_fp16, var_6428_cast_fp16, var_6449_cast_fp16, var_6470_cast_fp16, var_6491_cast_fp16, var_6512_cast_fp16, var_6533_cast_fp16, var_6554_cast_fp16, var_6575_cast_fp16, var_6596_cast_fp16, var_6617_cast_fp16, var_6638_cast_fp16, var_6659_cast_fp16, var_6680_cast_fp16, var_6701_cast_fp16, var_6722_cast_fp16, var_6743_cast_fp16, var_6764_cast_fp16))[name = string("op_6766_cast_fp16")]; + tensor var_6767 = const()[name = string("op_6767"), val = tensor([32, 2048, 128])]; + tensor var_6768_cast_fp16 = reshape(shape = var_6767, x = var_6766_cast_fp16)[name = string("op_6768_cast_fp16")]; + tensor var_6769 = const()[name = string("op_6769"), val = tensor([1, 0, 2])]; + tensor input_95_begin_0 = const()[name = string("input_95_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor input_95_end_0 = const()[name = string("input_95_end_0"), val = tensor([1, 2048, 32, 128])]; + tensor input_95_end_mask_0 = const()[name = string("input_95_end_mask_0"), val = tensor([false, true, true, true])]; + tensor input_95_squeeze_mask_0 = const()[name = string("input_95_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor input_95_cast_fp16 = slice_by_index(begin = input_95_begin_0, end = input_95_end_0, end_mask = input_95_end_mask_0, squeeze_mask = input_95_squeeze_mask_0, x = z_cast_fp16)[name = string("input_95_cast_fp16")]; + fp16 var_5749_promoted_1_to_fp16 = const()[name = string("op_5749_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor x_169_cast_fp16 = transpose(perm = var_6769, x = var_6768_cast_fp16)[name = string("transpose_4")]; + tensor var_6773_cast_fp16 = pow(x = x_169_cast_fp16, y = var_5749_promoted_1_to_fp16)[name = string("op_6773_cast_fp16")]; + tensor var_6775_axes_0 = const()[name = string("op_6775_axes_0"), val = tensor([-1])]; + bool var_6775_keep_dims_0 = const()[name = string("op_6775_keep_dims_0"), val = bool(true)]; + tensor var_6775_cast_fp16 = reduce_mean(axes = var_6775_axes_0, keep_dims = var_6775_keep_dims_0, x = var_6773_cast_fp16)[name = string("op_6775_cast_fp16")]; + fp16 var_6776_to_fp16 = const()[name = string("op_6776_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_6777_cast_fp16 = add(x = var_6775_cast_fp16, y = var_6776_to_fp16)[name = string("op_6777_cast_fp16")]; + fp32 var_6778_epsilon_0 = const()[name = string("op_6778_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6778_cast_fp16 = rsqrt(epsilon = var_6778_epsilon_0, x = var_6777_cast_fp16)[name = string("op_6778_cast_fp16")]; + tensor x_171_cast_fp16 = mul(x = x_169_cast_fp16, y = var_6778_cast_fp16)[name = string("x_171_cast_fp16")]; + tensor layers_6_linear_attn_norm_weight_to_fp16 = const()[name = string("layers_6_linear_attn_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719253760)))]; + tensor var_6780_cast_fp16 = mul(x = layers_6_linear_attn_norm_weight_to_fp16, y = x_171_cast_fp16)[name = string("op_6780_cast_fp16")]; + tensor var_6781_cast_fp16 = silu(x = input_95_cast_fp16)[name = string("op_6781_cast_fp16")]; + tensor core_cast_fp16 = mul(x = var_6780_cast_fp16, y = var_6781_cast_fp16)[name = string("core_cast_fp16")]; + tensor var_6783 = const()[name = string("op_6783"), val = tensor([1, 2048, 4096])]; + tensor input_97_cast_fp16 = reshape(shape = var_6783, x = core_cast_fp16)[name = string("input_97_cast_fp16")]; + tensor layers_6_linear_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719254080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729739904))))[name = string("layers_6_linear_attn_out_proj_weight_to_fp16_quantized")]; + tensor linear_51_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_linear_attn_out_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_51_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_151_cast_fp16, y = linear_51_cast_fp16)[name = string("x_173_cast_fp16")]; + fp16 var_5749_promoted_2_to_fp16 = const()[name = string("op_5749_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor var_6789_cast_fp16 = pow(x = x_173_cast_fp16, y = var_5749_promoted_2_to_fp16)[name = string("op_6789_cast_fp16")]; + tensor var_6791_axes_0 = const()[name = string("op_6791_axes_0"), val = tensor([-1])]; + bool var_6791_keep_dims_0 = const()[name = string("op_6791_keep_dims_0"), val = bool(true)]; + tensor var_6791_cast_fp16 = reduce_mean(axes = var_6791_axes_0, keep_dims = var_6791_keep_dims_0, x = var_6789_cast_fp16)[name = string("op_6791_cast_fp16")]; + fp16 var_6792_to_fp16 = const()[name = string("op_6792_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_6793_cast_fp16 = add(x = var_6791_cast_fp16, y = var_6792_to_fp16)[name = string("op_6793_cast_fp16")]; + fp32 var_6794_epsilon_0 = const()[name = string("op_6794_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6794_cast_fp16 = rsqrt(epsilon = var_6794_epsilon_0, x = var_6793_cast_fp16)[name = string("op_6794_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_173_cast_fp16, y = var_6794_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor var_6797_to_fp16 = const()[name = string("op_6797_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729745088)))]; + tensor input_99_cast_fp16 = mul(x = x_175_cast_fp16, y = var_6797_to_fp16)[name = string("input_99_cast_fp16")]; + tensor layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729750272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753343296))))[name = string("layers_6_mlp_gate_proj_weight_to_fp16_quantized")]; + tensor linear_52_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor var_6804_cast_fp16 = silu(x = linear_52_cast_fp16)[name = string("op_6804_cast_fp16")]; + tensor layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753361792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(776954816))))[name = string("layers_6_mlp_up_proj_weight_to_fp16_quantized")]; + tensor linear_53_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_103_cast_fp16 = mul(x = var_6804_cast_fp16, y = linear_53_cast_fp16)[name = string("input_103_cast_fp16")]; + tensor layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(776973312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800566336))))[name = string("layers_6_mlp_down_proj_weight_to_fp16_quantized")]; + tensor linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_54_cast_fp16")]; + tensor x_177_cast_fp16 = add(x = x_173_cast_fp16, y = linear_54_cast_fp16)[name = string("x_177_cast_fp16")]; + int32 var_6824 = const()[name = string("op_6824"), val = int32(-1)]; + fp16 var_6823_promoted_to_fp16 = const()[name = string("op_6823_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_6833_cast_fp16 = pow(x = x_177_cast_fp16, y = var_6823_promoted_to_fp16)[name = string("op_6833_cast_fp16")]; + tensor var_6835_axes_0 = const()[name = string("op_6835_axes_0"), val = tensor([-1])]; + bool var_6835_keep_dims_0 = const()[name = string("op_6835_keep_dims_0"), val = bool(true)]; + tensor var_6835_cast_fp16 = reduce_mean(axes = var_6835_axes_0, keep_dims = var_6835_keep_dims_0, x = var_6833_cast_fp16)[name = string("op_6835_cast_fp16")]; + fp16 var_6836_to_fp16 = const()[name = string("op_6836_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_6837_cast_fp16 = add(x = var_6835_cast_fp16, y = var_6836_to_fp16)[name = string("op_6837_cast_fp16")]; + fp32 var_6838_epsilon_0 = const()[name = string("op_6838_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6838_cast_fp16 = rsqrt(epsilon = var_6838_epsilon_0, x = var_6837_cast_fp16)[name = string("op_6838_cast_fp16")]; + tensor x_179_cast_fp16 = mul(x = x_177_cast_fp16, y = var_6838_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor var_6841_to_fp16 = const()[name = string("op_6841_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800571520)))]; + tensor x_181_cast_fp16 = mul(x = x_179_cast_fp16, y = var_6841_to_fp16)[name = string("x_181_cast_fp16")]; + tensor layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800576704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(821548288))))[name = string("layers_7_self_attn_q_proj_weight_to_fp16_quantized")]; + tensor linear_55_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = x_181_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor var_6852 = const()[name = string("op_6852"), val = tensor([1, 2048, 16, 512])]; + tensor qg_cast_fp16 = reshape(shape = var_6852, x = linear_55_cast_fp16)[name = string("qg_cast_fp16")]; + tensor x_183_begin_0 = const()[name = string("x_183_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_183_end_0 = const()[name = string("x_183_end_0"), val = tensor([1, 2048, 16, 256])]; + tensor x_183_end_mask_0 = const()[name = string("x_183_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_183_cast_fp16 = slice_by_index(begin = x_183_begin_0, end = x_183_end_0, end_mask = x_183_end_mask_0, x = qg_cast_fp16)[name = string("x_183_cast_fp16")]; + tensor gate_5_begin_0 = const()[name = string("gate_5_begin_0"), val = tensor([0, 0, 0, 256])]; + tensor gate_5_end_0 = const()[name = string("gate_5_end_0"), val = tensor([1, 2048, 16, 512])]; + tensor gate_5_end_mask_0 = const()[name = string("gate_5_end_mask_0"), val = tensor([true, true, true, true])]; + tensor gate_5_cast_fp16 = slice_by_index(begin = gate_5_begin_0, end = gate_5_end_0, end_mask = gate_5_end_mask_0, x = qg_cast_fp16)[name = string("gate_5_cast_fp16")]; + tensor var_6856 = const()[name = string("op_6856"), val = tensor([1, 2048, 4096])]; + tensor gate_cast_fp16 = reshape(shape = var_6856, x = gate_5_cast_fp16)[name = string("gate_cast_fp16")]; + fp16 var_6823_promoted_1_to_fp16 = const()[name = string("op_6823_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor var_6859_cast_fp16 = pow(x = x_183_cast_fp16, y = var_6823_promoted_1_to_fp16)[name = string("op_6859_cast_fp16")]; + tensor var_6861_axes_0 = const()[name = string("op_6861_axes_0"), val = tensor([-1])]; + bool var_6861_keep_dims_0 = const()[name = string("op_6861_keep_dims_0"), val = bool(true)]; + tensor var_6861_cast_fp16 = reduce_mean(axes = var_6861_axes_0, keep_dims = var_6861_keep_dims_0, x = var_6859_cast_fp16)[name = string("op_6861_cast_fp16")]; + fp16 var_6862_to_fp16 = const()[name = string("op_6862_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_6863_cast_fp16 = add(x = var_6861_cast_fp16, y = var_6862_to_fp16)[name = string("op_6863_cast_fp16")]; + fp32 var_6864_epsilon_0 = const()[name = string("op_6864_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6864_cast_fp16 = rsqrt(epsilon = var_6864_epsilon_0, x = var_6863_cast_fp16)[name = string("op_6864_cast_fp16")]; + tensor x_185_cast_fp16 = mul(x = x_183_cast_fp16, y = var_6864_cast_fp16)[name = string("x_185_cast_fp16")]; + tensor var_6867_to_fp16 = const()[name = string("op_6867_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(821564736)))]; + tensor var_6868_cast_fp16 = mul(x = x_185_cast_fp16, y = var_6867_to_fp16)[name = string("op_6868_cast_fp16")]; + tensor q_65_perm_0 = const()[name = string("q_65_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(821565312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(824186816))))[name = string("layers_7_self_attn_k_proj_weight_to_fp16_quantized")]; + tensor linear_56_cast_fp16 = linear(bias = linear_25_bias_0_to_fp16, weight = layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = x_181_cast_fp16)[name = string("linear_56_cast_fp16")]; + tensor var_6872 = const()[name = string("op_6872"), val = tensor([1, 2048, 4, 256])]; + tensor x_187_cast_fp16 = reshape(shape = var_6872, x = linear_56_cast_fp16)[name = string("x_187_cast_fp16")]; + fp16 var_6823_promoted_2_to_fp16 = const()[name = string("op_6823_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor var_6875_cast_fp16 = pow(x = x_187_cast_fp16, y = var_6823_promoted_2_to_fp16)[name = string("op_6875_cast_fp16")]; + tensor var_6877_axes_0 = const()[name = string("op_6877_axes_0"), val = tensor([-1])]; + bool var_6877_keep_dims_0 = const()[name = string("op_6877_keep_dims_0"), val = bool(true)]; + tensor var_6877_cast_fp16 = reduce_mean(axes = var_6877_axes_0, keep_dims = var_6877_keep_dims_0, x = var_6875_cast_fp16)[name = string("op_6877_cast_fp16")]; + fp16 var_6878_to_fp16 = const()[name = string("op_6878_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_6879_cast_fp16 = add(x = var_6877_cast_fp16, y = var_6878_to_fp16)[name = string("op_6879_cast_fp16")]; + fp32 var_6880_epsilon_0 = const()[name = string("op_6880_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6880_cast_fp16 = rsqrt(epsilon = var_6880_epsilon_0, x = var_6879_cast_fp16)[name = string("op_6880_cast_fp16")]; + tensor x_189_cast_fp16 = mul(x = x_187_cast_fp16, y = var_6880_cast_fp16)[name = string("x_189_cast_fp16")]; + tensor var_6883_to_fp16 = const()[name = string("op_6883_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(824188928)))]; + tensor var_6884_cast_fp16 = mul(x = x_189_cast_fp16, y = var_6883_to_fp16)[name = string("op_6884_cast_fp16")]; + tensor k_55_perm_0 = const()[name = string("k_55_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(824189504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(826811008))))[name = string("layers_7_self_attn_v_proj_weight_to_fp16_quantized")]; + tensor linear_57_cast_fp16 = linear(bias = linear_25_bias_0_to_fp16, weight = layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = x_181_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_6888 = const()[name = string("op_6888"), val = tensor([1, 2048, 4, 256])]; + tensor var_6889_cast_fp16 = reshape(shape = var_6888, x = linear_57_cast_fp16)[name = string("op_6889_cast_fp16")]; + tensor v_29_perm_0 = const()[name = string("v_29_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_6895_begin_0 = const()[name = string("op_6895_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6895_end_0 = const()[name = string("op_6895_end_0"), val = tensor([1, 16, 2048, 64])]; + tensor var_6895_end_mask_0 = const()[name = string("op_6895_end_mask_0"), val = tensor([true, true, true, false])]; + tensor q_65_cast_fp16 = transpose(perm = q_65_perm_0, x = var_6868_cast_fp16)[name = string("transpose_3")]; + tensor var_6895_cast_fp16 = slice_by_index(begin = var_6895_begin_0, end = var_6895_end_0, end_mask = var_6895_end_mask_0, x = q_65_cast_fp16)[name = string("op_6895_cast_fp16")]; + tensor var_6896_cast_fp16 = mul(x = var_6895_cast_fp16, y = c_1_cast_fp16)[name = string("op_6896_cast_fp16")]; + tensor var_6903_begin_0 = const()[name = string("op_6903_begin_0"), val = tensor([0, 0, 0, 32])]; + tensor var_6903_end_0 = const()[name = string("op_6903_end_0"), val = tensor([1, 16, 2048, 64])]; + tensor var_6903_end_mask_0 = const()[name = string("op_6903_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6903_cast_fp16 = slice_by_index(begin = var_6903_begin_0, end = var_6903_end_0, end_mask = var_6903_end_mask_0, x = var_6895_cast_fp16)[name = string("op_6903_cast_fp16")]; + fp16 const_79_promoted_to_fp16 = const()[name = string("const_79_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor var_6904_cast_fp16 = mul(x = var_6903_cast_fp16, y = const_79_promoted_to_fp16)[name = string("op_6904_cast_fp16")]; + tensor var_6905_begin_0 = const()[name = string("op_6905_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6905_end_0 = const()[name = string("op_6905_end_0"), val = tensor([1, 16, 2048, 32])]; + tensor var_6905_end_mask_0 = const()[name = string("op_6905_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_6905_cast_fp16 = slice_by_index(begin = var_6905_begin_0, end = var_6905_end_0, end_mask = var_6905_end_mask_0, x = var_6895_cast_fp16)[name = string("op_6905_cast_fp16")]; + bool var_6907_interleave_0 = const()[name = string("op_6907_interleave_0"), val = bool(false)]; + tensor var_6907_cast_fp16 = concat(axis = var_6824, interleave = var_6907_interleave_0, values = (var_6904_cast_fp16, var_6905_cast_fp16))[name = string("op_6907_cast_fp16")]; + tensor var_6908_cast_fp16 = mul(x = var_6907_cast_fp16, y = s_1_cast_fp16)[name = string("op_6908_cast_fp16")]; + tensor var_6909_cast_fp16 = add(x = var_6896_cast_fp16, y = var_6908_cast_fp16)[name = string("op_6909_cast_fp16")]; + tensor var_6910_begin_0 = const()[name = string("op_6910_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor var_6910_end_0 = const()[name = string("op_6910_end_0"), val = tensor([1, 16, 2048, 256])]; + tensor var_6910_end_mask_0 = const()[name = string("op_6910_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6910_cast_fp16 = slice_by_index(begin = var_6910_begin_0, end = var_6910_end_0, end_mask = var_6910_end_mask_0, x = q_65_cast_fp16)[name = string("op_6910_cast_fp16")]; + bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)]; + tensor q_cast_fp16 = concat(axis = var_6824, interleave = q_interleave_0, values = (var_6909_cast_fp16, var_6910_cast_fp16))[name = string("q_cast_fp16")]; + tensor var_6913_begin_0 = const()[name = string("op_6913_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6913_end_0 = const()[name = string("op_6913_end_0"), val = tensor([1, 4, 2048, 64])]; + tensor var_6913_end_mask_0 = const()[name = string("op_6913_end_mask_0"), val = tensor([true, true, true, false])]; + tensor k_55_cast_fp16 = transpose(perm = k_55_perm_0, x = var_6884_cast_fp16)[name = string("transpose_2")]; + tensor var_6913_cast_fp16 = slice_by_index(begin = var_6913_begin_0, end = var_6913_end_0, end_mask = var_6913_end_mask_0, x = k_55_cast_fp16)[name = string("op_6913_cast_fp16")]; + tensor var_6914_cast_fp16 = mul(x = var_6913_cast_fp16, y = c_1_cast_fp16)[name = string("op_6914_cast_fp16")]; + tensor var_6921_begin_0 = const()[name = string("op_6921_begin_0"), val = tensor([0, 0, 0, 32])]; + tensor var_6921_end_0 = const()[name = string("op_6921_end_0"), val = tensor([1, 4, 2048, 64])]; + tensor var_6921_end_mask_0 = const()[name = string("op_6921_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6921_cast_fp16 = slice_by_index(begin = var_6921_begin_0, end = var_6921_end_0, end_mask = var_6921_end_mask_0, x = var_6913_cast_fp16)[name = string("op_6921_cast_fp16")]; + fp16 const_81_promoted_to_fp16 = const()[name = string("const_81_promoted_to_fp16"), val = fp16(-0x1p+0)]; + tensor var_6922_cast_fp16 = mul(x = var_6921_cast_fp16, y = const_81_promoted_to_fp16)[name = string("op_6922_cast_fp16")]; + tensor var_6923_begin_0 = const()[name = string("op_6923_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6923_end_0 = const()[name = string("op_6923_end_0"), val = tensor([1, 4, 2048, 32])]; + tensor var_6923_end_mask_0 = const()[name = string("op_6923_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_6923_cast_fp16 = slice_by_index(begin = var_6923_begin_0, end = var_6923_end_0, end_mask = var_6923_end_mask_0, x = var_6913_cast_fp16)[name = string("op_6923_cast_fp16")]; + bool var_6925_interleave_0 = const()[name = string("op_6925_interleave_0"), val = bool(false)]; + tensor var_6925_cast_fp16 = concat(axis = var_6824, interleave = var_6925_interleave_0, values = (var_6922_cast_fp16, var_6923_cast_fp16))[name = string("op_6925_cast_fp16")]; + tensor var_6926_cast_fp16 = mul(x = var_6925_cast_fp16, y = s_1_cast_fp16)[name = string("op_6926_cast_fp16")]; + tensor var_6927_cast_fp16 = add(x = var_6914_cast_fp16, y = var_6926_cast_fp16)[name = string("op_6927_cast_fp16")]; + tensor var_6928_begin_0 = const()[name = string("op_6928_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor var_6928_end_0 = const()[name = string("op_6928_end_0"), val = tensor([1, 4, 2048, 256])]; + tensor var_6928_end_mask_0 = const()[name = string("op_6928_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6928_cast_fp16 = slice_by_index(begin = var_6928_begin_0, end = var_6928_end_0, end_mask = var_6928_end_mask_0, x = k_55_cast_fp16)[name = string("op_6928_cast_fp16")]; + bool k_57_interleave_0 = const()[name = string("k_57_interleave_0"), val = bool(false)]; + tensor k_57_cast_fp16 = concat(axis = var_6824, interleave = k_57_interleave_0, values = (var_6927_cast_fp16, var_6928_cast_fp16))[name = string("k_57_cast_fp16")]; + tensor var_6933_axes_0 = const()[name = string("op_6933_axes_0"), val = tensor([2])]; + tensor var_6933_cast_fp16 = expand_dims(axes = var_6933_axes_0, x = k_57_cast_fp16)[name = string("op_6933_cast_fp16")]; + tensor var_6935_reps_0 = const()[name = string("op_6935_reps_0"), val = tensor([1, 1, 4, 1, 1])]; + tensor var_6935_cast_fp16 = tile(reps = var_6935_reps_0, x = var_6933_cast_fp16)[name = string("op_6935_cast_fp16")]; + tensor var_6936 = const()[name = string("op_6936"), val = tensor([1, 16, 2048, 256])]; + tensor k_cast_fp16 = reshape(shape = var_6936, x = var_6935_cast_fp16)[name = string("k_cast_fp16")]; + tensor var_6940_axes_0 = const()[name = string("op_6940_axes_0"), val = tensor([2])]; + tensor v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_6889_cast_fp16)[name = string("transpose_1")]; + tensor var_6940_cast_fp16 = expand_dims(axes = var_6940_axes_0, x = v_29_cast_fp16)[name = string("op_6940_cast_fp16")]; + tensor var_6942_reps_0 = const()[name = string("op_6942_reps_0"), val = tensor([1, 1, 4, 1, 1])]; + tensor var_6942_cast_fp16 = tile(reps = var_6942_reps_0, x = var_6940_cast_fp16)[name = string("op_6942_cast_fp16")]; + tensor var_6943 = const()[name = string("op_6943"), val = tensor([1, 16, 2048, 256])]; + tensor v_cast_fp16 = reshape(shape = var_6943, x = var_6942_cast_fp16)[name = string("v_cast_fp16")]; + bool var_6946_transpose_x_1 = const()[name = string("op_6946_transpose_x_1"), val = bool(false)]; + bool var_6946_transpose_y_1 = const()[name = string("op_6946_transpose_y_1"), val = bool(true)]; + tensor var_6946_cast_fp16 = matmul(transpose_x = var_6946_transpose_x_1, transpose_y = var_6946_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_6946_cast_fp16")]; + fp16 var_6947_to_fp16 = const()[name = string("op_6947_to_fp16"), val = fp16(0x1p-4)]; + tensor var_6948_cast_fp16 = mul(x = var_6946_cast_fp16, y = var_6947_to_fp16)[name = string("op_6948_cast_fp16")]; + tensor scores_cast_fp16 = add(x = var_6948_cast_fp16, y = mask_to_fp16)[name = string("scores_cast_fp16")]; + tensor var_6950_cast_fp16 = softmax(axis = var_6824, x = scores_cast_fp16)[name = string("op_6950_cast_fp16")]; + bool out_5_transpose_x_0 = const()[name = string("out_5_transpose_x_0"), val = bool(false)]; + bool out_5_transpose_y_0 = const()[name = string("out_5_transpose_y_0"), val = bool(false)]; + tensor out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = var_6950_cast_fp16, y = v_cast_fp16)[name = string("out_5_cast_fp16")]; + tensor var_6952_perm_0 = const()[name = string("op_6952_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_6953 = const()[name = string("op_6953"), val = tensor([1, 2048, 4096])]; + tensor var_6952_cast_fp16 = transpose(perm = var_6952_perm_0, x = out_5_cast_fp16)[name = string("transpose_0")]; + tensor out_cast_fp16 = reshape(shape = var_6953, x = var_6952_cast_fp16)[name = string("out_cast_fp16")]; + tensor var_6955_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_6955_cast_fp16")]; + tensor input_105_cast_fp16 = mul(x = out_cast_fp16, y = var_6955_cast_fp16)[name = string("input_105_cast_fp16")]; + tensor layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(826813120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(837298944))))[name = string("layers_7_self_attn_o_proj_weight_to_fp16_quantized")]; + tensor linear_58_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor x_195_cast_fp16 = add(x = x_177_cast_fp16, y = linear_58_cast_fp16)[name = string("x_195_cast_fp16")]; + fp16 var_6823_promoted_3_to_fp16 = const()[name = string("op_6823_promoted_3_to_fp16"), val = fp16(0x1p+1)]; + tensor var_6961_cast_fp16 = pow(x = x_195_cast_fp16, y = var_6823_promoted_3_to_fp16)[name = string("op_6961_cast_fp16")]; + tensor var_6963_axes_0 = const()[name = string("op_6963_axes_0"), val = tensor([-1])]; + bool var_6963_keep_dims_0 = const()[name = string("op_6963_keep_dims_0"), val = bool(true)]; + tensor var_6963_cast_fp16 = reduce_mean(axes = var_6963_axes_0, keep_dims = var_6963_keep_dims_0, x = var_6961_cast_fp16)[name = string("op_6963_cast_fp16")]; + fp16 var_6964_to_fp16 = const()[name = string("op_6964_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_6965_cast_fp16 = add(x = var_6963_cast_fp16, y = var_6964_to_fp16)[name = string("op_6965_cast_fp16")]; + fp32 var_6966_epsilon_0 = const()[name = string("op_6966_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6966_cast_fp16 = rsqrt(epsilon = var_6966_epsilon_0, x = var_6965_cast_fp16)[name = string("op_6966_cast_fp16")]; + tensor x_cast_fp16 = mul(x = x_195_cast_fp16, y = var_6966_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_6969_to_fp16 = const()[name = string("op_6969_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(837304128)))]; + tensor input_107_cast_fp16 = mul(x = x_cast_fp16, y = var_6969_to_fp16)[name = string("input_107_cast_fp16")]; + tensor layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(837309312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(860902336))))[name = string("layers_7_mlp_gate_proj_weight_to_fp16_quantized")]; + tensor linear_59_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_6976_cast_fp16 = silu(x = linear_59_cast_fp16)[name = string("op_6976_cast_fp16")]; + tensor layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(860920832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(884513856))))[name = string("layers_7_mlp_up_proj_weight_to_fp16_quantized")]; + tensor linear_60_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_60_cast_fp16")]; + tensor input_cast_fp16 = mul(x = var_6976_cast_fp16, y = linear_60_cast_fp16)[name = string("input_cast_fp16")]; + tensor layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(884532352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(908125376))))[name = string("layers_7_mlp_down_proj_weight_to_fp16_quantized")]; + tensor linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor hidden_out = add(x = x_195_cast_fp16, y = linear_61_cast_fp16)[name = string("op_6982_cast_fp16")]; + } -> (hidden_out); +} \ No newline at end of file