diff --git "a/OmniASR_CTC_300M_fp16.mlmodelc/model.mil" "b/OmniASR_CTC_300M_fp16.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/OmniASR_CTC_300M_fp16.mlmodelc/model.mil" @@ -0,0 +1,1510 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor audio) [FlexibleShapeInformation = tuple, dict, tensor>>, tuple, dict, dict, tensor>>>>((("DefaultShapes", {{"audio", [1, 160000]}}), ("EnumeratedShapes", {{"audio_1_1_1_1_112000_", {{"audio", [1, 112000]}}}, {"audio_1_1_1_1_128000_", {{"audio", [1, 128000]}}}, {"audio_1_1_1_1_144000_", {{"audio", [1, 144000]}}}, {"audio_1_1_1_1_160000_", {{"audio", [1, 160000]}}}, {"audio_1_1_1_1_16000_", {{"audio", [1, 16000]}}}, {"audio_1_1_1_1_176000_", {{"audio", [1, 176000]}}}, {"audio_1_1_1_1_192000_", {{"audio", [1, 192000]}}}, {"audio_1_1_1_1_208000_", {{"audio", [1, 208000]}}}, {"audio_1_1_1_1_224000_", {{"audio", [1, 224000]}}}, {"audio_1_1_1_1_240000_", {{"audio", [1, 240000]}}}, {"audio_1_1_1_1_256000_", {{"audio", [1, 256000]}}}, {"audio_1_1_1_1_272000_", {{"audio", [1, 272000]}}}, {"audio_1_1_1_1_288000_", {{"audio", [1, 288000]}}}, {"audio_1_1_1_1_304000_", {{"audio", [1, 304000]}}}, {"audio_1_1_1_1_320000_", {{"audio", [1, 320000]}}}, {"audio_1_1_1_1_32000_", {{"audio", [1, 32000]}}}, {"audio_1_1_1_1_336000_", {{"audio", [1, 336000]}}}, {"audio_1_1_1_1_352000_", {{"audio", [1, 352000]}}}, {"audio_1_1_1_1_368000_", {{"audio", [1, 368000]}}}, {"audio_1_1_1_1_384000_", {{"audio", [1, 384000]}}}, {"audio_1_1_1_1_400000_", {{"audio", [1, 400000]}}}, {"audio_1_1_1_1_416000_", {{"audio", [1, 416000]}}}, {"audio_1_1_1_1_432000_", {{"audio", [1, 432000]}}}, {"audio_1_1_1_1_448000_", {{"audio", [1, 448000]}}}, {"audio_1_1_1_1_464000_", {{"audio", [1, 464000]}}}, {"audio_1_1_1_1_480000_", {{"audio", [1, 480000]}}}, {"audio_1_1_1_1_48000_", {{"audio", [1, 48000]}}}, {"audio_1_1_1_1_496000_", {{"audio", [1, 496000]}}}, {"audio_1_1_1_1_512000_", {{"audio", [1, 512000]}}}, {"audio_1_1_1_1_528000_", {{"audio", [1, 528000]}}}, {"audio_1_1_1_1_544000_", {{"audio", [1, 544000]}}}, {"audio_1_1_1_1_560000_", {{"audio", [1, 560000]}}}, {"audio_1_1_1_1_576000_", {{"audio", [1, 576000]}}}, {"audio_1_1_1_1_592000_", {{"audio", [1, 592000]}}}, {"audio_1_1_1_1_608000_", {{"audio", [1, 608000]}}}, {"audio_1_1_1_1_624000_", {{"audio", [1, 624000]}}}, {"audio_1_1_1_1_640000_", {{"audio", [1, 640000]}}}, {"audio_1_1_1_1_64000_", {{"audio", [1, 64000]}}}, {"audio_1_1_1_1_80000_", {{"audio", [1, 80000]}}}, {"audio_1_1_1_1_96000_", {{"audio", [1, 96000]}}}})))] { + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([1])]; + tensor input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = audio)[name = tensor("input_1_cast_fp16")]; + tensor seqs_1_pad_type_0 = const()[name = tensor("seqs_1_pad_type_0"), val = tensor("valid")]; + tensor seqs_1_strides_0 = const()[name = tensor("seqs_1_strides_0"), val = tensor([5])]; + tensor seqs_1_pad_0 = const()[name = tensor("seqs_1_pad_0"), val = tensor([0, 0])]; + tensor seqs_1_dilations_0 = const()[name = tensor("seqs_1_dilations_0"), val = tensor([1])]; + tensor seqs_1_groups_0 = const()[name = tensor("seqs_1_groups_0"), val = tensor(1)]; + tensor model_encoder_frontend_feature_extractor_layers_0_conv_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_0_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor model_encoder_frontend_feature_extractor_layers_0_conv_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_0_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10368)))]; + tensor seqs_1_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_0_conv_bias_to_fp16, dilations = seqs_1_dilations_0, groups = seqs_1_groups_0, pad = seqs_1_pad_0, pad_type = seqs_1_pad_type_0, strides = seqs_1_strides_0, weight = model_encoder_frontend_feature_extractor_layers_0_conv_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("seqs_1_cast_fp16")]; + tensor x_1_perm_0 = const()[name = tensor("x_1_perm_0"), val = tensor([0, 2, 1])]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11456)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12544)))]; + tensor var_18_to_fp16 = const()[name = tensor("op_18_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = seqs_1_cast_fp16)[name = tensor("transpose_254")]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = const_1_to_fp16, epsilon = var_18_to_fp16, gamma = const_0_to_fp16, x = x_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor input_5_perm_0 = const()[name = tensor("input_5_perm_0"), val = tensor([0, 2, 1])]; + tensor input_7_mode_0 = const()[name = tensor("input_7_mode_0"), val = tensor("EXACT")]; + tensor input_5_cast_fp16 = transpose(perm = input_5_perm_0, x = x_3_cast_fp16)[name = tensor("transpose_253")]; + tensor input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor seqs_5_pad_type_0 = const()[name = tensor("seqs_5_pad_type_0"), val = tensor("valid")]; + tensor seqs_5_strides_0 = const()[name = tensor("seqs_5_strides_0"), val = tensor([2])]; + tensor seqs_5_pad_0 = const()[name = tensor("seqs_5_pad_0"), val = tensor([0, 0])]; + tensor seqs_5_dilations_0 = const()[name = tensor("seqs_5_dilations_0"), val = tensor([1])]; + tensor seqs_5_groups_0 = const()[name = tensor("seqs_5_groups_0"), val = tensor(1)]; + tensor model_encoder_frontend_feature_extractor_layers_1_conv_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_1_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13632)))]; + tensor model_encoder_frontend_feature_extractor_layers_1_conv_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_1_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1586560)))]; + tensor seqs_5_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_1_conv_bias_to_fp16, dilations = seqs_5_dilations_0, groups = seqs_5_groups_0, pad = seqs_5_pad_0, pad_type = seqs_5_pad_type_0, strides = seqs_5_strides_0, weight = model_encoder_frontend_feature_extractor_layers_1_conv_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("seqs_5_cast_fp16")]; + tensor x_5_perm_0 = const()[name = tensor("x_5_perm_0"), val = tensor([0, 2, 1])]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1587648)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1588736)))]; + tensor x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = seqs_5_cast_fp16)[name = tensor("transpose_252")]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = const_3_to_fp16, epsilon = var_18_to_fp16, gamma = const_2_to_fp16, x = x_5_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor input_11_perm_0 = const()[name = tensor("input_11_perm_0"), val = tensor([0, 2, 1])]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = transpose(perm = input_11_perm_0, x = x_7_cast_fp16)[name = tensor("transpose_251")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = input_11_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor seqs_9_pad_type_0 = const()[name = tensor("seqs_9_pad_type_0"), val = tensor("valid")]; + tensor seqs_9_strides_0 = const()[name = tensor("seqs_9_strides_0"), val = tensor([2])]; + tensor seqs_9_pad_0 = const()[name = tensor("seqs_9_pad_0"), val = tensor([0, 0])]; + tensor seqs_9_dilations_0 = const()[name = tensor("seqs_9_dilations_0"), val = tensor([1])]; + tensor seqs_9_groups_0 = const()[name = tensor("seqs_9_groups_0"), val = tensor(1)]; + tensor model_encoder_frontend_feature_extractor_layers_2_conv_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_2_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1589824)))]; + tensor model_encoder_frontend_feature_extractor_layers_2_conv_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_2_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3162752)))]; + tensor seqs_9_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_2_conv_bias_to_fp16, dilations = seqs_9_dilations_0, groups = seqs_9_groups_0, pad = seqs_9_pad_0, pad_type = seqs_9_pad_type_0, strides = seqs_9_strides_0, weight = model_encoder_frontend_feature_extractor_layers_2_conv_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("seqs_9_cast_fp16")]; + tensor x_9_perm_0 = const()[name = tensor("x_9_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor const_4_to_fp16 = const()[name = tensor("const_4_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3163840)))]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3164928)))]; + tensor x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = seqs_9_cast_fp16)[name = tensor("transpose_250")]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = const_5_to_fp16, epsilon = var_18_to_fp16, gamma = const_4_to_fp16, x = x_9_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor input_17_perm_0 = const()[name = tensor("input_17_perm_0"), val = tensor([0, 2, 1])]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_17_cast_fp16 = transpose(perm = input_17_perm_0, x = x_11_cast_fp16)[name = tensor("transpose_249")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor seqs_13_pad_type_0 = const()[name = tensor("seqs_13_pad_type_0"), val = tensor("valid")]; + tensor seqs_13_strides_0 = const()[name = tensor("seqs_13_strides_0"), val = tensor([2])]; + tensor seqs_13_pad_0 = const()[name = tensor("seqs_13_pad_0"), val = tensor([0, 0])]; + tensor seqs_13_dilations_0 = const()[name = tensor("seqs_13_dilations_0"), val = tensor([1])]; + tensor seqs_13_groups_0 = const()[name = tensor("seqs_13_groups_0"), val = tensor(1)]; + tensor model_encoder_frontend_feature_extractor_layers_3_conv_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_3_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3166016)))]; + tensor model_encoder_frontend_feature_extractor_layers_3_conv_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_3_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4738944)))]; + tensor seqs_13_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_3_conv_bias_to_fp16, dilations = seqs_13_dilations_0, groups = seqs_13_groups_0, pad = seqs_13_pad_0, pad_type = seqs_13_pad_type_0, strides = seqs_13_strides_0, weight = model_encoder_frontend_feature_extractor_layers_3_conv_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("seqs_13_cast_fp16")]; + tensor x_13_perm_0 = const()[name = tensor("x_13_perm_0"), val = tensor([0, 2, 1])]; + tensor x_15_axes_0 = const()[name = tensor("x_15_axes_0"), val = tensor([-1])]; + tensor const_6_to_fp16 = const()[name = tensor("const_6_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4740032)))]; + tensor const_7_to_fp16 = const()[name = tensor("const_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4741120)))]; + tensor x_13_cast_fp16 = transpose(perm = x_13_perm_0, x = seqs_13_cast_fp16)[name = tensor("transpose_248")]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = const_7_to_fp16, epsilon = var_18_to_fp16, gamma = const_6_to_fp16, x = x_13_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = transpose(perm = input_23_perm_0, x = x_15_cast_fp16)[name = tensor("transpose_247")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = input_23_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor seqs_17_pad_type_0 = const()[name = tensor("seqs_17_pad_type_0"), val = tensor("valid")]; + tensor seqs_17_strides_0 = const()[name = tensor("seqs_17_strides_0"), val = tensor([2])]; + tensor seqs_17_pad_0 = const()[name = tensor("seqs_17_pad_0"), val = tensor([0, 0])]; + tensor seqs_17_dilations_0 = const()[name = tensor("seqs_17_dilations_0"), val = tensor([1])]; + tensor seqs_17_groups_0 = const()[name = tensor("seqs_17_groups_0"), val = tensor(1)]; + tensor model_encoder_frontend_feature_extractor_layers_4_conv_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_4_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4742208)))]; + tensor model_encoder_frontend_feature_extractor_layers_4_conv_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_4_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6315136)))]; + tensor seqs_17_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_4_conv_bias_to_fp16, dilations = seqs_17_dilations_0, groups = seqs_17_groups_0, pad = seqs_17_pad_0, pad_type = seqs_17_pad_type_0, strides = seqs_17_strides_0, weight = model_encoder_frontend_feature_extractor_layers_4_conv_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("seqs_17_cast_fp16")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor const_8_to_fp16 = const()[name = tensor("const_8_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6316224)))]; + tensor const_9_to_fp16 = const()[name = tensor("const_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6317312)))]; + tensor x_17_cast_fp16 = transpose(perm = x_17_perm_0, x = seqs_17_cast_fp16)[name = tensor("transpose_246")]; + tensor x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = const_9_to_fp16, epsilon = var_18_to_fp16, gamma = const_8_to_fp16, x = x_17_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor input_29_perm_0 = const()[name = tensor("input_29_perm_0"), val = tensor([0, 2, 1])]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_29_cast_fp16 = transpose(perm = input_29_perm_0, x = x_19_cast_fp16)[name = tensor("transpose_245")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor seqs_21_pad_type_0 = const()[name = tensor("seqs_21_pad_type_0"), val = tensor("valid")]; + tensor seqs_21_strides_0 = const()[name = tensor("seqs_21_strides_0"), val = tensor([2])]; + tensor seqs_21_pad_0 = const()[name = tensor("seqs_21_pad_0"), val = tensor([0, 0])]; + tensor seqs_21_dilations_0 = const()[name = tensor("seqs_21_dilations_0"), val = tensor([1])]; + tensor seqs_21_groups_0 = const()[name = tensor("seqs_21_groups_0"), val = tensor(1)]; + tensor model_encoder_frontend_feature_extractor_layers_5_conv_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_5_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6318400)))]; + tensor model_encoder_frontend_feature_extractor_layers_5_conv_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_5_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7367040)))]; + tensor seqs_21_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_5_conv_bias_to_fp16, dilations = seqs_21_dilations_0, groups = seqs_21_groups_0, pad = seqs_21_pad_0, pad_type = seqs_21_pad_type_0, strides = seqs_21_strides_0, weight = model_encoder_frontend_feature_extractor_layers_5_conv_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("seqs_21_cast_fp16")]; + tensor x_21_perm_0 = const()[name = tensor("x_21_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor const_10_to_fp16 = const()[name = tensor("const_10_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7368128)))]; + tensor const_11_to_fp16 = const()[name = tensor("const_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7369216)))]; + tensor x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = seqs_21_cast_fp16)[name = tensor("transpose_244")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = const_11_to_fp16, epsilon = var_18_to_fp16, gamma = const_10_to_fp16, x = x_21_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor input_35_perm_0 = const()[name = tensor("input_35_perm_0"), val = tensor([0, 2, 1])]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = transpose(perm = input_35_perm_0, x = x_23_cast_fp16)[name = tensor("transpose_243")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor seqs_25_pad_type_0 = const()[name = tensor("seqs_25_pad_type_0"), val = tensor("valid")]; + tensor seqs_25_strides_0 = const()[name = tensor("seqs_25_strides_0"), val = tensor([2])]; + tensor seqs_25_pad_0 = const()[name = tensor("seqs_25_pad_0"), val = tensor([0, 0])]; + tensor seqs_25_dilations_0 = const()[name = tensor("seqs_25_dilations_0"), val = tensor([1])]; + tensor seqs_25_groups_0 = const()[name = tensor("seqs_25_groups_0"), val = tensor(1)]; + tensor model_encoder_frontend_feature_extractor_layers_6_conv_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_6_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7370304)))]; + tensor model_encoder_frontend_feature_extractor_layers_6_conv_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_feature_extractor_layers_6_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8418944)))]; + tensor seqs_25_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_6_conv_bias_to_fp16, dilations = seqs_25_dilations_0, groups = seqs_25_groups_0, pad = seqs_25_pad_0, pad_type = seqs_25_pad_type_0, strides = seqs_25_strides_0, weight = model_encoder_frontend_feature_extractor_layers_6_conv_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("seqs_25_cast_fp16")]; + tensor x_25_perm_0 = const()[name = tensor("x_25_perm_0"), val = tensor([0, 2, 1])]; + tensor x_27_axes_0 = const()[name = tensor("x_27_axes_0"), val = tensor([-1])]; + tensor const_12_to_fp16 = const()[name = tensor("const_12_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8420032)))]; + tensor const_13_to_fp16 = const()[name = tensor("const_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8421120)))]; + tensor x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = seqs_25_cast_fp16)[name = tensor("transpose_242")]; + tensor x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = const_13_to_fp16, epsilon = var_18_to_fp16, gamma = const_12_to_fp16, x = x_25_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor seqs_29_mode_0 = const()[name = tensor("seqs_29_mode_0"), val = tensor("EXACT")]; + tensor seqs_29_cast_fp16 = gelu(mode = seqs_29_mode_0, x = x_27_cast_fp16)[name = tensor("seqs_29_cast_fp16")]; + tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; + tensor model_encoder_frontend_post_extract_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_post_extract_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8422208)))]; + tensor model_encoder_frontend_post_extract_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_post_extract_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8423296)))]; + tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = model_encoder_frontend_post_extract_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_frontend_post_extract_layer_norm_weight_to_fp16, x = seqs_29_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor model_encoder_frontend_model_dim_proj_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_model_dim_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8424384)))]; + tensor model_encoder_frontend_model_dim_proj_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_model_dim_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9473024)))]; + tensor linear_0_cast_fp16 = linear(bias = model_encoder_frontend_model_dim_proj_bias_to_fp16, weight = model_encoder_frontend_model_dim_proj_weight_to_fp16, x = x_29_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_43_perm_0 = const()[name = tensor("input_43_perm_0"), val = tensor([0, 2, 1])]; + tensor encodings_1_pad_type_0 = const()[name = tensor("encodings_1_pad_type_0"), val = tensor("custom")]; + tensor encodings_1_pad_0 = const()[name = tensor("encodings_1_pad_0"), val = tensor([64, 64])]; + tensor encodings_1_groups_0 = const()[name = tensor("encodings_1_groups_0"), val = tensor(16)]; + tensor encodings_1_strides_0 = const()[name = tensor("encodings_1_strides_0"), val = tensor([1])]; + tensor encodings_1_dilations_0 = const()[name = tensor("encodings_1_dilations_0"), val = tensor([1])]; + tensor model_encoder_frontend_pos_encoder_conv_weight_to_fp16 = const()[name = tensor("model_encoder_frontend_pos_encoder_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9475136)))]; + tensor model_encoder_frontend_pos_encoder_conv_bias_to_fp16 = const()[name = tensor("model_encoder_frontend_pos_encoder_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26252416)))]; + tensor input_43_cast_fp16 = transpose(perm = input_43_perm_0, x = linear_0_cast_fp16)[name = tensor("transpose_241")]; + tensor encodings_1_cast_fp16 = conv(bias = model_encoder_frontend_pos_encoder_conv_bias_to_fp16, dilations = encodings_1_dilations_0, groups = encodings_1_groups_0, pad = encodings_1_pad_0, pad_type = encodings_1_pad_type_0, strides = encodings_1_strides_0, weight = model_encoder_frontend_pos_encoder_conv_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("encodings_1_cast_fp16")]; + tensor input_45_begin_0 = const()[name = tensor("input_45_begin_0"), val = tensor([0, 0, 0])]; + tensor input_45_end_0 = const()[name = tensor("input_45_end_0"), val = tensor([1, 1024, -1])]; + tensor input_45_end_mask_0 = const()[name = tensor("input_45_end_mask_0"), val = tensor([true, true, false])]; + tensor input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = encodings_1_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor encodings_3_mode_0 = const()[name = tensor("encodings_3_mode_0"), val = tensor("EXACT")]; + tensor encodings_3_cast_fp16 = gelu(mode = encodings_3_mode_0, x = input_45_cast_fp16)[name = tensor("encodings_3_cast_fp16")]; + tensor encodings_perm_0 = const()[name = tensor("encodings_perm_0"), val = tensor([0, 2, 1])]; + tensor encodings_cast_fp16 = transpose(perm = encodings_perm_0, x = encodings_3_cast_fp16)[name = tensor("transpose_240")]; + tensor input_47_cast_fp16 = add(x = linear_0_cast_fp16, y = encodings_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor x_31_axes_0 = const()[name = tensor("x_31_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_0_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26254528)))]; + tensor model_encoder_layers_0_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26256640)))]; + tensor x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = model_encoder_layers_0_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_0_self_attn_layer_norm_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor model_encoder_layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26258752)))]; + tensor model_encoder_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28355968)))]; + tensor linear_1_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_q_proj_weight_to_fp16, x = x_31_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor concat_0x = const()[name = tensor("concat_0x"), val = tensor([1, -1, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = concat_0x, x = linear_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor model_encoder_layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28358080)))]; + tensor model_encoder_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30455296)))]; + tensor linear_2_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_k_proj_weight_to_fp16, x = x_31_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor model_encoder_layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30457408)))]; + tensor model_encoder_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554624)))]; + tensor linear_3_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_v_proj_weight_to_fp16, x = x_31_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor concat_1x = const()[name = tensor("concat_1x"), val = tensor([1, -1, 16, 64])]; + tensor k_1_cast_fp16 = reshape(shape = concat_1x, x = linear_2_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor concat_2x = const()[name = tensor("concat_2x"), val = tensor([1, -1, 16, 64])]; + tensor v_1_cast_fp16 = reshape(shape = concat_2x, x = linear_3_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_0_y_0_to_fp16 = const()[name = tensor("mul_0_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_0_cast_fp16 = mul(x = q_1_cast_fp16, y = mul_0_y_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_237")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = mul_0_cast_fp16)[name = tensor("transpose_238")]; + tensor matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("matmul_0_cast_fp16")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = matmul_0_cast_fp16)[name = tensor("softmax_0_cast_fp16")]; + tensor attns_1_transpose_x_0 = const()[name = tensor("attns_1_transpose_x_0"), val = tensor(false)]; + tensor attns_1_transpose_y_0 = const()[name = tensor("attns_1_transpose_y_0"), val = tensor(false)]; + tensor v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = v_1_cast_fp16)[name = tensor("transpose_239")]; + tensor attns_1_cast_fp16 = matmul(transpose_x = attns_1_transpose_x_0, transpose_y = attns_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_3_cast_fp16)[name = tensor("attns_1_cast_fp16")]; + tensor attns_3_perm_0 = const()[name = tensor("attns_3_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_3x = const()[name = tensor("concat_3x"), val = tensor([1, -1, 1024])]; + tensor attns_3_cast_fp16 = transpose(perm = attns_3_perm_0, x = attns_1_cast_fp16)[name = tensor("transpose_236")]; + tensor x_33_cast_fp16 = reshape(shape = concat_3x, x = attns_3_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor model_encoder_layers_0_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556736)))]; + tensor model_encoder_layers_0_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_0_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34653952)))]; + tensor linear_4_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_output_proj_weight_to_fp16, x = x_33_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = linear_4_cast_fp16, y = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_0_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_0_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34656064)))]; + tensor model_encoder_layers_0_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_0_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34658176)))]; + tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = model_encoder_layers_0_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_0_ffn_layer_norm_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor model_encoder_layers_0_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_0_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34660288)))]; + tensor model_encoder_layers_0_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_0_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43048960)))]; + tensor linear_5_cast_fp16 = linear(bias = model_encoder_layers_0_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_0_ffn_inner_proj_weight_to_fp16, x = x_35_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_53_mode_0 = const()[name = tensor("input_53_mode_0"), val = tensor("EXACT")]; + tensor input_53_cast_fp16 = gelu(mode = input_53_mode_0, x = linear_5_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor model_encoder_layers_0_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_0_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43057216)))]; + tensor model_encoder_layers_0_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_0_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51445888)))]; + tensor linear_6_cast_fp16 = linear(bias = model_encoder_layers_0_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_0_ffn_output_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_55_cast_fp16 = add(x = linear_6_cast_fp16, y = input_49_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor x_39_axes_0 = const()[name = tensor("x_39_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_1_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51448000)))]; + tensor model_encoder_layers_1_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51450112)))]; + tensor x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = model_encoder_layers_1_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_1_self_attn_layer_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor model_encoder_layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51452224)))]; + tensor model_encoder_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53549440)))]; + tensor linear_7_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_q_proj_weight_to_fp16, x = x_39_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor concat_4x = const()[name = tensor("concat_4x"), val = tensor([1, -1, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = concat_4x, x = linear_7_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor model_encoder_layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53551552)))]; + tensor model_encoder_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55648768)))]; + tensor linear_8_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_k_proj_weight_to_fp16, x = x_39_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor model_encoder_layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55650880)))]; + tensor model_encoder_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57748096)))]; + tensor linear_9_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_v_proj_weight_to_fp16, x = x_39_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor concat_5x = const()[name = tensor("concat_5x"), val = tensor([1, -1, 16, 64])]; + tensor k_5_cast_fp16 = reshape(shape = concat_5x, x = linear_8_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor concat_6x = const()[name = tensor("concat_6x"), val = tensor([1, -1, 16, 64])]; + tensor v_5_cast_fp16 = reshape(shape = concat_6x, x = linear_9_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_1_y_0_to_fp16 = const()[name = tensor("mul_1_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_1_cast_fp16 = mul(x = q_5_cast_fp16, y = mul_1_y_0_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_233")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = mul_1_cast_fp16)[name = tensor("transpose_234")]; + tensor matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("matmul_1_cast_fp16")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = matmul_1_cast_fp16)[name = tensor("softmax_1_cast_fp16")]; + tensor attns_5_transpose_x_0 = const()[name = tensor("attns_5_transpose_x_0"), val = tensor(false)]; + tensor attns_5_transpose_y_0 = const()[name = tensor("attns_5_transpose_y_0"), val = tensor(false)]; + tensor v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = v_5_cast_fp16)[name = tensor("transpose_235")]; + tensor attns_5_cast_fp16 = matmul(transpose_x = attns_5_transpose_x_0, transpose_y = attns_5_transpose_y_0, x = softmax_1_cast_fp16, y = v_7_cast_fp16)[name = tensor("attns_5_cast_fp16")]; + tensor attns_7_perm_0 = const()[name = tensor("attns_7_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_7x = const()[name = tensor("concat_7x"), val = tensor([1, -1, 1024])]; + tensor attns_7_cast_fp16 = transpose(perm = attns_7_perm_0, x = attns_5_cast_fp16)[name = tensor("transpose_232")]; + tensor x_41_cast_fp16 = reshape(shape = concat_7x, x = attns_7_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor model_encoder_layers_1_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57750208)))]; + tensor model_encoder_layers_1_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_1_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59847424)))]; + tensor linear_10_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_output_proj_weight_to_fp16, x = x_41_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = linear_10_cast_fp16, y = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_1_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_1_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59849536)))]; + tensor model_encoder_layers_1_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_1_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59851648)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = model_encoder_layers_1_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_1_ffn_layer_norm_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor model_encoder_layers_1_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_1_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59853760)))]; + tensor model_encoder_layers_1_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_1_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68242432)))]; + tensor linear_11_cast_fp16 = linear(bias = model_encoder_layers_1_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_1_ffn_inner_proj_weight_to_fp16, x = x_43_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_11_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor model_encoder_layers_1_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_1_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68250688)))]; + tensor model_encoder_layers_1_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_1_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76639360)))]; + tensor linear_12_cast_fp16 = linear(bias = model_encoder_layers_1_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_1_ffn_output_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = linear_12_cast_fp16, y = input_57_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor x_47_axes_0 = const()[name = tensor("x_47_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_2_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76641472)))]; + tensor model_encoder_layers_2_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76643584)))]; + tensor x_47_cast_fp16 = layer_norm(axes = x_47_axes_0, beta = model_encoder_layers_2_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_2_self_attn_layer_norm_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor model_encoder_layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76645696)))]; + tensor model_encoder_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78742912)))]; + tensor linear_13_cast_fp16 = linear(bias = model_encoder_layers_2_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_2_self_attn_q_proj_weight_to_fp16, x = x_47_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor concat_8x = const()[name = tensor("concat_8x"), val = tensor([1, -1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = concat_8x, x = linear_13_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor model_encoder_layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78745024)))]; + tensor model_encoder_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80842240)))]; + tensor linear_14_cast_fp16 = linear(bias = model_encoder_layers_2_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_2_self_attn_k_proj_weight_to_fp16, x = x_47_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor model_encoder_layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80844352)))]; + tensor model_encoder_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82941568)))]; + tensor linear_15_cast_fp16 = linear(bias = model_encoder_layers_2_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_2_self_attn_v_proj_weight_to_fp16, x = x_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor concat_9x = const()[name = tensor("concat_9x"), val = tensor([1, -1, 16, 64])]; + tensor k_9_cast_fp16 = reshape(shape = concat_9x, x = linear_14_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor concat_10x = const()[name = tensor("concat_10x"), val = tensor([1, -1, 16, 64])]; + tensor v_9_cast_fp16 = reshape(shape = concat_10x, x = linear_15_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor v_11_perm_0 = const()[name = tensor("v_11_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_2_y_0_to_fp16 = const()[name = tensor("mul_2_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_2_cast_fp16 = mul(x = q_9_cast_fp16, y = mul_2_y_0_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(true)]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_229")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = mul_2_cast_fp16)[name = tensor("transpose_230")]; + tensor matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("matmul_2_cast_fp16")]; + tensor softmax_2_axis_0 = const()[name = tensor("softmax_2_axis_0"), val = tensor(-1)]; + tensor softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = matmul_2_cast_fp16)[name = tensor("softmax_2_cast_fp16")]; + tensor attns_9_transpose_x_0 = const()[name = tensor("attns_9_transpose_x_0"), val = tensor(false)]; + tensor attns_9_transpose_y_0 = const()[name = tensor("attns_9_transpose_y_0"), val = tensor(false)]; + tensor v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = v_9_cast_fp16)[name = tensor("transpose_231")]; + tensor attns_9_cast_fp16 = matmul(transpose_x = attns_9_transpose_x_0, transpose_y = attns_9_transpose_y_0, x = softmax_2_cast_fp16, y = v_11_cast_fp16)[name = tensor("attns_9_cast_fp16")]; + tensor attns_11_perm_0 = const()[name = tensor("attns_11_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_11x = const()[name = tensor("concat_11x"), val = tensor([1, -1, 1024])]; + tensor attns_11_cast_fp16 = transpose(perm = attns_11_perm_0, x = attns_9_cast_fp16)[name = tensor("transpose_228")]; + tensor x_49_cast_fp16 = reshape(shape = concat_11x, x = attns_11_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor model_encoder_layers_2_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82943680)))]; + tensor model_encoder_layers_2_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_2_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85040896)))]; + tensor linear_16_cast_fp16 = linear(bias = model_encoder_layers_2_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_2_self_attn_output_proj_weight_to_fp16, x = x_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = linear_16_cast_fp16, y = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor x_51_axes_0 = const()[name = tensor("x_51_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_2_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_2_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85043008)))]; + tensor model_encoder_layers_2_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_2_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85045120)))]; + tensor x_51_cast_fp16 = layer_norm(axes = x_51_axes_0, beta = model_encoder_layers_2_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_2_ffn_layer_norm_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor model_encoder_layers_2_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_2_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85047232)))]; + tensor model_encoder_layers_2_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_2_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93435904)))]; + tensor linear_17_cast_fp16 = linear(bias = model_encoder_layers_2_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_2_ffn_inner_proj_weight_to_fp16, x = x_51_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor input_69_mode_0 = const()[name = tensor("input_69_mode_0"), val = tensor("EXACT")]; + tensor input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = linear_17_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor model_encoder_layers_2_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_2_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93444160)))]; + tensor model_encoder_layers_2_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_2_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832832)))]; + tensor linear_18_cast_fp16 = linear(bias = model_encoder_layers_2_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_2_ffn_output_proj_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_71_cast_fp16 = add(x = linear_18_cast_fp16, y = input_65_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor x_55_axes_0 = const()[name = tensor("x_55_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_3_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834944)))]; + tensor model_encoder_layers_3_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101837056)))]; + tensor x_55_cast_fp16 = layer_norm(axes = x_55_axes_0, beta = model_encoder_layers_3_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_3_self_attn_layer_norm_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor model_encoder_layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101839168)))]; + tensor model_encoder_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103936384)))]; + tensor linear_19_cast_fp16 = linear(bias = model_encoder_layers_3_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_3_self_attn_q_proj_weight_to_fp16, x = x_55_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor concat_12x = const()[name = tensor("concat_12x"), val = tensor([1, -1, 16, 64])]; + tensor q_13_cast_fp16 = reshape(shape = concat_12x, x = linear_19_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor model_encoder_layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103938496)))]; + tensor model_encoder_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106035712)))]; + tensor linear_20_cast_fp16 = linear(bias = model_encoder_layers_3_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_3_self_attn_k_proj_weight_to_fp16, x = x_55_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor model_encoder_layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106037824)))]; + tensor model_encoder_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108135040)))]; + tensor linear_21_cast_fp16 = linear(bias = model_encoder_layers_3_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_3_self_attn_v_proj_weight_to_fp16, x = x_55_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor concat_13x = const()[name = tensor("concat_13x"), val = tensor([1, -1, 16, 64])]; + tensor k_13_cast_fp16 = reshape(shape = concat_13x, x = linear_20_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor concat_14x = const()[name = tensor("concat_14x"), val = tensor([1, -1, 16, 64])]; + tensor v_13_cast_fp16 = reshape(shape = concat_14x, x = linear_21_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor v_15_perm_0 = const()[name = tensor("v_15_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_3_y_0_to_fp16 = const()[name = tensor("mul_3_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_3_cast_fp16 = mul(x = q_13_cast_fp16, y = mul_3_y_0_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor matmul_3_transpose_y_0 = const()[name = tensor("matmul_3_transpose_y_0"), val = tensor(true)]; + tensor matmul_3_transpose_x_0 = const()[name = tensor("matmul_3_transpose_x_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_225")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = mul_3_cast_fp16)[name = tensor("transpose_226")]; + tensor matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("matmul_3_cast_fp16")]; + tensor softmax_3_axis_0 = const()[name = tensor("softmax_3_axis_0"), val = tensor(-1)]; + tensor softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = matmul_3_cast_fp16)[name = tensor("softmax_3_cast_fp16")]; + tensor attns_13_transpose_x_0 = const()[name = tensor("attns_13_transpose_x_0"), val = tensor(false)]; + tensor attns_13_transpose_y_0 = const()[name = tensor("attns_13_transpose_y_0"), val = tensor(false)]; + tensor v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = v_13_cast_fp16)[name = tensor("transpose_227")]; + tensor attns_13_cast_fp16 = matmul(transpose_x = attns_13_transpose_x_0, transpose_y = attns_13_transpose_y_0, x = softmax_3_cast_fp16, y = v_15_cast_fp16)[name = tensor("attns_13_cast_fp16")]; + tensor attns_15_perm_0 = const()[name = tensor("attns_15_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_15x = const()[name = tensor("concat_15x"), val = tensor([1, -1, 1024])]; + tensor attns_15_cast_fp16 = transpose(perm = attns_15_perm_0, x = attns_13_cast_fp16)[name = tensor("transpose_224")]; + tensor x_57_cast_fp16 = reshape(shape = concat_15x, x = attns_15_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor model_encoder_layers_3_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108137152)))]; + tensor model_encoder_layers_3_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_3_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110234368)))]; + tensor linear_22_cast_fp16 = linear(bias = model_encoder_layers_3_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_3_self_attn_output_proj_weight_to_fp16, x = x_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = linear_22_cast_fp16, y = input_71_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor x_59_axes_0 = const()[name = tensor("x_59_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_3_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_3_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110236480)))]; + tensor model_encoder_layers_3_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_3_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110238592)))]; + tensor x_59_cast_fp16 = layer_norm(axes = x_59_axes_0, beta = model_encoder_layers_3_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_3_ffn_layer_norm_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor model_encoder_layers_3_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_3_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110240704)))]; + tensor model_encoder_layers_3_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_3_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118629376)))]; + tensor linear_23_cast_fp16 = linear(bias = model_encoder_layers_3_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_3_ffn_inner_proj_weight_to_fp16, x = x_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_77_mode_0 = const()[name = tensor("input_77_mode_0"), val = tensor("EXACT")]; + tensor input_77_cast_fp16 = gelu(mode = input_77_mode_0, x = linear_23_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor model_encoder_layers_3_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_3_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118637632)))]; + tensor model_encoder_layers_3_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_3_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127026304)))]; + tensor linear_24_cast_fp16 = linear(bias = model_encoder_layers_3_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_3_ffn_output_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_79_cast_fp16 = add(x = linear_24_cast_fp16, y = input_73_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor x_63_axes_0 = const()[name = tensor("x_63_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_4_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127028416)))]; + tensor model_encoder_layers_4_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127030528)))]; + tensor x_63_cast_fp16 = layer_norm(axes = x_63_axes_0, beta = model_encoder_layers_4_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_4_self_attn_layer_norm_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor model_encoder_layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127032640)))]; + tensor model_encoder_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129129856)))]; + tensor linear_25_cast_fp16 = linear(bias = model_encoder_layers_4_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_4_self_attn_q_proj_weight_to_fp16, x = x_63_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor concat_16x = const()[name = tensor("concat_16x"), val = tensor([1, -1, 16, 64])]; + tensor q_17_cast_fp16 = reshape(shape = concat_16x, x = linear_25_cast_fp16)[name = tensor("q_17_cast_fp16")]; + tensor model_encoder_layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129131968)))]; + tensor model_encoder_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131229184)))]; + tensor linear_26_cast_fp16 = linear(bias = model_encoder_layers_4_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_4_self_attn_k_proj_weight_to_fp16, x = x_63_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor model_encoder_layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131231296)))]; + tensor model_encoder_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133328512)))]; + tensor linear_27_cast_fp16 = linear(bias = model_encoder_layers_4_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_4_self_attn_v_proj_weight_to_fp16, x = x_63_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor concat_17x = const()[name = tensor("concat_17x"), val = tensor([1, -1, 16, 64])]; + tensor k_17_cast_fp16 = reshape(shape = concat_17x, x = linear_26_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor concat_18x = const()[name = tensor("concat_18x"), val = tensor([1, -1, 16, 64])]; + tensor v_17_cast_fp16 = reshape(shape = concat_18x, x = linear_27_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor v_19_perm_0 = const()[name = tensor("v_19_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_4_y_0_to_fp16 = const()[name = tensor("mul_4_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_4_cast_fp16 = mul(x = q_17_cast_fp16, y = mul_4_y_0_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor matmul_4_transpose_y_0 = const()[name = tensor("matmul_4_transpose_y_0"), val = tensor(true)]; + tensor matmul_4_transpose_x_0 = const()[name = tensor("matmul_4_transpose_x_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_221")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = mul_4_cast_fp16)[name = tensor("transpose_222")]; + tensor matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("matmul_4_cast_fp16")]; + tensor softmax_4_axis_0 = const()[name = tensor("softmax_4_axis_0"), val = tensor(-1)]; + tensor softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = matmul_4_cast_fp16)[name = tensor("softmax_4_cast_fp16")]; + tensor attns_17_transpose_x_0 = const()[name = tensor("attns_17_transpose_x_0"), val = tensor(false)]; + tensor attns_17_transpose_y_0 = const()[name = tensor("attns_17_transpose_y_0"), val = tensor(false)]; + tensor v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = v_17_cast_fp16)[name = tensor("transpose_223")]; + tensor attns_17_cast_fp16 = matmul(transpose_x = attns_17_transpose_x_0, transpose_y = attns_17_transpose_y_0, x = softmax_4_cast_fp16, y = v_19_cast_fp16)[name = tensor("attns_17_cast_fp16")]; + tensor attns_19_perm_0 = const()[name = tensor("attns_19_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_19x = const()[name = tensor("concat_19x"), val = tensor([1, -1, 1024])]; + tensor attns_19_cast_fp16 = transpose(perm = attns_19_perm_0, x = attns_17_cast_fp16)[name = tensor("transpose_220")]; + tensor x_65_cast_fp16 = reshape(shape = concat_19x, x = attns_19_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor model_encoder_layers_4_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133330624)))]; + tensor model_encoder_layers_4_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_4_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135427840)))]; + tensor linear_28_cast_fp16 = linear(bias = model_encoder_layers_4_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_4_self_attn_output_proj_weight_to_fp16, x = x_65_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = linear_28_cast_fp16, y = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor x_67_axes_0 = const()[name = tensor("x_67_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_4_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_4_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135429952)))]; + tensor model_encoder_layers_4_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_4_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135432064)))]; + tensor x_67_cast_fp16 = layer_norm(axes = x_67_axes_0, beta = model_encoder_layers_4_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_4_ffn_layer_norm_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor model_encoder_layers_4_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_4_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135434176)))]; + tensor model_encoder_layers_4_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_4_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143822848)))]; + tensor linear_29_cast_fp16 = linear(bias = model_encoder_layers_4_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_4_ffn_inner_proj_weight_to_fp16, x = x_67_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_29_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor model_encoder_layers_4_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_4_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143831104)))]; + tensor model_encoder_layers_4_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_4_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152219776)))]; + tensor linear_30_cast_fp16 = linear(bias = model_encoder_layers_4_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_4_ffn_output_proj_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = linear_30_cast_fp16, y = input_81_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor x_71_axes_0 = const()[name = tensor("x_71_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_5_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152221888)))]; + tensor model_encoder_layers_5_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152224000)))]; + tensor x_71_cast_fp16 = layer_norm(axes = x_71_axes_0, beta = model_encoder_layers_5_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_5_self_attn_layer_norm_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor model_encoder_layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152226112)))]; + tensor model_encoder_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154323328)))]; + tensor linear_31_cast_fp16 = linear(bias = model_encoder_layers_5_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_5_self_attn_q_proj_weight_to_fp16, x = x_71_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor concat_20x = const()[name = tensor("concat_20x"), val = tensor([1, -1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = concat_20x, x = linear_31_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor model_encoder_layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154325440)))]; + tensor model_encoder_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156422656)))]; + tensor linear_32_cast_fp16 = linear(bias = model_encoder_layers_5_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_5_self_attn_k_proj_weight_to_fp16, x = x_71_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor model_encoder_layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156424768)))]; + tensor model_encoder_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158521984)))]; + tensor linear_33_cast_fp16 = linear(bias = model_encoder_layers_5_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_5_self_attn_v_proj_weight_to_fp16, x = x_71_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor concat_21x = const()[name = tensor("concat_21x"), val = tensor([1, -1, 16, 64])]; + tensor k_21_cast_fp16 = reshape(shape = concat_21x, x = linear_32_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor concat_22x = const()[name = tensor("concat_22x"), val = tensor([1, -1, 16, 64])]; + tensor v_21_cast_fp16 = reshape(shape = concat_22x, x = linear_33_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor v_23_perm_0 = const()[name = tensor("v_23_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_5_y_0_to_fp16 = const()[name = tensor("mul_5_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_5_cast_fp16 = mul(x = q_21_cast_fp16, y = mul_5_y_0_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor matmul_5_transpose_y_0 = const()[name = tensor("matmul_5_transpose_y_0"), val = tensor(true)]; + tensor matmul_5_transpose_x_0 = const()[name = tensor("matmul_5_transpose_x_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_217")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = mul_5_cast_fp16)[name = tensor("transpose_218")]; + tensor matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("matmul_5_cast_fp16")]; + tensor softmax_5_axis_0 = const()[name = tensor("softmax_5_axis_0"), val = tensor(-1)]; + tensor softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = matmul_5_cast_fp16)[name = tensor("softmax_5_cast_fp16")]; + tensor attns_21_transpose_x_0 = const()[name = tensor("attns_21_transpose_x_0"), val = tensor(false)]; + tensor attns_21_transpose_y_0 = const()[name = tensor("attns_21_transpose_y_0"), val = tensor(false)]; + tensor v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = v_21_cast_fp16)[name = tensor("transpose_219")]; + tensor attns_21_cast_fp16 = matmul(transpose_x = attns_21_transpose_x_0, transpose_y = attns_21_transpose_y_0, x = softmax_5_cast_fp16, y = v_23_cast_fp16)[name = tensor("attns_21_cast_fp16")]; + tensor attns_23_perm_0 = const()[name = tensor("attns_23_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_23x = const()[name = tensor("concat_23x"), val = tensor([1, -1, 1024])]; + tensor attns_23_cast_fp16 = transpose(perm = attns_23_perm_0, x = attns_21_cast_fp16)[name = tensor("transpose_216")]; + tensor x_73_cast_fp16 = reshape(shape = concat_23x, x = attns_23_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor model_encoder_layers_5_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158524096)))]; + tensor model_encoder_layers_5_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_5_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160621312)))]; + tensor linear_34_cast_fp16 = linear(bias = model_encoder_layers_5_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_5_self_attn_output_proj_weight_to_fp16, x = x_73_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = linear_34_cast_fp16, y = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor x_75_axes_0 = const()[name = tensor("x_75_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_5_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_5_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160623424)))]; + tensor model_encoder_layers_5_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_5_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160625536)))]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = model_encoder_layers_5_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_5_ffn_layer_norm_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor model_encoder_layers_5_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_5_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160627648)))]; + tensor model_encoder_layers_5_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_5_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169016320)))]; + tensor linear_35_cast_fp16 = linear(bias = model_encoder_layers_5_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_5_ffn_inner_proj_weight_to_fp16, x = x_75_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_93_mode_0 = const()[name = tensor("input_93_mode_0"), val = tensor("EXACT")]; + tensor input_93_cast_fp16 = gelu(mode = input_93_mode_0, x = linear_35_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor model_encoder_layers_5_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_5_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169024576)))]; + tensor model_encoder_layers_5_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_5_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177413248)))]; + tensor linear_36_cast_fp16 = linear(bias = model_encoder_layers_5_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_5_ffn_output_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_95_cast_fp16 = add(x = linear_36_cast_fp16, y = input_89_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor x_79_axes_0 = const()[name = tensor("x_79_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_6_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177415360)))]; + tensor model_encoder_layers_6_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177417472)))]; + tensor x_79_cast_fp16 = layer_norm(axes = x_79_axes_0, beta = model_encoder_layers_6_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_6_self_attn_layer_norm_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor model_encoder_layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177419584)))]; + tensor model_encoder_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179516800)))]; + tensor linear_37_cast_fp16 = linear(bias = model_encoder_layers_6_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_6_self_attn_q_proj_weight_to_fp16, x = x_79_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor concat_24x = const()[name = tensor("concat_24x"), val = tensor([1, -1, 16, 64])]; + tensor q_25_cast_fp16 = reshape(shape = concat_24x, x = linear_37_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor model_encoder_layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179518912)))]; + tensor model_encoder_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181616128)))]; + tensor linear_38_cast_fp16 = linear(bias = model_encoder_layers_6_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_6_self_attn_k_proj_weight_to_fp16, x = x_79_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor model_encoder_layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181618240)))]; + tensor model_encoder_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183715456)))]; + tensor linear_39_cast_fp16 = linear(bias = model_encoder_layers_6_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_6_self_attn_v_proj_weight_to_fp16, x = x_79_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor concat_25x = const()[name = tensor("concat_25x"), val = tensor([1, -1, 16, 64])]; + tensor k_25_cast_fp16 = reshape(shape = concat_25x, x = linear_38_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor concat_26x = const()[name = tensor("concat_26x"), val = tensor([1, -1, 16, 64])]; + tensor v_25_cast_fp16 = reshape(shape = concat_26x, x = linear_39_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor v_27_perm_0 = const()[name = tensor("v_27_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_6_y_0_to_fp16 = const()[name = tensor("mul_6_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_6_cast_fp16 = mul(x = q_25_cast_fp16, y = mul_6_y_0_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor matmul_6_transpose_y_0 = const()[name = tensor("matmul_6_transpose_y_0"), val = tensor(true)]; + tensor matmul_6_transpose_x_0 = const()[name = tensor("matmul_6_transpose_x_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = mul_6_cast_fp16)[name = tensor("transpose_214")]; + tensor matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("matmul_6_cast_fp16")]; + tensor softmax_6_axis_0 = const()[name = tensor("softmax_6_axis_0"), val = tensor(-1)]; + tensor softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = matmul_6_cast_fp16)[name = tensor("softmax_6_cast_fp16")]; + tensor attns_25_transpose_x_0 = const()[name = tensor("attns_25_transpose_x_0"), val = tensor(false)]; + tensor attns_25_transpose_y_0 = const()[name = tensor("attns_25_transpose_y_0"), val = tensor(false)]; + tensor v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = v_25_cast_fp16)[name = tensor("transpose_215")]; + tensor attns_25_cast_fp16 = matmul(transpose_x = attns_25_transpose_x_0, transpose_y = attns_25_transpose_y_0, x = softmax_6_cast_fp16, y = v_27_cast_fp16)[name = tensor("attns_25_cast_fp16")]; + tensor attns_27_perm_0 = const()[name = tensor("attns_27_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_27x = const()[name = tensor("concat_27x"), val = tensor([1, -1, 1024])]; + tensor attns_27_cast_fp16 = transpose(perm = attns_27_perm_0, x = attns_25_cast_fp16)[name = tensor("transpose_212")]; + tensor x_81_cast_fp16 = reshape(shape = concat_27x, x = attns_27_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor model_encoder_layers_6_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183717568)))]; + tensor model_encoder_layers_6_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_6_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185814784)))]; + tensor linear_40_cast_fp16 = linear(bias = model_encoder_layers_6_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_6_self_attn_output_proj_weight_to_fp16, x = x_81_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = linear_40_cast_fp16, y = input_95_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor x_83_axes_0 = const()[name = tensor("x_83_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_6_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_6_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185816896)))]; + tensor model_encoder_layers_6_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_6_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185819008)))]; + tensor x_83_cast_fp16 = layer_norm(axes = x_83_axes_0, beta = model_encoder_layers_6_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_6_ffn_layer_norm_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor model_encoder_layers_6_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_6_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185821120)))]; + tensor model_encoder_layers_6_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_6_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194209792)))]; + tensor linear_41_cast_fp16 = linear(bias = model_encoder_layers_6_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_6_ffn_inner_proj_weight_to_fp16, x = x_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor model_encoder_layers_6_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_6_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194218048)))]; + tensor model_encoder_layers_6_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_6_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202606720)))]; + tensor linear_42_cast_fp16 = linear(bias = model_encoder_layers_6_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_6_ffn_output_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_103_cast_fp16 = add(x = linear_42_cast_fp16, y = input_97_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor x_87_axes_0 = const()[name = tensor("x_87_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_7_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202608832)))]; + tensor model_encoder_layers_7_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202610944)))]; + tensor x_87_cast_fp16 = layer_norm(axes = x_87_axes_0, beta = model_encoder_layers_7_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_7_self_attn_layer_norm_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor model_encoder_layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202613056)))]; + tensor model_encoder_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204710272)))]; + tensor linear_43_cast_fp16 = linear(bias = model_encoder_layers_7_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_7_self_attn_q_proj_weight_to_fp16, x = x_87_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor concat_28x = const()[name = tensor("concat_28x"), val = tensor([1, -1, 16, 64])]; + tensor q_29_cast_fp16 = reshape(shape = concat_28x, x = linear_43_cast_fp16)[name = tensor("q_29_cast_fp16")]; + tensor model_encoder_layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204712384)))]; + tensor model_encoder_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206809600)))]; + tensor linear_44_cast_fp16 = linear(bias = model_encoder_layers_7_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_7_self_attn_k_proj_weight_to_fp16, x = x_87_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor model_encoder_layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206811712)))]; + tensor model_encoder_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208908928)))]; + tensor linear_45_cast_fp16 = linear(bias = model_encoder_layers_7_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_7_self_attn_v_proj_weight_to_fp16, x = x_87_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor concat_29x = const()[name = tensor("concat_29x"), val = tensor([1, -1, 16, 64])]; + tensor k_29_cast_fp16 = reshape(shape = concat_29x, x = linear_44_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor concat_30x = const()[name = tensor("concat_30x"), val = tensor([1, -1, 16, 64])]; + tensor v_29_cast_fp16 = reshape(shape = concat_30x, x = linear_45_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor v_31_perm_0 = const()[name = tensor("v_31_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_7_y_0_to_fp16 = const()[name = tensor("mul_7_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_7_cast_fp16 = mul(x = q_29_cast_fp16, y = mul_7_y_0_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor matmul_7_transpose_y_0 = const()[name = tensor("matmul_7_transpose_y_0"), val = tensor(true)]; + tensor matmul_7_transpose_x_0 = const()[name = tensor("matmul_7_transpose_x_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = mul_7_cast_fp16)[name = tensor("transpose_210")]; + tensor matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("matmul_7_cast_fp16")]; + tensor softmax_7_axis_0 = const()[name = tensor("softmax_7_axis_0"), val = tensor(-1)]; + tensor softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = matmul_7_cast_fp16)[name = tensor("softmax_7_cast_fp16")]; + tensor attns_29_transpose_x_0 = const()[name = tensor("attns_29_transpose_x_0"), val = tensor(false)]; + tensor attns_29_transpose_y_0 = const()[name = tensor("attns_29_transpose_y_0"), val = tensor(false)]; + tensor v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = v_29_cast_fp16)[name = tensor("transpose_211")]; + tensor attns_29_cast_fp16 = matmul(transpose_x = attns_29_transpose_x_0, transpose_y = attns_29_transpose_y_0, x = softmax_7_cast_fp16, y = v_31_cast_fp16)[name = tensor("attns_29_cast_fp16")]; + tensor attns_31_perm_0 = const()[name = tensor("attns_31_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_31x = const()[name = tensor("concat_31x"), val = tensor([1, -1, 1024])]; + tensor attns_31_cast_fp16 = transpose(perm = attns_31_perm_0, x = attns_29_cast_fp16)[name = tensor("transpose_208")]; + tensor x_89_cast_fp16 = reshape(shape = concat_31x, x = attns_31_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor model_encoder_layers_7_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208911040)))]; + tensor model_encoder_layers_7_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_7_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211008256)))]; + tensor linear_46_cast_fp16 = linear(bias = model_encoder_layers_7_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_7_self_attn_output_proj_weight_to_fp16, x = x_89_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_105_cast_fp16 = add(x = linear_46_cast_fp16, y = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor x_91_axes_0 = const()[name = tensor("x_91_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_7_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_7_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211010368)))]; + tensor model_encoder_layers_7_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_7_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211012480)))]; + tensor x_91_cast_fp16 = layer_norm(axes = x_91_axes_0, beta = model_encoder_layers_7_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_7_ffn_layer_norm_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor model_encoder_layers_7_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_7_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211014592)))]; + tensor model_encoder_layers_7_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_7_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219403264)))]; + tensor linear_47_cast_fp16 = linear(bias = model_encoder_layers_7_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_7_ffn_inner_proj_weight_to_fp16, x = x_91_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_47_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor model_encoder_layers_7_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_7_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219411520)))]; + tensor model_encoder_layers_7_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_7_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227800192)))]; + tensor linear_48_cast_fp16 = linear(bias = model_encoder_layers_7_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_7_ffn_output_proj_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = linear_48_cast_fp16, y = input_105_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor x_95_axes_0 = const()[name = tensor("x_95_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_8_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227802304)))]; + tensor model_encoder_layers_8_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227804416)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = model_encoder_layers_8_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_8_self_attn_layer_norm_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor model_encoder_layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227806528)))]; + tensor model_encoder_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229903744)))]; + tensor linear_49_cast_fp16 = linear(bias = model_encoder_layers_8_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_8_self_attn_q_proj_weight_to_fp16, x = x_95_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor concat_32x = const()[name = tensor("concat_32x"), val = tensor([1, -1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = concat_32x, x = linear_49_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor model_encoder_layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229905856)))]; + tensor model_encoder_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232003072)))]; + tensor linear_50_cast_fp16 = linear(bias = model_encoder_layers_8_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_8_self_attn_k_proj_weight_to_fp16, x = x_95_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor model_encoder_layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232005184)))]; + tensor model_encoder_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234102400)))]; + tensor linear_51_cast_fp16 = linear(bias = model_encoder_layers_8_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_8_self_attn_v_proj_weight_to_fp16, x = x_95_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor concat_33x = const()[name = tensor("concat_33x"), val = tensor([1, -1, 16, 64])]; + tensor k_33_cast_fp16 = reshape(shape = concat_33x, x = linear_50_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor concat_34x = const()[name = tensor("concat_34x"), val = tensor([1, -1, 16, 64])]; + tensor v_33_cast_fp16 = reshape(shape = concat_34x, x = linear_51_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor v_35_perm_0 = const()[name = tensor("v_35_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_8_y_0_to_fp16 = const()[name = tensor("mul_8_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_8_cast_fp16 = mul(x = q_33_cast_fp16, y = mul_8_y_0_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor matmul_8_transpose_y_0 = const()[name = tensor("matmul_8_transpose_y_0"), val = tensor(true)]; + tensor matmul_8_transpose_x_0 = const()[name = tensor("matmul_8_transpose_x_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = mul_8_cast_fp16)[name = tensor("transpose_206")]; + tensor matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("matmul_8_cast_fp16")]; + tensor softmax_8_axis_0 = const()[name = tensor("softmax_8_axis_0"), val = tensor(-1)]; + tensor softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = matmul_8_cast_fp16)[name = tensor("softmax_8_cast_fp16")]; + tensor attns_33_transpose_x_0 = const()[name = tensor("attns_33_transpose_x_0"), val = tensor(false)]; + tensor attns_33_transpose_y_0 = const()[name = tensor("attns_33_transpose_y_0"), val = tensor(false)]; + tensor v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = v_33_cast_fp16)[name = tensor("transpose_207")]; + tensor attns_33_cast_fp16 = matmul(transpose_x = attns_33_transpose_x_0, transpose_y = attns_33_transpose_y_0, x = softmax_8_cast_fp16, y = v_35_cast_fp16)[name = tensor("attns_33_cast_fp16")]; + tensor attns_35_perm_0 = const()[name = tensor("attns_35_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_35x = const()[name = tensor("concat_35x"), val = tensor([1, -1, 1024])]; + tensor attns_35_cast_fp16 = transpose(perm = attns_35_perm_0, x = attns_33_cast_fp16)[name = tensor("transpose_204")]; + tensor x_97_cast_fp16 = reshape(shape = concat_35x, x = attns_35_cast_fp16)[name = tensor("x_97_cast_fp16")]; + tensor model_encoder_layers_8_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234104512)))]; + tensor model_encoder_layers_8_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_8_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236201728)))]; + tensor linear_52_cast_fp16 = linear(bias = model_encoder_layers_8_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_8_self_attn_output_proj_weight_to_fp16, x = x_97_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_113_cast_fp16 = add(x = linear_52_cast_fp16, y = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor x_99_axes_0 = const()[name = tensor("x_99_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_8_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_8_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236203840)))]; + tensor model_encoder_layers_8_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_8_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236205952)))]; + tensor x_99_cast_fp16 = layer_norm(axes = x_99_axes_0, beta = model_encoder_layers_8_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_8_ffn_layer_norm_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor model_encoder_layers_8_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_8_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236208064)))]; + tensor model_encoder_layers_8_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_8_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244596736)))]; + tensor linear_53_cast_fp16 = linear(bias = model_encoder_layers_8_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_8_ffn_inner_proj_weight_to_fp16, x = x_99_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor input_117_mode_0 = const()[name = tensor("input_117_mode_0"), val = tensor("EXACT")]; + tensor input_117_cast_fp16 = gelu(mode = input_117_mode_0, x = linear_53_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor model_encoder_layers_8_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_8_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244604992)))]; + tensor model_encoder_layers_8_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_8_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252993664)))]; + tensor linear_54_cast_fp16 = linear(bias = model_encoder_layers_8_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_8_ffn_output_proj_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_119_cast_fp16 = add(x = linear_54_cast_fp16, y = input_113_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor x_103_axes_0 = const()[name = tensor("x_103_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_9_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252995776)))]; + tensor model_encoder_layers_9_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252997888)))]; + tensor x_103_cast_fp16 = layer_norm(axes = x_103_axes_0, beta = model_encoder_layers_9_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_9_self_attn_layer_norm_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("x_103_cast_fp16")]; + tensor model_encoder_layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253000000)))]; + tensor model_encoder_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255097216)))]; + tensor linear_55_cast_fp16 = linear(bias = model_encoder_layers_9_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_9_self_attn_q_proj_weight_to_fp16, x = x_103_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor concat_36x = const()[name = tensor("concat_36x"), val = tensor([1, -1, 16, 64])]; + tensor q_37_cast_fp16 = reshape(shape = concat_36x, x = linear_55_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor model_encoder_layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255099328)))]; + tensor model_encoder_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257196544)))]; + tensor linear_56_cast_fp16 = linear(bias = model_encoder_layers_9_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_9_self_attn_k_proj_weight_to_fp16, x = x_103_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor model_encoder_layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257198656)))]; + tensor model_encoder_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259295872)))]; + tensor linear_57_cast_fp16 = linear(bias = model_encoder_layers_9_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_9_self_attn_v_proj_weight_to_fp16, x = x_103_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor concat_37x = const()[name = tensor("concat_37x"), val = tensor([1, -1, 16, 64])]; + tensor k_37_cast_fp16 = reshape(shape = concat_37x, x = linear_56_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor concat_38x = const()[name = tensor("concat_38x"), val = tensor([1, -1, 16, 64])]; + tensor v_37_cast_fp16 = reshape(shape = concat_38x, x = linear_57_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor v_39_perm_0 = const()[name = tensor("v_39_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_9_y_0_to_fp16 = const()[name = tensor("mul_9_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_9_cast_fp16 = mul(x = q_37_cast_fp16, y = mul_9_y_0_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor matmul_9_transpose_y_0 = const()[name = tensor("matmul_9_transpose_y_0"), val = tensor(true)]; + tensor matmul_9_transpose_x_0 = const()[name = tensor("matmul_9_transpose_x_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = mul_9_cast_fp16)[name = tensor("transpose_202")]; + tensor matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("matmul_9_cast_fp16")]; + tensor softmax_9_axis_0 = const()[name = tensor("softmax_9_axis_0"), val = tensor(-1)]; + tensor softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = matmul_9_cast_fp16)[name = tensor("softmax_9_cast_fp16")]; + tensor attns_37_transpose_x_0 = const()[name = tensor("attns_37_transpose_x_0"), val = tensor(false)]; + tensor attns_37_transpose_y_0 = const()[name = tensor("attns_37_transpose_y_0"), val = tensor(false)]; + tensor v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = v_37_cast_fp16)[name = tensor("transpose_203")]; + tensor attns_37_cast_fp16 = matmul(transpose_x = attns_37_transpose_x_0, transpose_y = attns_37_transpose_y_0, x = softmax_9_cast_fp16, y = v_39_cast_fp16)[name = tensor("attns_37_cast_fp16")]; + tensor attns_39_perm_0 = const()[name = tensor("attns_39_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_39x = const()[name = tensor("concat_39x"), val = tensor([1, -1, 1024])]; + tensor attns_39_cast_fp16 = transpose(perm = attns_39_perm_0, x = attns_37_cast_fp16)[name = tensor("transpose_200")]; + tensor x_105_cast_fp16 = reshape(shape = concat_39x, x = attns_39_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor model_encoder_layers_9_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259297984)))]; + tensor model_encoder_layers_9_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_9_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261395200)))]; + tensor linear_58_cast_fp16 = linear(bias = model_encoder_layers_9_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_9_self_attn_output_proj_weight_to_fp16, x = x_105_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = linear_58_cast_fp16, y = input_119_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor x_107_axes_0 = const()[name = tensor("x_107_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_9_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_9_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261397312)))]; + tensor model_encoder_layers_9_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_9_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261399424)))]; + tensor x_107_cast_fp16 = layer_norm(axes = x_107_axes_0, beta = model_encoder_layers_9_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_9_ffn_layer_norm_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor model_encoder_layers_9_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_9_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261401536)))]; + tensor model_encoder_layers_9_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_9_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269790208)))]; + tensor linear_59_cast_fp16 = linear(bias = model_encoder_layers_9_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_9_ffn_inner_proj_weight_to_fp16, x = x_107_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_125_mode_0 = const()[name = tensor("input_125_mode_0"), val = tensor("EXACT")]; + tensor input_125_cast_fp16 = gelu(mode = input_125_mode_0, x = linear_59_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor model_encoder_layers_9_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_9_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269798464)))]; + tensor model_encoder_layers_9_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_9_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278187136)))]; + tensor linear_60_cast_fp16 = linear(bias = model_encoder_layers_9_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_9_ffn_output_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = linear_60_cast_fp16, y = input_121_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor x_111_axes_0 = const()[name = tensor("x_111_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_10_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278189248)))]; + tensor model_encoder_layers_10_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278191360)))]; + tensor x_111_cast_fp16 = layer_norm(axes = x_111_axes_0, beta = model_encoder_layers_10_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_10_self_attn_layer_norm_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor model_encoder_layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278193472)))]; + tensor model_encoder_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280290688)))]; + tensor linear_61_cast_fp16 = linear(bias = model_encoder_layers_10_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_10_self_attn_q_proj_weight_to_fp16, x = x_111_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor concat_40x = const()[name = tensor("concat_40x"), val = tensor([1, -1, 16, 64])]; + tensor q_41_cast_fp16 = reshape(shape = concat_40x, x = linear_61_cast_fp16)[name = tensor("q_41_cast_fp16")]; + tensor model_encoder_layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280292800)))]; + tensor model_encoder_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282390016)))]; + tensor linear_62_cast_fp16 = linear(bias = model_encoder_layers_10_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_10_self_attn_k_proj_weight_to_fp16, x = x_111_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor model_encoder_layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282392128)))]; + tensor model_encoder_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284489344)))]; + tensor linear_63_cast_fp16 = linear(bias = model_encoder_layers_10_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_10_self_attn_v_proj_weight_to_fp16, x = x_111_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor concat_41x = const()[name = tensor("concat_41x"), val = tensor([1, -1, 16, 64])]; + tensor k_41_cast_fp16 = reshape(shape = concat_41x, x = linear_62_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor concat_42x = const()[name = tensor("concat_42x"), val = tensor([1, -1, 16, 64])]; + tensor v_41_cast_fp16 = reshape(shape = concat_42x, x = linear_63_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor v_43_perm_0 = const()[name = tensor("v_43_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_10_y_0_to_fp16 = const()[name = tensor("mul_10_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_10_cast_fp16 = mul(x = q_41_cast_fp16, y = mul_10_y_0_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor matmul_10_transpose_y_0 = const()[name = tensor("matmul_10_transpose_y_0"), val = tensor(true)]; + tensor matmul_10_transpose_x_0 = const()[name = tensor("matmul_10_transpose_x_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = mul_10_cast_fp16)[name = tensor("transpose_198")]; + tensor matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("matmul_10_cast_fp16")]; + tensor softmax_10_axis_0 = const()[name = tensor("softmax_10_axis_0"), val = tensor(-1)]; + tensor softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = matmul_10_cast_fp16)[name = tensor("softmax_10_cast_fp16")]; + tensor attns_41_transpose_x_0 = const()[name = tensor("attns_41_transpose_x_0"), val = tensor(false)]; + tensor attns_41_transpose_y_0 = const()[name = tensor("attns_41_transpose_y_0"), val = tensor(false)]; + tensor v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = v_41_cast_fp16)[name = tensor("transpose_199")]; + tensor attns_41_cast_fp16 = matmul(transpose_x = attns_41_transpose_x_0, transpose_y = attns_41_transpose_y_0, x = softmax_10_cast_fp16, y = v_43_cast_fp16)[name = tensor("attns_41_cast_fp16")]; + tensor attns_43_perm_0 = const()[name = tensor("attns_43_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_43x = const()[name = tensor("concat_43x"), val = tensor([1, -1, 1024])]; + tensor attns_43_cast_fp16 = transpose(perm = attns_43_perm_0, x = attns_41_cast_fp16)[name = tensor("transpose_196")]; + tensor x_113_cast_fp16 = reshape(shape = concat_43x, x = attns_43_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor model_encoder_layers_10_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284491456)))]; + tensor model_encoder_layers_10_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_10_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286588672)))]; + tensor linear_64_cast_fp16 = linear(bias = model_encoder_layers_10_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_10_self_attn_output_proj_weight_to_fp16, x = x_113_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_129_cast_fp16 = add(x = linear_64_cast_fp16, y = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor x_115_axes_0 = const()[name = tensor("x_115_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_10_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_10_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286590784)))]; + tensor model_encoder_layers_10_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_10_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286592896)))]; + tensor x_115_cast_fp16 = layer_norm(axes = x_115_axes_0, beta = model_encoder_layers_10_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_10_ffn_layer_norm_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("x_115_cast_fp16")]; + tensor model_encoder_layers_10_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_10_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286595008)))]; + tensor model_encoder_layers_10_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_10_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294983680)))]; + tensor linear_65_cast_fp16 = linear(bias = model_encoder_layers_10_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_10_ffn_inner_proj_weight_to_fp16, x = x_115_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_65_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor model_encoder_layers_10_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_10_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294991936)))]; + tensor model_encoder_layers_10_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_10_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303380608)))]; + tensor linear_66_cast_fp16 = linear(bias = model_encoder_layers_10_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_10_ffn_output_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = linear_66_cast_fp16, y = input_129_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor x_119_axes_0 = const()[name = tensor("x_119_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_11_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303382720)))]; + tensor model_encoder_layers_11_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303384832)))]; + tensor x_119_cast_fp16 = layer_norm(axes = x_119_axes_0, beta = model_encoder_layers_11_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_11_self_attn_layer_norm_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor model_encoder_layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303386944)))]; + tensor model_encoder_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305484160)))]; + tensor linear_67_cast_fp16 = linear(bias = model_encoder_layers_11_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_11_self_attn_q_proj_weight_to_fp16, x = x_119_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor concat_44x = const()[name = tensor("concat_44x"), val = tensor([1, -1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = concat_44x, x = linear_67_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor model_encoder_layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305486272)))]; + tensor model_encoder_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307583488)))]; + tensor linear_68_cast_fp16 = linear(bias = model_encoder_layers_11_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_11_self_attn_k_proj_weight_to_fp16, x = x_119_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor model_encoder_layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307585600)))]; + tensor model_encoder_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309682816)))]; + tensor linear_69_cast_fp16 = linear(bias = model_encoder_layers_11_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_11_self_attn_v_proj_weight_to_fp16, x = x_119_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor concat_45x = const()[name = tensor("concat_45x"), val = tensor([1, -1, 16, 64])]; + tensor k_45_cast_fp16 = reshape(shape = concat_45x, x = linear_68_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor concat_46x = const()[name = tensor("concat_46x"), val = tensor([1, -1, 16, 64])]; + tensor v_45_cast_fp16 = reshape(shape = concat_46x, x = linear_69_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor v_47_perm_0 = const()[name = tensor("v_47_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_11_y_0_to_fp16 = const()[name = tensor("mul_11_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_11_cast_fp16 = mul(x = q_45_cast_fp16, y = mul_11_y_0_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor matmul_11_transpose_y_0 = const()[name = tensor("matmul_11_transpose_y_0"), val = tensor(true)]; + tensor matmul_11_transpose_x_0 = const()[name = tensor("matmul_11_transpose_x_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = mul_11_cast_fp16)[name = tensor("transpose_194")]; + tensor matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("matmul_11_cast_fp16")]; + tensor softmax_11_axis_0 = const()[name = tensor("softmax_11_axis_0"), val = tensor(-1)]; + tensor softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = matmul_11_cast_fp16)[name = tensor("softmax_11_cast_fp16")]; + tensor attns_45_transpose_x_0 = const()[name = tensor("attns_45_transpose_x_0"), val = tensor(false)]; + tensor attns_45_transpose_y_0 = const()[name = tensor("attns_45_transpose_y_0"), val = tensor(false)]; + tensor v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = v_45_cast_fp16)[name = tensor("transpose_195")]; + tensor attns_45_cast_fp16 = matmul(transpose_x = attns_45_transpose_x_0, transpose_y = attns_45_transpose_y_0, x = softmax_11_cast_fp16, y = v_47_cast_fp16)[name = tensor("attns_45_cast_fp16")]; + tensor attns_47_perm_0 = const()[name = tensor("attns_47_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_47x = const()[name = tensor("concat_47x"), val = tensor([1, -1, 1024])]; + tensor attns_47_cast_fp16 = transpose(perm = attns_47_perm_0, x = attns_45_cast_fp16)[name = tensor("transpose_192")]; + tensor x_121_cast_fp16 = reshape(shape = concat_47x, x = attns_47_cast_fp16)[name = tensor("x_121_cast_fp16")]; + tensor model_encoder_layers_11_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309684928)))]; + tensor model_encoder_layers_11_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_11_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311782144)))]; + tensor linear_70_cast_fp16 = linear(bias = model_encoder_layers_11_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_11_self_attn_output_proj_weight_to_fp16, x = x_121_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_137_cast_fp16 = add(x = linear_70_cast_fp16, y = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor x_123_axes_0 = const()[name = tensor("x_123_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_11_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_11_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311784256)))]; + tensor model_encoder_layers_11_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_11_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311786368)))]; + tensor x_123_cast_fp16 = layer_norm(axes = x_123_axes_0, beta = model_encoder_layers_11_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_11_ffn_layer_norm_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor model_encoder_layers_11_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_11_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311788480)))]; + tensor model_encoder_layers_11_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_11_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320177152)))]; + tensor linear_71_cast_fp16 = linear(bias = model_encoder_layers_11_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_11_ffn_inner_proj_weight_to_fp16, x = x_123_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_71_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor model_encoder_layers_11_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_11_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320185408)))]; + tensor model_encoder_layers_11_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_11_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328574080)))]; + tensor linear_72_cast_fp16 = linear(bias = model_encoder_layers_11_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_11_ffn_output_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_143_cast_fp16 = add(x = linear_72_cast_fp16, y = input_137_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor x_127_axes_0 = const()[name = tensor("x_127_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_12_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328576192)))]; + tensor model_encoder_layers_12_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328578304)))]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = model_encoder_layers_12_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_12_self_attn_layer_norm_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("x_127_cast_fp16")]; + tensor model_encoder_layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328580416)))]; + tensor model_encoder_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330677632)))]; + tensor linear_73_cast_fp16 = linear(bias = model_encoder_layers_12_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_12_self_attn_q_proj_weight_to_fp16, x = x_127_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor concat_48x = const()[name = tensor("concat_48x"), val = tensor([1, -1, 16, 64])]; + tensor q_49_cast_fp16 = reshape(shape = concat_48x, x = linear_73_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor model_encoder_layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330679744)))]; + tensor model_encoder_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332776960)))]; + tensor linear_74_cast_fp16 = linear(bias = model_encoder_layers_12_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_12_self_attn_k_proj_weight_to_fp16, x = x_127_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor model_encoder_layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332779072)))]; + tensor model_encoder_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334876288)))]; + tensor linear_75_cast_fp16 = linear(bias = model_encoder_layers_12_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_12_self_attn_v_proj_weight_to_fp16, x = x_127_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor concat_49x = const()[name = tensor("concat_49x"), val = tensor([1, -1, 16, 64])]; + tensor k_49_cast_fp16 = reshape(shape = concat_49x, x = linear_74_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor concat_50x = const()[name = tensor("concat_50x"), val = tensor([1, -1, 16, 64])]; + tensor v_49_cast_fp16 = reshape(shape = concat_50x, x = linear_75_cast_fp16)[name = tensor("v_49_cast_fp16")]; + tensor v_51_perm_0 = const()[name = tensor("v_51_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_12_y_0_to_fp16 = const()[name = tensor("mul_12_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_12_cast_fp16 = mul(x = q_49_cast_fp16, y = mul_12_y_0_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor matmul_12_transpose_y_0 = const()[name = tensor("matmul_12_transpose_y_0"), val = tensor(true)]; + tensor matmul_12_transpose_x_0 = const()[name = tensor("matmul_12_transpose_x_0"), val = tensor(false)]; + tensor transpose_120_perm_0 = const()[name = tensor("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = tensor("transpose_121_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = mul_12_cast_fp16)[name = tensor("transpose_190")]; + tensor matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = transpose_120, y = transpose_121)[name = tensor("matmul_12_cast_fp16")]; + tensor softmax_12_axis_0 = const()[name = tensor("softmax_12_axis_0"), val = tensor(-1)]; + tensor softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = matmul_12_cast_fp16)[name = tensor("softmax_12_cast_fp16")]; + tensor attns_49_transpose_x_0 = const()[name = tensor("attns_49_transpose_x_0"), val = tensor(false)]; + tensor attns_49_transpose_y_0 = const()[name = tensor("attns_49_transpose_y_0"), val = tensor(false)]; + tensor v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = v_49_cast_fp16)[name = tensor("transpose_191")]; + tensor attns_49_cast_fp16 = matmul(transpose_x = attns_49_transpose_x_0, transpose_y = attns_49_transpose_y_0, x = softmax_12_cast_fp16, y = v_51_cast_fp16)[name = tensor("attns_49_cast_fp16")]; + tensor attns_51_perm_0 = const()[name = tensor("attns_51_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_51x = const()[name = tensor("concat_51x"), val = tensor([1, -1, 1024])]; + tensor attns_51_cast_fp16 = transpose(perm = attns_51_perm_0, x = attns_49_cast_fp16)[name = tensor("transpose_188")]; + tensor x_129_cast_fp16 = reshape(shape = concat_51x, x = attns_51_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor model_encoder_layers_12_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334878400)))]; + tensor model_encoder_layers_12_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_12_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336975616)))]; + tensor linear_76_cast_fp16 = linear(bias = model_encoder_layers_12_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_12_self_attn_output_proj_weight_to_fp16, x = x_129_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = linear_76_cast_fp16, y = input_143_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor x_131_axes_0 = const()[name = tensor("x_131_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_12_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_12_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336977728)))]; + tensor model_encoder_layers_12_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_12_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336979840)))]; + tensor x_131_cast_fp16 = layer_norm(axes = x_131_axes_0, beta = model_encoder_layers_12_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_12_ffn_layer_norm_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor model_encoder_layers_12_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_12_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336981952)))]; + tensor model_encoder_layers_12_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_12_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345370624)))]; + tensor linear_77_cast_fp16 = linear(bias = model_encoder_layers_12_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_12_ffn_inner_proj_weight_to_fp16, x = x_131_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor input_149_mode_0 = const()[name = tensor("input_149_mode_0"), val = tensor("EXACT")]; + tensor input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = linear_77_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor model_encoder_layers_12_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_12_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345378880)))]; + tensor model_encoder_layers_12_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_12_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353767552)))]; + tensor linear_78_cast_fp16 = linear(bias = model_encoder_layers_12_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_12_ffn_output_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = linear_78_cast_fp16, y = input_145_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor x_135_axes_0 = const()[name = tensor("x_135_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_13_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353769664)))]; + tensor model_encoder_layers_13_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353771776)))]; + tensor x_135_cast_fp16 = layer_norm(axes = x_135_axes_0, beta = model_encoder_layers_13_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_13_self_attn_layer_norm_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("x_135_cast_fp16")]; + tensor model_encoder_layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353773888)))]; + tensor model_encoder_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355871104)))]; + tensor linear_79_cast_fp16 = linear(bias = model_encoder_layers_13_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_13_self_attn_q_proj_weight_to_fp16, x = x_135_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor concat_52x = const()[name = tensor("concat_52x"), val = tensor([1, -1, 16, 64])]; + tensor q_53_cast_fp16 = reshape(shape = concat_52x, x = linear_79_cast_fp16)[name = tensor("q_53_cast_fp16")]; + tensor model_encoder_layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355873216)))]; + tensor model_encoder_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(357970432)))]; + tensor linear_80_cast_fp16 = linear(bias = model_encoder_layers_13_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_13_self_attn_k_proj_weight_to_fp16, x = x_135_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor model_encoder_layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(357972544)))]; + tensor model_encoder_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360069760)))]; + tensor linear_81_cast_fp16 = linear(bias = model_encoder_layers_13_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_13_self_attn_v_proj_weight_to_fp16, x = x_135_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor concat_53x = const()[name = tensor("concat_53x"), val = tensor([1, -1, 16, 64])]; + tensor k_53_cast_fp16 = reshape(shape = concat_53x, x = linear_80_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor concat_54x = const()[name = tensor("concat_54x"), val = tensor([1, -1, 16, 64])]; + tensor v_53_cast_fp16 = reshape(shape = concat_54x, x = linear_81_cast_fp16)[name = tensor("v_53_cast_fp16")]; + tensor v_55_perm_0 = const()[name = tensor("v_55_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_13_y_0_to_fp16 = const()[name = tensor("mul_13_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_13_cast_fp16 = mul(x = q_53_cast_fp16, y = mul_13_y_0_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor matmul_13_transpose_y_0 = const()[name = tensor("matmul_13_transpose_y_0"), val = tensor(true)]; + tensor matmul_13_transpose_x_0 = const()[name = tensor("matmul_13_transpose_x_0"), val = tensor(false)]; + tensor transpose_122_perm_0 = const()[name = tensor("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = tensor("transpose_123_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = mul_13_cast_fp16)[name = tensor("transpose_186")]; + tensor matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = transpose_122, y = transpose_123)[name = tensor("matmul_13_cast_fp16")]; + tensor softmax_13_axis_0 = const()[name = tensor("softmax_13_axis_0"), val = tensor(-1)]; + tensor softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = matmul_13_cast_fp16)[name = tensor("softmax_13_cast_fp16")]; + tensor attns_53_transpose_x_0 = const()[name = tensor("attns_53_transpose_x_0"), val = tensor(false)]; + tensor attns_53_transpose_y_0 = const()[name = tensor("attns_53_transpose_y_0"), val = tensor(false)]; + tensor v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = v_53_cast_fp16)[name = tensor("transpose_187")]; + tensor attns_53_cast_fp16 = matmul(transpose_x = attns_53_transpose_x_0, transpose_y = attns_53_transpose_y_0, x = softmax_13_cast_fp16, y = v_55_cast_fp16)[name = tensor("attns_53_cast_fp16")]; + tensor attns_55_perm_0 = const()[name = tensor("attns_55_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_55x = const()[name = tensor("concat_55x"), val = tensor([1, -1, 1024])]; + tensor attns_55_cast_fp16 = transpose(perm = attns_55_perm_0, x = attns_53_cast_fp16)[name = tensor("transpose_184")]; + tensor x_137_cast_fp16 = reshape(shape = concat_55x, x = attns_55_cast_fp16)[name = tensor("x_137_cast_fp16")]; + tensor model_encoder_layers_13_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360071872)))]; + tensor model_encoder_layers_13_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_13_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362169088)))]; + tensor linear_82_cast_fp16 = linear(bias = model_encoder_layers_13_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_13_self_attn_output_proj_weight_to_fp16, x = x_137_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = linear_82_cast_fp16, y = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor x_139_axes_0 = const()[name = tensor("x_139_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_13_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_13_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362171200)))]; + tensor model_encoder_layers_13_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_13_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362173312)))]; + tensor x_139_cast_fp16 = layer_norm(axes = x_139_axes_0, beta = model_encoder_layers_13_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_13_ffn_layer_norm_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor model_encoder_layers_13_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_13_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362175424)))]; + tensor model_encoder_layers_13_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_13_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370564096)))]; + tensor linear_83_cast_fp16 = linear(bias = model_encoder_layers_13_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_13_ffn_inner_proj_weight_to_fp16, x = x_139_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_83_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor model_encoder_layers_13_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_13_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370572352)))]; + tensor model_encoder_layers_13_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_13_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378961024)))]; + tensor linear_84_cast_fp16 = linear(bias = model_encoder_layers_13_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_13_ffn_output_proj_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = linear_84_cast_fp16, y = input_153_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor x_143_axes_0 = const()[name = tensor("x_143_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_14_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378963136)))]; + tensor model_encoder_layers_14_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378965248)))]; + tensor x_143_cast_fp16 = layer_norm(axes = x_143_axes_0, beta = model_encoder_layers_14_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_14_self_attn_layer_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor model_encoder_layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378967360)))]; + tensor model_encoder_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381064576)))]; + tensor linear_85_cast_fp16 = linear(bias = model_encoder_layers_14_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_14_self_attn_q_proj_weight_to_fp16, x = x_143_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor concat_56x = const()[name = tensor("concat_56x"), val = tensor([1, -1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = concat_56x, x = linear_85_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor model_encoder_layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381066688)))]; + tensor model_encoder_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383163904)))]; + tensor linear_86_cast_fp16 = linear(bias = model_encoder_layers_14_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_14_self_attn_k_proj_weight_to_fp16, x = x_143_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor model_encoder_layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383166016)))]; + tensor model_encoder_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385263232)))]; + tensor linear_87_cast_fp16 = linear(bias = model_encoder_layers_14_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_14_self_attn_v_proj_weight_to_fp16, x = x_143_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor concat_57x = const()[name = tensor("concat_57x"), val = tensor([1, -1, 16, 64])]; + tensor k_57_cast_fp16 = reshape(shape = concat_57x, x = linear_86_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor concat_58x = const()[name = tensor("concat_58x"), val = tensor([1, -1, 16, 64])]; + tensor v_57_cast_fp16 = reshape(shape = concat_58x, x = linear_87_cast_fp16)[name = tensor("v_57_cast_fp16")]; + tensor v_59_perm_0 = const()[name = tensor("v_59_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_14_y_0_to_fp16 = const()[name = tensor("mul_14_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_14_cast_fp16 = mul(x = q_57_cast_fp16, y = mul_14_y_0_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor matmul_14_transpose_y_0 = const()[name = tensor("matmul_14_transpose_y_0"), val = tensor(true)]; + tensor matmul_14_transpose_x_0 = const()[name = tensor("matmul_14_transpose_x_0"), val = tensor(false)]; + tensor transpose_124_perm_0 = const()[name = tensor("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = tensor("transpose_125_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = mul_14_cast_fp16)[name = tensor("transpose_182")]; + tensor matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = transpose_124, y = transpose_125)[name = tensor("matmul_14_cast_fp16")]; + tensor softmax_14_axis_0 = const()[name = tensor("softmax_14_axis_0"), val = tensor(-1)]; + tensor softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = matmul_14_cast_fp16)[name = tensor("softmax_14_cast_fp16")]; + tensor attns_57_transpose_x_0 = const()[name = tensor("attns_57_transpose_x_0"), val = tensor(false)]; + tensor attns_57_transpose_y_0 = const()[name = tensor("attns_57_transpose_y_0"), val = tensor(false)]; + tensor v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = v_57_cast_fp16)[name = tensor("transpose_183")]; + tensor attns_57_cast_fp16 = matmul(transpose_x = attns_57_transpose_x_0, transpose_y = attns_57_transpose_y_0, x = softmax_14_cast_fp16, y = v_59_cast_fp16)[name = tensor("attns_57_cast_fp16")]; + tensor attns_59_perm_0 = const()[name = tensor("attns_59_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_59x = const()[name = tensor("concat_59x"), val = tensor([1, -1, 1024])]; + tensor attns_59_cast_fp16 = transpose(perm = attns_59_perm_0, x = attns_57_cast_fp16)[name = tensor("transpose_180")]; + tensor x_145_cast_fp16 = reshape(shape = concat_59x, x = attns_59_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor model_encoder_layers_14_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385265344)))]; + tensor model_encoder_layers_14_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_14_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387362560)))]; + tensor linear_88_cast_fp16 = linear(bias = model_encoder_layers_14_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_14_self_attn_output_proj_weight_to_fp16, x = x_145_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = linear_88_cast_fp16, y = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor x_147_axes_0 = const()[name = tensor("x_147_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_14_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_14_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387364672)))]; + tensor model_encoder_layers_14_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_14_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387366784)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = model_encoder_layers_14_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_14_ffn_layer_norm_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("x_147_cast_fp16")]; + tensor model_encoder_layers_14_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_14_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387368896)))]; + tensor model_encoder_layers_14_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_14_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395757568)))]; + tensor linear_89_cast_fp16 = linear(bias = model_encoder_layers_14_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_14_ffn_inner_proj_weight_to_fp16, x = x_147_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor input_165_mode_0 = const()[name = tensor("input_165_mode_0"), val = tensor("EXACT")]; + tensor input_165_cast_fp16 = gelu(mode = input_165_mode_0, x = linear_89_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor model_encoder_layers_14_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_14_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395765824)))]; + tensor model_encoder_layers_14_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_14_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404154496)))]; + tensor linear_90_cast_fp16 = linear(bias = model_encoder_layers_14_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_14_ffn_output_proj_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_167_cast_fp16 = add(x = linear_90_cast_fp16, y = input_161_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor x_151_axes_0 = const()[name = tensor("x_151_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_15_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404156608)))]; + tensor model_encoder_layers_15_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404158720)))]; + tensor x_151_cast_fp16 = layer_norm(axes = x_151_axes_0, beta = model_encoder_layers_15_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_15_self_attn_layer_norm_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor model_encoder_layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404160832)))]; + tensor model_encoder_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406258048)))]; + tensor linear_91_cast_fp16 = linear(bias = model_encoder_layers_15_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_15_self_attn_q_proj_weight_to_fp16, x = x_151_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor concat_60x = const()[name = tensor("concat_60x"), val = tensor([1, -1, 16, 64])]; + tensor q_61_cast_fp16 = reshape(shape = concat_60x, x = linear_91_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor model_encoder_layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406260160)))]; + tensor model_encoder_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408357376)))]; + tensor linear_92_cast_fp16 = linear(bias = model_encoder_layers_15_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_15_self_attn_k_proj_weight_to_fp16, x = x_151_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor model_encoder_layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408359488)))]; + tensor model_encoder_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410456704)))]; + tensor linear_93_cast_fp16 = linear(bias = model_encoder_layers_15_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_15_self_attn_v_proj_weight_to_fp16, x = x_151_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor concat_61x = const()[name = tensor("concat_61x"), val = tensor([1, -1, 16, 64])]; + tensor k_61_cast_fp16 = reshape(shape = concat_61x, x = linear_92_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor concat_62x = const()[name = tensor("concat_62x"), val = tensor([1, -1, 16, 64])]; + tensor v_61_cast_fp16 = reshape(shape = concat_62x, x = linear_93_cast_fp16)[name = tensor("v_61_cast_fp16")]; + tensor v_63_perm_0 = const()[name = tensor("v_63_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_15_y_0_to_fp16 = const()[name = tensor("mul_15_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_15_cast_fp16 = mul(x = q_61_cast_fp16, y = mul_15_y_0_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor matmul_15_transpose_y_0 = const()[name = tensor("matmul_15_transpose_y_0"), val = tensor(true)]; + tensor matmul_15_transpose_x_0 = const()[name = tensor("matmul_15_transpose_x_0"), val = tensor(false)]; + tensor transpose_126_perm_0 = const()[name = tensor("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = tensor("transpose_127_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = mul_15_cast_fp16)[name = tensor("transpose_178")]; + tensor matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = transpose_126, y = transpose_127)[name = tensor("matmul_15_cast_fp16")]; + tensor softmax_15_axis_0 = const()[name = tensor("softmax_15_axis_0"), val = tensor(-1)]; + tensor softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = matmul_15_cast_fp16)[name = tensor("softmax_15_cast_fp16")]; + tensor attns_61_transpose_x_0 = const()[name = tensor("attns_61_transpose_x_0"), val = tensor(false)]; + tensor attns_61_transpose_y_0 = const()[name = tensor("attns_61_transpose_y_0"), val = tensor(false)]; + tensor v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = v_61_cast_fp16)[name = tensor("transpose_179")]; + tensor attns_61_cast_fp16 = matmul(transpose_x = attns_61_transpose_x_0, transpose_y = attns_61_transpose_y_0, x = softmax_15_cast_fp16, y = v_63_cast_fp16)[name = tensor("attns_61_cast_fp16")]; + tensor attns_63_perm_0 = const()[name = tensor("attns_63_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_63x = const()[name = tensor("concat_63x"), val = tensor([1, -1, 1024])]; + tensor attns_63_cast_fp16 = transpose(perm = attns_63_perm_0, x = attns_61_cast_fp16)[name = tensor("transpose_176")]; + tensor x_153_cast_fp16 = reshape(shape = concat_63x, x = attns_63_cast_fp16)[name = tensor("x_153_cast_fp16")]; + tensor model_encoder_layers_15_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410458816)))]; + tensor model_encoder_layers_15_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_15_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412556032)))]; + tensor linear_94_cast_fp16 = linear(bias = model_encoder_layers_15_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_15_self_attn_output_proj_weight_to_fp16, x = x_153_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = linear_94_cast_fp16, y = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor x_155_axes_0 = const()[name = tensor("x_155_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_15_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_15_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412558144)))]; + tensor model_encoder_layers_15_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_15_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412560256)))]; + tensor x_155_cast_fp16 = layer_norm(axes = x_155_axes_0, beta = model_encoder_layers_15_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_15_ffn_layer_norm_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor model_encoder_layers_15_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_15_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412562368)))]; + tensor model_encoder_layers_15_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_15_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420951040)))]; + tensor linear_95_cast_fp16 = linear(bias = model_encoder_layers_15_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_15_ffn_inner_proj_weight_to_fp16, x = x_155_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_173_mode_0 = const()[name = tensor("input_173_mode_0"), val = tensor("EXACT")]; + tensor input_173_cast_fp16 = gelu(mode = input_173_mode_0, x = linear_95_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor model_encoder_layers_15_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_15_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420959296)))]; + tensor model_encoder_layers_15_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_15_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429347968)))]; + tensor linear_96_cast_fp16 = linear(bias = model_encoder_layers_15_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_15_ffn_output_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_175_cast_fp16 = add(x = linear_96_cast_fp16, y = input_169_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor x_159_axes_0 = const()[name = tensor("x_159_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_16_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429350080)))]; + tensor model_encoder_layers_16_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429352192)))]; + tensor x_159_cast_fp16 = layer_norm(axes = x_159_axes_0, beta = model_encoder_layers_16_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_16_self_attn_layer_norm_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("x_159_cast_fp16")]; + tensor model_encoder_layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429354304)))]; + tensor model_encoder_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431451520)))]; + tensor linear_97_cast_fp16 = linear(bias = model_encoder_layers_16_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_16_self_attn_q_proj_weight_to_fp16, x = x_159_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor concat_64x = const()[name = tensor("concat_64x"), val = tensor([1, -1, 16, 64])]; + tensor q_65_cast_fp16 = reshape(shape = concat_64x, x = linear_97_cast_fp16)[name = tensor("q_65_cast_fp16")]; + tensor model_encoder_layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431453632)))]; + tensor model_encoder_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433550848)))]; + tensor linear_98_cast_fp16 = linear(bias = model_encoder_layers_16_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_16_self_attn_k_proj_weight_to_fp16, x = x_159_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor model_encoder_layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433552960)))]; + tensor model_encoder_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435650176)))]; + tensor linear_99_cast_fp16 = linear(bias = model_encoder_layers_16_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_16_self_attn_v_proj_weight_to_fp16, x = x_159_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor concat_65x = const()[name = tensor("concat_65x"), val = tensor([1, -1, 16, 64])]; + tensor k_65_cast_fp16 = reshape(shape = concat_65x, x = linear_98_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor concat_66x = const()[name = tensor("concat_66x"), val = tensor([1, -1, 16, 64])]; + tensor v_65_cast_fp16 = reshape(shape = concat_66x, x = linear_99_cast_fp16)[name = tensor("v_65_cast_fp16")]; + tensor v_67_perm_0 = const()[name = tensor("v_67_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_16_y_0_to_fp16 = const()[name = tensor("mul_16_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_16_cast_fp16 = mul(x = q_65_cast_fp16, y = mul_16_y_0_to_fp16)[name = tensor("mul_16_cast_fp16")]; + tensor matmul_16_transpose_y_0 = const()[name = tensor("matmul_16_transpose_y_0"), val = tensor(true)]; + tensor matmul_16_transpose_x_0 = const()[name = tensor("matmul_16_transpose_x_0"), val = tensor(false)]; + tensor transpose_128_perm_0 = const()[name = tensor("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = tensor("transpose_129_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = mul_16_cast_fp16)[name = tensor("transpose_174")]; + tensor matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = transpose_128, y = transpose_129)[name = tensor("matmul_16_cast_fp16")]; + tensor softmax_16_axis_0 = const()[name = tensor("softmax_16_axis_0"), val = tensor(-1)]; + tensor softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = matmul_16_cast_fp16)[name = tensor("softmax_16_cast_fp16")]; + tensor attns_65_transpose_x_0 = const()[name = tensor("attns_65_transpose_x_0"), val = tensor(false)]; + tensor attns_65_transpose_y_0 = const()[name = tensor("attns_65_transpose_y_0"), val = tensor(false)]; + tensor v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = v_65_cast_fp16)[name = tensor("transpose_175")]; + tensor attns_65_cast_fp16 = matmul(transpose_x = attns_65_transpose_x_0, transpose_y = attns_65_transpose_y_0, x = softmax_16_cast_fp16, y = v_67_cast_fp16)[name = tensor("attns_65_cast_fp16")]; + tensor attns_67_perm_0 = const()[name = tensor("attns_67_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_67x = const()[name = tensor("concat_67x"), val = tensor([1, -1, 1024])]; + tensor attns_67_cast_fp16 = transpose(perm = attns_67_perm_0, x = attns_65_cast_fp16)[name = tensor("transpose_172")]; + tensor x_161_cast_fp16 = reshape(shape = concat_67x, x = attns_67_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor model_encoder_layers_16_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435652288)))]; + tensor model_encoder_layers_16_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_16_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437749504)))]; + tensor linear_100_cast_fp16 = linear(bias = model_encoder_layers_16_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_16_self_attn_output_proj_weight_to_fp16, x = x_161_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = linear_100_cast_fp16, y = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor x_163_axes_0 = const()[name = tensor("x_163_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_16_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_16_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437751616)))]; + tensor model_encoder_layers_16_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_16_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437753728)))]; + tensor x_163_cast_fp16 = layer_norm(axes = x_163_axes_0, beta = model_encoder_layers_16_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_16_ffn_layer_norm_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor model_encoder_layers_16_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_16_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437755840)))]; + tensor model_encoder_layers_16_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_16_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446144512)))]; + tensor linear_101_cast_fp16 = linear(bias = model_encoder_layers_16_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_16_ffn_inner_proj_weight_to_fp16, x = x_163_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_101_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor model_encoder_layers_16_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_16_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446152768)))]; + tensor model_encoder_layers_16_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_16_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454541440)))]; + tensor linear_102_cast_fp16 = linear(bias = model_encoder_layers_16_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_16_ffn_output_proj_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = linear_102_cast_fp16, y = input_177_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor x_167_axes_0 = const()[name = tensor("x_167_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_17_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454543552)))]; + tensor model_encoder_layers_17_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454545664)))]; + tensor x_167_cast_fp16 = layer_norm(axes = x_167_axes_0, beta = model_encoder_layers_17_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_17_self_attn_layer_norm_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor model_encoder_layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454547776)))]; + tensor model_encoder_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456644992)))]; + tensor linear_103_cast_fp16 = linear(bias = model_encoder_layers_17_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_17_self_attn_q_proj_weight_to_fp16, x = x_167_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor concat_68x = const()[name = tensor("concat_68x"), val = tensor([1, -1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = concat_68x, x = linear_103_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor model_encoder_layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456647104)))]; + tensor model_encoder_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458744320)))]; + tensor linear_104_cast_fp16 = linear(bias = model_encoder_layers_17_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_17_self_attn_k_proj_weight_to_fp16, x = x_167_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor model_encoder_layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458746432)))]; + tensor model_encoder_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460843648)))]; + tensor linear_105_cast_fp16 = linear(bias = model_encoder_layers_17_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_17_self_attn_v_proj_weight_to_fp16, x = x_167_cast_fp16)[name = tensor("linear_105_cast_fp16")]; + tensor concat_69x = const()[name = tensor("concat_69x"), val = tensor([1, -1, 16, 64])]; + tensor k_69_cast_fp16 = reshape(shape = concat_69x, x = linear_104_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor concat_70x = const()[name = tensor("concat_70x"), val = tensor([1, -1, 16, 64])]; + tensor v_69_cast_fp16 = reshape(shape = concat_70x, x = linear_105_cast_fp16)[name = tensor("v_69_cast_fp16")]; + tensor v_71_perm_0 = const()[name = tensor("v_71_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_17_y_0_to_fp16 = const()[name = tensor("mul_17_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_17_cast_fp16 = mul(x = q_69_cast_fp16, y = mul_17_y_0_to_fp16)[name = tensor("mul_17_cast_fp16")]; + tensor matmul_17_transpose_y_0 = const()[name = tensor("matmul_17_transpose_y_0"), val = tensor(true)]; + tensor matmul_17_transpose_x_0 = const()[name = tensor("matmul_17_transpose_x_0"), val = tensor(false)]; + tensor transpose_130_perm_0 = const()[name = tensor("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = tensor("transpose_131_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = mul_17_cast_fp16)[name = tensor("transpose_170")]; + tensor matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = transpose_130, y = transpose_131)[name = tensor("matmul_17_cast_fp16")]; + tensor softmax_17_axis_0 = const()[name = tensor("softmax_17_axis_0"), val = tensor(-1)]; + tensor softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = matmul_17_cast_fp16)[name = tensor("softmax_17_cast_fp16")]; + tensor attns_69_transpose_x_0 = const()[name = tensor("attns_69_transpose_x_0"), val = tensor(false)]; + tensor attns_69_transpose_y_0 = const()[name = tensor("attns_69_transpose_y_0"), val = tensor(false)]; + tensor v_71_cast_fp16 = transpose(perm = v_71_perm_0, x = v_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attns_69_cast_fp16 = matmul(transpose_x = attns_69_transpose_x_0, transpose_y = attns_69_transpose_y_0, x = softmax_17_cast_fp16, y = v_71_cast_fp16)[name = tensor("attns_69_cast_fp16")]; + tensor attns_71_perm_0 = const()[name = tensor("attns_71_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_71x = const()[name = tensor("concat_71x"), val = tensor([1, -1, 1024])]; + tensor attns_71_cast_fp16 = transpose(perm = attns_71_perm_0, x = attns_69_cast_fp16)[name = tensor("transpose_168")]; + tensor x_169_cast_fp16 = reshape(shape = concat_71x, x = attns_71_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor model_encoder_layers_17_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460845760)))]; + tensor model_encoder_layers_17_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_17_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462942976)))]; + tensor linear_106_cast_fp16 = linear(bias = model_encoder_layers_17_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_17_self_attn_output_proj_weight_to_fp16, x = x_169_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = linear_106_cast_fp16, y = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor x_171_axes_0 = const()[name = tensor("x_171_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_17_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_17_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462945088)))]; + tensor model_encoder_layers_17_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_17_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462947200)))]; + tensor x_171_cast_fp16 = layer_norm(axes = x_171_axes_0, beta = model_encoder_layers_17_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_17_ffn_layer_norm_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("x_171_cast_fp16")]; + tensor model_encoder_layers_17_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_17_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462949312)))]; + tensor model_encoder_layers_17_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_17_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471337984)))]; + tensor linear_107_cast_fp16 = linear(bias = model_encoder_layers_17_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_17_ffn_inner_proj_weight_to_fp16, x = x_171_cast_fp16)[name = tensor("linear_107_cast_fp16")]; + tensor input_189_mode_0 = const()[name = tensor("input_189_mode_0"), val = tensor("EXACT")]; + tensor input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = linear_107_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor model_encoder_layers_17_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_17_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471346240)))]; + tensor model_encoder_layers_17_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_17_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479734912)))]; + tensor linear_108_cast_fp16 = linear(bias = model_encoder_layers_17_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_17_ffn_output_proj_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor input_191_cast_fp16 = add(x = linear_108_cast_fp16, y = input_185_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor x_175_axes_0 = const()[name = tensor("x_175_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_18_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479737024)))]; + tensor model_encoder_layers_18_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479739136)))]; + tensor x_175_cast_fp16 = layer_norm(axes = x_175_axes_0, beta = model_encoder_layers_18_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_18_self_attn_layer_norm_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor model_encoder_layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479741248)))]; + tensor model_encoder_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481838464)))]; + tensor linear_109_cast_fp16 = linear(bias = model_encoder_layers_18_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_18_self_attn_q_proj_weight_to_fp16, x = x_175_cast_fp16)[name = tensor("linear_109_cast_fp16")]; + tensor concat_72x = const()[name = tensor("concat_72x"), val = tensor([1, -1, 16, 64])]; + tensor q_73_cast_fp16 = reshape(shape = concat_72x, x = linear_109_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor model_encoder_layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481840576)))]; + tensor model_encoder_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483937792)))]; + tensor linear_110_cast_fp16 = linear(bias = model_encoder_layers_18_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_18_self_attn_k_proj_weight_to_fp16, x = x_175_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor model_encoder_layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483939904)))]; + tensor model_encoder_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486037120)))]; + tensor linear_111_cast_fp16 = linear(bias = model_encoder_layers_18_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_18_self_attn_v_proj_weight_to_fp16, x = x_175_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor concat_73x = const()[name = tensor("concat_73x"), val = tensor([1, -1, 16, 64])]; + tensor k_73_cast_fp16 = reshape(shape = concat_73x, x = linear_110_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor concat_74x = const()[name = tensor("concat_74x"), val = tensor([1, -1, 16, 64])]; + tensor v_73_cast_fp16 = reshape(shape = concat_74x, x = linear_111_cast_fp16)[name = tensor("v_73_cast_fp16")]; + tensor v_75_perm_0 = const()[name = tensor("v_75_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_18_y_0_to_fp16 = const()[name = tensor("mul_18_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_18_cast_fp16 = mul(x = q_73_cast_fp16, y = mul_18_y_0_to_fp16)[name = tensor("mul_18_cast_fp16")]; + tensor matmul_18_transpose_y_0 = const()[name = tensor("matmul_18_transpose_y_0"), val = tensor(true)]; + tensor matmul_18_transpose_x_0 = const()[name = tensor("matmul_18_transpose_x_0"), val = tensor(false)]; + tensor transpose_132_perm_0 = const()[name = tensor("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = tensor("transpose_133_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = mul_18_cast_fp16)[name = tensor("transpose_166")]; + tensor matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = transpose_132, y = transpose_133)[name = tensor("matmul_18_cast_fp16")]; + tensor softmax_18_axis_0 = const()[name = tensor("softmax_18_axis_0"), val = tensor(-1)]; + tensor softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = matmul_18_cast_fp16)[name = tensor("softmax_18_cast_fp16")]; + tensor attns_73_transpose_x_0 = const()[name = tensor("attns_73_transpose_x_0"), val = tensor(false)]; + tensor attns_73_transpose_y_0 = const()[name = tensor("attns_73_transpose_y_0"), val = tensor(false)]; + tensor v_75_cast_fp16 = transpose(perm = v_75_perm_0, x = v_73_cast_fp16)[name = tensor("transpose_167")]; + tensor attns_73_cast_fp16 = matmul(transpose_x = attns_73_transpose_x_0, transpose_y = attns_73_transpose_y_0, x = softmax_18_cast_fp16, y = v_75_cast_fp16)[name = tensor("attns_73_cast_fp16")]; + tensor attns_75_perm_0 = const()[name = tensor("attns_75_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_75x = const()[name = tensor("concat_75x"), val = tensor([1, -1, 1024])]; + tensor attns_75_cast_fp16 = transpose(perm = attns_75_perm_0, x = attns_73_cast_fp16)[name = tensor("transpose_164")]; + tensor x_177_cast_fp16 = reshape(shape = concat_75x, x = attns_75_cast_fp16)[name = tensor("x_177_cast_fp16")]; + tensor model_encoder_layers_18_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486039232)))]; + tensor model_encoder_layers_18_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_18_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488136448)))]; + tensor linear_112_cast_fp16 = linear(bias = model_encoder_layers_18_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_18_self_attn_output_proj_weight_to_fp16, x = x_177_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = linear_112_cast_fp16, y = input_191_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor x_179_axes_0 = const()[name = tensor("x_179_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_18_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_18_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488138560)))]; + tensor model_encoder_layers_18_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_18_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488140672)))]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = model_encoder_layers_18_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_18_ffn_layer_norm_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("x_179_cast_fp16")]; + tensor model_encoder_layers_18_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_18_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488142784)))]; + tensor model_encoder_layers_18_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_18_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496531456)))]; + tensor linear_113_cast_fp16 = linear(bias = model_encoder_layers_18_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_18_ffn_inner_proj_weight_to_fp16, x = x_179_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor input_197_mode_0 = const()[name = tensor("input_197_mode_0"), val = tensor("EXACT")]; + tensor input_197_cast_fp16 = gelu(mode = input_197_mode_0, x = linear_113_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor model_encoder_layers_18_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_18_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496539712)))]; + tensor model_encoder_layers_18_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_18_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504928384)))]; + tensor linear_114_cast_fp16 = linear(bias = model_encoder_layers_18_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_18_ffn_output_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_199_cast_fp16 = add(x = linear_114_cast_fp16, y = input_193_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor x_183_axes_0 = const()[name = tensor("x_183_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_19_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504930496)))]; + tensor model_encoder_layers_19_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504932608)))]; + tensor x_183_cast_fp16 = layer_norm(axes = x_183_axes_0, beta = model_encoder_layers_19_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_19_self_attn_layer_norm_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor model_encoder_layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504934720)))]; + tensor model_encoder_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507031936)))]; + tensor linear_115_cast_fp16 = linear(bias = model_encoder_layers_19_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_19_self_attn_q_proj_weight_to_fp16, x = x_183_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor concat_76x = const()[name = tensor("concat_76x"), val = tensor([1, -1, 16, 64])]; + tensor q_77_cast_fp16 = reshape(shape = concat_76x, x = linear_115_cast_fp16)[name = tensor("q_77_cast_fp16")]; + tensor model_encoder_layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507034048)))]; + tensor model_encoder_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509131264)))]; + tensor linear_116_cast_fp16 = linear(bias = model_encoder_layers_19_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_19_self_attn_k_proj_weight_to_fp16, x = x_183_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor model_encoder_layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509133376)))]; + tensor model_encoder_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511230592)))]; + tensor linear_117_cast_fp16 = linear(bias = model_encoder_layers_19_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_19_self_attn_v_proj_weight_to_fp16, x = x_183_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor concat_77x = const()[name = tensor("concat_77x"), val = tensor([1, -1, 16, 64])]; + tensor k_77_cast_fp16 = reshape(shape = concat_77x, x = linear_116_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor concat_78x = const()[name = tensor("concat_78x"), val = tensor([1, -1, 16, 64])]; + tensor v_77_cast_fp16 = reshape(shape = concat_78x, x = linear_117_cast_fp16)[name = tensor("v_77_cast_fp16")]; + tensor v_79_perm_0 = const()[name = tensor("v_79_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_19_y_0_to_fp16 = const()[name = tensor("mul_19_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_19_cast_fp16 = mul(x = q_77_cast_fp16, y = mul_19_y_0_to_fp16)[name = tensor("mul_19_cast_fp16")]; + tensor matmul_19_transpose_y_0 = const()[name = tensor("matmul_19_transpose_y_0"), val = tensor(true)]; + tensor matmul_19_transpose_x_0 = const()[name = tensor("matmul_19_transpose_x_0"), val = tensor(false)]; + tensor transpose_134_perm_0 = const()[name = tensor("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = tensor("transpose_135_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = mul_19_cast_fp16)[name = tensor("transpose_162")]; + tensor matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = transpose_134, y = transpose_135)[name = tensor("matmul_19_cast_fp16")]; + tensor softmax_19_axis_0 = const()[name = tensor("softmax_19_axis_0"), val = tensor(-1)]; + tensor softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = matmul_19_cast_fp16)[name = tensor("softmax_19_cast_fp16")]; + tensor attns_77_transpose_x_0 = const()[name = tensor("attns_77_transpose_x_0"), val = tensor(false)]; + tensor attns_77_transpose_y_0 = const()[name = tensor("attns_77_transpose_y_0"), val = tensor(false)]; + tensor v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = v_77_cast_fp16)[name = tensor("transpose_163")]; + tensor attns_77_cast_fp16 = matmul(transpose_x = attns_77_transpose_x_0, transpose_y = attns_77_transpose_y_0, x = softmax_19_cast_fp16, y = v_79_cast_fp16)[name = tensor("attns_77_cast_fp16")]; + tensor attns_79_perm_0 = const()[name = tensor("attns_79_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_79x = const()[name = tensor("concat_79x"), val = tensor([1, -1, 1024])]; + tensor attns_79_cast_fp16 = transpose(perm = attns_79_perm_0, x = attns_77_cast_fp16)[name = tensor("transpose_160")]; + tensor x_185_cast_fp16 = reshape(shape = concat_79x, x = attns_79_cast_fp16)[name = tensor("x_185_cast_fp16")]; + tensor model_encoder_layers_19_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511232704)))]; + tensor model_encoder_layers_19_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_19_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513329920)))]; + tensor linear_118_cast_fp16 = linear(bias = model_encoder_layers_19_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_19_self_attn_output_proj_weight_to_fp16, x = x_185_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = linear_118_cast_fp16, y = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor x_187_axes_0 = const()[name = tensor("x_187_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_19_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_19_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513332032)))]; + tensor model_encoder_layers_19_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_19_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513334144)))]; + tensor x_187_cast_fp16 = layer_norm(axes = x_187_axes_0, beta = model_encoder_layers_19_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_19_ffn_layer_norm_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor model_encoder_layers_19_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_19_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513336256)))]; + tensor model_encoder_layers_19_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_19_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521724928)))]; + tensor linear_119_cast_fp16 = linear(bias = model_encoder_layers_19_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_19_ffn_inner_proj_weight_to_fp16, x = x_187_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_119_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor model_encoder_layers_19_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_19_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521733184)))]; + tensor model_encoder_layers_19_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_19_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530121856)))]; + tensor linear_120_cast_fp16 = linear(bias = model_encoder_layers_19_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_19_ffn_output_proj_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = linear_120_cast_fp16, y = input_201_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor x_191_axes_0 = const()[name = tensor("x_191_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_20_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530123968)))]; + tensor model_encoder_layers_20_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530126080)))]; + tensor x_191_cast_fp16 = layer_norm(axes = x_191_axes_0, beta = model_encoder_layers_20_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_20_self_attn_layer_norm_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("x_191_cast_fp16")]; + tensor model_encoder_layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530128192)))]; + tensor model_encoder_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532225408)))]; + tensor linear_121_cast_fp16 = linear(bias = model_encoder_layers_20_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_20_self_attn_q_proj_weight_to_fp16, x = x_191_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor concat_80x = const()[name = tensor("concat_80x"), val = tensor([1, -1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = concat_80x, x = linear_121_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor model_encoder_layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532227520)))]; + tensor model_encoder_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534324736)))]; + tensor linear_122_cast_fp16 = linear(bias = model_encoder_layers_20_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_20_self_attn_k_proj_weight_to_fp16, x = x_191_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor model_encoder_layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534326848)))]; + tensor model_encoder_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536424064)))]; + tensor linear_123_cast_fp16 = linear(bias = model_encoder_layers_20_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_20_self_attn_v_proj_weight_to_fp16, x = x_191_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor concat_81x = const()[name = tensor("concat_81x"), val = tensor([1, -1, 16, 64])]; + tensor k_81_cast_fp16 = reshape(shape = concat_81x, x = linear_122_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor concat_82x = const()[name = tensor("concat_82x"), val = tensor([1, -1, 16, 64])]; + tensor v_81_cast_fp16 = reshape(shape = concat_82x, x = linear_123_cast_fp16)[name = tensor("v_81_cast_fp16")]; + tensor v_83_perm_0 = const()[name = tensor("v_83_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_20_y_0_to_fp16 = const()[name = tensor("mul_20_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_20_cast_fp16 = mul(x = q_81_cast_fp16, y = mul_20_y_0_to_fp16)[name = tensor("mul_20_cast_fp16")]; + tensor matmul_20_transpose_y_0 = const()[name = tensor("matmul_20_transpose_y_0"), val = tensor(true)]; + tensor matmul_20_transpose_x_0 = const()[name = tensor("matmul_20_transpose_x_0"), val = tensor(false)]; + tensor transpose_136_perm_0 = const()[name = tensor("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = tensor("transpose_137_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = mul_20_cast_fp16)[name = tensor("transpose_158")]; + tensor matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = transpose_136, y = transpose_137)[name = tensor("matmul_20_cast_fp16")]; + tensor softmax_20_axis_0 = const()[name = tensor("softmax_20_axis_0"), val = tensor(-1)]; + tensor softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = matmul_20_cast_fp16)[name = tensor("softmax_20_cast_fp16")]; + tensor attns_81_transpose_x_0 = const()[name = tensor("attns_81_transpose_x_0"), val = tensor(false)]; + tensor attns_81_transpose_y_0 = const()[name = tensor("attns_81_transpose_y_0"), val = tensor(false)]; + tensor v_83_cast_fp16 = transpose(perm = v_83_perm_0, x = v_81_cast_fp16)[name = tensor("transpose_159")]; + tensor attns_81_cast_fp16 = matmul(transpose_x = attns_81_transpose_x_0, transpose_y = attns_81_transpose_y_0, x = softmax_20_cast_fp16, y = v_83_cast_fp16)[name = tensor("attns_81_cast_fp16")]; + tensor attns_83_perm_0 = const()[name = tensor("attns_83_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_83x = const()[name = tensor("concat_83x"), val = tensor([1, -1, 1024])]; + tensor attns_83_cast_fp16 = transpose(perm = attns_83_perm_0, x = attns_81_cast_fp16)[name = tensor("transpose_156")]; + tensor x_193_cast_fp16 = reshape(shape = concat_83x, x = attns_83_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor model_encoder_layers_20_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536426176)))]; + tensor model_encoder_layers_20_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_20_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538523392)))]; + tensor linear_124_cast_fp16 = linear(bias = model_encoder_layers_20_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_20_self_attn_output_proj_weight_to_fp16, x = x_193_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor input_209_cast_fp16 = add(x = linear_124_cast_fp16, y = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor x_195_axes_0 = const()[name = tensor("x_195_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_20_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_20_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538525504)))]; + tensor model_encoder_layers_20_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_20_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538527616)))]; + tensor x_195_cast_fp16 = layer_norm(axes = x_195_axes_0, beta = model_encoder_layers_20_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_20_ffn_layer_norm_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor model_encoder_layers_20_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_20_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538529728)))]; + tensor model_encoder_layers_20_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_20_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546918400)))]; + tensor linear_125_cast_fp16 = linear(bias = model_encoder_layers_20_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_20_ffn_inner_proj_weight_to_fp16, x = x_195_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor input_213_mode_0 = const()[name = tensor("input_213_mode_0"), val = tensor("EXACT")]; + tensor input_213_cast_fp16 = gelu(mode = input_213_mode_0, x = linear_125_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor model_encoder_layers_20_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_20_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546926656)))]; + tensor model_encoder_layers_20_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_20_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555315328)))]; + tensor linear_126_cast_fp16 = linear(bias = model_encoder_layers_20_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_20_ffn_output_proj_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_215_cast_fp16 = add(x = linear_126_cast_fp16, y = input_209_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor x_199_axes_0 = const()[name = tensor("x_199_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_21_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555317440)))]; + tensor model_encoder_layers_21_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555319552)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = model_encoder_layers_21_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_21_self_attn_layer_norm_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor model_encoder_layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555321664)))]; + tensor model_encoder_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(557418880)))]; + tensor linear_127_cast_fp16 = linear(bias = model_encoder_layers_21_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_21_self_attn_q_proj_weight_to_fp16, x = x_199_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor concat_84x = const()[name = tensor("concat_84x"), val = tensor([1, -1, 16, 64])]; + tensor q_85_cast_fp16 = reshape(shape = concat_84x, x = linear_127_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor model_encoder_layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(557420992)))]; + tensor model_encoder_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559518208)))]; + tensor linear_128_cast_fp16 = linear(bias = model_encoder_layers_21_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_21_self_attn_k_proj_weight_to_fp16, x = x_199_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor model_encoder_layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559520320)))]; + tensor model_encoder_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561617536)))]; + tensor linear_129_cast_fp16 = linear(bias = model_encoder_layers_21_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_21_self_attn_v_proj_weight_to_fp16, x = x_199_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor concat_85x = const()[name = tensor("concat_85x"), val = tensor([1, -1, 16, 64])]; + tensor k_85_cast_fp16 = reshape(shape = concat_85x, x = linear_128_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor concat_86x = const()[name = tensor("concat_86x"), val = tensor([1, -1, 16, 64])]; + tensor v_85_cast_fp16 = reshape(shape = concat_86x, x = linear_129_cast_fp16)[name = tensor("v_85_cast_fp16")]; + tensor v_87_perm_0 = const()[name = tensor("v_87_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_21_y_0_to_fp16 = const()[name = tensor("mul_21_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_21_cast_fp16 = mul(x = q_85_cast_fp16, y = mul_21_y_0_to_fp16)[name = tensor("mul_21_cast_fp16")]; + tensor matmul_21_transpose_y_0 = const()[name = tensor("matmul_21_transpose_y_0"), val = tensor(true)]; + tensor matmul_21_transpose_x_0 = const()[name = tensor("matmul_21_transpose_x_0"), val = tensor(false)]; + tensor transpose_138_perm_0 = const()[name = tensor("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = tensor("transpose_139_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = mul_21_cast_fp16)[name = tensor("transpose_154")]; + tensor matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = transpose_138, y = transpose_139)[name = tensor("matmul_21_cast_fp16")]; + tensor softmax_21_axis_0 = const()[name = tensor("softmax_21_axis_0"), val = tensor(-1)]; + tensor softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = matmul_21_cast_fp16)[name = tensor("softmax_21_cast_fp16")]; + tensor attns_85_transpose_x_0 = const()[name = tensor("attns_85_transpose_x_0"), val = tensor(false)]; + tensor attns_85_transpose_y_0 = const()[name = tensor("attns_85_transpose_y_0"), val = tensor(false)]; + tensor v_87_cast_fp16 = transpose(perm = v_87_perm_0, x = v_85_cast_fp16)[name = tensor("transpose_155")]; + tensor attns_85_cast_fp16 = matmul(transpose_x = attns_85_transpose_x_0, transpose_y = attns_85_transpose_y_0, x = softmax_21_cast_fp16, y = v_87_cast_fp16)[name = tensor("attns_85_cast_fp16")]; + tensor attns_87_perm_0 = const()[name = tensor("attns_87_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_87x = const()[name = tensor("concat_87x"), val = tensor([1, -1, 1024])]; + tensor attns_87_cast_fp16 = transpose(perm = attns_87_perm_0, x = attns_85_cast_fp16)[name = tensor("transpose_152")]; + tensor x_201_cast_fp16 = reshape(shape = concat_87x, x = attns_87_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor model_encoder_layers_21_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561619648)))]; + tensor model_encoder_layers_21_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_21_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563716864)))]; + tensor linear_130_cast_fp16 = linear(bias = model_encoder_layers_21_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_21_self_attn_output_proj_weight_to_fp16, x = x_201_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = linear_130_cast_fp16, y = input_215_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor x_203_axes_0 = const()[name = tensor("x_203_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_21_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_21_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563718976)))]; + tensor model_encoder_layers_21_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_21_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563721088)))]; + tensor x_203_cast_fp16 = layer_norm(axes = x_203_axes_0, beta = model_encoder_layers_21_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_21_ffn_layer_norm_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("x_203_cast_fp16")]; + tensor model_encoder_layers_21_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_21_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563723200)))]; + tensor model_encoder_layers_21_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_21_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572111872)))]; + tensor linear_131_cast_fp16 = linear(bias = model_encoder_layers_21_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_21_ffn_inner_proj_weight_to_fp16, x = x_203_cast_fp16)[name = tensor("linear_131_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = linear_131_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor model_encoder_layers_21_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_21_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572120128)))]; + tensor model_encoder_layers_21_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_21_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580508800)))]; + tensor linear_132_cast_fp16 = linear(bias = model_encoder_layers_21_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_21_ffn_output_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor input_223_cast_fp16 = add(x = linear_132_cast_fp16, y = input_217_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor x_207_axes_0 = const()[name = tensor("x_207_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_22_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580510912)))]; + tensor model_encoder_layers_22_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580513024)))]; + tensor x_207_cast_fp16 = layer_norm(axes = x_207_axes_0, beta = model_encoder_layers_22_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_22_self_attn_layer_norm_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor model_encoder_layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580515136)))]; + tensor model_encoder_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582612352)))]; + tensor linear_133_cast_fp16 = linear(bias = model_encoder_layers_22_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_22_self_attn_q_proj_weight_to_fp16, x = x_207_cast_fp16)[name = tensor("linear_133_cast_fp16")]; + tensor concat_88x = const()[name = tensor("concat_88x"), val = tensor([1, -1, 16, 64])]; + tensor q_89_cast_fp16 = reshape(shape = concat_88x, x = linear_133_cast_fp16)[name = tensor("q_89_cast_fp16")]; + tensor model_encoder_layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582614464)))]; + tensor model_encoder_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584711680)))]; + tensor linear_134_cast_fp16 = linear(bias = model_encoder_layers_22_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_22_self_attn_k_proj_weight_to_fp16, x = x_207_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor model_encoder_layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584713792)))]; + tensor model_encoder_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586811008)))]; + tensor linear_135_cast_fp16 = linear(bias = model_encoder_layers_22_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_22_self_attn_v_proj_weight_to_fp16, x = x_207_cast_fp16)[name = tensor("linear_135_cast_fp16")]; + tensor concat_89x = const()[name = tensor("concat_89x"), val = tensor([1, -1, 16, 64])]; + tensor k_89_cast_fp16 = reshape(shape = concat_89x, x = linear_134_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor concat_90x = const()[name = tensor("concat_90x"), val = tensor([1, -1, 16, 64])]; + tensor v_89_cast_fp16 = reshape(shape = concat_90x, x = linear_135_cast_fp16)[name = tensor("v_89_cast_fp16")]; + tensor v_91_perm_0 = const()[name = tensor("v_91_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_22_y_0_to_fp16 = const()[name = tensor("mul_22_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_22_cast_fp16 = mul(x = q_89_cast_fp16, y = mul_22_y_0_to_fp16)[name = tensor("mul_22_cast_fp16")]; + tensor matmul_22_transpose_y_0 = const()[name = tensor("matmul_22_transpose_y_0"), val = tensor(true)]; + tensor matmul_22_transpose_x_0 = const()[name = tensor("matmul_22_transpose_x_0"), val = tensor(false)]; + tensor transpose_140_perm_0 = const()[name = tensor("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = tensor("transpose_141_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = mul_22_cast_fp16)[name = tensor("transpose_150")]; + tensor matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = transpose_140, y = transpose_141)[name = tensor("matmul_22_cast_fp16")]; + tensor softmax_22_axis_0 = const()[name = tensor("softmax_22_axis_0"), val = tensor(-1)]; + tensor softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = matmul_22_cast_fp16)[name = tensor("softmax_22_cast_fp16")]; + tensor attns_89_transpose_x_0 = const()[name = tensor("attns_89_transpose_x_0"), val = tensor(false)]; + tensor attns_89_transpose_y_0 = const()[name = tensor("attns_89_transpose_y_0"), val = tensor(false)]; + tensor v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = v_89_cast_fp16)[name = tensor("transpose_151")]; + tensor attns_89_cast_fp16 = matmul(transpose_x = attns_89_transpose_x_0, transpose_y = attns_89_transpose_y_0, x = softmax_22_cast_fp16, y = v_91_cast_fp16)[name = tensor("attns_89_cast_fp16")]; + tensor attns_91_perm_0 = const()[name = tensor("attns_91_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_91x = const()[name = tensor("concat_91x"), val = tensor([1, -1, 1024])]; + tensor attns_91_cast_fp16 = transpose(perm = attns_91_perm_0, x = attns_89_cast_fp16)[name = tensor("transpose_148")]; + tensor x_209_cast_fp16 = reshape(shape = concat_91x, x = attns_91_cast_fp16)[name = tensor("x_209_cast_fp16")]; + tensor model_encoder_layers_22_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586813120)))]; + tensor model_encoder_layers_22_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_22_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588910336)))]; + tensor linear_136_cast_fp16 = linear(bias = model_encoder_layers_22_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_22_self_attn_output_proj_weight_to_fp16, x = x_209_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor input_225_cast_fp16 = add(x = linear_136_cast_fp16, y = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor x_211_axes_0 = const()[name = tensor("x_211_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_22_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_22_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588912448)))]; + tensor model_encoder_layers_22_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_22_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588914560)))]; + tensor x_211_cast_fp16 = layer_norm(axes = x_211_axes_0, beta = model_encoder_layers_22_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_22_ffn_layer_norm_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor model_encoder_layers_22_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_22_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588916672)))]; + tensor model_encoder_layers_22_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_22_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597305344)))]; + tensor linear_137_cast_fp16 = linear(bias = model_encoder_layers_22_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_22_ffn_inner_proj_weight_to_fp16, x = x_211_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_137_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor model_encoder_layers_22_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_22_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597313600)))]; + tensor model_encoder_layers_22_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_22_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605702272)))]; + tensor linear_138_cast_fp16 = linear(bias = model_encoder_layers_22_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_22_ffn_output_proj_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = linear_138_cast_fp16, y = input_225_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor x_215_axes_0 = const()[name = tensor("x_215_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_23_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605704384)))]; + tensor model_encoder_layers_23_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605706496)))]; + tensor x_215_cast_fp16 = layer_norm(axes = x_215_axes_0, beta = model_encoder_layers_23_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_23_self_attn_layer_norm_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("x_215_cast_fp16")]; + tensor model_encoder_layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605708608)))]; + tensor model_encoder_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607805824)))]; + tensor linear_139_cast_fp16 = linear(bias = model_encoder_layers_23_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_23_self_attn_q_proj_weight_to_fp16, x = x_215_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor concat_92x = const()[name = tensor("concat_92x"), val = tensor([1, -1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = concat_92x, x = linear_139_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor model_encoder_layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607807936)))]; + tensor model_encoder_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(609905152)))]; + tensor linear_140_cast_fp16 = linear(bias = model_encoder_layers_23_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_23_self_attn_k_proj_weight_to_fp16, x = x_215_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor model_encoder_layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(609907264)))]; + tensor model_encoder_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(612004480)))]; + tensor linear_141_cast_fp16 = linear(bias = model_encoder_layers_23_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_23_self_attn_v_proj_weight_to_fp16, x = x_215_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor concat_93x = const()[name = tensor("concat_93x"), val = tensor([1, -1, 16, 64])]; + tensor k_93_cast_fp16 = reshape(shape = concat_93x, x = linear_140_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor concat_94x = const()[name = tensor("concat_94x"), val = tensor([1, -1, 16, 64])]; + tensor v_93_cast_fp16 = reshape(shape = concat_94x, x = linear_141_cast_fp16)[name = tensor("v_93_cast_fp16")]; + tensor v_perm_0 = const()[name = tensor("v_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor mul_23_y_0_to_fp16 = const()[name = tensor("mul_23_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_23_cast_fp16 = mul(x = q_93_cast_fp16, y = mul_23_y_0_to_fp16)[name = tensor("mul_23_cast_fp16")]; + tensor matmul_23_transpose_y_0 = const()[name = tensor("matmul_23_transpose_y_0"), val = tensor(true)]; + tensor matmul_23_transpose_x_0 = const()[name = tensor("matmul_23_transpose_x_0"), val = tensor(false)]; + tensor transpose_142_perm_0 = const()[name = tensor("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = tensor("transpose_143_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = mul_23_cast_fp16)[name = tensor("transpose_146")]; + tensor matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = transpose_142, y = transpose_143)[name = tensor("matmul_23_cast_fp16")]; + tensor softmax_23_axis_0 = const()[name = tensor("softmax_23_axis_0"), val = tensor(-1)]; + tensor softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = matmul_23_cast_fp16)[name = tensor("softmax_23_cast_fp16")]; + tensor attns_93_transpose_x_0 = const()[name = tensor("attns_93_transpose_x_0"), val = tensor(false)]; + tensor attns_93_transpose_y_0 = const()[name = tensor("attns_93_transpose_y_0"), val = tensor(false)]; + tensor v_cast_fp16 = transpose(perm = v_perm_0, x = v_93_cast_fp16)[name = tensor("transpose_147")]; + tensor attns_93_cast_fp16 = matmul(transpose_x = attns_93_transpose_x_0, transpose_y = attns_93_transpose_y_0, x = softmax_23_cast_fp16, y = v_cast_fp16)[name = tensor("attns_93_cast_fp16")]; + tensor attns_perm_0 = const()[name = tensor("attns_perm_0"), val = tensor([0, -2, -3, 3])]; + tensor concat_95x = const()[name = tensor("concat_95x"), val = tensor([1, -1, 1024])]; + tensor attns_cast_fp16 = transpose(perm = attns_perm_0, x = attns_93_cast_fp16)[name = tensor("transpose_144")]; + tensor x_217_cast_fp16 = reshape(shape = concat_95x, x = attns_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor model_encoder_layers_23_self_attn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(612006592)))]; + tensor model_encoder_layers_23_self_attn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_23_self_attn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614103808)))]; + tensor linear_142_cast_fp16 = linear(bias = model_encoder_layers_23_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_23_self_attn_output_proj_weight_to_fp16, x = x_217_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor input_233_cast_fp16 = add(x = linear_142_cast_fp16, y = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor x_219_axes_0 = const()[name = tensor("x_219_axes_0"), val = tensor([-1])]; + tensor model_encoder_layers_23_ffn_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layers_23_ffn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614105920)))]; + tensor model_encoder_layers_23_ffn_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layers_23_ffn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614108032)))]; + tensor x_219_cast_fp16 = layer_norm(axes = x_219_axes_0, beta = model_encoder_layers_23_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_23_ffn_layer_norm_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor model_encoder_layers_23_ffn_inner_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_23_ffn_inner_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614110144)))]; + tensor model_encoder_layers_23_ffn_inner_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_23_ffn_inner_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622498816)))]; + tensor linear_143_cast_fp16 = linear(bias = model_encoder_layers_23_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_23_ffn_inner_proj_weight_to_fp16, x = x_219_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_237_mode_0 = const()[name = tensor("input_237_mode_0"), val = tensor("EXACT")]; + tensor input_237_cast_fp16 = gelu(mode = input_237_mode_0, x = linear_143_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor model_encoder_layers_23_ffn_output_proj_weight_to_fp16 = const()[name = tensor("model_encoder_layers_23_ffn_output_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622507072)))]; + tensor model_encoder_layers_23_ffn_output_proj_bias_to_fp16 = const()[name = tensor("model_encoder_layers_23_ffn_output_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630895744)))]; + tensor linear_144_cast_fp16 = linear(bias = model_encoder_layers_23_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_23_ffn_output_proj_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor input_cast_fp16 = add(x = linear_144_cast_fp16, y = input_233_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor model_encoder_layer_norm_weight_to_fp16 = const()[name = tensor("model_encoder_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630897856)))]; + tensor model_encoder_layer_norm_bias_to_fp16 = const()[name = tensor("model_encoder_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630899968)))]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = model_encoder_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layer_norm_weight_to_fp16, x = input_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor model_final_proj_weight_to_fp16 = const()[name = tensor("model_final_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630902080)))]; + tensor model_final_proj_bias_to_fp16 = const()[name = tensor("model_final_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650997120)))]; + tensor logits = linear(bias = model_final_proj_bias_to_fp16, weight = model_final_proj_weight_to_fp16, x = x_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + } -> (logits); +} \ No newline at end of file