| program(1.0) |
| [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}})] |
| { |
| func main<ios16>(tensor<fp32, [1, 3, 224, 224]> pixel_values) { |
| tensor<fp32, []> pixel_values__scaled___y_0 = const()[name = tensor<string, []>("pixel_values__scaled___y_0"), val = tensor<fp32, []>(0x1.010102p-8)]; |
| tensor<fp32, [1, 3, 224, 224]> pixel_values__scaled__ = mul(x = pixel_values, y = pixel_values__scaled___y_0)[name = tensor<string, []>("pixel_values__scaled__")]; |
| tensor<int32, []> var_7 = const()[name = tensor<string, []>("op_7"), val = tensor<int32, []>(1)]; |
| tensor<string, []> var_35_pad_type_0 = const()[name = tensor<string, []>("op_35_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> var_35_strides_0 = const()[name = tensor<string, []>("op_35_strides_0"), val = tensor<int32, [2]>([16, 16])]; |
| tensor<int32, [4]> var_35_pad_0 = const()[name = tensor<string, []>("op_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_35_dilations_0 = const()[name = tensor<string, []>("op_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> var_35_groups_0 = const()[name = tensor<string, []>("op_35_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<string, []> pixel_values_to_fp16_dtype_0 = const()[name = tensor<string, []>("pixel_values_to_fp16_dtype_0"), val = tensor<string, []>("fp16")]; |
| tensor<fp16, [768, 3, 16, 16]> model_vit_embeddings_patch_embeddings_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_embeddings_patch_embeddings_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3, 16, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590784))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_embeddings_patch_embeddings_projection_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_embeddings_patch_embeddings_projection_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(592384)))]; |
| tensor<fp16, [1, 3, 224, 224]> pixel_values_to_fp16 = cast(dtype = pixel_values_to_fp16_dtype_0, x = pixel_values__scaled__)[name = tensor<string, []>("cast_1")]; |
| tensor<fp16, [1, 768, 14, 14]> var_35_cast_fp16 = conv(bias = model_vit_embeddings_patch_embeddings_projection_bias_to_fp16, dilations = var_35_dilations_0, groups = var_35_groups_0, pad = var_35_pad_0, pad_type = var_35_pad_type_0, strides = var_35_strides_0, weight = model_vit_embeddings_patch_embeddings_projection_weight_to_fp16_quantized, x = pixel_values_to_fp16)[name = tensor<string, []>("op_35_cast_fp16")]; |
| tensor<int32, [3]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<int32, [3]>([1, 768, 196])]; |
| tensor<fp16, [1, 768, 196]> var_36_cast_fp16 = reshape(shape = concat_0, x = var_35_cast_fp16)[name = tensor<string, []>("op_36_cast_fp16")]; |
| tensor<int32, [3]> embeddings_1_perm_0 = const()[name = tensor<string, []>("embeddings_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<bool, []> embeddings_3_interleave_0 = const()[name = tensor<string, []>("embeddings_3_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 1, 768]> model_vit_embeddings_cls_token_to_fp16 = const()[name = tensor<string, []>("model_vit_embeddings_cls_token_to_fp16"), val = tensor<fp16, [1, 1, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(593984)))]; |
| tensor<fp16, [1, 196, 768]> embeddings_1_cast_fp16 = transpose(perm = embeddings_1_perm_0, x = var_36_cast_fp16)[name = tensor<string, []>("transpose_120")]; |
| tensor<fp16, [1, 197, 768]> embeddings_3_cast_fp16 = concat(axis = var_7, interleave = embeddings_3_interleave_0, values = (model_vit_embeddings_cls_token_to_fp16, embeddings_1_cast_fp16))[name = tensor<string, []>("embeddings_3_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> model_vit_embeddings_position_embeddings_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_embeddings_position_embeddings_to_fp16_quantized"), quantized_data = tensor<int8, [1, 197, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(595584))), scale = tensor<fp16, []>(0x1.164p-7), zero_point = tensor<int8, []>(0)]; |
| tensor<fp16, [1, 197, 768]> input_1_cast_fp16 = add(x = embeddings_3_cast_fp16, y = model_vit_embeddings_position_embeddings_to_fp16_quantized)[name = tensor<string, []>("input_1_cast_fp16")]; |
| tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(746944)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748544)))]; |
| tensor<fp16, []> var_12_to_fp16 = const()[name = tensor<string, []>("op_12_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; |
| tensor<fp16, [1, 197, 768]> input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = model_vit_encoder_layer_0_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_0_layernorm_before_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_0_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_0_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(750144))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1340032))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1341632)))]; |
| tensor<fp16, [1, 197, 768]> linear_0_cast_fp16 = linear(bias = model_vit_encoder_layer_0_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_0_attention_attention_query_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_0_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_0_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1343232))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1933120))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1934720)))]; |
| tensor<fp16, [1, 197, 768]> linear_1_cast_fp16 = linear(bias = model_vit_encoder_layer_0_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_0_attention_attention_key_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")]; |
| tensor<int32, [4]> var_90 = const()[name = tensor<string, []>("op_90"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_3_cast_fp16 = reshape(shape = var_90, x = linear_1_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_0_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_0_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1936320))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2526208))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2527808)))]; |
| tensor<fp16, [1, 197, 768]> linear_2_cast_fp16 = linear(bias = model_vit_encoder_layer_0_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_0_attention_attention_value_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")]; |
| tensor<int32, [4]> var_99 = const()[name = tensor<string, []>("op_99"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_7_cast_fp16 = reshape(shape = var_99, x = linear_2_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")]; |
| tensor<int32, [4]> var_101 = const()[name = tensor<string, []>("op_101"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_105 = const()[name = tensor<string, []>("op_105"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_11_cast_fp16 = reshape(shape = var_105, x = linear_0_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")]; |
| tensor<fp16, []> mul_0_y_0_to_fp16 = const()[name = tensor<string, []>("mul_0_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_0_cast_fp16 = mul(x = x_11_cast_fp16, y = mul_0_y_0_to_fp16)[name = tensor<string, []>("mul_0_cast_fp16")]; |
| tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_48_perm_0 = const()[name = tensor<string, []>("transpose_48_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_49_perm_0 = const()[name = tensor<string, []>("transpose_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_49 = transpose(perm = transpose_49_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_118")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_48 = transpose(perm = transpose_48_perm_0, x = mul_0_cast_fp16)[name = tensor<string, []>("transpose_119")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = transpose_48, y = transpose_49)[name = tensor<string, []>("matmul_0_cast_fp16")]; |
| tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = matmul_0_cast_fp16)[name = tensor<string, []>("softmax_0_cast_fp16")]; |
| tensor<bool, []> context_layer_1_transpose_x_0 = const()[name = tensor<string, []>("context_layer_1_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_1_transpose_y_0 = const()[name = tensor<string, []>("context_layer_1_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_1_cast_fp16 = transpose(perm = var_101, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_117")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_layer_1_cast_fp16)[name = tensor<string, []>("context_layer_1_cast_fp16")]; |
| tensor<int32, [4]> var_110 = const()[name = tensor<string, []>("op_110"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_115 = const()[name = tensor<string, []>("op_115"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_111_cast_fp16 = transpose(perm = var_110, x = context_layer_1_cast_fp16)[name = tensor<string, []>("transpose_116")]; |
| tensor<fp16, [1, 197, 768]> input_7_cast_fp16 = reshape(shape = var_115, x = var_111_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_0_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_0_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2529408))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3119296))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3120896)))]; |
| tensor<fp16, [1, 197, 768]> linear_3_cast_fp16 = linear(bias = model_vit_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_0_attention_output_dense_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_11_cast_fp16 = add(x = linear_3_cast_fp16, y = input_1_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")]; |
| tensor<int32, [1]> input_13_axes_0 = const()[name = tensor<string, []>("input_13_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3122496)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3124096)))]; |
| tensor<fp16, [1, 197, 768]> input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = model_vit_encoder_layer_0_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_0_layernorm_after_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_0_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_0_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3125696))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5488192))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_0_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5494400)))]; |
| tensor<fp16, [1, 197, 3072]> linear_4_cast_fp16 = linear(bias = model_vit_encoder_layer_0_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_0_intermediate_dense_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")]; |
| tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_17_cast_fp16 = gelu(mode = input_17_mode_0, x = linear_4_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_0_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_0_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5500608))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7859968))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_0_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_0_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7861568)))]; |
| tensor<fp16, [1, 197, 768]> linear_5_cast_fp16 = linear(bias = model_vit_encoder_layer_0_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_0_output_dense_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_21_cast_fp16 = add(x = linear_5_cast_fp16, y = input_11_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")]; |
| tensor<int32, [1]> input_23_axes_0 = const()[name = tensor<string, []>("input_23_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7863168)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7864768)))]; |
| tensor<fp16, [1, 197, 768]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = model_vit_encoder_layer_1_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_1_layernorm_before_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_1_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_1_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7866368))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8456256))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8457856)))]; |
| tensor<fp16, [1, 197, 768]> linear_6_cast_fp16 = linear(bias = model_vit_encoder_layer_1_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_1_attention_attention_query_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_1_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_1_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8459456))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9049344))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9050944)))]; |
| tensor<fp16, [1, 197, 768]> linear_7_cast_fp16 = linear(bias = model_vit_encoder_layer_1_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_1_attention_attention_key_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")]; |
| tensor<int32, [4]> var_160 = const()[name = tensor<string, []>("op_160"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_15_cast_fp16 = reshape(shape = var_160, x = linear_7_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_1_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_1_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9052544))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9642432))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9644032)))]; |
| tensor<fp16, [1, 197, 768]> linear_8_cast_fp16 = linear(bias = model_vit_encoder_layer_1_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_1_attention_attention_value_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")]; |
| tensor<int32, [4]> var_169 = const()[name = tensor<string, []>("op_169"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_19_cast_fp16 = reshape(shape = var_169, x = linear_8_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")]; |
| tensor<int32, [4]> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_175 = const()[name = tensor<string, []>("op_175"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_23_cast_fp16 = reshape(shape = var_175, x = linear_6_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")]; |
| tensor<fp16, []> mul_1_y_0_to_fp16 = const()[name = tensor<string, []>("mul_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_1_cast_fp16 = mul(x = x_23_cast_fp16, y = mul_1_y_0_to_fp16)[name = tensor<string, []>("mul_1_cast_fp16")]; |
| tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_50_perm_0 = const()[name = tensor<string, []>("transpose_50_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_51_perm_0 = const()[name = tensor<string, []>("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_51 = transpose(perm = transpose_51_perm_0, x = x_15_cast_fp16)[name = tensor<string, []>("transpose_114")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_50 = transpose(perm = transpose_50_perm_0, x = mul_1_cast_fp16)[name = tensor<string, []>("transpose_115")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_50, y = transpose_51)[name = tensor<string, []>("matmul_1_cast_fp16")]; |
| tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = matmul_1_cast_fp16)[name = tensor<string, []>("softmax_1_cast_fp16")]; |
| tensor<bool, []> context_layer_5_transpose_x_0 = const()[name = tensor<string, []>("context_layer_5_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_5_transpose_y_0 = const()[name = tensor<string, []>("context_layer_5_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_3_cast_fp16 = transpose(perm = var_171, x = x_19_cast_fp16)[name = tensor<string, []>("transpose_113")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_layer_3_cast_fp16)[name = tensor<string, []>("context_layer_5_cast_fp16")]; |
| tensor<int32, [4]> var_180 = const()[name = tensor<string, []>("op_180"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_185 = const()[name = tensor<string, []>("op_185"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_181_cast_fp16 = transpose(perm = var_180, x = context_layer_5_cast_fp16)[name = tensor<string, []>("transpose_112")]; |
| tensor<fp16, [1, 197, 768]> input_25_cast_fp16 = reshape(shape = var_185, x = var_181_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_1_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_1_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9645632))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10235520))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10237120)))]; |
| tensor<fp16, [1, 197, 768]> linear_9_cast_fp16 = linear(bias = model_vit_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_1_attention_output_dense_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_29_cast_fp16 = add(x = linear_9_cast_fp16, y = input_21_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")]; |
| tensor<int32, [1]> input_31_axes_0 = const()[name = tensor<string, []>("input_31_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10238720)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10240320)))]; |
| tensor<fp16, [1, 197, 768]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = model_vit_encoder_layer_1_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_1_layernorm_after_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_1_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_1_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10241920))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12601280))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_1_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12607488)))]; |
| tensor<fp16, [1, 197, 3072]> linear_10_cast_fp16 = linear(bias = model_vit_encoder_layer_1_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_1_intermediate_dense_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")]; |
| tensor<string, []> input_35_mode_0 = const()[name = tensor<string, []>("input_35_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_1_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_1_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12613696))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14973056))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_1_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_1_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14974656)))]; |
| tensor<fp16, [1, 197, 768]> linear_11_cast_fp16 = linear(bias = model_vit_encoder_layer_1_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_1_output_dense_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_39_cast_fp16 = add(x = linear_11_cast_fp16, y = input_29_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")]; |
| tensor<int32, [1]> input_41_axes_0 = const()[name = tensor<string, []>("input_41_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14976256)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14977856)))]; |
| tensor<fp16, [1, 197, 768]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = model_vit_encoder_layer_2_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_2_layernorm_before_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_2_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_2_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14979456))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15569344))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15570944)))]; |
| tensor<fp16, [1, 197, 768]> linear_12_cast_fp16 = linear(bias = model_vit_encoder_layer_2_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_2_attention_attention_query_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_2_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_2_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15572544))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16162432))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16164032)))]; |
| tensor<fp16, [1, 197, 768]> linear_13_cast_fp16 = linear(bias = model_vit_encoder_layer_2_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_2_attention_attention_key_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")]; |
| tensor<int32, [4]> var_230 = const()[name = tensor<string, []>("op_230"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_27_cast_fp16 = reshape(shape = var_230, x = linear_13_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_2_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_2_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16165632))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16755520))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16757120)))]; |
| tensor<fp16, [1, 197, 768]> linear_14_cast_fp16 = linear(bias = model_vit_encoder_layer_2_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_2_attention_attention_value_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")]; |
| tensor<int32, [4]> var_239 = const()[name = tensor<string, []>("op_239"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_31_cast_fp16 = reshape(shape = var_239, x = linear_14_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")]; |
| tensor<int32, [4]> var_241 = const()[name = tensor<string, []>("op_241"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_245 = const()[name = tensor<string, []>("op_245"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_35_cast_fp16 = reshape(shape = var_245, x = linear_12_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")]; |
| tensor<fp16, []> mul_2_y_0_to_fp16 = const()[name = tensor<string, []>("mul_2_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_2_cast_fp16 = mul(x = x_35_cast_fp16, y = mul_2_y_0_to_fp16)[name = tensor<string, []>("mul_2_cast_fp16")]; |
| tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_52_perm_0 = const()[name = tensor<string, []>("transpose_52_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_53_perm_0 = const()[name = tensor<string, []>("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_53 = transpose(perm = transpose_53_perm_0, x = x_27_cast_fp16)[name = tensor<string, []>("transpose_110")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_52 = transpose(perm = transpose_52_perm_0, x = mul_2_cast_fp16)[name = tensor<string, []>("transpose_111")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = transpose_52, y = transpose_53)[name = tensor<string, []>("matmul_2_cast_fp16")]; |
| tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = matmul_2_cast_fp16)[name = tensor<string, []>("softmax_2_cast_fp16")]; |
| tensor<bool, []> context_layer_9_transpose_x_0 = const()[name = tensor<string, []>("context_layer_9_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_9_transpose_y_0 = const()[name = tensor<string, []>("context_layer_9_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_5_cast_fp16 = transpose(perm = var_241, x = x_31_cast_fp16)[name = tensor<string, []>("transpose_109")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_layer_5_cast_fp16)[name = tensor<string, []>("context_layer_9_cast_fp16")]; |
| tensor<int32, [4]> var_250 = const()[name = tensor<string, []>("op_250"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_255 = const()[name = tensor<string, []>("op_255"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_251_cast_fp16 = transpose(perm = var_250, x = context_layer_9_cast_fp16)[name = tensor<string, []>("transpose_108")]; |
| tensor<fp16, [1, 197, 768]> input_43_cast_fp16 = reshape(shape = var_255, x = var_251_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_2_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_2_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16758720))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17348608))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17350208)))]; |
| tensor<fp16, [1, 197, 768]> linear_15_cast_fp16 = linear(bias = model_vit_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_2_attention_output_dense_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_47_cast_fp16 = add(x = linear_15_cast_fp16, y = input_39_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")]; |
| tensor<int32, [1]> input_49_axes_0 = const()[name = tensor<string, []>("input_49_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17351808)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17353408)))]; |
| tensor<fp16, [1, 197, 768]> input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = model_vit_encoder_layer_2_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_2_layernorm_after_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_2_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_2_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17355008))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19714368))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_2_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19720576)))]; |
| tensor<fp16, [1, 197, 3072]> linear_16_cast_fp16 = linear(bias = model_vit_encoder_layer_2_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_2_intermediate_dense_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")]; |
| tensor<string, []> input_53_mode_0 = const()[name = tensor<string, []>("input_53_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_53_cast_fp16 = gelu(mode = input_53_mode_0, x = linear_16_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_2_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_2_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19726784))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22086144))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_2_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_2_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22087744)))]; |
| tensor<fp16, [1, 197, 768]> linear_17_cast_fp16 = linear(bias = model_vit_encoder_layer_2_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_2_output_dense_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_57_cast_fp16 = add(x = linear_17_cast_fp16, y = input_47_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")]; |
| tensor<int32, [1]> input_59_axes_0 = const()[name = tensor<string, []>("input_59_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22089344)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22090944)))]; |
| tensor<fp16, [1, 197, 768]> input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = model_vit_encoder_layer_3_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_3_layernorm_before_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_3_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_3_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22092544))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22682432))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22684032)))]; |
| tensor<fp16, [1, 197, 768]> linear_18_cast_fp16 = linear(bias = model_vit_encoder_layer_3_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_3_attention_attention_query_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_3_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_3_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22685632))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23275520))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23277120)))]; |
| tensor<fp16, [1, 197, 768]> linear_19_cast_fp16 = linear(bias = model_vit_encoder_layer_3_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_3_attention_attention_key_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")]; |
| tensor<int32, [4]> var_300 = const()[name = tensor<string, []>("op_300"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_39_cast_fp16 = reshape(shape = var_300, x = linear_19_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_3_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_3_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23278720))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23868608))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23870208)))]; |
| tensor<fp16, [1, 197, 768]> linear_20_cast_fp16 = linear(bias = model_vit_encoder_layer_3_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_3_attention_attention_value_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")]; |
| tensor<int32, [4]> var_309 = const()[name = tensor<string, []>("op_309"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_43_cast_fp16 = reshape(shape = var_309, x = linear_20_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")]; |
| tensor<int32, [4]> var_311 = const()[name = tensor<string, []>("op_311"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_315 = const()[name = tensor<string, []>("op_315"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_47_cast_fp16 = reshape(shape = var_315, x = linear_18_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")]; |
| tensor<fp16, []> mul_3_y_0_to_fp16 = const()[name = tensor<string, []>("mul_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_3_cast_fp16 = mul(x = x_47_cast_fp16, y = mul_3_y_0_to_fp16)[name = tensor<string, []>("mul_3_cast_fp16")]; |
| tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_54_perm_0 = const()[name = tensor<string, []>("transpose_54_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_55_perm_0 = const()[name = tensor<string, []>("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_55 = transpose(perm = transpose_55_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_106")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_54 = transpose(perm = transpose_54_perm_0, x = mul_3_cast_fp16)[name = tensor<string, []>("transpose_107")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = transpose_54, y = transpose_55)[name = tensor<string, []>("matmul_3_cast_fp16")]; |
| tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = matmul_3_cast_fp16)[name = tensor<string, []>("softmax_3_cast_fp16")]; |
| tensor<bool, []> context_layer_13_transpose_x_0 = const()[name = tensor<string, []>("context_layer_13_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_13_transpose_y_0 = const()[name = tensor<string, []>("context_layer_13_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_7_cast_fp16 = transpose(perm = var_311, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_105")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_layer_7_cast_fp16)[name = tensor<string, []>("context_layer_13_cast_fp16")]; |
| tensor<int32, [4]> var_320 = const()[name = tensor<string, []>("op_320"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_325 = const()[name = tensor<string, []>("op_325"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_321_cast_fp16 = transpose(perm = var_320, x = context_layer_13_cast_fp16)[name = tensor<string, []>("transpose_104")]; |
| tensor<fp16, [1, 197, 768]> input_61_cast_fp16 = reshape(shape = var_325, x = var_321_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_3_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_3_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23871808))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24461696))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24463296)))]; |
| tensor<fp16, [1, 197, 768]> linear_21_cast_fp16 = linear(bias = model_vit_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_3_attention_output_dense_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_65_cast_fp16 = add(x = linear_21_cast_fp16, y = input_57_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")]; |
| tensor<int32, [1]> input_67_axes_0 = const()[name = tensor<string, []>("input_67_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24464896)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24466496)))]; |
| tensor<fp16, [1, 197, 768]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = model_vit_encoder_layer_3_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_3_layernorm_after_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_3_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_3_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24468096))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26827456))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_3_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26833664)))]; |
| tensor<fp16, [1, 197, 3072]> linear_22_cast_fp16 = linear(bias = model_vit_encoder_layer_3_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_3_intermediate_dense_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")]; |
| tensor<string, []> input_71_mode_0 = const()[name = tensor<string, []>("input_71_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_3_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_3_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26839872))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29199232))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_3_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_3_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29200832)))]; |
| tensor<fp16, [1, 197, 768]> linear_23_cast_fp16 = linear(bias = model_vit_encoder_layer_3_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_3_output_dense_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_75_cast_fp16 = add(x = linear_23_cast_fp16, y = input_65_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")]; |
| tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29202432)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29204032)))]; |
| tensor<fp16, [1, 197, 768]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = model_vit_encoder_layer_4_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_4_layernorm_before_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_4_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_4_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29205632))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29795520))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29797120)))]; |
| tensor<fp16, [1, 197, 768]> linear_24_cast_fp16 = linear(bias = model_vit_encoder_layer_4_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_4_attention_attention_query_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_4_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_4_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29798720))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30388608))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30390208)))]; |
| tensor<fp16, [1, 197, 768]> linear_25_cast_fp16 = linear(bias = model_vit_encoder_layer_4_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_4_attention_attention_key_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")]; |
| tensor<int32, [4]> var_370 = const()[name = tensor<string, []>("op_370"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_51_cast_fp16 = reshape(shape = var_370, x = linear_25_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_4_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_4_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30391808))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30981696))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30983296)))]; |
| tensor<fp16, [1, 197, 768]> linear_26_cast_fp16 = linear(bias = model_vit_encoder_layer_4_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_4_attention_attention_value_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")]; |
| tensor<int32, [4]> var_379 = const()[name = tensor<string, []>("op_379"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_55_cast_fp16 = reshape(shape = var_379, x = linear_26_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")]; |
| tensor<int32, [4]> var_381 = const()[name = tensor<string, []>("op_381"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_385 = const()[name = tensor<string, []>("op_385"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_59_cast_fp16 = reshape(shape = var_385, x = linear_24_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")]; |
| tensor<fp16, []> mul_4_y_0_to_fp16 = const()[name = tensor<string, []>("mul_4_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_4_cast_fp16 = mul(x = x_59_cast_fp16, y = mul_4_y_0_to_fp16)[name = tensor<string, []>("mul_4_cast_fp16")]; |
| tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_56_perm_0 = const()[name = tensor<string, []>("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_57_perm_0 = const()[name = tensor<string, []>("transpose_57_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_57 = transpose(perm = transpose_57_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_102")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_56 = transpose(perm = transpose_56_perm_0, x = mul_4_cast_fp16)[name = tensor<string, []>("transpose_103")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = transpose_56, y = transpose_57)[name = tensor<string, []>("matmul_4_cast_fp16")]; |
| tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = matmul_4_cast_fp16)[name = tensor<string, []>("softmax_4_cast_fp16")]; |
| tensor<bool, []> context_layer_17_transpose_x_0 = const()[name = tensor<string, []>("context_layer_17_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_17_transpose_y_0 = const()[name = tensor<string, []>("context_layer_17_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_9_cast_fp16 = transpose(perm = var_381, x = x_55_cast_fp16)[name = tensor<string, []>("transpose_101")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_17_cast_fp16 = matmul(transpose_x = context_layer_17_transpose_x_0, transpose_y = context_layer_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_layer_9_cast_fp16)[name = tensor<string, []>("context_layer_17_cast_fp16")]; |
| tensor<int32, [4]> var_390 = const()[name = tensor<string, []>("op_390"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_395 = const()[name = tensor<string, []>("op_395"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_391_cast_fp16 = transpose(perm = var_390, x = context_layer_17_cast_fp16)[name = tensor<string, []>("transpose_100")]; |
| tensor<fp16, [1, 197, 768]> input_79_cast_fp16 = reshape(shape = var_395, x = var_391_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_4_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_4_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30984896))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31574784))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31576384)))]; |
| tensor<fp16, [1, 197, 768]> linear_27_cast_fp16 = linear(bias = model_vit_encoder_layer_4_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_4_attention_output_dense_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_83_cast_fp16 = add(x = linear_27_cast_fp16, y = input_75_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")]; |
| tensor<int32, [1]> input_85_axes_0 = const()[name = tensor<string, []>("input_85_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31577984)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31579584)))]; |
| tensor<fp16, [1, 197, 768]> input_85_cast_fp16 = layer_norm(axes = input_85_axes_0, beta = model_vit_encoder_layer_4_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_4_layernorm_after_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_4_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_4_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31581184))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33940544))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_4_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33946752)))]; |
| tensor<fp16, [1, 197, 3072]> linear_28_cast_fp16 = linear(bias = model_vit_encoder_layer_4_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_4_intermediate_dense_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")]; |
| tensor<string, []> input_89_mode_0 = const()[name = tensor<string, []>("input_89_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = linear_28_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_4_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_4_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33952960))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36312320))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_4_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_4_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36313920)))]; |
| tensor<fp16, [1, 197, 768]> linear_29_cast_fp16 = linear(bias = model_vit_encoder_layer_4_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_4_output_dense_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_93_cast_fp16 = add(x = linear_29_cast_fp16, y = input_83_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")]; |
| tensor<int32, [1]> input_95_axes_0 = const()[name = tensor<string, []>("input_95_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36315520)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36317120)))]; |
| tensor<fp16, [1, 197, 768]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = model_vit_encoder_layer_5_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_5_layernorm_before_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_5_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_5_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36318720))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36908608))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36910208)))]; |
| tensor<fp16, [1, 197, 768]> linear_30_cast_fp16 = linear(bias = model_vit_encoder_layer_5_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_5_attention_attention_query_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_5_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_5_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36911808))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37501696))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37503296)))]; |
| tensor<fp16, [1, 197, 768]> linear_31_cast_fp16 = linear(bias = model_vit_encoder_layer_5_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_5_attention_attention_key_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")]; |
| tensor<int32, [4]> var_440 = const()[name = tensor<string, []>("op_440"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_63_cast_fp16 = reshape(shape = var_440, x = linear_31_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_5_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_5_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37504896))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38094784))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38096384)))]; |
| tensor<fp16, [1, 197, 768]> linear_32_cast_fp16 = linear(bias = model_vit_encoder_layer_5_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_5_attention_attention_value_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")]; |
| tensor<int32, [4]> var_449 = const()[name = tensor<string, []>("op_449"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_67_cast_fp16 = reshape(shape = var_449, x = linear_32_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")]; |
| tensor<int32, [4]> var_451 = const()[name = tensor<string, []>("op_451"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_455 = const()[name = tensor<string, []>("op_455"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_71_cast_fp16 = reshape(shape = var_455, x = linear_30_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")]; |
| tensor<fp16, []> mul_5_y_0_to_fp16 = const()[name = tensor<string, []>("mul_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_5_cast_fp16 = mul(x = x_71_cast_fp16, y = mul_5_y_0_to_fp16)[name = tensor<string, []>("mul_5_cast_fp16")]; |
| tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_58_perm_0 = const()[name = tensor<string, []>("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_59_perm_0 = const()[name = tensor<string, []>("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_59 = transpose(perm = transpose_59_perm_0, x = x_63_cast_fp16)[name = tensor<string, []>("transpose_98")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_58 = transpose(perm = transpose_58_perm_0, x = mul_5_cast_fp16)[name = tensor<string, []>("transpose_99")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = transpose_58, y = transpose_59)[name = tensor<string, []>("matmul_5_cast_fp16")]; |
| tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = matmul_5_cast_fp16)[name = tensor<string, []>("softmax_5_cast_fp16")]; |
| tensor<bool, []> context_layer_21_transpose_x_0 = const()[name = tensor<string, []>("context_layer_21_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_21_transpose_y_0 = const()[name = tensor<string, []>("context_layer_21_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_11_cast_fp16 = transpose(perm = var_451, x = x_67_cast_fp16)[name = tensor<string, []>("transpose_97")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_21_cast_fp16 = matmul(transpose_x = context_layer_21_transpose_x_0, transpose_y = context_layer_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_layer_11_cast_fp16)[name = tensor<string, []>("context_layer_21_cast_fp16")]; |
| tensor<int32, [4]> var_460 = const()[name = tensor<string, []>("op_460"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_465 = const()[name = tensor<string, []>("op_465"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_461_cast_fp16 = transpose(perm = var_460, x = context_layer_21_cast_fp16)[name = tensor<string, []>("transpose_96")]; |
| tensor<fp16, [1, 197, 768]> input_97_cast_fp16 = reshape(shape = var_465, x = var_461_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_5_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_5_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38097984))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38687872))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38689472)))]; |
| tensor<fp16, [1, 197, 768]> linear_33_cast_fp16 = linear(bias = model_vit_encoder_layer_5_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_5_attention_output_dense_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_101_cast_fp16 = add(x = linear_33_cast_fp16, y = input_93_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")]; |
| tensor<int32, [1]> input_103_axes_0 = const()[name = tensor<string, []>("input_103_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38691072)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38692672)))]; |
| tensor<fp16, [1, 197, 768]> input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = model_vit_encoder_layer_5_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_5_layernorm_after_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_5_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_5_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38694272))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41053632))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_5_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41059840)))]; |
| tensor<fp16, [1, 197, 3072]> linear_34_cast_fp16 = linear(bias = model_vit_encoder_layer_5_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_5_intermediate_dense_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")]; |
| tensor<string, []> input_107_mode_0 = const()[name = tensor<string, []>("input_107_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_5_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_5_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41066048))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43425408))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_5_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_5_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43427008)))]; |
| tensor<fp16, [1, 197, 768]> linear_35_cast_fp16 = linear(bias = model_vit_encoder_layer_5_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_5_output_dense_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_111_cast_fp16 = add(x = linear_35_cast_fp16, y = input_101_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")]; |
| tensor<int32, [1]> input_113_axes_0 = const()[name = tensor<string, []>("input_113_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43428608)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43430208)))]; |
| tensor<fp16, [1, 197, 768]> input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = model_vit_encoder_layer_6_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_6_layernorm_before_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_6_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_6_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43431808))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44021696))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44023296)))]; |
| tensor<fp16, [1, 197, 768]> linear_36_cast_fp16 = linear(bias = model_vit_encoder_layer_6_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_6_attention_attention_query_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_6_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_6_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44024896))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44614784))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44616384)))]; |
| tensor<fp16, [1, 197, 768]> linear_37_cast_fp16 = linear(bias = model_vit_encoder_layer_6_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_6_attention_attention_key_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")]; |
| tensor<int32, [4]> var_510 = const()[name = tensor<string, []>("op_510"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_75_cast_fp16 = reshape(shape = var_510, x = linear_37_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_6_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_6_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44617984))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45207872))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45209472)))]; |
| tensor<fp16, [1, 197, 768]> linear_38_cast_fp16 = linear(bias = model_vit_encoder_layer_6_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_6_attention_attention_value_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")]; |
| tensor<int32, [4]> var_519 = const()[name = tensor<string, []>("op_519"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_79_cast_fp16 = reshape(shape = var_519, x = linear_38_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")]; |
| tensor<int32, [4]> var_521 = const()[name = tensor<string, []>("op_521"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_525 = const()[name = tensor<string, []>("op_525"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_83_cast_fp16 = reshape(shape = var_525, x = linear_36_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")]; |
| tensor<fp16, []> mul_6_y_0_to_fp16 = const()[name = tensor<string, []>("mul_6_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_6_cast_fp16 = mul(x = x_83_cast_fp16, y = mul_6_y_0_to_fp16)[name = tensor<string, []>("mul_6_cast_fp16")]; |
| tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_60_perm_0 = const()[name = tensor<string, []>("transpose_60_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_61_perm_0 = const()[name = tensor<string, []>("transpose_61_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_61 = transpose(perm = transpose_61_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_94")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_60 = transpose(perm = transpose_60_perm_0, x = mul_6_cast_fp16)[name = tensor<string, []>("transpose_95")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = transpose_60, y = transpose_61)[name = tensor<string, []>("matmul_6_cast_fp16")]; |
| tensor<int32, []> softmax_6_axis_0 = const()[name = tensor<string, []>("softmax_6_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = matmul_6_cast_fp16)[name = tensor<string, []>("softmax_6_cast_fp16")]; |
| tensor<bool, []> context_layer_25_transpose_x_0 = const()[name = tensor<string, []>("context_layer_25_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_25_transpose_y_0 = const()[name = tensor<string, []>("context_layer_25_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_13_cast_fp16 = transpose(perm = var_521, x = x_79_cast_fp16)[name = tensor<string, []>("transpose_93")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_25_cast_fp16 = matmul(transpose_x = context_layer_25_transpose_x_0, transpose_y = context_layer_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_layer_13_cast_fp16)[name = tensor<string, []>("context_layer_25_cast_fp16")]; |
| tensor<int32, [4]> var_530 = const()[name = tensor<string, []>("op_530"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_535 = const()[name = tensor<string, []>("op_535"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_531_cast_fp16 = transpose(perm = var_530, x = context_layer_25_cast_fp16)[name = tensor<string, []>("transpose_92")]; |
| tensor<fp16, [1, 197, 768]> input_115_cast_fp16 = reshape(shape = var_535, x = var_531_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_6_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_6_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45211072))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45800960))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45802560)))]; |
| tensor<fp16, [1, 197, 768]> linear_39_cast_fp16 = linear(bias = model_vit_encoder_layer_6_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_6_attention_output_dense_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_119_cast_fp16 = add(x = linear_39_cast_fp16, y = input_111_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")]; |
| tensor<int32, [1]> input_121_axes_0 = const()[name = tensor<string, []>("input_121_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45804160)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45805760)))]; |
| tensor<fp16, [1, 197, 768]> input_121_cast_fp16 = layer_norm(axes = input_121_axes_0, beta = model_vit_encoder_layer_6_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_6_layernorm_after_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_6_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_6_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45807360))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48166720))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_6_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48172928)))]; |
| tensor<fp16, [1, 197, 3072]> linear_40_cast_fp16 = linear(bias = model_vit_encoder_layer_6_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_6_intermediate_dense_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")]; |
| tensor<string, []> input_125_mode_0 = const()[name = tensor<string, []>("input_125_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_125_cast_fp16 = gelu(mode = input_125_mode_0, x = linear_40_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_6_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_6_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48179136))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50538496))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_6_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_6_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50540096)))]; |
| tensor<fp16, [1, 197, 768]> linear_41_cast_fp16 = linear(bias = model_vit_encoder_layer_6_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_6_output_dense_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_129_cast_fp16 = add(x = linear_41_cast_fp16, y = input_119_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")]; |
| tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50541696)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50543296)))]; |
| tensor<fp16, [1, 197, 768]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = model_vit_encoder_layer_7_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_7_layernorm_before_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_7_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_7_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50544896))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51134784))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51136384)))]; |
| tensor<fp16, [1, 197, 768]> linear_42_cast_fp16 = linear(bias = model_vit_encoder_layer_7_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_7_attention_attention_query_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_7_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_7_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51137984))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51727872))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51729472)))]; |
| tensor<fp16, [1, 197, 768]> linear_43_cast_fp16 = linear(bias = model_vit_encoder_layer_7_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_7_attention_attention_key_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")]; |
| tensor<int32, [4]> var_580 = const()[name = tensor<string, []>("op_580"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_87_cast_fp16 = reshape(shape = var_580, x = linear_43_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_7_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_7_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51731072))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52320960))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52322560)))]; |
| tensor<fp16, [1, 197, 768]> linear_44_cast_fp16 = linear(bias = model_vit_encoder_layer_7_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_7_attention_attention_value_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")]; |
| tensor<int32, [4]> var_589 = const()[name = tensor<string, []>("op_589"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_91_cast_fp16 = reshape(shape = var_589, x = linear_44_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")]; |
| tensor<int32, [4]> var_591 = const()[name = tensor<string, []>("op_591"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_595 = const()[name = tensor<string, []>("op_595"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_95_cast_fp16 = reshape(shape = var_595, x = linear_42_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")]; |
| tensor<fp16, []> mul_7_y_0_to_fp16 = const()[name = tensor<string, []>("mul_7_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_7_cast_fp16 = mul(x = x_95_cast_fp16, y = mul_7_y_0_to_fp16)[name = tensor<string, []>("mul_7_cast_fp16")]; |
| tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_62_perm_0 = const()[name = tensor<string, []>("transpose_62_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_63_perm_0 = const()[name = tensor<string, []>("transpose_63_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_63 = transpose(perm = transpose_63_perm_0, x = x_87_cast_fp16)[name = tensor<string, []>("transpose_90")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_62 = transpose(perm = transpose_62_perm_0, x = mul_7_cast_fp16)[name = tensor<string, []>("transpose_91")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = transpose_62, y = transpose_63)[name = tensor<string, []>("matmul_7_cast_fp16")]; |
| tensor<int32, []> softmax_7_axis_0 = const()[name = tensor<string, []>("softmax_7_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = matmul_7_cast_fp16)[name = tensor<string, []>("softmax_7_cast_fp16")]; |
| tensor<bool, []> context_layer_29_transpose_x_0 = const()[name = tensor<string, []>("context_layer_29_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_29_transpose_y_0 = const()[name = tensor<string, []>("context_layer_29_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_15_cast_fp16 = transpose(perm = var_591, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_89")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_29_cast_fp16 = matmul(transpose_x = context_layer_29_transpose_x_0, transpose_y = context_layer_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_layer_15_cast_fp16)[name = tensor<string, []>("context_layer_29_cast_fp16")]; |
| tensor<int32, [4]> var_600 = const()[name = tensor<string, []>("op_600"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_605 = const()[name = tensor<string, []>("op_605"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_601_cast_fp16 = transpose(perm = var_600, x = context_layer_29_cast_fp16)[name = tensor<string, []>("transpose_88")]; |
| tensor<fp16, [1, 197, 768]> input_133_cast_fp16 = reshape(shape = var_605, x = var_601_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_7_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_7_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52324160))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52914048))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52915648)))]; |
| tensor<fp16, [1, 197, 768]> linear_45_cast_fp16 = linear(bias = model_vit_encoder_layer_7_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_7_attention_output_dense_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_137_cast_fp16 = add(x = linear_45_cast_fp16, y = input_129_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")]; |
| tensor<int32, [1]> input_139_axes_0 = const()[name = tensor<string, []>("input_139_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52917248)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52918848)))]; |
| tensor<fp16, [1, 197, 768]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = model_vit_encoder_layer_7_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_7_layernorm_after_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_7_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_7_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52920448))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55279808))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_7_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55286016)))]; |
| tensor<fp16, [1, 197, 3072]> linear_46_cast_fp16 = linear(bias = model_vit_encoder_layer_7_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_7_intermediate_dense_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")]; |
| tensor<string, []> input_143_mode_0 = const()[name = tensor<string, []>("input_143_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_7_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_7_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55292224))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57651584))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_7_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_7_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57653184)))]; |
| tensor<fp16, [1, 197, 768]> linear_47_cast_fp16 = linear(bias = model_vit_encoder_layer_7_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_7_output_dense_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_147_cast_fp16 = add(x = linear_47_cast_fp16, y = input_137_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")]; |
| tensor<int32, [1]> input_149_axes_0 = const()[name = tensor<string, []>("input_149_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57654784)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57656384)))]; |
| tensor<fp16, [1, 197, 768]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = model_vit_encoder_layer_8_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_8_layernorm_before_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_8_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_8_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57657984))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58247872))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58249472)))]; |
| tensor<fp16, [1, 197, 768]> linear_48_cast_fp16 = linear(bias = model_vit_encoder_layer_8_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_8_attention_attention_query_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_8_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_8_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58251072))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58840960))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58842560)))]; |
| tensor<fp16, [1, 197, 768]> linear_49_cast_fp16 = linear(bias = model_vit_encoder_layer_8_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_8_attention_attention_key_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")]; |
| tensor<int32, [4]> var_650 = const()[name = tensor<string, []>("op_650"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_99_cast_fp16 = reshape(shape = var_650, x = linear_49_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_8_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_8_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58844160))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59434048))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59435648)))]; |
| tensor<fp16, [1, 197, 768]> linear_50_cast_fp16 = linear(bias = model_vit_encoder_layer_8_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_8_attention_attention_value_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")]; |
| tensor<int32, [4]> var_659 = const()[name = tensor<string, []>("op_659"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_103_cast_fp16 = reshape(shape = var_659, x = linear_50_cast_fp16)[name = tensor<string, []>("x_103_cast_fp16")]; |
| tensor<int32, [4]> var_661 = const()[name = tensor<string, []>("op_661"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_665 = const()[name = tensor<string, []>("op_665"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_107_cast_fp16 = reshape(shape = var_665, x = linear_48_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")]; |
| tensor<fp16, []> mul_8_y_0_to_fp16 = const()[name = tensor<string, []>("mul_8_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_8_cast_fp16 = mul(x = x_107_cast_fp16, y = mul_8_y_0_to_fp16)[name = tensor<string, []>("mul_8_cast_fp16")]; |
| tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_64_perm_0 = const()[name = tensor<string, []>("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_65_perm_0 = const()[name = tensor<string, []>("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_65 = transpose(perm = transpose_65_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_86")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_64 = transpose(perm = transpose_64_perm_0, x = mul_8_cast_fp16)[name = tensor<string, []>("transpose_87")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = transpose_64, y = transpose_65)[name = tensor<string, []>("matmul_8_cast_fp16")]; |
| tensor<int32, []> softmax_8_axis_0 = const()[name = tensor<string, []>("softmax_8_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = matmul_8_cast_fp16)[name = tensor<string, []>("softmax_8_cast_fp16")]; |
| tensor<bool, []> context_layer_33_transpose_x_0 = const()[name = tensor<string, []>("context_layer_33_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_33_transpose_y_0 = const()[name = tensor<string, []>("context_layer_33_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_17_cast_fp16 = transpose(perm = var_661, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_85")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_33_cast_fp16 = matmul(transpose_x = context_layer_33_transpose_x_0, transpose_y = context_layer_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_layer_17_cast_fp16)[name = tensor<string, []>("context_layer_33_cast_fp16")]; |
| tensor<int32, [4]> var_670 = const()[name = tensor<string, []>("op_670"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_675 = const()[name = tensor<string, []>("op_675"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_671_cast_fp16 = transpose(perm = var_670, x = context_layer_33_cast_fp16)[name = tensor<string, []>("transpose_84")]; |
| tensor<fp16, [1, 197, 768]> input_151_cast_fp16 = reshape(shape = var_675, x = var_671_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_8_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_8_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59437248))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60027136))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60028736)))]; |
| tensor<fp16, [1, 197, 768]> linear_51_cast_fp16 = linear(bias = model_vit_encoder_layer_8_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_8_attention_output_dense_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_155_cast_fp16 = add(x = linear_51_cast_fp16, y = input_147_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")]; |
| tensor<int32, [1]> input_157_axes_0 = const()[name = tensor<string, []>("input_157_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60030336)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60031936)))]; |
| tensor<fp16, [1, 197, 768]> input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, beta = model_vit_encoder_layer_8_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_8_layernorm_after_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_8_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_8_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60033536))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62392896))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_8_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62399104)))]; |
| tensor<fp16, [1, 197, 3072]> linear_52_cast_fp16 = linear(bias = model_vit_encoder_layer_8_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_8_intermediate_dense_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")]; |
| tensor<string, []> input_161_mode_0 = const()[name = tensor<string, []>("input_161_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = linear_52_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_8_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_8_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62405312))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64764672))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_8_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_8_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64766272)))]; |
| tensor<fp16, [1, 197, 768]> linear_53_cast_fp16 = linear(bias = model_vit_encoder_layer_8_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_8_output_dense_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_165_cast_fp16 = add(x = linear_53_cast_fp16, y = input_155_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")]; |
| tensor<int32, [1]> input_167_axes_0 = const()[name = tensor<string, []>("input_167_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64767872)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64769472)))]; |
| tensor<fp16, [1, 197, 768]> input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = model_vit_encoder_layer_9_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_9_layernorm_before_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_9_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_9_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64771072))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65360960))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65362560)))]; |
| tensor<fp16, [1, 197, 768]> linear_54_cast_fp16 = linear(bias = model_vit_encoder_layer_9_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_9_attention_attention_query_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_9_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_9_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65364160))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65954048))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65955648)))]; |
| tensor<fp16, [1, 197, 768]> linear_55_cast_fp16 = linear(bias = model_vit_encoder_layer_9_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_9_attention_attention_key_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")]; |
| tensor<int32, [4]> var_720 = const()[name = tensor<string, []>("op_720"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_111_cast_fp16 = reshape(shape = var_720, x = linear_55_cast_fp16)[name = tensor<string, []>("x_111_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_9_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_9_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65957248))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66547136))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66548736)))]; |
| tensor<fp16, [1, 197, 768]> linear_56_cast_fp16 = linear(bias = model_vit_encoder_layer_9_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_9_attention_attention_value_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")]; |
| tensor<int32, [4]> var_729 = const()[name = tensor<string, []>("op_729"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_115_cast_fp16 = reshape(shape = var_729, x = linear_56_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")]; |
| tensor<int32, [4]> var_731 = const()[name = tensor<string, []>("op_731"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_735 = const()[name = tensor<string, []>("op_735"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_119_cast_fp16 = reshape(shape = var_735, x = linear_54_cast_fp16)[name = tensor<string, []>("x_119_cast_fp16")]; |
| tensor<fp16, []> mul_9_y_0_to_fp16 = const()[name = tensor<string, []>("mul_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_9_cast_fp16 = mul(x = x_119_cast_fp16, y = mul_9_y_0_to_fp16)[name = tensor<string, []>("mul_9_cast_fp16")]; |
| tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_66_perm_0 = const()[name = tensor<string, []>("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_67_perm_0 = const()[name = tensor<string, []>("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_67 = transpose(perm = transpose_67_perm_0, x = x_111_cast_fp16)[name = tensor<string, []>("transpose_82")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_66 = transpose(perm = transpose_66_perm_0, x = mul_9_cast_fp16)[name = tensor<string, []>("transpose_83")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = transpose_66, y = transpose_67)[name = tensor<string, []>("matmul_9_cast_fp16")]; |
| tensor<int32, []> softmax_9_axis_0 = const()[name = tensor<string, []>("softmax_9_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = matmul_9_cast_fp16)[name = tensor<string, []>("softmax_9_cast_fp16")]; |
| tensor<bool, []> context_layer_37_transpose_x_0 = const()[name = tensor<string, []>("context_layer_37_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_37_transpose_y_0 = const()[name = tensor<string, []>("context_layer_37_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_19_cast_fp16 = transpose(perm = var_731, x = x_115_cast_fp16)[name = tensor<string, []>("transpose_81")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_37_cast_fp16 = matmul(transpose_x = context_layer_37_transpose_x_0, transpose_y = context_layer_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_layer_19_cast_fp16)[name = tensor<string, []>("context_layer_37_cast_fp16")]; |
| tensor<int32, [4]> var_740 = const()[name = tensor<string, []>("op_740"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_745 = const()[name = tensor<string, []>("op_745"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_741_cast_fp16 = transpose(perm = var_740, x = context_layer_37_cast_fp16)[name = tensor<string, []>("transpose_80")]; |
| tensor<fp16, [1, 197, 768]> input_169_cast_fp16 = reshape(shape = var_745, x = var_741_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_9_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_9_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66550336))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67140224))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67141824)))]; |
| tensor<fp16, [1, 197, 768]> linear_57_cast_fp16 = linear(bias = model_vit_encoder_layer_9_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_9_attention_output_dense_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_173_cast_fp16 = add(x = linear_57_cast_fp16, y = input_165_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")]; |
| tensor<int32, [1]> input_175_axes_0 = const()[name = tensor<string, []>("input_175_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67143424)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67145024)))]; |
| tensor<fp16, [1, 197, 768]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, beta = model_vit_encoder_layer_9_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_9_layernorm_after_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_9_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_9_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67146624))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69505984))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_9_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69512192)))]; |
| tensor<fp16, [1, 197, 3072]> linear_58_cast_fp16 = linear(bias = model_vit_encoder_layer_9_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_9_intermediate_dense_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")]; |
| tensor<string, []> input_179_mode_0 = const()[name = tensor<string, []>("input_179_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_9_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_9_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69518400))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71877760))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_9_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_9_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71879360)))]; |
| tensor<fp16, [1, 197, 768]> linear_59_cast_fp16 = linear(bias = model_vit_encoder_layer_9_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_9_output_dense_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_183_cast_fp16 = add(x = linear_59_cast_fp16, y = input_173_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")]; |
| tensor<int32, [1]> input_185_axes_0 = const()[name = tensor<string, []>("input_185_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71880960)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71882560)))]; |
| tensor<fp16, [1, 197, 768]> input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = model_vit_encoder_layer_10_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_10_layernorm_before_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_10_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_10_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71884160))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72474048))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72475648)))]; |
| tensor<fp16, [1, 197, 768]> linear_60_cast_fp16 = linear(bias = model_vit_encoder_layer_10_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_10_attention_attention_query_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = tensor<string, []>("linear_60_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_10_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_10_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72477248))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73067136))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73068736)))]; |
| tensor<fp16, [1, 197, 768]> linear_61_cast_fp16 = linear(bias = model_vit_encoder_layer_10_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_10_attention_attention_key_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")]; |
| tensor<int32, [4]> var_790 = const()[name = tensor<string, []>("op_790"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_123_cast_fp16 = reshape(shape = var_790, x = linear_61_cast_fp16)[name = tensor<string, []>("x_123_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_10_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_10_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73070336))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73660224))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73661824)))]; |
| tensor<fp16, [1, 197, 768]> linear_62_cast_fp16 = linear(bias = model_vit_encoder_layer_10_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_10_attention_attention_value_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")]; |
| tensor<int32, [4]> var_799 = const()[name = tensor<string, []>("op_799"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_127_cast_fp16 = reshape(shape = var_799, x = linear_62_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")]; |
| tensor<int32, [4]> var_801 = const()[name = tensor<string, []>("op_801"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_805 = const()[name = tensor<string, []>("op_805"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_131_cast_fp16 = reshape(shape = var_805, x = linear_60_cast_fp16)[name = tensor<string, []>("x_131_cast_fp16")]; |
| tensor<fp16, []> mul_10_y_0_to_fp16 = const()[name = tensor<string, []>("mul_10_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_10_cast_fp16 = mul(x = x_131_cast_fp16, y = mul_10_y_0_to_fp16)[name = tensor<string, []>("mul_10_cast_fp16")]; |
| tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_68_perm_0 = const()[name = tensor<string, []>("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_69_perm_0 = const()[name = tensor<string, []>("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_69 = transpose(perm = transpose_69_perm_0, x = x_123_cast_fp16)[name = tensor<string, []>("transpose_78")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_68 = transpose(perm = transpose_68_perm_0, x = mul_10_cast_fp16)[name = tensor<string, []>("transpose_79")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = transpose_68, y = transpose_69)[name = tensor<string, []>("matmul_10_cast_fp16")]; |
| tensor<int32, []> softmax_10_axis_0 = const()[name = tensor<string, []>("softmax_10_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = matmul_10_cast_fp16)[name = tensor<string, []>("softmax_10_cast_fp16")]; |
| tensor<bool, []> context_layer_41_transpose_x_0 = const()[name = tensor<string, []>("context_layer_41_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_41_transpose_y_0 = const()[name = tensor<string, []>("context_layer_41_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_21_cast_fp16 = transpose(perm = var_801, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_77")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_41_cast_fp16 = matmul(transpose_x = context_layer_41_transpose_x_0, transpose_y = context_layer_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_layer_21_cast_fp16)[name = tensor<string, []>("context_layer_41_cast_fp16")]; |
| tensor<int32, [4]> var_810 = const()[name = tensor<string, []>("op_810"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_815 = const()[name = tensor<string, []>("op_815"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_811_cast_fp16 = transpose(perm = var_810, x = context_layer_41_cast_fp16)[name = tensor<string, []>("transpose_76")]; |
| tensor<fp16, [1, 197, 768]> input_187_cast_fp16 = reshape(shape = var_815, x = var_811_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_10_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_10_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73663424))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74253312))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74254912)))]; |
| tensor<fp16, [1, 197, 768]> linear_63_cast_fp16 = linear(bias = model_vit_encoder_layer_10_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_10_attention_output_dense_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_191_cast_fp16 = add(x = linear_63_cast_fp16, y = input_183_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")]; |
| tensor<int32, [1]> input_193_axes_0 = const()[name = tensor<string, []>("input_193_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74256512)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74258112)))]; |
| tensor<fp16, [1, 197, 768]> input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, beta = model_vit_encoder_layer_10_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_10_layernorm_after_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_10_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_10_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74259712))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76619072))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_10_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76625280)))]; |
| tensor<fp16, [1, 197, 3072]> linear_64_cast_fp16 = linear(bias = model_vit_encoder_layer_10_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_10_intermediate_dense_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")]; |
| tensor<string, []> input_197_mode_0 = const()[name = tensor<string, []>("input_197_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_197_cast_fp16 = gelu(mode = input_197_mode_0, x = linear_64_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_10_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_10_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76631488))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78990848))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_10_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_10_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78992448)))]; |
| tensor<fp16, [1, 197, 768]> linear_65_cast_fp16 = linear(bias = model_vit_encoder_layer_10_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_10_output_dense_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_201_cast_fp16 = add(x = linear_65_cast_fp16, y = input_191_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")]; |
| tensor<int32, [1]> input_203_axes_0 = const()[name = tensor<string, []>("input_203_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_layernorm_before_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_layernorm_before_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78994048)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_layernorm_before_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_layernorm_before_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78995648)))]; |
| tensor<fp16, [1, 197, 768]> input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = model_vit_encoder_layer_11_layernorm_before_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_11_layernorm_before_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_11_attention_attention_query_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_11_attention_attention_query_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78997248))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79587136))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_attention_attention_query_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_attention_attention_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79588736)))]; |
| tensor<fp16, [1, 197, 768]> linear_66_cast_fp16 = linear(bias = model_vit_encoder_layer_11_attention_attention_query_bias_to_fp16, weight = model_vit_encoder_layer_11_attention_attention_query_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_11_attention_attention_key_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_11_attention_attention_key_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79590336))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80180224))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_attention_attention_key_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_attention_attention_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80181824)))]; |
| tensor<fp16, [1, 197, 768]> linear_67_cast_fp16 = linear(bias = model_vit_encoder_layer_11_attention_attention_key_bias_to_fp16, weight = model_vit_encoder_layer_11_attention_attention_key_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")]; |
| tensor<int32, [4]> var_860 = const()[name = tensor<string, []>("op_860"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_135_cast_fp16 = reshape(shape = var_860, x = linear_67_cast_fp16)[name = tensor<string, []>("x_135_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_11_attention_attention_value_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_11_attention_attention_value_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80183424))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80773312))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_attention_attention_value_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_attention_attention_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80774912)))]; |
| tensor<fp16, [1, 197, 768]> linear_68_cast_fp16 = linear(bias = model_vit_encoder_layer_11_attention_attention_value_bias_to_fp16, weight = model_vit_encoder_layer_11_attention_attention_value_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")]; |
| tensor<int32, [4]> var_869 = const()[name = tensor<string, []>("op_869"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_139_cast_fp16 = reshape(shape = var_869, x = linear_68_cast_fp16)[name = tensor<string, []>("x_139_cast_fp16")]; |
| tensor<int32, [4]> var_871 = const()[name = tensor<string, []>("op_871"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> var_875 = const()[name = tensor<string, []>("op_875"), val = tensor<int32, [4]>([1, 197, 12, 64])]; |
| tensor<fp16, [1, 197, 12, 64]> x_cast_fp16 = reshape(shape = var_875, x = linear_66_cast_fp16)[name = tensor<string, []>("x_cast_fp16")]; |
| tensor<fp16, []> mul_11_y_0_to_fp16 = const()[name = tensor<string, []>("mul_11_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 197, 12, 64]> mul_11_cast_fp16 = mul(x = x_cast_fp16, y = mul_11_y_0_to_fp16)[name = tensor<string, []>("mul_11_cast_fp16")]; |
| tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_70_perm_0 = const()[name = tensor<string, []>("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_71_perm_0 = const()[name = tensor<string, []>("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_71 = transpose(perm = transpose_71_perm_0, x = x_135_cast_fp16)[name = tensor<string, []>("transpose_74")]; |
| tensor<fp16, [1, 12, 197, 64]> transpose_70 = transpose(perm = transpose_70_perm_0, x = mul_11_cast_fp16)[name = tensor<string, []>("transpose_75")]; |
| tensor<fp16, [1, 12, 197, 197]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = transpose_70, y = transpose_71)[name = tensor<string, []>("matmul_11_cast_fp16")]; |
| tensor<int32, []> softmax_11_axis_0 = const()[name = tensor<string, []>("softmax_11_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 12, 197, 197]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = matmul_11_cast_fp16)[name = tensor<string, []>("softmax_11_cast_fp16")]; |
| tensor<bool, []> context_layer_45_transpose_x_0 = const()[name = tensor<string, []>("context_layer_45_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> context_layer_45_transpose_y_0 = const()[name = tensor<string, []>("context_layer_45_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 197, 64]> value_layer_cast_fp16 = transpose(perm = var_871, x = x_139_cast_fp16)[name = tensor<string, []>("transpose_73")]; |
| tensor<fp16, [1, 12, 197, 64]> context_layer_45_cast_fp16 = matmul(transpose_x = context_layer_45_transpose_x_0, transpose_y = context_layer_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_layer_cast_fp16)[name = tensor<string, []>("context_layer_45_cast_fp16")]; |
| tensor<int32, [4]> var_880 = const()[name = tensor<string, []>("op_880"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_885 = const()[name = tensor<string, []>("op_885"), val = tensor<int32, [3]>([1, 197, 768])]; |
| tensor<fp16, [1, 197, 12, 64]> var_881_cast_fp16 = transpose(perm = var_880, x = context_layer_45_cast_fp16)[name = tensor<string, []>("transpose_72")]; |
| tensor<fp16, [1, 197, 768]> input_205_cast_fp16 = reshape(shape = var_885, x = var_881_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")]; |
| tensor<fp16, [768, 768]> model_vit_encoder_layer_11_attention_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_11_attention_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80776512))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81366400))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81368000)))]; |
| tensor<fp16, [1, 197, 768]> linear_69_cast_fp16 = linear(bias = model_vit_encoder_layer_11_attention_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_11_attention_output_dense_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = tensor<string, []>("linear_69_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_209_cast_fp16 = add(x = linear_69_cast_fp16, y = input_201_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")]; |
| tensor<int32, [1]> input_211_axes_0 = const()[name = tensor<string, []>("input_211_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_layernorm_after_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_layernorm_after_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81369600)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_layernorm_after_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_layernorm_after_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81371200)))]; |
| tensor<fp16, [1, 197, 768]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, beta = model_vit_encoder_layer_11_layernorm_after_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_encoder_layer_11_layernorm_after_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")]; |
| tensor<fp16, [3072, 768]> model_vit_encoder_layer_11_intermediate_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_11_intermediate_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81372800))), scale = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83732160))), zero_point = tensor<int8, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5485056)))]; |
| tensor<fp16, [3072]> model_vit_encoder_layer_11_intermediate_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83738368)))]; |
| tensor<fp16, [1, 197, 3072]> linear_70_cast_fp16 = linear(bias = model_vit_encoder_layer_11_intermediate_dense_bias_to_fp16, weight = model_vit_encoder_layer_11_intermediate_dense_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")]; |
| tensor<string, []> input_215_mode_0 = const()[name = tensor<string, []>("input_215_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 197, 3072]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")]; |
| tensor<fp16, [768, 3072]> model_vit_encoder_layer_11_output_dense_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_vit_encoder_layer_11_output_dense_weight_to_fp16_quantized"), quantized_data = tensor<int8, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83744576))), scale = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86103936))), zero_point = tensor<int8, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589952)))]; |
| tensor<fp16, [768]> model_vit_encoder_layer_11_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_encoder_layer_11_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86105536)))]; |
| tensor<fp16, [1, 197, 768]> linear_71_cast_fp16 = linear(bias = model_vit_encoder_layer_11_output_dense_bias_to_fp16, weight = model_vit_encoder_layer_11_output_dense_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")]; |
| tensor<fp16, [1, 197, 768]> input_219_cast_fp16 = add(x = linear_71_cast_fp16, y = input_209_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")]; |
| tensor<int32, [1]> sequence_output_axes_0 = const()[name = tensor<string, []>("sequence_output_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> model_vit_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("model_vit_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86107136)))]; |
| tensor<fp16, [768]> model_vit_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("model_vit_layernorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86108736)))]; |
| tensor<fp16, [1, 197, 768]> sequence_output_cast_fp16 = layer_norm(axes = sequence_output_axes_0, beta = model_vit_layernorm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = model_vit_layernorm_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("sequence_output_cast_fp16")]; |
| tensor<int32, [3]> var_913_begin_0 = const()[name = tensor<string, []>("op_913_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> var_913_end_0 = const()[name = tensor<string, []>("op_913_end_0"), val = tensor<int32, [3]>([1, 1, 768])]; |
| tensor<bool, [3]> var_913_end_mask_0 = const()[name = tensor<string, []>("op_913_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; |
| tensor<bool, [3]> var_913_squeeze_mask_0 = const()[name = tensor<string, []>("op_913_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])]; |
| tensor<fp16, [1, 768]> var_913_cast_fp16 = slice_by_index(begin = var_913_begin_0, end = var_913_end_0, end_mask = var_913_end_mask_0, squeeze_mask = var_913_squeeze_mask_0, x = sequence_output_cast_fp16)[name = tensor<string, []>("op_913_cast_fp16")]; |
| tensor<fp16, [1000, 768]> model_classifier_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_classifier_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1000, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86110336))), scale = tensor<fp16, [1000]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86879488))), zero_point = tensor<int8, [1000]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86878400)))]; |
| tensor<fp16, [1000]> model_classifier_bias_to_fp16 = const()[name = tensor<string, []>("model_classifier_bias_to_fp16"), val = tensor<fp16, [1000]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86881600)))]; |
| tensor<fp16, [1, 1000]> var_917 = linear(bias = model_classifier_bias_to_fp16, weight = model_classifier_weight_to_fp16_quantized, x = var_913_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")]; |
| tensor<string, []> var_917_cast_to_fp32_dtype_0 = const()[name = tensor<string, []>("var_917_cast_to_fp32_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1000]> var_917_cast_to_fp32 = cast(dtype = var_917_cast_to_fp32_dtype_0, x = var_917)[name = tensor<string, []>("cast_0")]; |
| } -> (var_917_cast_to_fp32); |
| } |
