diff --git "a/coreml_minicpmv40_vit_f16.mlmodelc/model.mil" "b/coreml_minicpmv40_vit_f16.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/coreml_minicpmv40_vit_f16.mlmodelc/model.mil" @@ -0,0 +1,1570 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor pixel_values, tensor pos_embed, tensor position_ids) { + tensor var_14 = const()[name = tensor("op_14"), val = tensor(1)]; + tensor var_18 = const()[name = tensor("op_18"), val = tensor(-1)]; + tensor var_32 = const()[name = tensor("op_32"), val = tensor([14, 14])]; + tensor var_33 = const()[name = tensor("op_33"), val = tensor([1, 1])]; + tensor patch_embeds_pad_type_0 = const()[name = tensor("patch_embeds_pad_type_0"), val = tensor("valid")]; + tensor patch_embeds_pad_0 = const()[name = tensor("patch_embeds_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor pixel_values_to_fp16_dtype_0 = const()[name = tensor("pixel_values_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor vpm_embeddings_patch_embedding_weight_to_fp16 = const()[name = tensor("vpm_embeddings_patch_embedding_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor vpm_embeddings_patch_embedding_bias_to_fp16 = const()[name = tensor("vpm_embeddings_patch_embedding_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1354880)))]; + tensor pixel_values_to_fp16 = cast(dtype = pixel_values_to_fp16_dtype_0, x = pixel_values)[name = tensor("cast_150")]; + tensor patch_embeds_cast_fp16 = conv(bias = vpm_embeddings_patch_embedding_bias_to_fp16, dilations = var_33, groups = var_14, pad = patch_embeds_pad_0, pad_type = patch_embeds_pad_type_0, strides = var_32, weight = vpm_embeddings_patch_embedding_weight_to_fp16, x = pixel_values_to_fp16)[name = tensor("patch_embeds_cast_fp16")]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1152, 1024])]; + tensor var_35_cast_fp16 = reshape(shape = concat_0, x = patch_embeds_cast_fp16)[name = tensor("op_35_cast_fp16")]; + tensor embeddings_1_perm_0 = const()[name = tensor("embeddings_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_40_axis_0 = const()[name = tensor("op_40_axis_0"), val = tensor(0)]; + tensor vpm_embeddings_position_embedding_weight_to_fp16 = const()[name = tensor("vpm_embeddings_position_embedding_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1357248)))]; + tensor var_40_cast_fp16 = gather(axis = var_40_axis_0, indices = position_ids, x = vpm_embeddings_position_embedding_weight_to_fp16)[name = tensor("op_40_cast_fp16")]; + tensor embeddings_1_cast_fp16 = transpose(perm = embeddings_1_perm_0, x = var_35_cast_fp16)[name = tensor("transpose_250")]; + tensor input_3_cast_fp16 = add(x = embeddings_1_cast_fp16, y = var_40_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor hidden_states_1_axes_0 = const()[name = tensor("hidden_states_1_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_0_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12646912)))]; + tensor vpm_encoder_layers_0_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12649280)))]; + tensor var_12_to_fp16 = const()[name = tensor("op_12_to_fp16"), val = tensor(0x1.1p-20)]; + tensor hidden_states_1_cast_fp16 = layer_norm(axes = hidden_states_1_axes_0, beta = vpm_encoder_layers_0_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_0_layer_norm1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor vpm_encoder_layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12651648)))]; + tensor vpm_encoder_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15305920)))]; + tensor linear_0_cast_fp16 = linear(bias = vpm_encoder_layers_0_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_0_self_attn_q_proj_weight_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor vpm_encoder_layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15308288)))]; + tensor vpm_encoder_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17962560)))]; + tensor linear_1_cast_fp16 = linear(bias = vpm_encoder_layers_0_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_0_self_attn_k_proj_weight_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor vpm_encoder_layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17964928)))]; + tensor vpm_encoder_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20619200)))]; + tensor linear_2_cast_fp16 = linear(bias = vpm_encoder_layers_0_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_0_self_attn_v_proj_weight_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_119 = const()[name = tensor("op_119"), val = tensor([1, 1024, 16, 72])]; + tensor var_120_cast_fp16 = reshape(shape = var_119, x = linear_0_cast_fp16)[name = tensor("op_120_cast_fp16")]; + tensor var_122 = const()[name = tensor("op_122"), val = tensor([1, 1024, 16, 72])]; + tensor var_123_cast_fp16 = reshape(shape = var_122, x = linear_1_cast_fp16)[name = tensor("op_123_cast_fp16")]; + tensor var_125 = const()[name = tensor("op_125"), val = tensor([1, 1024, 16, 72])]; + tensor var_126_cast_fp16 = reshape(shape = var_125, x = linear_2_cast_fp16)[name = tensor("op_126_cast_fp16")]; + tensor value_states_3_perm_0 = const()[name = tensor("value_states_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_129_transpose_x_0 = const()[name = tensor("op_129_transpose_x_0"), val = tensor(false)]; + tensor var_129_transpose_y_0 = const()[name = tensor("op_129_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = var_123_cast_fp16)[name = tensor("transpose_247")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = var_120_cast_fp16)[name = tensor("transpose_248")]; + tensor var_129_cast_fp16 = matmul(transpose_x = var_129_transpose_x_0, transpose_y = var_129_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_129_cast_fp16")]; + tensor var_130_to_fp16 = const()[name = tensor("op_130_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_1_cast_fp16 = mul(x = var_129_cast_fp16, y = var_130_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_132_cast_fp16 = softmax(axis = var_18, x = attn_weights_1_cast_fp16)[name = tensor("op_132_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_126_cast_fp16)[name = tensor("transpose_249")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_132_cast_fp16, y = value_states_3_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_136_perm_0 = const()[name = tensor("op_136_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_138 = const()[name = tensor("op_138"), val = tensor([1, 1024, 1152])]; + tensor var_136_cast_fp16 = transpose(perm = var_136_perm_0, x = attn_output_1_cast_fp16)[name = tensor("transpose_246")]; + tensor input_7_cast_fp16 = reshape(shape = var_138, x = var_136_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor vpm_encoder_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20621568)))]; + tensor vpm_encoder_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23275840)))]; + tensor linear_3_cast_fp16 = linear(bias = vpm_encoder_layers_0_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_0_self_attn_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_9_cast_fp16 = add(x = input_3_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_0_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23278208)))]; + tensor vpm_encoder_layers_0_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23280576)))]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = vpm_encoder_layers_0_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_0_layer_norm2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor vpm_encoder_layers_0_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23282944)))]; + tensor vpm_encoder_layers_0_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33199424)))]; + tensor linear_4_cast_fp16 = linear(bias = vpm_encoder_layers_0_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_0_mlp_fc1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_mode_0 = const()[name = tensor("input_15_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor vpm_encoder_layers_0_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33208128)))]; + tensor vpm_encoder_layers_0_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_0_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43124608)))]; + tensor linear_5_cast_fp16 = linear(bias = vpm_encoder_layers_0_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_0_mlp_fc2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = input_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor hidden_states_7_axes_0 = const()[name = tensor("hidden_states_7_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_1_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43126976)))]; + tensor vpm_encoder_layers_1_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43129344)))]; + tensor hidden_states_7_cast_fp16 = layer_norm(axes = hidden_states_7_axes_0, beta = vpm_encoder_layers_1_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_1_layer_norm1_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor vpm_encoder_layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43131712)))]; + tensor vpm_encoder_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45785984)))]; + tensor linear_6_cast_fp16 = linear(bias = vpm_encoder_layers_1_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_1_self_attn_q_proj_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor vpm_encoder_layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45788352)))]; + tensor vpm_encoder_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48442624)))]; + tensor linear_7_cast_fp16 = linear(bias = vpm_encoder_layers_1_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_1_self_attn_k_proj_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor vpm_encoder_layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48444992)))]; + tensor vpm_encoder_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51099264)))]; + tensor linear_8_cast_fp16 = linear(bias = vpm_encoder_layers_1_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_1_self_attn_v_proj_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_181 = const()[name = tensor("op_181"), val = tensor([1, 1024, 16, 72])]; + tensor var_182_cast_fp16 = reshape(shape = var_181, x = linear_6_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor var_184 = const()[name = tensor("op_184"), val = tensor([1, 1024, 16, 72])]; + tensor var_185_cast_fp16 = reshape(shape = var_184, x = linear_7_cast_fp16)[name = tensor("op_185_cast_fp16")]; + tensor var_187 = const()[name = tensor("op_187"), val = tensor([1, 1024, 16, 72])]; + tensor var_188_cast_fp16 = reshape(shape = var_187, x = linear_8_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor value_states_7_perm_0 = const()[name = tensor("value_states_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_191_transpose_x_0 = const()[name = tensor("op_191_transpose_x_0"), val = tensor(false)]; + tensor var_191_transpose_y_0 = const()[name = tensor("op_191_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = var_185_cast_fp16)[name = tensor("transpose_243")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = var_182_cast_fp16)[name = tensor("transpose_244")]; + tensor var_191_cast_fp16 = matmul(transpose_x = var_191_transpose_x_0, transpose_y = var_191_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_191_cast_fp16")]; + tensor var_192_to_fp16 = const()[name = tensor("op_192_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_5_cast_fp16 = mul(x = var_191_cast_fp16, y = var_192_to_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor var_194_cast_fp16 = softmax(axis = var_18, x = attn_weights_5_cast_fp16)[name = tensor("op_194_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor value_states_7_cast_fp16 = transpose(perm = value_states_7_perm_0, x = var_188_cast_fp16)[name = tensor("transpose_245")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = var_194_cast_fp16, y = value_states_7_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_198_perm_0 = const()[name = tensor("op_198_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1, 1024, 1152])]; + tensor var_198_cast_fp16 = transpose(perm = var_198_perm_0, x = attn_output_5_cast_fp16)[name = tensor("transpose_242")]; + tensor input_21_cast_fp16 = reshape(shape = var_200, x = var_198_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor vpm_encoder_layers_1_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51101632)))]; + tensor vpm_encoder_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53755904)))]; + tensor linear_9_cast_fp16 = linear(bias = vpm_encoder_layers_1_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_1_self_attn_out_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor input_25_axes_0 = const()[name = tensor("input_25_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_1_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53758272)))]; + tensor vpm_encoder_layers_1_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53760640)))]; + tensor input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = vpm_encoder_layers_1_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_1_layer_norm2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor vpm_encoder_layers_1_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53763008)))]; + tensor vpm_encoder_layers_1_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63679488)))]; + tensor linear_10_cast_fp16 = linear(bias = vpm_encoder_layers_1_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_1_mlp_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor vpm_encoder_layers_1_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63688192)))]; + tensor vpm_encoder_layers_1_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_1_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73604672)))]; + tensor linear_11_cast_fp16 = linear(bias = vpm_encoder_layers_1_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_1_mlp_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_cast_fp16 = add(x = input_23_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor hidden_states_13_axes_0 = const()[name = tensor("hidden_states_13_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_2_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73607040)))]; + tensor vpm_encoder_layers_2_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73609408)))]; + tensor hidden_states_13_cast_fp16 = layer_norm(axes = hidden_states_13_axes_0, beta = vpm_encoder_layers_2_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_2_layer_norm1_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor vpm_encoder_layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73611776)))]; + tensor vpm_encoder_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76266048)))]; + tensor linear_12_cast_fp16 = linear(bias = vpm_encoder_layers_2_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_2_self_attn_q_proj_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor vpm_encoder_layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76268416)))]; + tensor vpm_encoder_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78922688)))]; + tensor linear_13_cast_fp16 = linear(bias = vpm_encoder_layers_2_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_2_self_attn_k_proj_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor vpm_encoder_layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78925056)))]; + tensor vpm_encoder_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81579328)))]; + tensor linear_14_cast_fp16 = linear(bias = vpm_encoder_layers_2_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_2_self_attn_v_proj_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_243 = const()[name = tensor("op_243"), val = tensor([1, 1024, 16, 72])]; + tensor var_244_cast_fp16 = reshape(shape = var_243, x = linear_12_cast_fp16)[name = tensor("op_244_cast_fp16")]; + tensor var_246 = const()[name = tensor("op_246"), val = tensor([1, 1024, 16, 72])]; + tensor var_247_cast_fp16 = reshape(shape = var_246, x = linear_13_cast_fp16)[name = tensor("op_247_cast_fp16")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor([1, 1024, 16, 72])]; + tensor var_250_cast_fp16 = reshape(shape = var_249, x = linear_14_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor value_states_11_perm_0 = const()[name = tensor("value_states_11_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_253_transpose_x_0 = const()[name = tensor("op_253_transpose_x_0"), val = tensor(false)]; + tensor var_253_transpose_y_0 = const()[name = tensor("op_253_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = var_247_cast_fp16)[name = tensor("transpose_239")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = var_244_cast_fp16)[name = tensor("transpose_240")]; + tensor var_253_cast_fp16 = matmul(transpose_x = var_253_transpose_x_0, transpose_y = var_253_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_253_cast_fp16")]; + tensor var_254_to_fp16 = const()[name = tensor("op_254_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_9_cast_fp16 = mul(x = var_253_cast_fp16, y = var_254_to_fp16)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_256_cast_fp16 = softmax(axis = var_18, x = attn_weights_9_cast_fp16)[name = tensor("op_256_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor value_states_11_cast_fp16 = transpose(perm = value_states_11_perm_0, x = var_250_cast_fp16)[name = tensor("transpose_241")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = var_256_cast_fp16, y = value_states_11_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_260_perm_0 = const()[name = tensor("op_260_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 1024, 1152])]; + tensor var_260_cast_fp16 = transpose(perm = var_260_perm_0, x = attn_output_9_cast_fp16)[name = tensor("transpose_238")]; + tensor input_35_cast_fp16 = reshape(shape = var_262, x = var_260_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor vpm_encoder_layers_2_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81581696)))]; + tensor vpm_encoder_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84235968)))]; + tensor linear_15_cast_fp16 = linear(bias = vpm_encoder_layers_2_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_2_self_attn_out_proj_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_2_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84238336)))]; + tensor vpm_encoder_layers_2_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84240704)))]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = vpm_encoder_layers_2_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_2_layer_norm2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor vpm_encoder_layers_2_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84243072)))]; + tensor vpm_encoder_layers_2_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94159552)))]; + tensor linear_16_cast_fp16 = linear(bias = vpm_encoder_layers_2_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_2_mlp_fc1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_mode_0 = const()[name = tensor("input_43_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_43_cast_fp16 = gelu(mode = input_43_mode_0, x = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor vpm_encoder_layers_2_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94168256)))]; + tensor vpm_encoder_layers_2_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_2_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104084736)))]; + tensor linear_17_cast_fp16 = linear(bias = vpm_encoder_layers_2_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_2_mlp_fc2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor input_45_cast_fp16 = add(x = input_37_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor hidden_states_19_axes_0 = const()[name = tensor("hidden_states_19_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_3_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104087104)))]; + tensor vpm_encoder_layers_3_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104089472)))]; + tensor hidden_states_19_cast_fp16 = layer_norm(axes = hidden_states_19_axes_0, beta = vpm_encoder_layers_3_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_3_layer_norm1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor vpm_encoder_layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104091840)))]; + tensor vpm_encoder_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106746112)))]; + tensor linear_18_cast_fp16 = linear(bias = vpm_encoder_layers_3_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_3_self_attn_q_proj_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor vpm_encoder_layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106748480)))]; + tensor vpm_encoder_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109402752)))]; + tensor linear_19_cast_fp16 = linear(bias = vpm_encoder_layers_3_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_3_self_attn_k_proj_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor vpm_encoder_layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109405120)))]; + tensor vpm_encoder_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112059392)))]; + tensor linear_20_cast_fp16 = linear(bias = vpm_encoder_layers_3_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_3_self_attn_v_proj_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_305 = const()[name = tensor("op_305"), val = tensor([1, 1024, 16, 72])]; + tensor var_306_cast_fp16 = reshape(shape = var_305, x = linear_18_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_308 = const()[name = tensor("op_308"), val = tensor([1, 1024, 16, 72])]; + tensor var_309_cast_fp16 = reshape(shape = var_308, x = linear_19_cast_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_311 = const()[name = tensor("op_311"), val = tensor([1, 1024, 16, 72])]; + tensor var_312_cast_fp16 = reshape(shape = var_311, x = linear_20_cast_fp16)[name = tensor("op_312_cast_fp16")]; + tensor value_states_15_perm_0 = const()[name = tensor("value_states_15_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_315_transpose_x_0 = const()[name = tensor("op_315_transpose_x_0"), val = tensor(false)]; + tensor var_315_transpose_y_0 = const()[name = tensor("op_315_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = var_309_cast_fp16)[name = tensor("transpose_235")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = var_306_cast_fp16)[name = tensor("transpose_236")]; + tensor var_315_cast_fp16 = matmul(transpose_x = var_315_transpose_x_0, transpose_y = var_315_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_315_cast_fp16")]; + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_13_cast_fp16 = mul(x = var_315_cast_fp16, y = var_316_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_318_cast_fp16 = softmax(axis = var_18, x = attn_weights_13_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_312_cast_fp16)[name = tensor("transpose_237")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = var_318_cast_fp16, y = value_states_15_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_322_perm_0 = const()[name = tensor("op_322_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_324 = const()[name = tensor("op_324"), val = tensor([1, 1024, 1152])]; + tensor var_322_cast_fp16 = transpose(perm = var_322_perm_0, x = attn_output_13_cast_fp16)[name = tensor("transpose_234")]; + tensor input_49_cast_fp16 = reshape(shape = var_324, x = var_322_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor vpm_encoder_layers_3_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112061760)))]; + tensor vpm_encoder_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114716032)))]; + tensor linear_21_cast_fp16 = linear(bias = vpm_encoder_layers_3_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_3_self_attn_out_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_3_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114718400)))]; + tensor vpm_encoder_layers_3_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114720768)))]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = vpm_encoder_layers_3_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_3_layer_norm2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor vpm_encoder_layers_3_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114723136)))]; + tensor vpm_encoder_layers_3_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124639616)))]; + tensor linear_22_cast_fp16 = linear(bias = vpm_encoder_layers_3_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_3_mlp_fc1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_57_mode_0 = const()[name = tensor("input_57_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_57_cast_fp16 = gelu(mode = input_57_mode_0, x = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor vpm_encoder_layers_3_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124648320)))]; + tensor vpm_encoder_layers_3_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_3_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134564800)))]; + tensor linear_23_cast_fp16 = linear(bias = vpm_encoder_layers_3_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_3_mlp_fc2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_59_cast_fp16 = add(x = input_51_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor hidden_states_25_axes_0 = const()[name = tensor("hidden_states_25_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_4_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134567168)))]; + tensor vpm_encoder_layers_4_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134569536)))]; + tensor hidden_states_25_cast_fp16 = layer_norm(axes = hidden_states_25_axes_0, beta = vpm_encoder_layers_4_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_4_layer_norm1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor vpm_encoder_layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134571904)))]; + tensor vpm_encoder_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137226176)))]; + tensor linear_24_cast_fp16 = linear(bias = vpm_encoder_layers_4_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_4_self_attn_q_proj_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor vpm_encoder_layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137228544)))]; + tensor vpm_encoder_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139882816)))]; + tensor linear_25_cast_fp16 = linear(bias = vpm_encoder_layers_4_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_4_self_attn_k_proj_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor vpm_encoder_layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139885184)))]; + tensor vpm_encoder_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142539456)))]; + tensor linear_26_cast_fp16 = linear(bias = vpm_encoder_layers_4_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_4_self_attn_v_proj_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1024, 16, 72])]; + tensor var_368_cast_fp16 = reshape(shape = var_367, x = linear_24_cast_fp16)[name = tensor("op_368_cast_fp16")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor([1, 1024, 16, 72])]; + tensor var_371_cast_fp16 = reshape(shape = var_370, x = linear_25_cast_fp16)[name = tensor("op_371_cast_fp16")]; + tensor var_373 = const()[name = tensor("op_373"), val = tensor([1, 1024, 16, 72])]; + tensor var_374_cast_fp16 = reshape(shape = var_373, x = linear_26_cast_fp16)[name = tensor("op_374_cast_fp16")]; + tensor value_states_19_perm_0 = const()[name = tensor("value_states_19_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_377_transpose_x_0 = const()[name = tensor("op_377_transpose_x_0"), val = tensor(false)]; + tensor var_377_transpose_y_0 = const()[name = tensor("op_377_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = var_371_cast_fp16)[name = tensor("transpose_231")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = var_368_cast_fp16)[name = tensor("transpose_232")]; + tensor var_377_cast_fp16 = matmul(transpose_x = var_377_transpose_x_0, transpose_y = var_377_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_377_cast_fp16")]; + tensor var_378_to_fp16 = const()[name = tensor("op_378_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_17_cast_fp16 = mul(x = var_377_cast_fp16, y = var_378_to_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor var_380_cast_fp16 = softmax(axis = var_18, x = attn_weights_17_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor value_states_19_cast_fp16 = transpose(perm = value_states_19_perm_0, x = var_374_cast_fp16)[name = tensor("transpose_233")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = var_380_cast_fp16, y = value_states_19_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_384_perm_0 = const()[name = tensor("op_384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 1024, 1152])]; + tensor var_384_cast_fp16 = transpose(perm = var_384_perm_0, x = attn_output_17_cast_fp16)[name = tensor("transpose_230")]; + tensor input_63_cast_fp16 = reshape(shape = var_386, x = var_384_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor vpm_encoder_layers_4_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142541824)))]; + tensor vpm_encoder_layers_4_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145196096)))]; + tensor linear_27_cast_fp16 = linear(bias = vpm_encoder_layers_4_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_4_self_attn_out_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = input_59_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor input_67_axes_0 = const()[name = tensor("input_67_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_4_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145198464)))]; + tensor vpm_encoder_layers_4_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145200832)))]; + tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = vpm_encoder_layers_4_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_4_layer_norm2_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor vpm_encoder_layers_4_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145203200)))]; + tensor vpm_encoder_layers_4_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155119680)))]; + tensor linear_28_cast_fp16 = linear(bias = vpm_encoder_layers_4_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_4_mlp_fc1_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_28_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor vpm_encoder_layers_4_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155128384)))]; + tensor vpm_encoder_layers_4_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_4_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165044864)))]; + tensor linear_29_cast_fp16 = linear(bias = vpm_encoder_layers_4_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_4_mlp_fc2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_65_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor hidden_states_31_axes_0 = const()[name = tensor("hidden_states_31_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_5_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165047232)))]; + tensor vpm_encoder_layers_5_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165049600)))]; + tensor hidden_states_31_cast_fp16 = layer_norm(axes = hidden_states_31_axes_0, beta = vpm_encoder_layers_5_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_5_layer_norm1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor vpm_encoder_layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165051968)))]; + tensor vpm_encoder_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167706240)))]; + tensor linear_30_cast_fp16 = linear(bias = vpm_encoder_layers_5_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_5_self_attn_q_proj_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor vpm_encoder_layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167708608)))]; + tensor vpm_encoder_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170362880)))]; + tensor linear_31_cast_fp16 = linear(bias = vpm_encoder_layers_5_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_5_self_attn_k_proj_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor vpm_encoder_layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170365248)))]; + tensor vpm_encoder_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173019520)))]; + tensor linear_32_cast_fp16 = linear(bias = vpm_encoder_layers_5_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_5_self_attn_v_proj_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 1024, 16, 72])]; + tensor var_430_cast_fp16 = reshape(shape = var_429, x = linear_30_cast_fp16)[name = tensor("op_430_cast_fp16")]; + tensor var_432 = const()[name = tensor("op_432"), val = tensor([1, 1024, 16, 72])]; + tensor var_433_cast_fp16 = reshape(shape = var_432, x = linear_31_cast_fp16)[name = tensor("op_433_cast_fp16")]; + tensor var_435 = const()[name = tensor("op_435"), val = tensor([1, 1024, 16, 72])]; + tensor var_436_cast_fp16 = reshape(shape = var_435, x = linear_32_cast_fp16)[name = tensor("op_436_cast_fp16")]; + tensor value_states_23_perm_0 = const()[name = tensor("value_states_23_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_439_transpose_x_0 = const()[name = tensor("op_439_transpose_x_0"), val = tensor(false)]; + tensor var_439_transpose_y_0 = const()[name = tensor("op_439_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = var_433_cast_fp16)[name = tensor("transpose_227")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = var_430_cast_fp16)[name = tensor("transpose_228")]; + tensor var_439_cast_fp16 = matmul(transpose_x = var_439_transpose_x_0, transpose_y = var_439_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_439_cast_fp16")]; + tensor var_440_to_fp16 = const()[name = tensor("op_440_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_21_cast_fp16 = mul(x = var_439_cast_fp16, y = var_440_to_fp16)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_442_cast_fp16 = softmax(axis = var_18, x = attn_weights_21_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor value_states_23_cast_fp16 = transpose(perm = value_states_23_perm_0, x = var_436_cast_fp16)[name = tensor("transpose_229")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = var_442_cast_fp16, y = value_states_23_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_446_perm_0 = const()[name = tensor("op_446_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([1, 1024, 1152])]; + tensor var_446_cast_fp16 = transpose(perm = var_446_perm_0, x = attn_output_21_cast_fp16)[name = tensor("transpose_226")]; + tensor input_77_cast_fp16 = reshape(shape = var_448, x = var_446_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor vpm_encoder_layers_5_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173021888)))]; + tensor vpm_encoder_layers_5_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175676160)))]; + tensor linear_33_cast_fp16 = linear(bias = vpm_encoder_layers_5_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_5_self_attn_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_5_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175678528)))]; + tensor vpm_encoder_layers_5_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175680896)))]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = vpm_encoder_layers_5_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_5_layer_norm2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor vpm_encoder_layers_5_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175683264)))]; + tensor vpm_encoder_layers_5_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185599744)))]; + tensor linear_34_cast_fp16 = linear(bias = vpm_encoder_layers_5_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_5_mlp_fc1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_34_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor vpm_encoder_layers_5_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185608448)))]; + tensor vpm_encoder_layers_5_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_5_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195524928)))]; + tensor linear_35_cast_fp16 = linear(bias = vpm_encoder_layers_5_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_5_mlp_fc2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_79_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor hidden_states_37_axes_0 = const()[name = tensor("hidden_states_37_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_6_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195527296)))]; + tensor vpm_encoder_layers_6_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195529664)))]; + tensor hidden_states_37_cast_fp16 = layer_norm(axes = hidden_states_37_axes_0, beta = vpm_encoder_layers_6_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_6_layer_norm1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor vpm_encoder_layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195532032)))]; + tensor vpm_encoder_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198186304)))]; + tensor linear_36_cast_fp16 = linear(bias = vpm_encoder_layers_6_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_6_self_attn_q_proj_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor vpm_encoder_layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198188672)))]; + tensor vpm_encoder_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200842944)))]; + tensor linear_37_cast_fp16 = linear(bias = vpm_encoder_layers_6_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_6_self_attn_k_proj_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor vpm_encoder_layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200845312)))]; + tensor vpm_encoder_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203499584)))]; + tensor linear_38_cast_fp16 = linear(bias = vpm_encoder_layers_6_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_6_self_attn_v_proj_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_491 = const()[name = tensor("op_491"), val = tensor([1, 1024, 16, 72])]; + tensor var_492_cast_fp16 = reshape(shape = var_491, x = linear_36_cast_fp16)[name = tensor("op_492_cast_fp16")]; + tensor var_494 = const()[name = tensor("op_494"), val = tensor([1, 1024, 16, 72])]; + tensor var_495_cast_fp16 = reshape(shape = var_494, x = linear_37_cast_fp16)[name = tensor("op_495_cast_fp16")]; + tensor var_497 = const()[name = tensor("op_497"), val = tensor([1, 1024, 16, 72])]; + tensor var_498_cast_fp16 = reshape(shape = var_497, x = linear_38_cast_fp16)[name = tensor("op_498_cast_fp16")]; + tensor value_states_27_perm_0 = const()[name = tensor("value_states_27_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_501_transpose_x_0 = const()[name = tensor("op_501_transpose_x_0"), val = tensor(false)]; + tensor var_501_transpose_y_0 = const()[name = tensor("op_501_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_495_cast_fp16)[name = tensor("transpose_223")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = var_492_cast_fp16)[name = tensor("transpose_224")]; + tensor var_501_cast_fp16 = matmul(transpose_x = var_501_transpose_x_0, transpose_y = var_501_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_501_cast_fp16")]; + tensor var_502_to_fp16 = const()[name = tensor("op_502_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_25_cast_fp16 = mul(x = var_501_cast_fp16, y = var_502_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_504_cast_fp16 = softmax(axis = var_18, x = attn_weights_25_cast_fp16)[name = tensor("op_504_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_498_cast_fp16)[name = tensor("transpose_225")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = var_504_cast_fp16, y = value_states_27_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_508_perm_0 = const()[name = tensor("op_508_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_510 = const()[name = tensor("op_510"), val = tensor([1, 1024, 1152])]; + tensor var_508_cast_fp16 = transpose(perm = var_508_perm_0, x = attn_output_25_cast_fp16)[name = tensor("transpose_222")]; + tensor input_91_cast_fp16 = reshape(shape = var_510, x = var_508_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor vpm_encoder_layers_6_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203501952)))]; + tensor vpm_encoder_layers_6_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206156224)))]; + tensor linear_39_cast_fp16 = linear(bias = vpm_encoder_layers_6_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_6_self_attn_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_6_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206158592)))]; + tensor vpm_encoder_layers_6_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206160960)))]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = vpm_encoder_layers_6_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_6_layer_norm2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor vpm_encoder_layers_6_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206163328)))]; + tensor vpm_encoder_layers_6_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216079808)))]; + tensor linear_40_cast_fp16 = linear(bias = vpm_encoder_layers_6_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_6_mlp_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_40_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor vpm_encoder_layers_6_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216088512)))]; + tensor vpm_encoder_layers_6_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_6_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226004992)))]; + tensor linear_41_cast_fp16 = linear(bias = vpm_encoder_layers_6_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_6_mlp_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = input_93_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor hidden_states_43_axes_0 = const()[name = tensor("hidden_states_43_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_7_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226007360)))]; + tensor vpm_encoder_layers_7_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226009728)))]; + tensor hidden_states_43_cast_fp16 = layer_norm(axes = hidden_states_43_axes_0, beta = vpm_encoder_layers_7_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_7_layer_norm1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor vpm_encoder_layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226012096)))]; + tensor vpm_encoder_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228666368)))]; + tensor linear_42_cast_fp16 = linear(bias = vpm_encoder_layers_7_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_7_self_attn_q_proj_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor vpm_encoder_layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228668736)))]; + tensor vpm_encoder_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231323008)))]; + tensor linear_43_cast_fp16 = linear(bias = vpm_encoder_layers_7_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_7_self_attn_k_proj_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor vpm_encoder_layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231325376)))]; + tensor vpm_encoder_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233979648)))]; + tensor linear_44_cast_fp16 = linear(bias = vpm_encoder_layers_7_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_7_self_attn_v_proj_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor([1, 1024, 16, 72])]; + tensor var_554_cast_fp16 = reshape(shape = var_553, x = linear_42_cast_fp16)[name = tensor("op_554_cast_fp16")]; + tensor var_556 = const()[name = tensor("op_556"), val = tensor([1, 1024, 16, 72])]; + tensor var_557_cast_fp16 = reshape(shape = var_556, x = linear_43_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 1024, 16, 72])]; + tensor var_560_cast_fp16 = reshape(shape = var_559, x = linear_44_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor value_states_31_perm_0 = const()[name = tensor("value_states_31_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_563_transpose_x_0 = const()[name = tensor("op_563_transpose_x_0"), val = tensor(false)]; + tensor var_563_transpose_y_0 = const()[name = tensor("op_563_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_557_cast_fp16)[name = tensor("transpose_219")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = var_554_cast_fp16)[name = tensor("transpose_220")]; + tensor var_563_cast_fp16 = matmul(transpose_x = var_563_transpose_x_0, transpose_y = var_563_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_563_cast_fp16")]; + tensor var_564_to_fp16 = const()[name = tensor("op_564_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_29_cast_fp16 = mul(x = var_563_cast_fp16, y = var_564_to_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor var_566_cast_fp16 = softmax(axis = var_18, x = attn_weights_29_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor value_states_31_cast_fp16 = transpose(perm = value_states_31_perm_0, x = var_560_cast_fp16)[name = tensor("transpose_221")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = var_566_cast_fp16, y = value_states_31_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_570_perm_0 = const()[name = tensor("op_570_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_572 = const()[name = tensor("op_572"), val = tensor([1, 1024, 1152])]; + tensor var_570_cast_fp16 = transpose(perm = var_570_perm_0, x = attn_output_29_cast_fp16)[name = tensor("transpose_218")]; + tensor input_105_cast_fp16 = reshape(shape = var_572, x = var_570_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor vpm_encoder_layers_7_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233982016)))]; + tensor vpm_encoder_layers_7_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236636288)))]; + tensor linear_45_cast_fp16 = linear(bias = vpm_encoder_layers_7_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_7_self_attn_out_proj_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_107_cast_fp16 = add(x = input_101_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_axes_0 = const()[name = tensor("input_109_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_7_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236638656)))]; + tensor vpm_encoder_layers_7_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236641024)))]; + tensor input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = vpm_encoder_layers_7_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_7_layer_norm2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor vpm_encoder_layers_7_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236643392)))]; + tensor vpm_encoder_layers_7_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246559872)))]; + tensor linear_46_cast_fp16 = linear(bias = vpm_encoder_layers_7_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_7_mlp_fc1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_113_mode_0 = const()[name = tensor("input_113_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_113_cast_fp16 = gelu(mode = input_113_mode_0, x = linear_46_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor vpm_encoder_layers_7_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246568576)))]; + tensor vpm_encoder_layers_7_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_7_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256485056)))]; + tensor linear_47_cast_fp16 = linear(bias = vpm_encoder_layers_7_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_7_mlp_fc2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_115_cast_fp16 = add(x = input_107_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor hidden_states_49_axes_0 = const()[name = tensor("hidden_states_49_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_8_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256487424)))]; + tensor vpm_encoder_layers_8_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256489792)))]; + tensor hidden_states_49_cast_fp16 = layer_norm(axes = hidden_states_49_axes_0, beta = vpm_encoder_layers_8_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_8_layer_norm1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor vpm_encoder_layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256492160)))]; + tensor vpm_encoder_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259146432)))]; + tensor linear_48_cast_fp16 = linear(bias = vpm_encoder_layers_8_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_8_self_attn_q_proj_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor vpm_encoder_layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259148800)))]; + tensor vpm_encoder_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261803072)))]; + tensor linear_49_cast_fp16 = linear(bias = vpm_encoder_layers_8_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_8_self_attn_k_proj_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor vpm_encoder_layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261805440)))]; + tensor vpm_encoder_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264459712)))]; + tensor linear_50_cast_fp16 = linear(bias = vpm_encoder_layers_8_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_8_self_attn_v_proj_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_615 = const()[name = tensor("op_615"), val = tensor([1, 1024, 16, 72])]; + tensor var_616_cast_fp16 = reshape(shape = var_615, x = linear_48_cast_fp16)[name = tensor("op_616_cast_fp16")]; + tensor var_618 = const()[name = tensor("op_618"), val = tensor([1, 1024, 16, 72])]; + tensor var_619_cast_fp16 = reshape(shape = var_618, x = linear_49_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 1024, 16, 72])]; + tensor var_622_cast_fp16 = reshape(shape = var_621, x = linear_50_cast_fp16)[name = tensor("op_622_cast_fp16")]; + tensor value_states_35_perm_0 = const()[name = tensor("value_states_35_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_625_transpose_x_0 = const()[name = tensor("op_625_transpose_x_0"), val = tensor(false)]; + tensor var_625_transpose_y_0 = const()[name = tensor("op_625_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_619_cast_fp16)[name = tensor("transpose_215")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = var_616_cast_fp16)[name = tensor("transpose_216")]; + tensor var_625_cast_fp16 = matmul(transpose_x = var_625_transpose_x_0, transpose_y = var_625_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_625_cast_fp16")]; + tensor var_626_to_fp16 = const()[name = tensor("op_626_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_33_cast_fp16 = mul(x = var_625_cast_fp16, y = var_626_to_fp16)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_628_cast_fp16 = softmax(axis = var_18, x = attn_weights_33_cast_fp16)[name = tensor("op_628_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor value_states_35_cast_fp16 = transpose(perm = value_states_35_perm_0, x = var_622_cast_fp16)[name = tensor("transpose_217")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = var_628_cast_fp16, y = value_states_35_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_632_perm_0 = const()[name = tensor("op_632_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_634 = const()[name = tensor("op_634"), val = tensor([1, 1024, 1152])]; + tensor var_632_cast_fp16 = transpose(perm = var_632_perm_0, x = attn_output_33_cast_fp16)[name = tensor("transpose_214")]; + tensor input_119_cast_fp16 = reshape(shape = var_634, x = var_632_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor vpm_encoder_layers_8_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264462080)))]; + tensor vpm_encoder_layers_8_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267116352)))]; + tensor linear_51_cast_fp16 = linear(bias = vpm_encoder_layers_8_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_8_self_attn_out_proj_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_8_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267118720)))]; + tensor vpm_encoder_layers_8_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267121088)))]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = vpm_encoder_layers_8_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_8_layer_norm2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor vpm_encoder_layers_8_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267123456)))]; + tensor vpm_encoder_layers_8_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277039936)))]; + tensor linear_52_cast_fp16 = linear(bias = vpm_encoder_layers_8_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_8_mlp_fc1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_127_mode_0 = const()[name = tensor("input_127_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = linear_52_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor vpm_encoder_layers_8_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277048640)))]; + tensor vpm_encoder_layers_8_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_8_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286965120)))]; + tensor linear_53_cast_fp16 = linear(bias = vpm_encoder_layers_8_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_8_mlp_fc2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor input_129_cast_fp16 = add(x = input_121_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor hidden_states_55_axes_0 = const()[name = tensor("hidden_states_55_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_9_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286967488)))]; + tensor vpm_encoder_layers_9_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286969856)))]; + tensor hidden_states_55_cast_fp16 = layer_norm(axes = hidden_states_55_axes_0, beta = vpm_encoder_layers_9_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_9_layer_norm1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("hidden_states_55_cast_fp16")]; + tensor vpm_encoder_layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286972224)))]; + tensor vpm_encoder_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289626496)))]; + tensor linear_54_cast_fp16 = linear(bias = vpm_encoder_layers_9_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_9_self_attn_q_proj_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor vpm_encoder_layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289628864)))]; + tensor vpm_encoder_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292283136)))]; + tensor linear_55_cast_fp16 = linear(bias = vpm_encoder_layers_9_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_9_self_attn_k_proj_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor vpm_encoder_layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292285504)))]; + tensor vpm_encoder_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294939776)))]; + tensor linear_56_cast_fp16 = linear(bias = vpm_encoder_layers_9_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_9_self_attn_v_proj_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1, 1024, 16, 72])]; + tensor var_678_cast_fp16 = reshape(shape = var_677, x = linear_54_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor var_680 = const()[name = tensor("op_680"), val = tensor([1, 1024, 16, 72])]; + tensor var_681_cast_fp16 = reshape(shape = var_680, x = linear_55_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor var_683 = const()[name = tensor("op_683"), val = tensor([1, 1024, 16, 72])]; + tensor var_684_cast_fp16 = reshape(shape = var_683, x = linear_56_cast_fp16)[name = tensor("op_684_cast_fp16")]; + tensor value_states_39_perm_0 = const()[name = tensor("value_states_39_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_687_transpose_x_0 = const()[name = tensor("op_687_transpose_x_0"), val = tensor(false)]; + tensor var_687_transpose_y_0 = const()[name = tensor("op_687_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_681_cast_fp16)[name = tensor("transpose_211")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = var_678_cast_fp16)[name = tensor("transpose_212")]; + tensor var_687_cast_fp16 = matmul(transpose_x = var_687_transpose_x_0, transpose_y = var_687_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_687_cast_fp16")]; + tensor var_688_to_fp16 = const()[name = tensor("op_688_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_37_cast_fp16 = mul(x = var_687_cast_fp16, y = var_688_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_690_cast_fp16 = softmax(axis = var_18, x = attn_weights_37_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_684_cast_fp16)[name = tensor("transpose_213")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = var_690_cast_fp16, y = value_states_39_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_694_perm_0 = const()[name = tensor("op_694_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_696 = const()[name = tensor("op_696"), val = tensor([1, 1024, 1152])]; + tensor var_694_cast_fp16 = transpose(perm = var_694_perm_0, x = attn_output_37_cast_fp16)[name = tensor("transpose_210")]; + tensor input_133_cast_fp16 = reshape(shape = var_696, x = var_694_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor vpm_encoder_layers_9_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294942144)))]; + tensor vpm_encoder_layers_9_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297596416)))]; + tensor linear_57_cast_fp16 = linear(bias = vpm_encoder_layers_9_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_9_self_attn_out_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_9_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297598784)))]; + tensor vpm_encoder_layers_9_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297601152)))]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = vpm_encoder_layers_9_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_9_layer_norm2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor vpm_encoder_layers_9_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297603520)))]; + tensor vpm_encoder_layers_9_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307520000)))]; + tensor linear_58_cast_fp16 = linear(bias = vpm_encoder_layers_9_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_9_mlp_fc1_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_58_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor vpm_encoder_layers_9_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307528704)))]; + tensor vpm_encoder_layers_9_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_9_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317445184)))]; + tensor linear_59_cast_fp16 = linear(bias = vpm_encoder_layers_9_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_9_mlp_fc2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_143_cast_fp16 = add(x = input_135_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor hidden_states_61_axes_0 = const()[name = tensor("hidden_states_61_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_10_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317447552)))]; + tensor vpm_encoder_layers_10_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317449920)))]; + tensor hidden_states_61_cast_fp16 = layer_norm(axes = hidden_states_61_axes_0, beta = vpm_encoder_layers_10_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_10_layer_norm1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("hidden_states_61_cast_fp16")]; + tensor vpm_encoder_layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317452288)))]; + tensor vpm_encoder_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320106560)))]; + tensor linear_60_cast_fp16 = linear(bias = vpm_encoder_layers_10_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_10_self_attn_q_proj_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor vpm_encoder_layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320108928)))]; + tensor vpm_encoder_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322763200)))]; + tensor linear_61_cast_fp16 = linear(bias = vpm_encoder_layers_10_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_10_self_attn_k_proj_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor vpm_encoder_layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322765568)))]; + tensor vpm_encoder_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325419840)))]; + tensor linear_62_cast_fp16 = linear(bias = vpm_encoder_layers_10_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_10_self_attn_v_proj_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_739 = const()[name = tensor("op_739"), val = tensor([1, 1024, 16, 72])]; + tensor var_740_cast_fp16 = reshape(shape = var_739, x = linear_60_cast_fp16)[name = tensor("op_740_cast_fp16")]; + tensor var_742 = const()[name = tensor("op_742"), val = tensor([1, 1024, 16, 72])]; + tensor var_743_cast_fp16 = reshape(shape = var_742, x = linear_61_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor var_745 = const()[name = tensor("op_745"), val = tensor([1, 1024, 16, 72])]; + tensor var_746_cast_fp16 = reshape(shape = var_745, x = linear_62_cast_fp16)[name = tensor("op_746_cast_fp16")]; + tensor value_states_43_perm_0 = const()[name = tensor("value_states_43_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_749_transpose_x_0 = const()[name = tensor("op_749_transpose_x_0"), val = tensor(false)]; + tensor var_749_transpose_y_0 = const()[name = tensor("op_749_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_743_cast_fp16)[name = tensor("transpose_207")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = var_740_cast_fp16)[name = tensor("transpose_208")]; + tensor var_749_cast_fp16 = matmul(transpose_x = var_749_transpose_x_0, transpose_y = var_749_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_749_cast_fp16")]; + tensor var_750_to_fp16 = const()[name = tensor("op_750_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_41_cast_fp16 = mul(x = var_749_cast_fp16, y = var_750_to_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor var_752_cast_fp16 = softmax(axis = var_18, x = attn_weights_41_cast_fp16)[name = tensor("op_752_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor value_states_43_cast_fp16 = transpose(perm = value_states_43_perm_0, x = var_746_cast_fp16)[name = tensor("transpose_209")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = var_752_cast_fp16, y = value_states_43_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_756_perm_0 = const()[name = tensor("op_756_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1024, 1152])]; + tensor var_756_cast_fp16 = transpose(perm = var_756_perm_0, x = attn_output_41_cast_fp16)[name = tensor("transpose_206")]; + tensor input_147_cast_fp16 = reshape(shape = var_758, x = var_756_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor vpm_encoder_layers_10_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325422208)))]; + tensor vpm_encoder_layers_10_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328076480)))]; + tensor linear_63_cast_fp16 = linear(bias = vpm_encoder_layers_10_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_10_self_attn_out_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = input_143_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_axes_0 = const()[name = tensor("input_151_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_10_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328078848)))]; + tensor vpm_encoder_layers_10_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328081216)))]; + tensor input_151_cast_fp16 = layer_norm(axes = input_151_axes_0, beta = vpm_encoder_layers_10_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_10_layer_norm2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor vpm_encoder_layers_10_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328083584)))]; + tensor vpm_encoder_layers_10_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338000064)))]; + tensor linear_64_cast_fp16 = linear(bias = vpm_encoder_layers_10_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_10_mlp_fc1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_64_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor vpm_encoder_layers_10_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338008768)))]; + tensor vpm_encoder_layers_10_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_10_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347925248)))]; + tensor linear_65_cast_fp16 = linear(bias = vpm_encoder_layers_10_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_10_mlp_fc2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_149_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor hidden_states_67_axes_0 = const()[name = tensor("hidden_states_67_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_11_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347927616)))]; + tensor vpm_encoder_layers_11_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347929984)))]; + tensor hidden_states_67_cast_fp16 = layer_norm(axes = hidden_states_67_axes_0, beta = vpm_encoder_layers_11_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_11_layer_norm1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("hidden_states_67_cast_fp16")]; + tensor vpm_encoder_layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347932352)))]; + tensor vpm_encoder_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350586624)))]; + tensor linear_66_cast_fp16 = linear(bias = vpm_encoder_layers_11_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_11_self_attn_q_proj_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor vpm_encoder_layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350588992)))]; + tensor vpm_encoder_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353243264)))]; + tensor linear_67_cast_fp16 = linear(bias = vpm_encoder_layers_11_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_11_self_attn_k_proj_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor vpm_encoder_layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353245632)))]; + tensor vpm_encoder_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355899904)))]; + tensor linear_68_cast_fp16 = linear(bias = vpm_encoder_layers_11_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_11_self_attn_v_proj_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 1024, 16, 72])]; + tensor var_802_cast_fp16 = reshape(shape = var_801, x = linear_66_cast_fp16)[name = tensor("op_802_cast_fp16")]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 1024, 16, 72])]; + tensor var_805_cast_fp16 = reshape(shape = var_804, x = linear_67_cast_fp16)[name = tensor("op_805_cast_fp16")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 1024, 16, 72])]; + tensor var_808_cast_fp16 = reshape(shape = var_807, x = linear_68_cast_fp16)[name = tensor("op_808_cast_fp16")]; + tensor value_states_47_perm_0 = const()[name = tensor("value_states_47_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; + tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_805_cast_fp16)[name = tensor("transpose_203")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = var_802_cast_fp16)[name = tensor("transpose_204")]; + tensor var_811_cast_fp16 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_811_cast_fp16")]; + tensor var_812_to_fp16 = const()[name = tensor("op_812_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_45_cast_fp16 = mul(x = var_811_cast_fp16, y = var_812_to_fp16)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_814_cast_fp16 = softmax(axis = var_18, x = attn_weights_45_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor value_states_47_cast_fp16 = transpose(perm = value_states_47_perm_0, x = var_808_cast_fp16)[name = tensor("transpose_205")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = var_814_cast_fp16, y = value_states_47_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_818_perm_0 = const()[name = tensor("op_818_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_820 = const()[name = tensor("op_820"), val = tensor([1, 1024, 1152])]; + tensor var_818_cast_fp16 = transpose(perm = var_818_perm_0, x = attn_output_45_cast_fp16)[name = tensor("transpose_202")]; + tensor input_161_cast_fp16 = reshape(shape = var_820, x = var_818_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor vpm_encoder_layers_11_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355902272)))]; + tensor vpm_encoder_layers_11_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358556544)))]; + tensor linear_69_cast_fp16 = linear(bias = vpm_encoder_layers_11_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_11_self_attn_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_11_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358558912)))]; + tensor vpm_encoder_layers_11_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358561280)))]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = vpm_encoder_layers_11_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_11_layer_norm2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor vpm_encoder_layers_11_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358563648)))]; + tensor vpm_encoder_layers_11_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368480128)))]; + tensor linear_70_cast_fp16 = linear(bias = vpm_encoder_layers_11_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_11_mlp_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_70_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor vpm_encoder_layers_11_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368488832)))]; + tensor vpm_encoder_layers_11_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_11_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378405312)))]; + tensor linear_71_cast_fp16 = linear(bias = vpm_encoder_layers_11_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_11_mlp_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_163_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor hidden_states_73_axes_0 = const()[name = tensor("hidden_states_73_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_12_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378407680)))]; + tensor vpm_encoder_layers_12_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378410048)))]; + tensor hidden_states_73_cast_fp16 = layer_norm(axes = hidden_states_73_axes_0, beta = vpm_encoder_layers_12_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_12_layer_norm1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("hidden_states_73_cast_fp16")]; + tensor vpm_encoder_layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378412416)))]; + tensor vpm_encoder_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381066688)))]; + tensor linear_72_cast_fp16 = linear(bias = vpm_encoder_layers_12_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_12_self_attn_q_proj_weight_to_fp16, x = hidden_states_73_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor vpm_encoder_layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381069056)))]; + tensor vpm_encoder_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383723328)))]; + tensor linear_73_cast_fp16 = linear(bias = vpm_encoder_layers_12_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_12_self_attn_k_proj_weight_to_fp16, x = hidden_states_73_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor vpm_encoder_layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383725696)))]; + tensor vpm_encoder_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386379968)))]; + tensor linear_74_cast_fp16 = linear(bias = vpm_encoder_layers_12_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_12_self_attn_v_proj_weight_to_fp16, x = hidden_states_73_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_863 = const()[name = tensor("op_863"), val = tensor([1, 1024, 16, 72])]; + tensor var_864_cast_fp16 = reshape(shape = var_863, x = linear_72_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1024, 16, 72])]; + tensor var_867_cast_fp16 = reshape(shape = var_866, x = linear_73_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor var_869 = const()[name = tensor("op_869"), val = tensor([1, 1024, 16, 72])]; + tensor var_870_cast_fp16 = reshape(shape = var_869, x = linear_74_cast_fp16)[name = tensor("op_870_cast_fp16")]; + tensor value_states_51_perm_0 = const()[name = tensor("value_states_51_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_873_transpose_x_0 = const()[name = tensor("op_873_transpose_x_0"), val = tensor(false)]; + tensor var_873_transpose_y_0 = const()[name = tensor("op_873_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_867_cast_fp16)[name = tensor("transpose_199")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = var_864_cast_fp16)[name = tensor("transpose_200")]; + tensor var_873_cast_fp16 = matmul(transpose_x = var_873_transpose_x_0, transpose_y = var_873_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_873_cast_fp16")]; + tensor var_874_to_fp16 = const()[name = tensor("op_874_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_49_cast_fp16 = mul(x = var_873_cast_fp16, y = var_874_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_876_cast_fp16 = softmax(axis = var_18, x = attn_weights_49_cast_fp16)[name = tensor("op_876_cast_fp16")]; + tensor attn_output_49_transpose_x_0 = const()[name = tensor("attn_output_49_transpose_x_0"), val = tensor(false)]; + tensor attn_output_49_transpose_y_0 = const()[name = tensor("attn_output_49_transpose_y_0"), val = tensor(false)]; + tensor value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_870_cast_fp16)[name = tensor("transpose_201")]; + tensor attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = var_876_cast_fp16, y = value_states_51_cast_fp16)[name = tensor("attn_output_49_cast_fp16")]; + tensor var_880_perm_0 = const()[name = tensor("op_880_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_882 = const()[name = tensor("op_882"), val = tensor([1, 1024, 1152])]; + tensor var_880_cast_fp16 = transpose(perm = var_880_perm_0, x = attn_output_49_cast_fp16)[name = tensor("transpose_198")]; + tensor input_175_cast_fp16 = reshape(shape = var_882, x = var_880_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor vpm_encoder_layers_12_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386382336)))]; + tensor vpm_encoder_layers_12_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389036608)))]; + tensor linear_75_cast_fp16 = linear(bias = vpm_encoder_layers_12_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_12_self_attn_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_12_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389038976)))]; + tensor vpm_encoder_layers_12_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389041344)))]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = vpm_encoder_layers_12_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_12_layer_norm2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor vpm_encoder_layers_12_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389043712)))]; + tensor vpm_encoder_layers_12_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398960192)))]; + tensor linear_76_cast_fp16 = linear(bias = vpm_encoder_layers_12_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_12_mlp_fc1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_183_mode_0 = const()[name = tensor("input_183_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_183_cast_fp16 = gelu(mode = input_183_mode_0, x = linear_76_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor vpm_encoder_layers_12_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398968896)))]; + tensor vpm_encoder_layers_12_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_12_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408885376)))]; + tensor linear_77_cast_fp16 = linear(bias = vpm_encoder_layers_12_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_12_mlp_fc2_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = input_177_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor hidden_states_79_axes_0 = const()[name = tensor("hidden_states_79_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_13_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408887744)))]; + tensor vpm_encoder_layers_13_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408890112)))]; + tensor hidden_states_79_cast_fp16 = layer_norm(axes = hidden_states_79_axes_0, beta = vpm_encoder_layers_13_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_13_layer_norm1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("hidden_states_79_cast_fp16")]; + tensor vpm_encoder_layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408892480)))]; + tensor vpm_encoder_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411546752)))]; + tensor linear_78_cast_fp16 = linear(bias = vpm_encoder_layers_13_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_13_self_attn_q_proj_weight_to_fp16, x = hidden_states_79_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor vpm_encoder_layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411549120)))]; + tensor vpm_encoder_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414203392)))]; + tensor linear_79_cast_fp16 = linear(bias = vpm_encoder_layers_13_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_13_self_attn_k_proj_weight_to_fp16, x = hidden_states_79_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor vpm_encoder_layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414205760)))]; + tensor vpm_encoder_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416860032)))]; + tensor linear_80_cast_fp16 = linear(bias = vpm_encoder_layers_13_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_13_self_attn_v_proj_weight_to_fp16, x = hidden_states_79_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_925 = const()[name = tensor("op_925"), val = tensor([1, 1024, 16, 72])]; + tensor var_926_cast_fp16 = reshape(shape = var_925, x = linear_78_cast_fp16)[name = tensor("op_926_cast_fp16")]; + tensor var_928 = const()[name = tensor("op_928"), val = tensor([1, 1024, 16, 72])]; + tensor var_929_cast_fp16 = reshape(shape = var_928, x = linear_79_cast_fp16)[name = tensor("op_929_cast_fp16")]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor([1, 1024, 16, 72])]; + tensor var_932_cast_fp16 = reshape(shape = var_931, x = linear_80_cast_fp16)[name = tensor("op_932_cast_fp16")]; + tensor value_states_55_perm_0 = const()[name = tensor("value_states_55_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_935_transpose_x_0 = const()[name = tensor("op_935_transpose_x_0"), val = tensor(false)]; + tensor var_935_transpose_y_0 = const()[name = tensor("op_935_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_929_cast_fp16)[name = tensor("transpose_195")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = var_926_cast_fp16)[name = tensor("transpose_196")]; + tensor var_935_cast_fp16 = matmul(transpose_x = var_935_transpose_x_0, transpose_y = var_935_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_935_cast_fp16")]; + tensor var_936_to_fp16 = const()[name = tensor("op_936_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_53_cast_fp16 = mul(x = var_935_cast_fp16, y = var_936_to_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor var_938_cast_fp16 = softmax(axis = var_18, x = attn_weights_53_cast_fp16)[name = tensor("op_938_cast_fp16")]; + tensor attn_output_53_transpose_x_0 = const()[name = tensor("attn_output_53_transpose_x_0"), val = tensor(false)]; + tensor attn_output_53_transpose_y_0 = const()[name = tensor("attn_output_53_transpose_y_0"), val = tensor(false)]; + tensor value_states_55_cast_fp16 = transpose(perm = value_states_55_perm_0, x = var_932_cast_fp16)[name = tensor("transpose_197")]; + tensor attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = var_938_cast_fp16, y = value_states_55_cast_fp16)[name = tensor("attn_output_53_cast_fp16")]; + tensor var_942_perm_0 = const()[name = tensor("op_942_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([1, 1024, 1152])]; + tensor var_942_cast_fp16 = transpose(perm = var_942_perm_0, x = attn_output_53_cast_fp16)[name = tensor("transpose_194")]; + tensor input_189_cast_fp16 = reshape(shape = var_944, x = var_942_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor vpm_encoder_layers_13_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416862400)))]; + tensor vpm_encoder_layers_13_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419516672)))]; + tensor linear_81_cast_fp16 = linear(bias = vpm_encoder_layers_13_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_13_self_attn_out_proj_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor input_191_cast_fp16 = add(x = input_185_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor input_193_axes_0 = const()[name = tensor("input_193_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_13_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419519040)))]; + tensor vpm_encoder_layers_13_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419521408)))]; + tensor input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, beta = vpm_encoder_layers_13_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_13_layer_norm2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor vpm_encoder_layers_13_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419523776)))]; + tensor vpm_encoder_layers_13_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429440256)))]; + tensor linear_82_cast_fp16 = linear(bias = vpm_encoder_layers_13_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_13_mlp_fc1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_197_mode_0 = const()[name = tensor("input_197_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_197_cast_fp16 = gelu(mode = input_197_mode_0, x = linear_82_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor vpm_encoder_layers_13_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429448960)))]; + tensor vpm_encoder_layers_13_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_13_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439365440)))]; + tensor linear_83_cast_fp16 = linear(bias = vpm_encoder_layers_13_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_13_mlp_fc2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_199_cast_fp16 = add(x = input_191_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor hidden_states_85_axes_0 = const()[name = tensor("hidden_states_85_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_14_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439367808)))]; + tensor vpm_encoder_layers_14_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439370176)))]; + tensor hidden_states_85_cast_fp16 = layer_norm(axes = hidden_states_85_axes_0, beta = vpm_encoder_layers_14_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_14_layer_norm1_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("hidden_states_85_cast_fp16")]; + tensor vpm_encoder_layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439372544)))]; + tensor vpm_encoder_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442026816)))]; + tensor linear_84_cast_fp16 = linear(bias = vpm_encoder_layers_14_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_14_self_attn_q_proj_weight_to_fp16, x = hidden_states_85_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor vpm_encoder_layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442029184)))]; + tensor vpm_encoder_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444683456)))]; + tensor linear_85_cast_fp16 = linear(bias = vpm_encoder_layers_14_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_14_self_attn_k_proj_weight_to_fp16, x = hidden_states_85_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor vpm_encoder_layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444685824)))]; + tensor vpm_encoder_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447340096)))]; + tensor linear_86_cast_fp16 = linear(bias = vpm_encoder_layers_14_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_14_self_attn_v_proj_weight_to_fp16, x = hidden_states_85_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor([1, 1024, 16, 72])]; + tensor var_988_cast_fp16 = reshape(shape = var_987, x = linear_84_cast_fp16)[name = tensor("op_988_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1024, 16, 72])]; + tensor var_991_cast_fp16 = reshape(shape = var_990, x = linear_85_cast_fp16)[name = tensor("op_991_cast_fp16")]; + tensor var_993 = const()[name = tensor("op_993"), val = tensor([1, 1024, 16, 72])]; + tensor var_994_cast_fp16 = reshape(shape = var_993, x = linear_86_cast_fp16)[name = tensor("op_994_cast_fp16")]; + tensor value_states_59_perm_0 = const()[name = tensor("value_states_59_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_997_transpose_x_0 = const()[name = tensor("op_997_transpose_x_0"), val = tensor(false)]; + tensor var_997_transpose_y_0 = const()[name = tensor("op_997_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_991_cast_fp16)[name = tensor("transpose_191")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = var_988_cast_fp16)[name = tensor("transpose_192")]; + tensor var_997_cast_fp16 = matmul(transpose_x = var_997_transpose_x_0, transpose_y = var_997_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_997_cast_fp16")]; + tensor var_998_to_fp16 = const()[name = tensor("op_998_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_57_cast_fp16 = mul(x = var_997_cast_fp16, y = var_998_to_fp16)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_1000_cast_fp16 = softmax(axis = var_18, x = attn_weights_57_cast_fp16)[name = tensor("op_1000_cast_fp16")]; + tensor attn_output_57_transpose_x_0 = const()[name = tensor("attn_output_57_transpose_x_0"), val = tensor(false)]; + tensor attn_output_57_transpose_y_0 = const()[name = tensor("attn_output_57_transpose_y_0"), val = tensor(false)]; + tensor value_states_59_cast_fp16 = transpose(perm = value_states_59_perm_0, x = var_994_cast_fp16)[name = tensor("transpose_193")]; + tensor attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = var_1000_cast_fp16, y = value_states_59_cast_fp16)[name = tensor("attn_output_57_cast_fp16")]; + tensor var_1004_perm_0 = const()[name = tensor("op_1004_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor([1, 1024, 1152])]; + tensor var_1004_cast_fp16 = transpose(perm = var_1004_perm_0, x = attn_output_57_cast_fp16)[name = tensor("transpose_190")]; + tensor input_203_cast_fp16 = reshape(shape = var_1006, x = var_1004_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor vpm_encoder_layers_14_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447342464)))]; + tensor vpm_encoder_layers_14_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449996736)))]; + tensor linear_87_cast_fp16 = linear(bias = vpm_encoder_layers_14_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_14_self_attn_out_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_14_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449999104)))]; + tensor vpm_encoder_layers_14_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450001472)))]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = vpm_encoder_layers_14_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_14_layer_norm2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor vpm_encoder_layers_14_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450003840)))]; + tensor vpm_encoder_layers_14_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459920320)))]; + tensor linear_88_cast_fp16 = linear(bias = vpm_encoder_layers_14_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_14_mlp_fc1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_88_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor vpm_encoder_layers_14_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459929024)))]; + tensor vpm_encoder_layers_14_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_14_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469845504)))]; + tensor linear_89_cast_fp16 = linear(bias = vpm_encoder_layers_14_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_14_mlp_fc2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor input_213_cast_fp16 = add(x = input_205_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor hidden_states_91_axes_0 = const()[name = tensor("hidden_states_91_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_15_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469847872)))]; + tensor vpm_encoder_layers_15_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469850240)))]; + tensor hidden_states_91_cast_fp16 = layer_norm(axes = hidden_states_91_axes_0, beta = vpm_encoder_layers_15_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_15_layer_norm1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("hidden_states_91_cast_fp16")]; + tensor vpm_encoder_layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469852608)))]; + tensor vpm_encoder_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(472506880)))]; + tensor linear_90_cast_fp16 = linear(bias = vpm_encoder_layers_15_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_15_self_attn_q_proj_weight_to_fp16, x = hidden_states_91_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor vpm_encoder_layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(472509248)))]; + tensor vpm_encoder_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475163520)))]; + tensor linear_91_cast_fp16 = linear(bias = vpm_encoder_layers_15_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_15_self_attn_k_proj_weight_to_fp16, x = hidden_states_91_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor vpm_encoder_layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475165888)))]; + tensor vpm_encoder_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477820160)))]; + tensor linear_92_cast_fp16 = linear(bias = vpm_encoder_layers_15_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_15_self_attn_v_proj_weight_to_fp16, x = hidden_states_91_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1024, 16, 72])]; + tensor var_1050_cast_fp16 = reshape(shape = var_1049, x = linear_90_cast_fp16)[name = tensor("op_1050_cast_fp16")]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1, 1024, 16, 72])]; + tensor var_1053_cast_fp16 = reshape(shape = var_1052, x = linear_91_cast_fp16)[name = tensor("op_1053_cast_fp16")]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor([1, 1024, 16, 72])]; + tensor var_1056_cast_fp16 = reshape(shape = var_1055, x = linear_92_cast_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor value_states_63_perm_0 = const()[name = tensor("value_states_63_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1059_transpose_x_0 = const()[name = tensor("op_1059_transpose_x_0"), val = tensor(false)]; + tensor var_1059_transpose_y_0 = const()[name = tensor("op_1059_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_1053_cast_fp16)[name = tensor("transpose_187")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = var_1050_cast_fp16)[name = tensor("transpose_188")]; + tensor var_1059_cast_fp16 = matmul(transpose_x = var_1059_transpose_x_0, transpose_y = var_1059_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_1059_cast_fp16")]; + tensor var_1060_to_fp16 = const()[name = tensor("op_1060_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_61_cast_fp16 = mul(x = var_1059_cast_fp16, y = var_1060_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_1062_cast_fp16 = softmax(axis = var_18, x = attn_weights_61_cast_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor attn_output_61_transpose_x_0 = const()[name = tensor("attn_output_61_transpose_x_0"), val = tensor(false)]; + tensor attn_output_61_transpose_y_0 = const()[name = tensor("attn_output_61_transpose_y_0"), val = tensor(false)]; + tensor value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_1056_cast_fp16)[name = tensor("transpose_189")]; + tensor attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = var_1062_cast_fp16, y = value_states_63_cast_fp16)[name = tensor("attn_output_61_cast_fp16")]; + tensor var_1066_perm_0 = const()[name = tensor("op_1066_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1, 1024, 1152])]; + tensor var_1066_cast_fp16 = transpose(perm = var_1066_perm_0, x = attn_output_61_cast_fp16)[name = tensor("transpose_186")]; + tensor input_217_cast_fp16 = reshape(shape = var_1068, x = var_1066_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor vpm_encoder_layers_15_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477822528)))]; + tensor vpm_encoder_layers_15_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480476800)))]; + tensor linear_93_cast_fp16 = linear(bias = vpm_encoder_layers_15_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_15_self_attn_out_proj_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_93_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_15_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480479168)))]; + tensor vpm_encoder_layers_15_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480481536)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = vpm_encoder_layers_15_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_15_layer_norm2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor vpm_encoder_layers_15_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480483904)))]; + tensor vpm_encoder_layers_15_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490400384)))]; + tensor linear_94_cast_fp16 = linear(bias = vpm_encoder_layers_15_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_15_mlp_fc1_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_225_mode_0 = const()[name = tensor("input_225_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_225_cast_fp16 = gelu(mode = input_225_mode_0, x = linear_94_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor vpm_encoder_layers_15_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490409088)))]; + tensor vpm_encoder_layers_15_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_15_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500325568)))]; + tensor linear_95_cast_fp16 = linear(bias = vpm_encoder_layers_15_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_15_mlp_fc2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_227_cast_fp16 = add(x = input_219_cast_fp16, y = linear_95_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor hidden_states_97_axes_0 = const()[name = tensor("hidden_states_97_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_16_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500327936)))]; + tensor vpm_encoder_layers_16_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500330304)))]; + tensor hidden_states_97_cast_fp16 = layer_norm(axes = hidden_states_97_axes_0, beta = vpm_encoder_layers_16_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_16_layer_norm1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("hidden_states_97_cast_fp16")]; + tensor vpm_encoder_layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500332672)))]; + tensor vpm_encoder_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(502986944)))]; + tensor linear_96_cast_fp16 = linear(bias = vpm_encoder_layers_16_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_16_self_attn_q_proj_weight_to_fp16, x = hidden_states_97_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor vpm_encoder_layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(502989312)))]; + tensor vpm_encoder_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505643584)))]; + tensor linear_97_cast_fp16 = linear(bias = vpm_encoder_layers_16_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_16_self_attn_k_proj_weight_to_fp16, x = hidden_states_97_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor vpm_encoder_layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505645952)))]; + tensor vpm_encoder_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508300224)))]; + tensor linear_98_cast_fp16 = linear(bias = vpm_encoder_layers_16_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_16_self_attn_v_proj_weight_to_fp16, x = hidden_states_97_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1111 = const()[name = tensor("op_1111"), val = tensor([1, 1024, 16, 72])]; + tensor var_1112_cast_fp16 = reshape(shape = var_1111, x = linear_96_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([1, 1024, 16, 72])]; + tensor var_1115_cast_fp16 = reshape(shape = var_1114, x = linear_97_cast_fp16)[name = tensor("op_1115_cast_fp16")]; + tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([1, 1024, 16, 72])]; + tensor var_1118_cast_fp16 = reshape(shape = var_1117, x = linear_98_cast_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor value_states_67_perm_0 = const()[name = tensor("value_states_67_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1121_transpose_x_0 = const()[name = tensor("op_1121_transpose_x_0"), val = tensor(false)]; + tensor var_1121_transpose_y_0 = const()[name = tensor("op_1121_transpose_y_0"), val = tensor(false)]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_1115_cast_fp16)[name = tensor("transpose_183")]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = var_1112_cast_fp16)[name = tensor("transpose_184")]; + tensor var_1121_cast_fp16 = matmul(transpose_x = var_1121_transpose_x_0, transpose_y = var_1121_transpose_y_0, x = transpose_115, y = transpose_116)[name = tensor("op_1121_cast_fp16")]; + tensor var_1122_to_fp16 = const()[name = tensor("op_1122_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_65_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1122_to_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor var_1124_cast_fp16 = softmax(axis = var_18, x = attn_weights_65_cast_fp16)[name = tensor("op_1124_cast_fp16")]; + tensor attn_output_65_transpose_x_0 = const()[name = tensor("attn_output_65_transpose_x_0"), val = tensor(false)]; + tensor attn_output_65_transpose_y_0 = const()[name = tensor("attn_output_65_transpose_y_0"), val = tensor(false)]; + tensor value_states_67_cast_fp16 = transpose(perm = value_states_67_perm_0, x = var_1118_cast_fp16)[name = tensor("transpose_185")]; + tensor attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = var_1124_cast_fp16, y = value_states_67_cast_fp16)[name = tensor("attn_output_65_cast_fp16")]; + tensor var_1128_perm_0 = const()[name = tensor("op_1128_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([1, 1024, 1152])]; + tensor var_1128_cast_fp16 = transpose(perm = var_1128_perm_0, x = attn_output_65_cast_fp16)[name = tensor("transpose_182")]; + tensor input_231_cast_fp16 = reshape(shape = var_1130, x = var_1128_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor vpm_encoder_layers_16_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508302592)))]; + tensor vpm_encoder_layers_16_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510956864)))]; + tensor linear_99_cast_fp16 = linear(bias = vpm_encoder_layers_16_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_16_self_attn_out_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor input_233_cast_fp16 = add(x = input_227_cast_fp16, y = linear_99_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor input_235_axes_0 = const()[name = tensor("input_235_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_16_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510959232)))]; + tensor vpm_encoder_layers_16_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510961600)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = vpm_encoder_layers_16_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_16_layer_norm2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor vpm_encoder_layers_16_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510963968)))]; + tensor vpm_encoder_layers_16_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520880448)))]; + tensor linear_100_cast_fp16 = linear(bias = vpm_encoder_layers_16_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_16_mlp_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_100_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor vpm_encoder_layers_16_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520889152)))]; + tensor vpm_encoder_layers_16_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_16_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530805632)))]; + tensor linear_101_cast_fp16 = linear(bias = vpm_encoder_layers_16_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_16_mlp_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_233_cast_fp16, y = linear_101_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor hidden_states_103_axes_0 = const()[name = tensor("hidden_states_103_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_17_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530808000)))]; + tensor vpm_encoder_layers_17_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530810368)))]; + tensor hidden_states_103_cast_fp16 = layer_norm(axes = hidden_states_103_axes_0, beta = vpm_encoder_layers_17_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_17_layer_norm1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("hidden_states_103_cast_fp16")]; + tensor vpm_encoder_layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530812736)))]; + tensor vpm_encoder_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533467008)))]; + tensor linear_102_cast_fp16 = linear(bias = vpm_encoder_layers_17_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_17_self_attn_q_proj_weight_to_fp16, x = hidden_states_103_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor vpm_encoder_layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533469376)))]; + tensor vpm_encoder_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536123648)))]; + tensor linear_103_cast_fp16 = linear(bias = vpm_encoder_layers_17_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_17_self_attn_k_proj_weight_to_fp16, x = hidden_states_103_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor vpm_encoder_layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536126016)))]; + tensor vpm_encoder_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538780288)))]; + tensor linear_104_cast_fp16 = linear(bias = vpm_encoder_layers_17_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_17_self_attn_v_proj_weight_to_fp16, x = hidden_states_103_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor var_1173 = const()[name = tensor("op_1173"), val = tensor([1, 1024, 16, 72])]; + tensor var_1174_cast_fp16 = reshape(shape = var_1173, x = linear_102_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([1, 1024, 16, 72])]; + tensor var_1177_cast_fp16 = reshape(shape = var_1176, x = linear_103_cast_fp16)[name = tensor("op_1177_cast_fp16")]; + tensor var_1179 = const()[name = tensor("op_1179"), val = tensor([1, 1024, 16, 72])]; + tensor var_1180_cast_fp16 = reshape(shape = var_1179, x = linear_104_cast_fp16)[name = tensor("op_1180_cast_fp16")]; + tensor value_states_71_perm_0 = const()[name = tensor("value_states_71_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_1177_cast_fp16)[name = tensor("transpose_179")]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = var_1174_cast_fp16)[name = tensor("transpose_180")]; + tensor var_1183_cast_fp16 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = transpose_117, y = transpose_118)[name = tensor("op_1183_cast_fp16")]; + tensor var_1184_to_fp16 = const()[name = tensor("op_1184_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_69_cast_fp16 = mul(x = var_1183_cast_fp16, y = var_1184_to_fp16)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_1186_cast_fp16 = softmax(axis = var_18, x = attn_weights_69_cast_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor attn_output_69_transpose_x_0 = const()[name = tensor("attn_output_69_transpose_x_0"), val = tensor(false)]; + tensor attn_output_69_transpose_y_0 = const()[name = tensor("attn_output_69_transpose_y_0"), val = tensor(false)]; + tensor value_states_71_cast_fp16 = transpose(perm = value_states_71_perm_0, x = var_1180_cast_fp16)[name = tensor("transpose_181")]; + tensor attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = var_1186_cast_fp16, y = value_states_71_cast_fp16)[name = tensor("attn_output_69_cast_fp16")]; + tensor var_1190_perm_0 = const()[name = tensor("op_1190_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 1024, 1152])]; + tensor var_1190_cast_fp16 = transpose(perm = var_1190_perm_0, x = attn_output_69_cast_fp16)[name = tensor("transpose_178")]; + tensor input_245_cast_fp16 = reshape(shape = var_1192, x = var_1190_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor vpm_encoder_layers_17_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538782656)))]; + tensor vpm_encoder_layers_17_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541436928)))]; + tensor linear_105_cast_fp16 = linear(bias = vpm_encoder_layers_17_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_17_self_attn_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_105_cast_fp16")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_105_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_17_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541439296)))]; + tensor vpm_encoder_layers_17_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541441664)))]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = vpm_encoder_layers_17_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_17_layer_norm2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor vpm_encoder_layers_17_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541444032)))]; + tensor vpm_encoder_layers_17_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551360512)))]; + tensor linear_106_cast_fp16 = linear(bias = vpm_encoder_layers_17_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_17_mlp_fc1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_106_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor vpm_encoder_layers_17_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551369216)))]; + tensor vpm_encoder_layers_17_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_17_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561285696)))]; + tensor linear_107_cast_fp16 = linear(bias = vpm_encoder_layers_17_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_17_mlp_fc2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_107_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_247_cast_fp16, y = linear_107_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor hidden_states_109_axes_0 = const()[name = tensor("hidden_states_109_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_18_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561288064)))]; + tensor vpm_encoder_layers_18_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561290432)))]; + tensor hidden_states_109_cast_fp16 = layer_norm(axes = hidden_states_109_axes_0, beta = vpm_encoder_layers_18_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_18_layer_norm1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("hidden_states_109_cast_fp16")]; + tensor vpm_encoder_layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561292800)))]; + tensor vpm_encoder_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563947072)))]; + tensor linear_108_cast_fp16 = linear(bias = vpm_encoder_layers_18_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_18_self_attn_q_proj_weight_to_fp16, x = hidden_states_109_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor vpm_encoder_layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563949440)))]; + tensor vpm_encoder_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566603712)))]; + tensor linear_109_cast_fp16 = linear(bias = vpm_encoder_layers_18_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_18_self_attn_k_proj_weight_to_fp16, x = hidden_states_109_cast_fp16)[name = tensor("linear_109_cast_fp16")]; + tensor vpm_encoder_layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566606080)))]; + tensor vpm_encoder_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569260352)))]; + tensor linear_110_cast_fp16 = linear(bias = vpm_encoder_layers_18_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_18_self_attn_v_proj_weight_to_fp16, x = hidden_states_109_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1235 = const()[name = tensor("op_1235"), val = tensor([1, 1024, 16, 72])]; + tensor var_1236_cast_fp16 = reshape(shape = var_1235, x = linear_108_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([1, 1024, 16, 72])]; + tensor var_1239_cast_fp16 = reshape(shape = var_1238, x = linear_109_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([1, 1024, 16, 72])]; + tensor var_1242_cast_fp16 = reshape(shape = var_1241, x = linear_110_cast_fp16)[name = tensor("op_1242_cast_fp16")]; + tensor value_states_75_perm_0 = const()[name = tensor("value_states_75_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1245_transpose_x_0 = const()[name = tensor("op_1245_transpose_x_0"), val = tensor(false)]; + tensor var_1245_transpose_y_0 = const()[name = tensor("op_1245_transpose_y_0"), val = tensor(false)]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_120_perm_0 = const()[name = tensor("transpose_120_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_1239_cast_fp16)[name = tensor("transpose_175")]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = var_1236_cast_fp16)[name = tensor("transpose_176")]; + tensor var_1245_cast_fp16 = matmul(transpose_x = var_1245_transpose_x_0, transpose_y = var_1245_transpose_y_0, x = transpose_119, y = transpose_120)[name = tensor("op_1245_cast_fp16")]; + tensor var_1246_to_fp16 = const()[name = tensor("op_1246_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_73_cast_fp16 = mul(x = var_1245_cast_fp16, y = var_1246_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_1248_cast_fp16 = softmax(axis = var_18, x = attn_weights_73_cast_fp16)[name = tensor("op_1248_cast_fp16")]; + tensor attn_output_73_transpose_x_0 = const()[name = tensor("attn_output_73_transpose_x_0"), val = tensor(false)]; + tensor attn_output_73_transpose_y_0 = const()[name = tensor("attn_output_73_transpose_y_0"), val = tensor(false)]; + tensor value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_1242_cast_fp16)[name = tensor("transpose_177")]; + tensor attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = var_1248_cast_fp16, y = value_states_75_cast_fp16)[name = tensor("attn_output_73_cast_fp16")]; + tensor var_1252_perm_0 = const()[name = tensor("op_1252_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 1024, 1152])]; + tensor var_1252_cast_fp16 = transpose(perm = var_1252_perm_0, x = attn_output_73_cast_fp16)[name = tensor("transpose_174")]; + tensor input_259_cast_fp16 = reshape(shape = var_1254, x = var_1252_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor vpm_encoder_layers_18_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569262720)))]; + tensor vpm_encoder_layers_18_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571916992)))]; + tensor linear_111_cast_fp16 = linear(bias = vpm_encoder_layers_18_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_18_self_attn_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_111_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_18_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571919360)))]; + tensor vpm_encoder_layers_18_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571921728)))]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = vpm_encoder_layers_18_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_18_layer_norm2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor vpm_encoder_layers_18_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571924096)))]; + tensor vpm_encoder_layers_18_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581840576)))]; + tensor linear_112_cast_fp16 = linear(bias = vpm_encoder_layers_18_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_18_mlp_fc1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor input_267_mode_0 = const()[name = tensor("input_267_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_267_cast_fp16 = gelu(mode = input_267_mode_0, x = linear_112_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor vpm_encoder_layers_18_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581849280)))]; + tensor vpm_encoder_layers_18_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_18_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591765760)))]; + tensor linear_113_cast_fp16 = linear(bias = vpm_encoder_layers_18_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_18_mlp_fc2_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = input_261_cast_fp16, y = linear_113_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor hidden_states_115_axes_0 = const()[name = tensor("hidden_states_115_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_19_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591768128)))]; + tensor vpm_encoder_layers_19_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591770496)))]; + tensor hidden_states_115_cast_fp16 = layer_norm(axes = hidden_states_115_axes_0, beta = vpm_encoder_layers_19_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_19_layer_norm1_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("hidden_states_115_cast_fp16")]; + tensor vpm_encoder_layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591772864)))]; + tensor vpm_encoder_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594427136)))]; + tensor linear_114_cast_fp16 = linear(bias = vpm_encoder_layers_19_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_19_self_attn_q_proj_weight_to_fp16, x = hidden_states_115_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor vpm_encoder_layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594429504)))]; + tensor vpm_encoder_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597083776)))]; + tensor linear_115_cast_fp16 = linear(bias = vpm_encoder_layers_19_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_19_self_attn_k_proj_weight_to_fp16, x = hidden_states_115_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor vpm_encoder_layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597086144)))]; + tensor vpm_encoder_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(599740416)))]; + tensor linear_116_cast_fp16 = linear(bias = vpm_encoder_layers_19_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_19_self_attn_v_proj_weight_to_fp16, x = hidden_states_115_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1297 = const()[name = tensor("op_1297"), val = tensor([1, 1024, 16, 72])]; + tensor var_1298_cast_fp16 = reshape(shape = var_1297, x = linear_114_cast_fp16)[name = tensor("op_1298_cast_fp16")]; + tensor var_1300 = const()[name = tensor("op_1300"), val = tensor([1, 1024, 16, 72])]; + tensor var_1301_cast_fp16 = reshape(shape = var_1300, x = linear_115_cast_fp16)[name = tensor("op_1301_cast_fp16")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 1024, 16, 72])]; + tensor var_1304_cast_fp16 = reshape(shape = var_1303, x = linear_116_cast_fp16)[name = tensor("op_1304_cast_fp16")]; + tensor value_states_79_perm_0 = const()[name = tensor("value_states_79_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1307_transpose_x_0 = const()[name = tensor("op_1307_transpose_x_0"), val = tensor(false)]; + tensor var_1307_transpose_y_0 = const()[name = tensor("op_1307_transpose_y_0"), val = tensor(false)]; + tensor transpose_121_perm_0 = const()[name = tensor("transpose_121_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_122_perm_0 = const()[name = tensor("transpose_122_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_1301_cast_fp16)[name = tensor("transpose_171")]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = var_1298_cast_fp16)[name = tensor("transpose_172")]; + tensor var_1307_cast_fp16 = matmul(transpose_x = var_1307_transpose_x_0, transpose_y = var_1307_transpose_y_0, x = transpose_121, y = transpose_122)[name = tensor("op_1307_cast_fp16")]; + tensor var_1308_to_fp16 = const()[name = tensor("op_1308_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_77_cast_fp16 = mul(x = var_1307_cast_fp16, y = var_1308_to_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor var_1310_cast_fp16 = softmax(axis = var_18, x = attn_weights_77_cast_fp16)[name = tensor("op_1310_cast_fp16")]; + tensor attn_output_77_transpose_x_0 = const()[name = tensor("attn_output_77_transpose_x_0"), val = tensor(false)]; + tensor attn_output_77_transpose_y_0 = const()[name = tensor("attn_output_77_transpose_y_0"), val = tensor(false)]; + tensor value_states_79_cast_fp16 = transpose(perm = value_states_79_perm_0, x = var_1304_cast_fp16)[name = tensor("transpose_173")]; + tensor attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = var_1310_cast_fp16, y = value_states_79_cast_fp16)[name = tensor("attn_output_77_cast_fp16")]; + tensor var_1314_perm_0 = const()[name = tensor("op_1314_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([1, 1024, 1152])]; + tensor var_1314_cast_fp16 = transpose(perm = var_1314_perm_0, x = attn_output_77_cast_fp16)[name = tensor("transpose_170")]; + tensor input_273_cast_fp16 = reshape(shape = var_1316, x = var_1314_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor vpm_encoder_layers_19_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(599742784)))]; + tensor vpm_encoder_layers_19_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(602397056)))]; + tensor linear_117_cast_fp16 = linear(bias = vpm_encoder_layers_19_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_19_self_attn_out_proj_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_275_cast_fp16 = add(x = input_269_cast_fp16, y = linear_117_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor input_277_axes_0 = const()[name = tensor("input_277_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_19_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(602399424)))]; + tensor vpm_encoder_layers_19_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(602401792)))]; + tensor input_277_cast_fp16 = layer_norm(axes = input_277_axes_0, beta = vpm_encoder_layers_19_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_19_layer_norm2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor vpm_encoder_layers_19_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(602404160)))]; + tensor vpm_encoder_layers_19_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(612320640)))]; + tensor linear_118_cast_fp16 = linear(bias = vpm_encoder_layers_19_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_19_mlp_fc1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor input_281_mode_0 = const()[name = tensor("input_281_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_281_cast_fp16 = gelu(mode = input_281_mode_0, x = linear_118_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor vpm_encoder_layers_19_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(612329344)))]; + tensor vpm_encoder_layers_19_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_19_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622245824)))]; + tensor linear_119_cast_fp16 = linear(bias = vpm_encoder_layers_19_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_19_mlp_fc2_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_275_cast_fp16, y = linear_119_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor hidden_states_121_axes_0 = const()[name = tensor("hidden_states_121_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_20_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622248192)))]; + tensor vpm_encoder_layers_20_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622250560)))]; + tensor hidden_states_121_cast_fp16 = layer_norm(axes = hidden_states_121_axes_0, beta = vpm_encoder_layers_20_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_20_layer_norm1_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("hidden_states_121_cast_fp16")]; + tensor vpm_encoder_layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622252928)))]; + tensor vpm_encoder_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(624907200)))]; + tensor linear_120_cast_fp16 = linear(bias = vpm_encoder_layers_20_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_20_self_attn_q_proj_weight_to_fp16, x = hidden_states_121_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor vpm_encoder_layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(624909568)))]; + tensor vpm_encoder_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627563840)))]; + tensor linear_121_cast_fp16 = linear(bias = vpm_encoder_layers_20_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_20_self_attn_k_proj_weight_to_fp16, x = hidden_states_121_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor vpm_encoder_layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627566208)))]; + tensor vpm_encoder_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630220480)))]; + tensor linear_122_cast_fp16 = linear(bias = vpm_encoder_layers_20_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_20_self_attn_v_proj_weight_to_fp16, x = hidden_states_121_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([1, 1024, 16, 72])]; + tensor var_1360_cast_fp16 = reshape(shape = var_1359, x = linear_120_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([1, 1024, 16, 72])]; + tensor var_1363_cast_fp16 = reshape(shape = var_1362, x = linear_121_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1365 = const()[name = tensor("op_1365"), val = tensor([1, 1024, 16, 72])]; + tensor var_1366_cast_fp16 = reshape(shape = var_1365, x = linear_122_cast_fp16)[name = tensor("op_1366_cast_fp16")]; + tensor value_states_83_perm_0 = const()[name = tensor("value_states_83_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1369_transpose_x_0 = const()[name = tensor("op_1369_transpose_x_0"), val = tensor(false)]; + tensor var_1369_transpose_y_0 = const()[name = tensor("op_1369_transpose_y_0"), val = tensor(false)]; + tensor transpose_123_perm_0 = const()[name = tensor("transpose_123_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_124_perm_0 = const()[name = tensor("transpose_124_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_1363_cast_fp16)[name = tensor("transpose_167")]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = var_1360_cast_fp16)[name = tensor("transpose_168")]; + tensor var_1369_cast_fp16 = matmul(transpose_x = var_1369_transpose_x_0, transpose_y = var_1369_transpose_y_0, x = transpose_123, y = transpose_124)[name = tensor("op_1369_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_81_cast_fp16 = mul(x = var_1369_cast_fp16, y = var_1370_to_fp16)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_1372_cast_fp16 = softmax(axis = var_18, x = attn_weights_81_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor attn_output_81_transpose_x_0 = const()[name = tensor("attn_output_81_transpose_x_0"), val = tensor(false)]; + tensor attn_output_81_transpose_y_0 = const()[name = tensor("attn_output_81_transpose_y_0"), val = tensor(false)]; + tensor value_states_83_cast_fp16 = transpose(perm = value_states_83_perm_0, x = var_1366_cast_fp16)[name = tensor("transpose_169")]; + tensor attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = var_1372_cast_fp16, y = value_states_83_cast_fp16)[name = tensor("attn_output_81_cast_fp16")]; + tensor var_1376_perm_0 = const()[name = tensor("op_1376_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([1, 1024, 1152])]; + tensor var_1376_cast_fp16 = transpose(perm = var_1376_perm_0, x = attn_output_81_cast_fp16)[name = tensor("transpose_166")]; + tensor input_287_cast_fp16 = reshape(shape = var_1378, x = var_1376_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor vpm_encoder_layers_20_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630222848)))]; + tensor vpm_encoder_layers_20_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632877120)))]; + tensor linear_123_cast_fp16 = linear(bias = vpm_encoder_layers_20_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_20_self_attn_out_proj_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = input_283_cast_fp16, y = linear_123_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor input_291_axes_0 = const()[name = tensor("input_291_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_20_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632879488)))]; + tensor vpm_encoder_layers_20_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632881856)))]; + tensor input_291_cast_fp16 = layer_norm(axes = input_291_axes_0, beta = vpm_encoder_layers_20_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_20_layer_norm2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor vpm_encoder_layers_20_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632884224)))]; + tensor vpm_encoder_layers_20_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(642800704)))]; + tensor linear_124_cast_fp16 = linear(bias = vpm_encoder_layers_20_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_20_mlp_fc1_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor input_295_mode_0 = const()[name = tensor("input_295_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_295_cast_fp16 = gelu(mode = input_295_mode_0, x = linear_124_cast_fp16)[name = tensor("input_295_cast_fp16")]; + tensor vpm_encoder_layers_20_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(642809408)))]; + tensor vpm_encoder_layers_20_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_20_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652725888)))]; + tensor linear_125_cast_fp16 = linear(bias = vpm_encoder_layers_20_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_20_mlp_fc2_weight_to_fp16, x = input_295_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_289_cast_fp16, y = linear_125_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor hidden_states_127_axes_0 = const()[name = tensor("hidden_states_127_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_21_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652728256)))]; + tensor vpm_encoder_layers_21_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652730624)))]; + tensor hidden_states_127_cast_fp16 = layer_norm(axes = hidden_states_127_axes_0, beta = vpm_encoder_layers_21_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_21_layer_norm1_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("hidden_states_127_cast_fp16")]; + tensor vpm_encoder_layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652732992)))]; + tensor vpm_encoder_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(655387264)))]; + tensor linear_126_cast_fp16 = linear(bias = vpm_encoder_layers_21_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_21_self_attn_q_proj_weight_to_fp16, x = hidden_states_127_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor vpm_encoder_layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(655389632)))]; + tensor vpm_encoder_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658043904)))]; + tensor linear_127_cast_fp16 = linear(bias = vpm_encoder_layers_21_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_21_self_attn_k_proj_weight_to_fp16, x = hidden_states_127_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor vpm_encoder_layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658046272)))]; + tensor vpm_encoder_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(660700544)))]; + tensor linear_128_cast_fp16 = linear(bias = vpm_encoder_layers_21_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_21_self_attn_v_proj_weight_to_fp16, x = hidden_states_127_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([1, 1024, 16, 72])]; + tensor var_1422_cast_fp16 = reshape(shape = var_1421, x = linear_126_cast_fp16)[name = tensor("op_1422_cast_fp16")]; + tensor var_1424 = const()[name = tensor("op_1424"), val = tensor([1, 1024, 16, 72])]; + tensor var_1425_cast_fp16 = reshape(shape = var_1424, x = linear_127_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1427 = const()[name = tensor("op_1427"), val = tensor([1, 1024, 16, 72])]; + tensor var_1428_cast_fp16 = reshape(shape = var_1427, x = linear_128_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor value_states_87_perm_0 = const()[name = tensor("value_states_87_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1431_transpose_x_0 = const()[name = tensor("op_1431_transpose_x_0"), val = tensor(false)]; + tensor var_1431_transpose_y_0 = const()[name = tensor("op_1431_transpose_y_0"), val = tensor(false)]; + tensor transpose_125_perm_0 = const()[name = tensor("transpose_125_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_126_perm_0 = const()[name = tensor("transpose_126_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_1425_cast_fp16)[name = tensor("transpose_163")]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = var_1422_cast_fp16)[name = tensor("transpose_164")]; + tensor var_1431_cast_fp16 = matmul(transpose_x = var_1431_transpose_x_0, transpose_y = var_1431_transpose_y_0, x = transpose_125, y = transpose_126)[name = tensor("op_1431_cast_fp16")]; + tensor var_1432_to_fp16 = const()[name = tensor("op_1432_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_85_cast_fp16 = mul(x = var_1431_cast_fp16, y = var_1432_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_1434_cast_fp16 = softmax(axis = var_18, x = attn_weights_85_cast_fp16)[name = tensor("op_1434_cast_fp16")]; + tensor attn_output_85_transpose_x_0 = const()[name = tensor("attn_output_85_transpose_x_0"), val = tensor(false)]; + tensor attn_output_85_transpose_y_0 = const()[name = tensor("attn_output_85_transpose_y_0"), val = tensor(false)]; + tensor value_states_87_cast_fp16 = transpose(perm = value_states_87_perm_0, x = var_1428_cast_fp16)[name = tensor("transpose_165")]; + tensor attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = var_1434_cast_fp16, y = value_states_87_cast_fp16)[name = tensor("attn_output_85_cast_fp16")]; + tensor var_1438_perm_0 = const()[name = tensor("op_1438_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1440 = const()[name = tensor("op_1440"), val = tensor([1, 1024, 1152])]; + tensor var_1438_cast_fp16 = transpose(perm = var_1438_perm_0, x = attn_output_85_cast_fp16)[name = tensor("transpose_162")]; + tensor input_301_cast_fp16 = reshape(shape = var_1440, x = var_1438_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor vpm_encoder_layers_21_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(660702912)))]; + tensor vpm_encoder_layers_21_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(663357184)))]; + tensor linear_129_cast_fp16 = linear(bias = vpm_encoder_layers_21_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_21_self_attn_out_proj_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor input_303_cast_fp16 = add(x = input_297_cast_fp16, y = linear_129_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor input_305_axes_0 = const()[name = tensor("input_305_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_21_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(663359552)))]; + tensor vpm_encoder_layers_21_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(663361920)))]; + tensor input_305_cast_fp16 = layer_norm(axes = input_305_axes_0, beta = vpm_encoder_layers_21_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_21_layer_norm2_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("input_305_cast_fp16")]; + tensor vpm_encoder_layers_21_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(663364288)))]; + tensor vpm_encoder_layers_21_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(673280768)))]; + tensor linear_130_cast_fp16 = linear(bias = vpm_encoder_layers_21_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_21_mlp_fc1_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_309_mode_0 = const()[name = tensor("input_309_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_309_cast_fp16 = gelu(mode = input_309_mode_0, x = linear_130_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor vpm_encoder_layers_21_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(673289472)))]; + tensor vpm_encoder_layers_21_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_21_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683205952)))]; + tensor linear_131_cast_fp16 = linear(bias = vpm_encoder_layers_21_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_21_mlp_fc2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_131_cast_fp16")]; + tensor input_311_cast_fp16 = add(x = input_303_cast_fp16, y = linear_131_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor hidden_states_133_axes_0 = const()[name = tensor("hidden_states_133_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_22_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683208320)))]; + tensor vpm_encoder_layers_22_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683210688)))]; + tensor hidden_states_133_cast_fp16 = layer_norm(axes = hidden_states_133_axes_0, beta = vpm_encoder_layers_22_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_22_layer_norm1_weight_to_fp16, x = input_311_cast_fp16)[name = tensor("hidden_states_133_cast_fp16")]; + tensor vpm_encoder_layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683213056)))]; + tensor vpm_encoder_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(685867328)))]; + tensor linear_132_cast_fp16 = linear(bias = vpm_encoder_layers_22_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_22_self_attn_q_proj_weight_to_fp16, x = hidden_states_133_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor vpm_encoder_layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(685869696)))]; + tensor vpm_encoder_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(688523968)))]; + tensor linear_133_cast_fp16 = linear(bias = vpm_encoder_layers_22_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_22_self_attn_k_proj_weight_to_fp16, x = hidden_states_133_cast_fp16)[name = tensor("linear_133_cast_fp16")]; + tensor vpm_encoder_layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(688526336)))]; + tensor vpm_encoder_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(691180608)))]; + tensor linear_134_cast_fp16 = linear(bias = vpm_encoder_layers_22_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_22_self_attn_v_proj_weight_to_fp16, x = hidden_states_133_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor var_1483 = const()[name = tensor("op_1483"), val = tensor([1, 1024, 16, 72])]; + tensor var_1484_cast_fp16 = reshape(shape = var_1483, x = linear_132_cast_fp16)[name = tensor("op_1484_cast_fp16")]; + tensor var_1486 = const()[name = tensor("op_1486"), val = tensor([1, 1024, 16, 72])]; + tensor var_1487_cast_fp16 = reshape(shape = var_1486, x = linear_133_cast_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor var_1489 = const()[name = tensor("op_1489"), val = tensor([1, 1024, 16, 72])]; + tensor var_1490_cast_fp16 = reshape(shape = var_1489, x = linear_134_cast_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor value_states_91_perm_0 = const()[name = tensor("value_states_91_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1493_transpose_x_0 = const()[name = tensor("op_1493_transpose_x_0"), val = tensor(false)]; + tensor var_1493_transpose_y_0 = const()[name = tensor("op_1493_transpose_y_0"), val = tensor(false)]; + tensor transpose_127_perm_0 = const()[name = tensor("transpose_127_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_128_perm_0 = const()[name = tensor("transpose_128_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_1487_cast_fp16)[name = tensor("transpose_159")]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = var_1484_cast_fp16)[name = tensor("transpose_160")]; + tensor var_1493_cast_fp16 = matmul(transpose_x = var_1493_transpose_x_0, transpose_y = var_1493_transpose_y_0, x = transpose_127, y = transpose_128)[name = tensor("op_1493_cast_fp16")]; + tensor var_1494_to_fp16 = const()[name = tensor("op_1494_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_89_cast_fp16 = mul(x = var_1493_cast_fp16, y = var_1494_to_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor var_1496_cast_fp16 = softmax(axis = var_18, x = attn_weights_89_cast_fp16)[name = tensor("op_1496_cast_fp16")]; + tensor attn_output_89_transpose_x_0 = const()[name = tensor("attn_output_89_transpose_x_0"), val = tensor(false)]; + tensor attn_output_89_transpose_y_0 = const()[name = tensor("attn_output_89_transpose_y_0"), val = tensor(false)]; + tensor value_states_91_cast_fp16 = transpose(perm = value_states_91_perm_0, x = var_1490_cast_fp16)[name = tensor("transpose_161")]; + tensor attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = var_1496_cast_fp16, y = value_states_91_cast_fp16)[name = tensor("attn_output_89_cast_fp16")]; + tensor var_1500_perm_0 = const()[name = tensor("op_1500_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1502 = const()[name = tensor("op_1502"), val = tensor([1, 1024, 1152])]; + tensor var_1500_cast_fp16 = transpose(perm = var_1500_perm_0, x = attn_output_89_cast_fp16)[name = tensor("transpose_158")]; + tensor input_315_cast_fp16 = reshape(shape = var_1502, x = var_1500_cast_fp16)[name = tensor("input_315_cast_fp16")]; + tensor vpm_encoder_layers_22_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(691182976)))]; + tensor vpm_encoder_layers_22_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693837248)))]; + tensor linear_135_cast_fp16 = linear(bias = vpm_encoder_layers_22_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_22_self_attn_out_proj_weight_to_fp16, x = input_315_cast_fp16)[name = tensor("linear_135_cast_fp16")]; + tensor input_317_cast_fp16 = add(x = input_311_cast_fp16, y = linear_135_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor input_319_axes_0 = const()[name = tensor("input_319_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_22_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693839616)))]; + tensor vpm_encoder_layers_22_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693841984)))]; + tensor input_319_cast_fp16 = layer_norm(axes = input_319_axes_0, beta = vpm_encoder_layers_22_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_22_layer_norm2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("input_319_cast_fp16")]; + tensor vpm_encoder_layers_22_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693844352)))]; + tensor vpm_encoder_layers_22_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(703760832)))]; + tensor linear_136_cast_fp16 = linear(bias = vpm_encoder_layers_22_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_22_mlp_fc1_weight_to_fp16, x = input_319_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor input_323_mode_0 = const()[name = tensor("input_323_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_323_cast_fp16 = gelu(mode = input_323_mode_0, x = linear_136_cast_fp16)[name = tensor("input_323_cast_fp16")]; + tensor vpm_encoder_layers_22_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(703769536)))]; + tensor vpm_encoder_layers_22_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_22_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(713686016)))]; + tensor linear_137_cast_fp16 = linear(bias = vpm_encoder_layers_22_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_22_mlp_fc2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_325_cast_fp16 = add(x = input_317_cast_fp16, y = linear_137_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor hidden_states_139_axes_0 = const()[name = tensor("hidden_states_139_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_23_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(713688384)))]; + tensor vpm_encoder_layers_23_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(713690752)))]; + tensor hidden_states_139_cast_fp16 = layer_norm(axes = hidden_states_139_axes_0, beta = vpm_encoder_layers_23_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_23_layer_norm1_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("hidden_states_139_cast_fp16")]; + tensor vpm_encoder_layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(713693120)))]; + tensor vpm_encoder_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(716347392)))]; + tensor linear_138_cast_fp16 = linear(bias = vpm_encoder_layers_23_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_23_self_attn_q_proj_weight_to_fp16, x = hidden_states_139_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor vpm_encoder_layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(716349760)))]; + tensor vpm_encoder_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(719004032)))]; + tensor linear_139_cast_fp16 = linear(bias = vpm_encoder_layers_23_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_23_self_attn_k_proj_weight_to_fp16, x = hidden_states_139_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor vpm_encoder_layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(719006400)))]; + tensor vpm_encoder_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(721660672)))]; + tensor linear_140_cast_fp16 = linear(bias = vpm_encoder_layers_23_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_23_self_attn_v_proj_weight_to_fp16, x = hidden_states_139_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor var_1545 = const()[name = tensor("op_1545"), val = tensor([1, 1024, 16, 72])]; + tensor var_1546_cast_fp16 = reshape(shape = var_1545, x = linear_138_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1548 = const()[name = tensor("op_1548"), val = tensor([1, 1024, 16, 72])]; + tensor var_1549_cast_fp16 = reshape(shape = var_1548, x = linear_139_cast_fp16)[name = tensor("op_1549_cast_fp16")]; + tensor var_1551 = const()[name = tensor("op_1551"), val = tensor([1, 1024, 16, 72])]; + tensor var_1552_cast_fp16 = reshape(shape = var_1551, x = linear_140_cast_fp16)[name = tensor("op_1552_cast_fp16")]; + tensor value_states_95_perm_0 = const()[name = tensor("value_states_95_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1555_transpose_x_0 = const()[name = tensor("op_1555_transpose_x_0"), val = tensor(false)]; + tensor var_1555_transpose_y_0 = const()[name = tensor("op_1555_transpose_y_0"), val = tensor(false)]; + tensor transpose_129_perm_0 = const()[name = tensor("transpose_129_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_130_perm_0 = const()[name = tensor("transpose_130_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_1549_cast_fp16)[name = tensor("transpose_155")]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = var_1546_cast_fp16)[name = tensor("transpose_156")]; + tensor var_1555_cast_fp16 = matmul(transpose_x = var_1555_transpose_x_0, transpose_y = var_1555_transpose_y_0, x = transpose_129, y = transpose_130)[name = tensor("op_1555_cast_fp16")]; + tensor var_1556_to_fp16 = const()[name = tensor("op_1556_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_93_cast_fp16 = mul(x = var_1555_cast_fp16, y = var_1556_to_fp16)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_1558_cast_fp16 = softmax(axis = var_18, x = attn_weights_93_cast_fp16)[name = tensor("op_1558_cast_fp16")]; + tensor attn_output_93_transpose_x_0 = const()[name = tensor("attn_output_93_transpose_x_0"), val = tensor(false)]; + tensor attn_output_93_transpose_y_0 = const()[name = tensor("attn_output_93_transpose_y_0"), val = tensor(false)]; + tensor value_states_95_cast_fp16 = transpose(perm = value_states_95_perm_0, x = var_1552_cast_fp16)[name = tensor("transpose_157")]; + tensor attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = var_1558_cast_fp16, y = value_states_95_cast_fp16)[name = tensor("attn_output_93_cast_fp16")]; + tensor var_1562_perm_0 = const()[name = tensor("op_1562_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1564 = const()[name = tensor("op_1564"), val = tensor([1, 1024, 1152])]; + tensor var_1562_cast_fp16 = transpose(perm = var_1562_perm_0, x = attn_output_93_cast_fp16)[name = tensor("transpose_154")]; + tensor input_329_cast_fp16 = reshape(shape = var_1564, x = var_1562_cast_fp16)[name = tensor("input_329_cast_fp16")]; + tensor vpm_encoder_layers_23_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(721663040)))]; + tensor vpm_encoder_layers_23_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724317312)))]; + tensor linear_141_cast_fp16 = linear(bias = vpm_encoder_layers_23_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_23_self_attn_out_proj_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor input_331_cast_fp16 = add(x = input_325_cast_fp16, y = linear_141_cast_fp16)[name = tensor("input_331_cast_fp16")]; + tensor input_333_axes_0 = const()[name = tensor("input_333_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_23_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724319680)))]; + tensor vpm_encoder_layers_23_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724322048)))]; + tensor input_333_cast_fp16 = layer_norm(axes = input_333_axes_0, beta = vpm_encoder_layers_23_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_23_layer_norm2_weight_to_fp16, x = input_331_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor vpm_encoder_layers_23_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724324416)))]; + tensor vpm_encoder_layers_23_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(734240896)))]; + tensor linear_142_cast_fp16 = linear(bias = vpm_encoder_layers_23_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_23_mlp_fc1_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor input_337_mode_0 = const()[name = tensor("input_337_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_337_cast_fp16 = gelu(mode = input_337_mode_0, x = linear_142_cast_fp16)[name = tensor("input_337_cast_fp16")]; + tensor vpm_encoder_layers_23_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(734249600)))]; + tensor vpm_encoder_layers_23_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_23_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(744166080)))]; + tensor linear_143_cast_fp16 = linear(bias = vpm_encoder_layers_23_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_23_mlp_fc2_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_331_cast_fp16, y = linear_143_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor hidden_states_145_axes_0 = const()[name = tensor("hidden_states_145_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_24_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(744168448)))]; + tensor vpm_encoder_layers_24_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(744170816)))]; + tensor hidden_states_145_cast_fp16 = layer_norm(axes = hidden_states_145_axes_0, beta = vpm_encoder_layers_24_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_24_layer_norm1_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("hidden_states_145_cast_fp16")]; + tensor vpm_encoder_layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(744173184)))]; + tensor vpm_encoder_layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(746827456)))]; + tensor linear_144_cast_fp16 = linear(bias = vpm_encoder_layers_24_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_24_self_attn_q_proj_weight_to_fp16, x = hidden_states_145_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor vpm_encoder_layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(746829824)))]; + tensor vpm_encoder_layers_24_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749484096)))]; + tensor linear_145_cast_fp16 = linear(bias = vpm_encoder_layers_24_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_24_self_attn_k_proj_weight_to_fp16, x = hidden_states_145_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor vpm_encoder_layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749486464)))]; + tensor vpm_encoder_layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(752140736)))]; + tensor linear_146_cast_fp16 = linear(bias = vpm_encoder_layers_24_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_24_self_attn_v_proj_weight_to_fp16, x = hidden_states_145_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor var_1607 = const()[name = tensor("op_1607"), val = tensor([1, 1024, 16, 72])]; + tensor var_1608_cast_fp16 = reshape(shape = var_1607, x = linear_144_cast_fp16)[name = tensor("op_1608_cast_fp16")]; + tensor var_1610 = const()[name = tensor("op_1610"), val = tensor([1, 1024, 16, 72])]; + tensor var_1611_cast_fp16 = reshape(shape = var_1610, x = linear_145_cast_fp16)[name = tensor("op_1611_cast_fp16")]; + tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([1, 1024, 16, 72])]; + tensor var_1614_cast_fp16 = reshape(shape = var_1613, x = linear_146_cast_fp16)[name = tensor("op_1614_cast_fp16")]; + tensor value_states_99_perm_0 = const()[name = tensor("value_states_99_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1617_transpose_x_0 = const()[name = tensor("op_1617_transpose_x_0"), val = tensor(false)]; + tensor var_1617_transpose_y_0 = const()[name = tensor("op_1617_transpose_y_0"), val = tensor(false)]; + tensor transpose_131_perm_0 = const()[name = tensor("transpose_131_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_132_perm_0 = const()[name = tensor("transpose_132_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_1611_cast_fp16)[name = tensor("transpose_151")]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = var_1608_cast_fp16)[name = tensor("transpose_152")]; + tensor var_1617_cast_fp16 = matmul(transpose_x = var_1617_transpose_x_0, transpose_y = var_1617_transpose_y_0, x = transpose_131, y = transpose_132)[name = tensor("op_1617_cast_fp16")]; + tensor var_1618_to_fp16 = const()[name = tensor("op_1618_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_97_cast_fp16 = mul(x = var_1617_cast_fp16, y = var_1618_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_1620_cast_fp16 = softmax(axis = var_18, x = attn_weights_97_cast_fp16)[name = tensor("op_1620_cast_fp16")]; + tensor attn_output_97_transpose_x_0 = const()[name = tensor("attn_output_97_transpose_x_0"), val = tensor(false)]; + tensor attn_output_97_transpose_y_0 = const()[name = tensor("attn_output_97_transpose_y_0"), val = tensor(false)]; + tensor value_states_99_cast_fp16 = transpose(perm = value_states_99_perm_0, x = var_1614_cast_fp16)[name = tensor("transpose_153")]; + tensor attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_0, transpose_y = attn_output_97_transpose_y_0, x = var_1620_cast_fp16, y = value_states_99_cast_fp16)[name = tensor("attn_output_97_cast_fp16")]; + tensor var_1624_perm_0 = const()[name = tensor("op_1624_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1626 = const()[name = tensor("op_1626"), val = tensor([1, 1024, 1152])]; + tensor var_1624_cast_fp16 = transpose(perm = var_1624_perm_0, x = attn_output_97_cast_fp16)[name = tensor("transpose_150")]; + tensor input_343_cast_fp16 = reshape(shape = var_1626, x = var_1624_cast_fp16)[name = tensor("input_343_cast_fp16")]; + tensor vpm_encoder_layers_24_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(752143104)))]; + tensor vpm_encoder_layers_24_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(754797376)))]; + tensor linear_147_cast_fp16 = linear(bias = vpm_encoder_layers_24_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_24_self_attn_out_proj_weight_to_fp16, x = input_343_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor input_345_cast_fp16 = add(x = input_339_cast_fp16, y = linear_147_cast_fp16)[name = tensor("input_345_cast_fp16")]; + tensor input_347_axes_0 = const()[name = tensor("input_347_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_24_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(754799744)))]; + tensor vpm_encoder_layers_24_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(754802112)))]; + tensor input_347_cast_fp16 = layer_norm(axes = input_347_axes_0, beta = vpm_encoder_layers_24_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_24_layer_norm2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("input_347_cast_fp16")]; + tensor vpm_encoder_layers_24_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(754804480)))]; + tensor vpm_encoder_layers_24_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(764720960)))]; + tensor linear_148_cast_fp16 = linear(bias = vpm_encoder_layers_24_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_24_mlp_fc1_weight_to_fp16, x = input_347_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor input_351_mode_0 = const()[name = tensor("input_351_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_351_cast_fp16 = gelu(mode = input_351_mode_0, x = linear_148_cast_fp16)[name = tensor("input_351_cast_fp16")]; + tensor vpm_encoder_layers_24_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(764729664)))]; + tensor vpm_encoder_layers_24_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_24_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(774646144)))]; + tensor linear_149_cast_fp16 = linear(bias = vpm_encoder_layers_24_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_24_mlp_fc2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = input_345_cast_fp16, y = linear_149_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor hidden_states_151_axes_0 = const()[name = tensor("hidden_states_151_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_25_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(774648512)))]; + tensor vpm_encoder_layers_25_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(774650880)))]; + tensor hidden_states_151_cast_fp16 = layer_norm(axes = hidden_states_151_axes_0, beta = vpm_encoder_layers_25_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_25_layer_norm1_weight_to_fp16, x = input_353_cast_fp16)[name = tensor("hidden_states_151_cast_fp16")]; + tensor vpm_encoder_layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(774653248)))]; + tensor vpm_encoder_layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(777307520)))]; + tensor linear_150_cast_fp16 = linear(bias = vpm_encoder_layers_25_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_25_self_attn_q_proj_weight_to_fp16, x = hidden_states_151_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor vpm_encoder_layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(777309888)))]; + tensor vpm_encoder_layers_25_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(779964160)))]; + tensor linear_151_cast_fp16 = linear(bias = vpm_encoder_layers_25_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_25_self_attn_k_proj_weight_to_fp16, x = hidden_states_151_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor vpm_encoder_layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(779966528)))]; + tensor vpm_encoder_layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(782620800)))]; + tensor linear_152_cast_fp16 = linear(bias = vpm_encoder_layers_25_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_25_self_attn_v_proj_weight_to_fp16, x = hidden_states_151_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor var_1669 = const()[name = tensor("op_1669"), val = tensor([1, 1024, 16, 72])]; + tensor var_1670_cast_fp16 = reshape(shape = var_1669, x = linear_150_cast_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 1024, 16, 72])]; + tensor var_1673_cast_fp16 = reshape(shape = var_1672, x = linear_151_cast_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1675 = const()[name = tensor("op_1675"), val = tensor([1, 1024, 16, 72])]; + tensor var_1676_cast_fp16 = reshape(shape = var_1675, x = linear_152_cast_fp16)[name = tensor("op_1676_cast_fp16")]; + tensor value_states_103_perm_0 = const()[name = tensor("value_states_103_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1679_transpose_x_0 = const()[name = tensor("op_1679_transpose_x_0"), val = tensor(false)]; + tensor var_1679_transpose_y_0 = const()[name = tensor("op_1679_transpose_y_0"), val = tensor(false)]; + tensor transpose_133_perm_0 = const()[name = tensor("transpose_133_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_134_perm_0 = const()[name = tensor("transpose_134_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_1673_cast_fp16)[name = tensor("transpose_147")]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = var_1670_cast_fp16)[name = tensor("transpose_148")]; + tensor var_1679_cast_fp16 = matmul(transpose_x = var_1679_transpose_x_0, transpose_y = var_1679_transpose_y_0, x = transpose_133, y = transpose_134)[name = tensor("op_1679_cast_fp16")]; + tensor var_1680_to_fp16 = const()[name = tensor("op_1680_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_101_cast_fp16 = mul(x = var_1679_cast_fp16, y = var_1680_to_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor var_1682_cast_fp16 = softmax(axis = var_18, x = attn_weights_101_cast_fp16)[name = tensor("op_1682_cast_fp16")]; + tensor attn_output_101_transpose_x_0 = const()[name = tensor("attn_output_101_transpose_x_0"), val = tensor(false)]; + tensor attn_output_101_transpose_y_0 = const()[name = tensor("attn_output_101_transpose_y_0"), val = tensor(false)]; + tensor value_states_103_cast_fp16 = transpose(perm = value_states_103_perm_0, x = var_1676_cast_fp16)[name = tensor("transpose_149")]; + tensor attn_output_101_cast_fp16 = matmul(transpose_x = attn_output_101_transpose_x_0, transpose_y = attn_output_101_transpose_y_0, x = var_1682_cast_fp16, y = value_states_103_cast_fp16)[name = tensor("attn_output_101_cast_fp16")]; + tensor var_1686_perm_0 = const()[name = tensor("op_1686_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1688 = const()[name = tensor("op_1688"), val = tensor([1, 1024, 1152])]; + tensor var_1686_cast_fp16 = transpose(perm = var_1686_perm_0, x = attn_output_101_cast_fp16)[name = tensor("transpose_146")]; + tensor input_357_cast_fp16 = reshape(shape = var_1688, x = var_1686_cast_fp16)[name = tensor("input_357_cast_fp16")]; + tensor vpm_encoder_layers_25_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(782623168)))]; + tensor vpm_encoder_layers_25_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(785277440)))]; + tensor linear_153_cast_fp16 = linear(bias = vpm_encoder_layers_25_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_25_self_attn_out_proj_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_353_cast_fp16, y = linear_153_cast_fp16)[name = tensor("input_359_cast_fp16")]; + tensor input_361_axes_0 = const()[name = tensor("input_361_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_25_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(785279808)))]; + tensor vpm_encoder_layers_25_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(785282176)))]; + tensor input_361_cast_fp16 = layer_norm(axes = input_361_axes_0, beta = vpm_encoder_layers_25_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_25_layer_norm2_weight_to_fp16, x = input_359_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor vpm_encoder_layers_25_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(785284544)))]; + tensor vpm_encoder_layers_25_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795201024)))]; + tensor linear_154_cast_fp16 = linear(bias = vpm_encoder_layers_25_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_25_mlp_fc1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor input_365_mode_0 = const()[name = tensor("input_365_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_365_cast_fp16 = gelu(mode = input_365_mode_0, x = linear_154_cast_fp16)[name = tensor("input_365_cast_fp16")]; + tensor vpm_encoder_layers_25_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795209728)))]; + tensor vpm_encoder_layers_25_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_25_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805126208)))]; + tensor linear_155_cast_fp16 = linear(bias = vpm_encoder_layers_25_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_25_mlp_fc2_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor input_367_cast_fp16 = add(x = input_359_cast_fp16, y = linear_155_cast_fp16)[name = tensor("input_367_cast_fp16")]; + tensor hidden_states_157_axes_0 = const()[name = tensor("hidden_states_157_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_26_layer_norm1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_layer_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805128576)))]; + tensor vpm_encoder_layers_26_layer_norm1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_layer_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805130944)))]; + tensor hidden_states_157_cast_fp16 = layer_norm(axes = hidden_states_157_axes_0, beta = vpm_encoder_layers_26_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_26_layer_norm1_weight_to_fp16, x = input_367_cast_fp16)[name = tensor("hidden_states_157_cast_fp16")]; + tensor vpm_encoder_layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805133312)))]; + tensor vpm_encoder_layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807787584)))]; + tensor linear_156_cast_fp16 = linear(bias = vpm_encoder_layers_26_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_26_self_attn_q_proj_weight_to_fp16, x = hidden_states_157_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor vpm_encoder_layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807789952)))]; + tensor vpm_encoder_layers_26_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(810444224)))]; + tensor linear_157_cast_fp16 = linear(bias = vpm_encoder_layers_26_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_26_self_attn_k_proj_weight_to_fp16, x = hidden_states_157_cast_fp16)[name = tensor("linear_157_cast_fp16")]; + tensor vpm_encoder_layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(810446592)))]; + tensor vpm_encoder_layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(813100864)))]; + tensor linear_158_cast_fp16 = linear(bias = vpm_encoder_layers_26_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_26_self_attn_v_proj_weight_to_fp16, x = hidden_states_157_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor var_1731 = const()[name = tensor("op_1731"), val = tensor([1, 1024, 16, 72])]; + tensor var_1732_cast_fp16 = reshape(shape = var_1731, x = linear_156_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor var_1734 = const()[name = tensor("op_1734"), val = tensor([1, 1024, 16, 72])]; + tensor var_1735_cast_fp16 = reshape(shape = var_1734, x = linear_157_cast_fp16)[name = tensor("op_1735_cast_fp16")]; + tensor var_1737 = const()[name = tensor("op_1737"), val = tensor([1, 1024, 16, 72])]; + tensor var_1738_cast_fp16 = reshape(shape = var_1737, x = linear_158_cast_fp16)[name = tensor("op_1738_cast_fp16")]; + tensor value_states_perm_0 = const()[name = tensor("value_states_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1741_transpose_x_0 = const()[name = tensor("op_1741_transpose_x_0"), val = tensor(false)]; + tensor var_1741_transpose_y_0 = const()[name = tensor("op_1741_transpose_y_0"), val = tensor(false)]; + tensor transpose_135_perm_0 = const()[name = tensor("transpose_135_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_136_perm_0 = const()[name = tensor("transpose_136_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_1735_cast_fp16)[name = tensor("transpose_143")]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = var_1732_cast_fp16)[name = tensor("transpose_144")]; + tensor var_1741_cast_fp16 = matmul(transpose_x = var_1741_transpose_x_0, transpose_y = var_1741_transpose_y_0, x = transpose_135, y = transpose_136)[name = tensor("op_1741_cast_fp16")]; + tensor var_1742_to_fp16 = const()[name = tensor("op_1742_to_fp16"), val = tensor(0x1.e2cp-4)]; + tensor attn_weights_105_cast_fp16 = mul(x = var_1741_cast_fp16, y = var_1742_to_fp16)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_1744_cast_fp16 = softmax(axis = var_18, x = attn_weights_105_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor attn_output_105_transpose_x_0 = const()[name = tensor("attn_output_105_transpose_x_0"), val = tensor(false)]; + tensor attn_output_105_transpose_y_0 = const()[name = tensor("attn_output_105_transpose_y_0"), val = tensor(false)]; + tensor value_states_cast_fp16 = transpose(perm = value_states_perm_0, x = var_1738_cast_fp16)[name = tensor("transpose_145")]; + tensor attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_0, transpose_y = attn_output_105_transpose_y_0, x = var_1744_cast_fp16, y = value_states_cast_fp16)[name = tensor("attn_output_105_cast_fp16")]; + tensor var_1748_perm_0 = const()[name = tensor("op_1748_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1750 = const()[name = tensor("op_1750"), val = tensor([1, 1024, 1152])]; + tensor var_1748_cast_fp16 = transpose(perm = var_1748_perm_0, x = attn_output_105_cast_fp16)[name = tensor("transpose_142")]; + tensor input_371_cast_fp16 = reshape(shape = var_1750, x = var_1748_cast_fp16)[name = tensor("input_371_cast_fp16")]; + tensor vpm_encoder_layers_26_self_attn_out_proj_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(813103232)))]; + tensor vpm_encoder_layers_26_self_attn_out_proj_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(815757504)))]; + tensor linear_159_cast_fp16 = linear(bias = vpm_encoder_layers_26_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_26_self_attn_out_proj_weight_to_fp16, x = input_371_cast_fp16)[name = tensor("linear_159_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = input_367_cast_fp16, y = linear_159_cast_fp16)[name = tensor("input_373_cast_fp16")]; + tensor input_375_axes_0 = const()[name = tensor("input_375_axes_0"), val = tensor([-1])]; + tensor vpm_encoder_layers_26_layer_norm2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_layer_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(815759872)))]; + tensor vpm_encoder_layers_26_layer_norm2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_layer_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(815762240)))]; + tensor input_375_cast_fp16 = layer_norm(axes = input_375_axes_0, beta = vpm_encoder_layers_26_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_26_layer_norm2_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("input_375_cast_fp16")]; + tensor vpm_encoder_layers_26_mlp_fc1_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(815764608)))]; + tensor vpm_encoder_layers_26_mlp_fc1_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(825681088)))]; + tensor linear_160_cast_fp16 = linear(bias = vpm_encoder_layers_26_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_26_mlp_fc1_weight_to_fp16, x = input_375_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor input_379_mode_0 = const()[name = tensor("input_379_mode_0"), val = tensor("TANH_APPROXIMATION")]; + tensor input_379_cast_fp16 = gelu(mode = input_379_mode_0, x = linear_160_cast_fp16)[name = tensor("input_379_cast_fp16")]; + tensor vpm_encoder_layers_26_mlp_fc2_weight_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(825689792)))]; + tensor vpm_encoder_layers_26_mlp_fc2_bias_to_fp16 = const()[name = tensor("vpm_encoder_layers_26_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(835606272)))]; + tensor linear_161_cast_fp16 = linear(bias = vpm_encoder_layers_26_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_26_mlp_fc2_weight_to_fp16, x = input_379_cast_fp16)[name = tensor("linear_161_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = input_373_cast_fp16, y = linear_161_cast_fp16)[name = tensor("input_381_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor vpm_post_layernorm_weight_to_fp16 = const()[name = tensor("vpm_post_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(835608640)))]; + tensor vpm_post_layernorm_bias_to_fp16 = const()[name = tensor("vpm_post_layernorm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(835611008)))]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = vpm_post_layernorm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_post_layernorm_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor var_1777 = const()[name = tensor("op_1777"), val = tensor(-1)]; + tensor resampler_kv_proj_weight_to_fp16 = const()[name = tensor("resampler_kv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(835613376)))]; + tensor linear_162_bias_0_to_fp16 = const()[name = tensor("linear_162_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841511680)))]; + tensor linear_162_cast_fp16 = linear(bias = linear_162_bias_0_to_fp16, weight = resampler_kv_proj_weight_to_fp16, x = x_1_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_1798_axes_0 = const()[name = tensor("op_1798_axes_0"), val = tensor([-1])]; + tensor resampler_ln_kv_weight_to_fp16 = const()[name = tensor("resampler_ln_kv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841516864)))]; + tensor resampler_ln_kv_bias_to_fp16 = const()[name = tensor("resampler_ln_kv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841522048)))]; + tensor var_1781_to_fp16 = const()[name = tensor("op_1781_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1798_cast_fp16 = layer_norm(axes = var_1798_axes_0, beta = resampler_ln_kv_bias_to_fp16, epsilon = var_1781_to_fp16, gamma = resampler_ln_kv_weight_to_fp16, x = linear_162_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1799 = const()[name = tensor("op_1799"), val = tensor([1, 0, 2])]; + tensor pos_embed_to_fp16_dtype_0 = const()[name = tensor("pos_embed_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor pos_embed_to_fp16 = cast(dtype = pos_embed_to_fp16_dtype_0, x = pos_embed)[name = tensor("cast_149")]; + tensor x_3_cast_fp16 = transpose(perm = var_1799, x = var_1798_cast_fp16)[name = tensor("transpose_141")]; + tensor key_cast_fp16 = add(x = x_3_cast_fp16, y = pos_embed_to_fp16)[name = tensor("key_cast_fp16")]; + tensor var_1822_1_to_fp16 = const()[name = tensor("op_1822_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841527232)))]; + tensor var_1825_1_to_fp16 = const()[name = tensor("op_1825_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854634496)))]; + tensor linear_164_cast_fp16 = linear(bias = var_1825_1_to_fp16, weight = var_1822_1_to_fp16, x = key_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_1822_2_to_fp16 = const()[name = tensor("op_1822_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854639680)))]; + tensor var_1825_2_to_fp16 = const()[name = tensor("op_1825_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867746944)))]; + tensor linear_165_cast_fp16 = linear(bias = var_1825_2_to_fp16, weight = var_1822_2_to_fp16, x = x_3_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_1841 = const()[name = tensor("op_1841"), val = tensor([1024, 20, 128])]; + tensor var_1842_cast_fp16 = reshape(shape = var_1841, x = linear_164_cast_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([1, 2, 0])]; + tensor var_1848 = const()[name = tensor("op_1848"), val = tensor([1024, 20, 128])]; + tensor var_1849_cast_fp16 = reshape(shape = var_1848, x = linear_165_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor v_perm_0 = const()[name = tensor("v_perm_0"), val = tensor([1, 0, 2])]; + tensor input_387_transpose_x_0 = const()[name = tensor("input_387_transpose_x_0"), val = tensor(false)]; + tensor input_387_transpose_y_0 = const()[name = tensor("input_387_transpose_y_0"), val = tensor(false)]; + tensor q_scaled_to_fp16 = const()[name = tensor("q_scaled_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867752128)))]; + tensor transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = var_1842_cast_fp16)[name = tensor("transpose_140")]; + tensor input_387_cast_fp16 = matmul(transpose_x = input_387_transpose_x_0, transpose_y = input_387_transpose_y_0, x = q_scaled_to_fp16, y = transpose_28_cast_fp16)[name = tensor("input_387_cast_fp16")]; + tensor attn_output_weights_1_cast_fp16 = softmax(axis = var_1777, x = input_387_cast_fp16)[name = tensor("attn_output_weights_1_cast_fp16")]; + tensor attn_output_109_transpose_x_0 = const()[name = tensor("attn_output_109_transpose_x_0"), val = tensor(false)]; + tensor attn_output_109_transpose_y_0 = const()[name = tensor("attn_output_109_transpose_y_0"), val = tensor(false)]; + tensor v_cast_fp16 = transpose(perm = v_perm_0, x = var_1849_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_109_cast_fp16 = matmul(transpose_x = attn_output_109_transpose_x_0, transpose_y = attn_output_109_transpose_y_0, x = attn_output_weights_1_cast_fp16, y = v_cast_fp16)[name = tensor("attn_output_109_cast_fp16")]; + tensor var_1857_perm_0 = const()[name = tensor("op_1857_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1861 = const()[name = tensor("op_1861"), val = tensor([64, 2560])]; + tensor var_1857_cast_fp16 = transpose(perm = var_1857_perm_0, x = attn_output_109_cast_fp16)[name = tensor("transpose_138")]; + tensor input_389_cast_fp16 = reshape(shape = var_1861, x = var_1857_cast_fp16)[name = tensor("input_389_cast_fp16")]; + tensor resampler_attn_out_proj_weight_to_fp16 = const()[name = tensor("resampler_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868079872)))]; + tensor resampler_attn_out_proj_bias_to_fp16 = const()[name = tensor("resampler_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881187136)))]; + tensor linear_166_cast_fp16 = linear(bias = resampler_attn_out_proj_bias_to_fp16, weight = resampler_attn_out_proj_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor var_1867 = const()[name = tensor("op_1867"), val = tensor([64, 1, 2560])]; + tensor out_cast_fp16 = reshape(shape = var_1867, x = linear_166_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor var_1869 = const()[name = tensor("op_1869"), val = tensor([1, 0, 2])]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor resampler_ln_post_weight_to_fp16 = const()[name = tensor("resampler_ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881192320)))]; + tensor resampler_ln_post_bias_to_fp16 = const()[name = tensor("resampler_ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881197504)))]; + tensor input_cast_fp16 = transpose(perm = var_1869, x = out_cast_fp16)[name = tensor("transpose_137")]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = resampler_ln_post_bias_to_fp16, epsilon = var_1781_to_fp16, gamma = resampler_ln_post_weight_to_fp16, x = input_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor transpose_0_to_fp16 = const()[name = tensor("transpose_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881202688)))]; + tensor var_1875_cast_fp16 = linear(bias = linear_162_bias_0_to_fp16, weight = transpose_0_to_fp16, x = x_cast_fp16)[name = tensor("op_1875_cast_fp16")]; + tensor var_1875_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1875_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_1875_cast_fp16_to_fp32_dtype_0, x = var_1875_cast_fp16)[name = tensor("cast_148")]; + } -> (output); +} \ No newline at end of file