| program(1.3) |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] |
| { |
| func main<ios18>(tensor<int32, [1, 512]> attention_mask, tensor<int32, [1, 512]> input_ids, tensor<int32, [1, 16, 2]> span_idx, tensor<int32, [1, 16]> span_mask, tensor<int32, [1, 1]> text_lengths, tensor<int32, [1, 512]> words_mask) { |
| tensor<int32, [1, 320]> word_positions = const()[name = string("word_positions"), val = tensor<int32, [1, 320]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; |
| int32 var_32 = const()[name = string("op_32"), val = int32(-1)]; |
| int32 inputs_embeds_axis_0 = const()[name = string("inputs_embeds_axis_0"), val = int32(0)]; |
| int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)]; |
| bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [128004, 768]> deberta_embeddings_word_embeddings_weight_to_fp16 = const()[name = string("deberta_embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [128004, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1408)))]; |
| tensor<fp16, [1, 512, 768]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids, validate_indices = inputs_embeds_validate_indices_0, x = deberta_embeddings_word_embeddings_weight_to_fp16)[name = string("inputs_embeds_cast_fp16")]; |
| tensor<int32, [1]> embeddings_1_axes_0 = const()[name = string("embeddings_1_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_embeddings_LayerNorm_weight_to_fp16 = const()[name = string("deberta_embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196615616)))]; |
| tensor<fp16, [768]> deberta_embeddings_LayerNorm_bias_to_fp16 = const()[name = string("deberta_embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196617216)))]; |
| fp16 var_51_to_fp16 = const()[name = string("op_51_to_fp16"), val = fp16(0x1p-23)]; |
| tensor<fp16, [1, 512, 768]> embeddings_1_cast_fp16 = layer_norm(axes = embeddings_1_axes_0, beta = deberta_embeddings_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_embeddings_LayerNorm_weight_to_fp16, x = inputs_embeds_cast_fp16)[name = string("embeddings_1_cast_fp16")]; |
| tensor<int32, [1]> mask_3_axes_0 = const()[name = string("mask_3_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [1, 512, 1]> mask_3 = expand_dims(axes = mask_3_axes_0, x = attention_mask)[name = string("mask_3")]; |
| string cast_3_to_fp16_dtype_0 = const()[name = string("cast_3_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1, 512, 1]> mask_3_to_fp16 = cast(dtype = cast_3_to_fp16_dtype_0, x = mask_3)[name = string("cast_295")]; |
| tensor<fp16, [1, 512, 768]> input_1_cast_fp16 = mul(x = embeddings_1_cast_fp16, y = mask_3_to_fp16)[name = string("input_1_cast_fp16")]; |
| tensor<int32, [1]> var_98_axes_0 = const()[name = string("op_98_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [1, 1, 512]> var_98 = expand_dims(axes = var_98_axes_0, x = attention_mask)[name = string("op_98")]; |
| tensor<int32, [1]> extended_attention_mask_axes_0 = const()[name = string("extended_attention_mask_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [1, 1, 1, 512]> extended_attention_mask = expand_dims(axes = extended_attention_mask_axes_0, x = var_98)[name = string("extended_attention_mask")]; |
| tensor<int32, [1]> var_100_axes_0 = const()[name = string("op_100_axes_0"), val = tensor<int32, [1]>([-2])]; |
| string extended_attention_mask_to_uint16_dtype_0 = const()[name = string("extended_attention_mask_to_uint16_dtype_0"), val = string("uint16")]; |
| tensor<uint16, [1, 1, 1, 512]> extended_attention_mask_to_uint16 = cast(dtype = extended_attention_mask_to_uint16_dtype_0, x = extended_attention_mask)[name = string("cast_294")]; |
| tensor<uint16, [1, 1, 512]> var_100_cast_int16 = squeeze(axes = var_100_axes_0, x = extended_attention_mask_to_uint16)[name = string("op_100_cast_int16")]; |
| string var_100_cast_int16_to_int32_dtype_0 = const()[name = string("op_100_cast_int16_to_int32_dtype_0"), val = string("int32")]; |
| tensor<int32, [1]> var_101_axes_0 = const()[name = string("op_101_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<int32, [1, 1, 512]> var_100_cast_int16_to_int32 = cast(dtype = var_100_cast_int16_to_int32_dtype_0, x = var_100_cast_int16)[name = string("cast_293")]; |
| tensor<int32, [1, 1, 512, 1]> var_101 = expand_dims(axes = var_101_axes_0, x = var_100_cast_int16_to_int32)[name = string("op_101")]; |
| tensor<int32, [1, 1, 512, 512]> attention_mask_3 = mul(x = extended_attention_mask, y = var_101)[name = string("attention_mask_3")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_0_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196618816)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197798528)))]; |
| tensor<fp16, [1, 512, 768]> linear_0_cast_fp16 = linear(bias = deberta_encoder_layer_0_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_0_attention_self_query_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; |
| tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_3_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("x_3_cast_fp16")]; |
| tensor<int32, [4]> var_160 = const()[name = string("op_160"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_165 = const()[name = string("op_165"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_161_cast_fp16 = transpose(perm = var_160, x = x_3_cast_fp16)[name = string("transpose_79")]; |
| tensor<fp16, [12, 512, 64]> query_layer_3_cast_fp16 = reshape(shape = var_165, x = var_161_cast_fp16)[name = string("query_layer_3_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_0_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197800128)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198979840)))]; |
| tensor<fp16, [1, 512, 768]> linear_1_cast_fp16 = linear(bias = deberta_encoder_layer_0_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_0_attention_self_key_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")]; |
| tensor<int32, [4]> var_172 = const()[name = string("op_172"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_7_cast_fp16 = reshape(shape = var_172, x = linear_1_cast_fp16)[name = string("x_7_cast_fp16")]; |
| tensor<int32, [4]> var_174 = const()[name = string("op_174"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_179 = const()[name = string("op_179"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_175_cast_fp16 = transpose(perm = var_174, x = x_7_cast_fp16)[name = string("transpose_78")]; |
| tensor<fp16, [12, 512, 64]> key_layer_1_cast_fp16 = reshape(shape = var_179, x = var_175_cast_fp16)[name = string("key_layer_1_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_0_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198981440)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200161152)))]; |
| tensor<fp16, [1, 512, 768]> linear_2_cast_fp16 = linear(bias = deberta_encoder_layer_0_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_0_attention_self_value_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")]; |
| tensor<int32, [4]> var_186 = const()[name = string("op_186"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_11_cast_fp16 = reshape(shape = var_186, x = linear_2_cast_fp16)[name = string("x_11_cast_fp16")]; |
| tensor<int32, [4]> var_188 = const()[name = string("op_188"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_193 = const()[name = string("op_193"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_189_cast_fp16 = transpose(perm = var_188, x = x_11_cast_fp16)[name = string("transpose_77")]; |
| tensor<fp16, [12, 512, 64]> value_layer_1_cast_fp16 = reshape(shape = var_193, x = var_189_cast_fp16)[name = string("value_layer_1_cast_fp16")]; |
| tensor<int32, [3]> var_203_perm_0 = const()[name = string("op_203_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_205_y_0_to_fp16 = const()[name = string("_inversed_205_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_203_cast_fp16 = transpose(perm = var_203_perm_0, x = key_layer_1_cast_fp16)[name = string("transpose_76")]; |
| tensor<fp16, [12, 64, 512]> _inversed_205_cast_fp16 = mul(x = var_203_cast_fp16, y = _inversed_205_y_0_to_fp16)[name = string("_inversed_205_cast_fp16")]; |
| bool attention_scores_1_transpose_x_0 = const()[name = string("attention_scores_1_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_1_transpose_y_0 = const()[name = string("attention_scores_1_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_0, transpose_y = attention_scores_1_transpose_y_0, x = query_layer_3_cast_fp16, y = _inversed_205_cast_fp16)[name = string("attention_scores_1_cast_fp16")]; |
| bool c2p_att_1_transpose_x_0 = const()[name = string("c2p_att_1_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_1_transpose_y_0 = const()[name = string("c2p_att_1_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200162752)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_1_cast_fp16 = matmul(transpose_x = c2p_att_1_transpose_x_0, transpose_y = c2p_att_1_transpose_y_0, x = query_layer_3_cast_fp16, y = var_261_to_fp16)[name = string("c2p_att_1_cast_fp16")]; |
| bool c2p_att_3_validate_indices_0 = const()[name = string("c2p_att_3_validate_indices_0"), val = bool(false)]; |
| tensor<uint16, [12, 512, 512]> var_271_to_uint16 = const()[name = string("op_271_to_uint16"), val = tensor<uint16, [12, 512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200949248)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_3_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_3_validate_indices_0, x = c2p_att_1_cast_fp16)[name = string("c2p_att_3_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_274_y_0_to_fp16 = const()[name = string("_inversed_274_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_274_cast_fp16 = mul(x = c2p_att_3_cast_fp16_cast_int16, y = _inversed_274_y_0_to_fp16)[name = string("_inversed_274_cast_fp16")]; |
| bool p2c_att_1_transpose_x_0 = const()[name = string("p2c_att_1_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_1_transpose_y_0 = const()[name = string("p2c_att_1_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_289_to_fp16 = const()[name = string("op_289_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207240768)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_1_cast_fp16 = matmul(transpose_x = p2c_att_1_transpose_x_0, transpose_y = p2c_att_1_transpose_y_0, x = key_layer_1_cast_fp16, y = var_289_to_fp16)[name = string("p2c_att_1_cast_fp16")]; |
| bool var_297_validate_indices_0 = const()[name = string("op_297_validate_indices_0"), val = bool(false)]; |
| tensor<uint16, [12, 512, 512]> var_296_to_uint16 = const()[name = string("op_296_to_uint16"), val = tensor<uint16, [12, 512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208027264)))]; |
| tensor<fp16, [12, 512, 512]> var_297_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_297_validate_indices_0, x = p2c_att_1_cast_fp16)[name = string("op_297_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_3_perm_0 = const()[name = string("p2c_att_3_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_300_y_0_to_fp16 = const()[name = string("_inversed_300_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_3_cast_fp16 = transpose(perm = p2c_att_3_perm_0, x = var_297_cast_fp16_cast_int16)[name = string("transpose_75")]; |
| tensor<fp16, [12, 512, 512]> _inversed_300_cast_fp16 = mul(x = p2c_att_3_cast_fp16, y = _inversed_300_y_0_to_fp16)[name = string("_inversed_300_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_1_cast_fp16 = add(x = _inversed_274_cast_fp16, y = _inversed_300_cast_fp16)[name = string("rel_att_1_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_3_cast_fp16 = add(x = attention_scores_1_cast_fp16, y = rel_att_1_cast_fp16)[name = string("attention_scores_3_cast_fp16")]; |
| tensor<int32, [4]> var_305 = const()[name = string("op_305"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_5_cast_fp16 = reshape(shape = var_305, x = attention_scores_3_cast_fp16)[name = string("attention_scores_5_cast_fp16")]; |
| string cast_27_dtype_0 = const()[name = string("cast_27_dtype_0"), val = string("bool")]; |
| tensor<bool, [1, 1, 512, 512]> cast_27 = cast(dtype = cast_27_dtype_0, x = attention_mask_3)[name = string("cast_292")]; |
| tensor<bool, [1, 1, 512, 512]> var_308 = logical_not(x = cast_27)[name = string("op_308")]; |
| fp16 var_41_to_fp16 = const()[name = string("op_41_to_fp16"), val = fp16(-inf)]; |
| tensor<fp16, [1, 12, 512, 512]> input_9_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_5_cast_fp16, cond = var_308)[name = string("input_9_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_11_cast_fp16 = softmax(axis = var_32, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")]; |
| tensor<int32, [3]> var_314 = const()[name = string("op_314"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_315_cast_fp16 = reshape(shape = var_314, x = input_11_cast_fp16)[name = string("op_315_cast_fp16")]; |
| bool context_layer_1_transpose_x_0 = const()[name = string("context_layer_1_transpose_x_0"), val = bool(false)]; |
| bool context_layer_1_transpose_y_0 = const()[name = string("context_layer_1_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = var_315_cast_fp16, y = value_layer_1_cast_fp16)[name = string("context_layer_1_cast_fp16")]; |
| tensor<int32, [4]> var_319 = const()[name = string("op_319"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_320_cast_fp16 = reshape(shape = var_319, x = context_layer_1_cast_fp16)[name = string("op_320_cast_fp16")]; |
| tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_326 = const()[name = string("op_326"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_322_cast_fp16 = transpose(perm = var_321, x = var_320_cast_fp16)[name = string("transpose_74")]; |
| tensor<fp16, [1, 512, 768]> input_13_cast_fp16 = reshape(shape = var_326, x = var_322_cast_fp16)[name = string("input_13_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_0_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214318784)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215498496)))]; |
| tensor<fp16, [1, 512, 768]> linear_5_cast_fp16 = linear(bias = deberta_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_5_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_17_cast_fp16 = add(x = linear_5_cast_fp16, y = input_1_cast_fp16)[name = string("input_17_cast_fp16")]; |
| tensor<int32, [1]> input_19_axes_0 = const()[name = string("input_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215500096)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215501696)))]; |
| tensor<fp16, [1, 512, 768]> input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = deberta_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_17_cast_fp16)[name = string("input_19_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_0_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_0_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215503296)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_0_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_0_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220221952)))]; |
| tensor<fp16, [1, 512, 3072]> linear_6_cast_fp16 = linear(bias = deberta_encoder_layer_0_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_0_intermediate_dense_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_6_cast_fp16")]; |
| string input_23_mode_0 = const()[name = string("input_23_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = string("input_23_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_0_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_0_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220228160)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_0_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224946816)))]; |
| tensor<fp16, [1, 512, 768]> linear_7_cast_fp16 = linear(bias = deberta_encoder_layer_0_output_dense_bias_to_fp16, weight = deberta_encoder_layer_0_output_dense_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_7_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_27_cast_fp16 = add(x = linear_7_cast_fp16, y = input_19_cast_fp16)[name = string("input_27_cast_fp16")]; |
| tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_0_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224948416)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_0_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_0_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224950016)))]; |
| tensor<fp16, [1, 512, 768]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = deberta_encoder_layer_0_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_0_output_LayerNorm_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_1_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224951616)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226131328)))]; |
| tensor<fp16, [1, 512, 768]> linear_8_cast_fp16 = linear(bias = deberta_encoder_layer_1_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_1_attention_self_query_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_8_cast_fp16")]; |
| tensor<int32, [4]> var_377 = const()[name = string("op_377"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_23_cast_fp16 = reshape(shape = var_377, x = linear_8_cast_fp16)[name = string("x_23_cast_fp16")]; |
| tensor<int32, [4]> var_379 = const()[name = string("op_379"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_384 = const()[name = string("op_384"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_380_cast_fp16 = transpose(perm = var_379, x = x_23_cast_fp16)[name = string("transpose_73")]; |
| tensor<fp16, [12, 512, 64]> query_layer_9_cast_fp16 = reshape(shape = var_384, x = var_380_cast_fp16)[name = string("query_layer_9_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_1_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226132928)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227312640)))]; |
| tensor<fp16, [1, 512, 768]> linear_9_cast_fp16 = linear(bias = deberta_encoder_layer_1_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_1_attention_self_key_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_9_cast_fp16")]; |
| tensor<int32, [4]> var_391 = const()[name = string("op_391"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_27_cast_fp16 = reshape(shape = var_391, x = linear_9_cast_fp16)[name = string("x_27_cast_fp16")]; |
| tensor<int32, [4]> var_393 = const()[name = string("op_393"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_398 = const()[name = string("op_398"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_394_cast_fp16 = transpose(perm = var_393, x = x_27_cast_fp16)[name = string("transpose_72")]; |
| tensor<fp16, [12, 512, 64]> key_layer_3_cast_fp16 = reshape(shape = var_398, x = var_394_cast_fp16)[name = string("key_layer_3_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_1_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227314240)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228493952)))]; |
| tensor<fp16, [1, 512, 768]> linear_10_cast_fp16 = linear(bias = deberta_encoder_layer_1_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_1_attention_self_value_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_10_cast_fp16")]; |
| tensor<int32, [4]> var_405 = const()[name = string("op_405"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_31_cast_fp16 = reshape(shape = var_405, x = linear_10_cast_fp16)[name = string("x_31_cast_fp16")]; |
| tensor<int32, [4]> var_407 = const()[name = string("op_407"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_412 = const()[name = string("op_412"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_408_cast_fp16 = transpose(perm = var_407, x = x_31_cast_fp16)[name = string("transpose_71")]; |
| tensor<fp16, [12, 512, 64]> value_layer_3_cast_fp16 = reshape(shape = var_412, x = var_408_cast_fp16)[name = string("value_layer_3_cast_fp16")]; |
| tensor<int32, [3]> var_422_perm_0 = const()[name = string("op_422_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_424_y_0_to_fp16 = const()[name = string("_inversed_424_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_422_cast_fp16 = transpose(perm = var_422_perm_0, x = key_layer_3_cast_fp16)[name = string("transpose_70")]; |
| tensor<fp16, [12, 64, 512]> _inversed_424_cast_fp16 = mul(x = var_422_cast_fp16, y = _inversed_424_y_0_to_fp16)[name = string("_inversed_424_cast_fp16")]; |
| bool attention_scores_7_transpose_x_0 = const()[name = string("attention_scores_7_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_7_transpose_y_0 = const()[name = string("attention_scores_7_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_7_cast_fp16 = matmul(transpose_x = attention_scores_7_transpose_x_0, transpose_y = attention_scores_7_transpose_y_0, x = query_layer_9_cast_fp16, y = _inversed_424_cast_fp16)[name = string("attention_scores_7_cast_fp16")]; |
| bool c2p_att_5_transpose_x_0 = const()[name = string("c2p_att_5_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_5_transpose_y_0 = const()[name = string("c2p_att_5_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_480_to_fp16 = const()[name = string("op_480_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228495552)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_5_cast_fp16 = matmul(transpose_x = c2p_att_5_transpose_x_0, transpose_y = c2p_att_5_transpose_y_0, x = query_layer_9_cast_fp16, y = var_480_to_fp16)[name = string("c2p_att_5_cast_fp16")]; |
| bool c2p_att_7_validate_indices_0 = const()[name = string("c2p_att_7_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_7_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_7_validate_indices_0, x = c2p_att_5_cast_fp16)[name = string("c2p_att_7_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_493_y_0_to_fp16 = const()[name = string("_inversed_493_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_493_cast_fp16 = mul(x = c2p_att_7_cast_fp16_cast_int16, y = _inversed_493_y_0_to_fp16)[name = string("_inversed_493_cast_fp16")]; |
| bool p2c_att_5_transpose_x_0 = const()[name = string("p2c_att_5_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_5_transpose_y_0 = const()[name = string("p2c_att_5_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_508_to_fp16 = const()[name = string("op_508_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229282048)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_5_cast_fp16 = matmul(transpose_x = p2c_att_5_transpose_x_0, transpose_y = p2c_att_5_transpose_y_0, x = key_layer_3_cast_fp16, y = var_508_to_fp16)[name = string("p2c_att_5_cast_fp16")]; |
| bool var_516_validate_indices_0 = const()[name = string("op_516_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_516_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_516_validate_indices_0, x = p2c_att_5_cast_fp16)[name = string("op_516_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_7_perm_0 = const()[name = string("p2c_att_7_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_519_y_0_to_fp16 = const()[name = string("_inversed_519_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_7_cast_fp16 = transpose(perm = p2c_att_7_perm_0, x = var_516_cast_fp16_cast_int16)[name = string("transpose_69")]; |
| tensor<fp16, [12, 512, 512]> _inversed_519_cast_fp16 = mul(x = p2c_att_7_cast_fp16, y = _inversed_519_y_0_to_fp16)[name = string("_inversed_519_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_3_cast_fp16 = add(x = _inversed_493_cast_fp16, y = _inversed_519_cast_fp16)[name = string("rel_att_3_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_9_cast_fp16 = add(x = attention_scores_7_cast_fp16, y = rel_att_3_cast_fp16)[name = string("attention_scores_9_cast_fp16")]; |
| tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_11_cast_fp16 = reshape(shape = var_524, x = attention_scores_9_cast_fp16)[name = string("attention_scores_11_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_33_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_11_cast_fp16, cond = var_308)[name = string("input_33_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_35_cast_fp16 = softmax(axis = var_32, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")]; |
| tensor<int32, [3]> var_533 = const()[name = string("op_533"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_534_cast_fp16 = reshape(shape = var_533, x = input_35_cast_fp16)[name = string("op_534_cast_fp16")]; |
| bool context_layer_5_transpose_x_0 = const()[name = string("context_layer_5_transpose_x_0"), val = bool(false)]; |
| bool context_layer_5_transpose_y_0 = const()[name = string("context_layer_5_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = var_534_cast_fp16, y = value_layer_3_cast_fp16)[name = string("context_layer_5_cast_fp16")]; |
| tensor<int32, [4]> var_538 = const()[name = string("op_538"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_539_cast_fp16 = reshape(shape = var_538, x = context_layer_5_cast_fp16)[name = string("op_539_cast_fp16")]; |
| tensor<int32, [4]> var_540 = const()[name = string("op_540"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_545 = const()[name = string("op_545"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_541_cast_fp16 = transpose(perm = var_540, x = var_539_cast_fp16)[name = string("transpose_68")]; |
| tensor<fp16, [1, 512, 768]> input_37_cast_fp16 = reshape(shape = var_545, x = var_541_cast_fp16)[name = string("input_37_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_1_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230068544)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231248256)))]; |
| tensor<fp16, [1, 512, 768]> linear_13_cast_fp16 = linear(bias = deberta_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_37_cast_fp16)[name = string("linear_13_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_41_cast_fp16 = add(x = linear_13_cast_fp16, y = input_29_cast_fp16)[name = string("input_41_cast_fp16")]; |
| tensor<int32, [1]> input_43_axes_0 = const()[name = string("input_43_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231249856)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231251456)))]; |
| tensor<fp16, [1, 512, 768]> input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = deberta_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_41_cast_fp16)[name = string("input_43_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_1_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_1_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231253056)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_1_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_1_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235971712)))]; |
| tensor<fp16, [1, 512, 3072]> linear_14_cast_fp16 = linear(bias = deberta_encoder_layer_1_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_1_intermediate_dense_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_14_cast_fp16")]; |
| string input_47_mode_0 = const()[name = string("input_47_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = string("input_47_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_1_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_1_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235977920)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_1_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240696576)))]; |
| tensor<fp16, [1, 512, 768]> linear_15_cast_fp16 = linear(bias = deberta_encoder_layer_1_output_dense_bias_to_fp16, weight = deberta_encoder_layer_1_output_dense_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_15_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_51_cast_fp16 = add(x = linear_15_cast_fp16, y = input_43_cast_fp16)[name = string("input_51_cast_fp16")]; |
| tensor<int32, [1]> input_53_axes_0 = const()[name = string("input_53_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_1_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240698176)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_1_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_1_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240699776)))]; |
| tensor<fp16, [1, 512, 768]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = deberta_encoder_layer_1_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_1_output_LayerNorm_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_2_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240701376)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241881088)))]; |
| tensor<fp16, [1, 512, 768]> linear_16_cast_fp16 = linear(bias = deberta_encoder_layer_2_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_2_attention_self_query_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_16_cast_fp16")]; |
| tensor<int32, [4]> var_596 = const()[name = string("op_596"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_43_cast_fp16 = reshape(shape = var_596, x = linear_16_cast_fp16)[name = string("x_43_cast_fp16")]; |
| tensor<int32, [4]> var_598 = const()[name = string("op_598"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_603 = const()[name = string("op_603"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_599_cast_fp16 = transpose(perm = var_598, x = x_43_cast_fp16)[name = string("transpose_67")]; |
| tensor<fp16, [12, 512, 64]> query_layer_15_cast_fp16 = reshape(shape = var_603, x = var_599_cast_fp16)[name = string("query_layer_15_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_2_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241882688)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243062400)))]; |
| tensor<fp16, [1, 512, 768]> linear_17_cast_fp16 = linear(bias = deberta_encoder_layer_2_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_2_attention_self_key_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_17_cast_fp16")]; |
| tensor<int32, [4]> var_610 = const()[name = string("op_610"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_47_cast_fp16 = reshape(shape = var_610, x = linear_17_cast_fp16)[name = string("x_47_cast_fp16")]; |
| tensor<int32, [4]> var_612 = const()[name = string("op_612"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_617 = const()[name = string("op_617"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_613_cast_fp16 = transpose(perm = var_612, x = x_47_cast_fp16)[name = string("transpose_66")]; |
| tensor<fp16, [12, 512, 64]> key_layer_5_cast_fp16 = reshape(shape = var_617, x = var_613_cast_fp16)[name = string("key_layer_5_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_2_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243064000)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244243712)))]; |
| tensor<fp16, [1, 512, 768]> linear_18_cast_fp16 = linear(bias = deberta_encoder_layer_2_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_2_attention_self_value_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_18_cast_fp16")]; |
| tensor<int32, [4]> var_624 = const()[name = string("op_624"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_51_cast_fp16 = reshape(shape = var_624, x = linear_18_cast_fp16)[name = string("x_51_cast_fp16")]; |
| tensor<int32, [4]> var_626 = const()[name = string("op_626"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_631 = const()[name = string("op_631"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_627_cast_fp16 = transpose(perm = var_626, x = x_51_cast_fp16)[name = string("transpose_65")]; |
| tensor<fp16, [12, 512, 64]> value_layer_5_cast_fp16 = reshape(shape = var_631, x = var_627_cast_fp16)[name = string("value_layer_5_cast_fp16")]; |
| tensor<int32, [3]> var_641_perm_0 = const()[name = string("op_641_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_643_y_0_to_fp16 = const()[name = string("_inversed_643_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_641_cast_fp16 = transpose(perm = var_641_perm_0, x = key_layer_5_cast_fp16)[name = string("transpose_64")]; |
| tensor<fp16, [12, 64, 512]> _inversed_643_cast_fp16 = mul(x = var_641_cast_fp16, y = _inversed_643_y_0_to_fp16)[name = string("_inversed_643_cast_fp16")]; |
| bool attention_scores_13_transpose_x_0 = const()[name = string("attention_scores_13_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_13_transpose_y_0 = const()[name = string("attention_scores_13_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_0, transpose_y = attention_scores_13_transpose_y_0, x = query_layer_15_cast_fp16, y = _inversed_643_cast_fp16)[name = string("attention_scores_13_cast_fp16")]; |
| bool c2p_att_9_transpose_x_0 = const()[name = string("c2p_att_9_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_9_transpose_y_0 = const()[name = string("c2p_att_9_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_699_to_fp16 = const()[name = string("op_699_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244245312)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_9_cast_fp16 = matmul(transpose_x = c2p_att_9_transpose_x_0, transpose_y = c2p_att_9_transpose_y_0, x = query_layer_15_cast_fp16, y = var_699_to_fp16)[name = string("c2p_att_9_cast_fp16")]; |
| bool c2p_att_11_validate_indices_0 = const()[name = string("c2p_att_11_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_11_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_11_validate_indices_0, x = c2p_att_9_cast_fp16)[name = string("c2p_att_11_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_712_y_0_to_fp16 = const()[name = string("_inversed_712_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_712_cast_fp16 = mul(x = c2p_att_11_cast_fp16_cast_int16, y = _inversed_712_y_0_to_fp16)[name = string("_inversed_712_cast_fp16")]; |
| bool p2c_att_9_transpose_x_0 = const()[name = string("p2c_att_9_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_9_transpose_y_0 = const()[name = string("p2c_att_9_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_727_to_fp16 = const()[name = string("op_727_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245031808)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_9_cast_fp16 = matmul(transpose_x = p2c_att_9_transpose_x_0, transpose_y = p2c_att_9_transpose_y_0, x = key_layer_5_cast_fp16, y = var_727_to_fp16)[name = string("p2c_att_9_cast_fp16")]; |
| bool var_735_validate_indices_0 = const()[name = string("op_735_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_735_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_735_validate_indices_0, x = p2c_att_9_cast_fp16)[name = string("op_735_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_11_perm_0 = const()[name = string("p2c_att_11_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_738_y_0_to_fp16 = const()[name = string("_inversed_738_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_11_cast_fp16 = transpose(perm = p2c_att_11_perm_0, x = var_735_cast_fp16_cast_int16)[name = string("transpose_63")]; |
| tensor<fp16, [12, 512, 512]> _inversed_738_cast_fp16 = mul(x = p2c_att_11_cast_fp16, y = _inversed_738_y_0_to_fp16)[name = string("_inversed_738_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_5_cast_fp16 = add(x = _inversed_712_cast_fp16, y = _inversed_738_cast_fp16)[name = string("rel_att_5_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_15_cast_fp16 = add(x = attention_scores_13_cast_fp16, y = rel_att_5_cast_fp16)[name = string("attention_scores_15_cast_fp16")]; |
| tensor<int32, [4]> var_743 = const()[name = string("op_743"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_17_cast_fp16 = reshape(shape = var_743, x = attention_scores_15_cast_fp16)[name = string("attention_scores_17_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_57_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_17_cast_fp16, cond = var_308)[name = string("input_57_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_59_cast_fp16 = softmax(axis = var_32, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; |
| tensor<int32, [3]> var_752 = const()[name = string("op_752"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_753_cast_fp16 = reshape(shape = var_752, x = input_59_cast_fp16)[name = string("op_753_cast_fp16")]; |
| bool context_layer_9_transpose_x_0 = const()[name = string("context_layer_9_transpose_x_0"), val = bool(false)]; |
| bool context_layer_9_transpose_y_0 = const()[name = string("context_layer_9_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = var_753_cast_fp16, y = value_layer_5_cast_fp16)[name = string("context_layer_9_cast_fp16")]; |
| tensor<int32, [4]> var_757 = const()[name = string("op_757"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_758_cast_fp16 = reshape(shape = var_757, x = context_layer_9_cast_fp16)[name = string("op_758_cast_fp16")]; |
| tensor<int32, [4]> var_759 = const()[name = string("op_759"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_764 = const()[name = string("op_764"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_760_cast_fp16 = transpose(perm = var_759, x = var_758_cast_fp16)[name = string("transpose_62")]; |
| tensor<fp16, [1, 512, 768]> input_61_cast_fp16 = reshape(shape = var_764, x = var_760_cast_fp16)[name = string("input_61_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_2_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245818304)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246998016)))]; |
| tensor<fp16, [1, 512, 768]> linear_21_cast_fp16 = linear(bias = deberta_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_21_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_65_cast_fp16 = add(x = linear_21_cast_fp16, y = input_53_cast_fp16)[name = string("input_65_cast_fp16")]; |
| tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246999616)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247001216)))]; |
| tensor<fp16, [1, 512, 768]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = deberta_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_2_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_2_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247002816)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_2_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_2_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251721472)))]; |
| tensor<fp16, [1, 512, 3072]> linear_22_cast_fp16 = linear(bias = deberta_encoder_layer_2_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_2_intermediate_dense_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_22_cast_fp16")]; |
| string input_71_mode_0 = const()[name = string("input_71_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = string("input_71_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_2_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_2_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251727680)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_2_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256446336)))]; |
| tensor<fp16, [1, 512, 768]> linear_23_cast_fp16 = linear(bias = deberta_encoder_layer_2_output_dense_bias_to_fp16, weight = deberta_encoder_layer_2_output_dense_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_23_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_75_cast_fp16 = add(x = linear_23_cast_fp16, y = input_67_cast_fp16)[name = string("input_75_cast_fp16")]; |
| tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_2_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256447936)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_2_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_2_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256449536)))]; |
| tensor<fp16, [1, 512, 768]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = deberta_encoder_layer_2_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_2_output_LayerNorm_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_3_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256451136)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257630848)))]; |
| tensor<fp16, [1, 512, 768]> linear_24_cast_fp16 = linear(bias = deberta_encoder_layer_3_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_3_attention_self_query_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_24_cast_fp16")]; |
| tensor<int32, [4]> var_815 = const()[name = string("op_815"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_63_cast_fp16 = reshape(shape = var_815, x = linear_24_cast_fp16)[name = string("x_63_cast_fp16")]; |
| tensor<int32, [4]> var_817 = const()[name = string("op_817"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_822 = const()[name = string("op_822"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_818_cast_fp16 = transpose(perm = var_817, x = x_63_cast_fp16)[name = string("transpose_61")]; |
| tensor<fp16, [12, 512, 64]> query_layer_21_cast_fp16 = reshape(shape = var_822, x = var_818_cast_fp16)[name = string("query_layer_21_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_3_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257632448)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258812160)))]; |
| tensor<fp16, [1, 512, 768]> linear_25_cast_fp16 = linear(bias = deberta_encoder_layer_3_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_3_attention_self_key_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_25_cast_fp16")]; |
| tensor<int32, [4]> var_829 = const()[name = string("op_829"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_67_cast_fp16 = reshape(shape = var_829, x = linear_25_cast_fp16)[name = string("x_67_cast_fp16")]; |
| tensor<int32, [4]> var_831 = const()[name = string("op_831"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_836 = const()[name = string("op_836"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_832_cast_fp16 = transpose(perm = var_831, x = x_67_cast_fp16)[name = string("transpose_60")]; |
| tensor<fp16, [12, 512, 64]> key_layer_7_cast_fp16 = reshape(shape = var_836, x = var_832_cast_fp16)[name = string("key_layer_7_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_3_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258813760)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259993472)))]; |
| tensor<fp16, [1, 512, 768]> linear_26_cast_fp16 = linear(bias = deberta_encoder_layer_3_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_3_attention_self_value_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_26_cast_fp16")]; |
| tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_71_cast_fp16 = reshape(shape = var_843, x = linear_26_cast_fp16)[name = string("x_71_cast_fp16")]; |
| tensor<int32, [4]> var_845 = const()[name = string("op_845"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_850 = const()[name = string("op_850"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_846_cast_fp16 = transpose(perm = var_845, x = x_71_cast_fp16)[name = string("transpose_59")]; |
| tensor<fp16, [12, 512, 64]> value_layer_7_cast_fp16 = reshape(shape = var_850, x = var_846_cast_fp16)[name = string("value_layer_7_cast_fp16")]; |
| tensor<int32, [3]> var_860_perm_0 = const()[name = string("op_860_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_862_y_0_to_fp16 = const()[name = string("_inversed_862_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_860_cast_fp16 = transpose(perm = var_860_perm_0, x = key_layer_7_cast_fp16)[name = string("transpose_58")]; |
| tensor<fp16, [12, 64, 512]> _inversed_862_cast_fp16 = mul(x = var_860_cast_fp16, y = _inversed_862_y_0_to_fp16)[name = string("_inversed_862_cast_fp16")]; |
| bool attention_scores_19_transpose_x_0 = const()[name = string("attention_scores_19_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_19_transpose_y_0 = const()[name = string("attention_scores_19_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_19_cast_fp16 = matmul(transpose_x = attention_scores_19_transpose_x_0, transpose_y = attention_scores_19_transpose_y_0, x = query_layer_21_cast_fp16, y = _inversed_862_cast_fp16)[name = string("attention_scores_19_cast_fp16")]; |
| bool c2p_att_13_transpose_x_0 = const()[name = string("c2p_att_13_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_13_transpose_y_0 = const()[name = string("c2p_att_13_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_918_to_fp16 = const()[name = string("op_918_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259995072)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_13_cast_fp16 = matmul(transpose_x = c2p_att_13_transpose_x_0, transpose_y = c2p_att_13_transpose_y_0, x = query_layer_21_cast_fp16, y = var_918_to_fp16)[name = string("c2p_att_13_cast_fp16")]; |
| bool c2p_att_15_validate_indices_0 = const()[name = string("c2p_att_15_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_15_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_15_validate_indices_0, x = c2p_att_13_cast_fp16)[name = string("c2p_att_15_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_931_y_0_to_fp16 = const()[name = string("_inversed_931_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_931_cast_fp16 = mul(x = c2p_att_15_cast_fp16_cast_int16, y = _inversed_931_y_0_to_fp16)[name = string("_inversed_931_cast_fp16")]; |
| bool p2c_att_13_transpose_x_0 = const()[name = string("p2c_att_13_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_13_transpose_y_0 = const()[name = string("p2c_att_13_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_946_to_fp16 = const()[name = string("op_946_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260781568)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_13_cast_fp16 = matmul(transpose_x = p2c_att_13_transpose_x_0, transpose_y = p2c_att_13_transpose_y_0, x = key_layer_7_cast_fp16, y = var_946_to_fp16)[name = string("p2c_att_13_cast_fp16")]; |
| bool var_954_validate_indices_0 = const()[name = string("op_954_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_954_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_954_validate_indices_0, x = p2c_att_13_cast_fp16)[name = string("op_954_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_15_perm_0 = const()[name = string("p2c_att_15_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_957_y_0_to_fp16 = const()[name = string("_inversed_957_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_15_cast_fp16 = transpose(perm = p2c_att_15_perm_0, x = var_954_cast_fp16_cast_int16)[name = string("transpose_57")]; |
| tensor<fp16, [12, 512, 512]> _inversed_957_cast_fp16 = mul(x = p2c_att_15_cast_fp16, y = _inversed_957_y_0_to_fp16)[name = string("_inversed_957_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_7_cast_fp16 = add(x = _inversed_931_cast_fp16, y = _inversed_957_cast_fp16)[name = string("rel_att_7_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_21_cast_fp16 = add(x = attention_scores_19_cast_fp16, y = rel_att_7_cast_fp16)[name = string("attention_scores_21_cast_fp16")]; |
| tensor<int32, [4]> var_962 = const()[name = string("op_962"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_23_cast_fp16 = reshape(shape = var_962, x = attention_scores_21_cast_fp16)[name = string("attention_scores_23_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_81_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_23_cast_fp16, cond = var_308)[name = string("input_81_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_83_cast_fp16 = softmax(axis = var_32, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")]; |
| tensor<int32, [3]> var_971 = const()[name = string("op_971"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_972_cast_fp16 = reshape(shape = var_971, x = input_83_cast_fp16)[name = string("op_972_cast_fp16")]; |
| bool context_layer_13_transpose_x_0 = const()[name = string("context_layer_13_transpose_x_0"), val = bool(false)]; |
| bool context_layer_13_transpose_y_0 = const()[name = string("context_layer_13_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = var_972_cast_fp16, y = value_layer_7_cast_fp16)[name = string("context_layer_13_cast_fp16")]; |
| tensor<int32, [4]> var_976 = const()[name = string("op_976"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_977_cast_fp16 = reshape(shape = var_976, x = context_layer_13_cast_fp16)[name = string("op_977_cast_fp16")]; |
| tensor<int32, [4]> var_978 = const()[name = string("op_978"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_983 = const()[name = string("op_983"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_979_cast_fp16 = transpose(perm = var_978, x = var_977_cast_fp16)[name = string("transpose_56")]; |
| tensor<fp16, [1, 512, 768]> input_85_cast_fp16 = reshape(shape = var_983, x = var_979_cast_fp16)[name = string("input_85_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_3_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261568064)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262747776)))]; |
| tensor<fp16, [1, 512, 768]> linear_29_cast_fp16 = linear(bias = deberta_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_29_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_89_cast_fp16 = add(x = linear_29_cast_fp16, y = input_77_cast_fp16)[name = string("input_89_cast_fp16")]; |
| tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262749376)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262750976)))]; |
| tensor<fp16, [1, 512, 768]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = deberta_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_3_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_3_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262752576)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_3_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_3_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267471232)))]; |
| tensor<fp16, [1, 512, 3072]> linear_30_cast_fp16 = linear(bias = deberta_encoder_layer_3_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_3_intermediate_dense_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_30_cast_fp16")]; |
| string input_95_mode_0 = const()[name = string("input_95_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = string("input_95_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_3_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_3_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267477440)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_3_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272196096)))]; |
| tensor<fp16, [1, 512, 768]> linear_31_cast_fp16 = linear(bias = deberta_encoder_layer_3_output_dense_bias_to_fp16, weight = deberta_encoder_layer_3_output_dense_weight_to_fp16, x = input_95_cast_fp16)[name = string("linear_31_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_99_cast_fp16 = add(x = linear_31_cast_fp16, y = input_91_cast_fp16)[name = string("input_99_cast_fp16")]; |
| tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_3_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272197696)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_3_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_3_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272199296)))]; |
| tensor<fp16, [1, 512, 768]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = deberta_encoder_layer_3_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_3_output_LayerNorm_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_4_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272200896)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273380608)))]; |
| tensor<fp16, [1, 512, 768]> linear_32_cast_fp16 = linear(bias = deberta_encoder_layer_4_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_4_attention_self_query_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_32_cast_fp16")]; |
| tensor<int32, [4]> var_1034 = const()[name = string("op_1034"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_83_cast_fp16 = reshape(shape = var_1034, x = linear_32_cast_fp16)[name = string("x_83_cast_fp16")]; |
| tensor<int32, [4]> var_1036 = const()[name = string("op_1036"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1041 = const()[name = string("op_1041"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1037_cast_fp16 = transpose(perm = var_1036, x = x_83_cast_fp16)[name = string("transpose_55")]; |
| tensor<fp16, [12, 512, 64]> query_layer_27_cast_fp16 = reshape(shape = var_1041, x = var_1037_cast_fp16)[name = string("query_layer_27_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_4_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273382208)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274561920)))]; |
| tensor<fp16, [1, 512, 768]> linear_33_cast_fp16 = linear(bias = deberta_encoder_layer_4_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_4_attention_self_key_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_33_cast_fp16")]; |
| tensor<int32, [4]> var_1048 = const()[name = string("op_1048"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_87_cast_fp16 = reshape(shape = var_1048, x = linear_33_cast_fp16)[name = string("x_87_cast_fp16")]; |
| tensor<int32, [4]> var_1050 = const()[name = string("op_1050"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1055 = const()[name = string("op_1055"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1051_cast_fp16 = transpose(perm = var_1050, x = x_87_cast_fp16)[name = string("transpose_54")]; |
| tensor<fp16, [12, 512, 64]> key_layer_9_cast_fp16 = reshape(shape = var_1055, x = var_1051_cast_fp16)[name = string("key_layer_9_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_4_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274563520)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275743232)))]; |
| tensor<fp16, [1, 512, 768]> linear_34_cast_fp16 = linear(bias = deberta_encoder_layer_4_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_4_attention_self_value_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_34_cast_fp16")]; |
| tensor<int32, [4]> var_1062 = const()[name = string("op_1062"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_91_cast_fp16 = reshape(shape = var_1062, x = linear_34_cast_fp16)[name = string("x_91_cast_fp16")]; |
| tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1069 = const()[name = string("op_1069"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1065_cast_fp16 = transpose(perm = var_1064, x = x_91_cast_fp16)[name = string("transpose_53")]; |
| tensor<fp16, [12, 512, 64]> value_layer_9_cast_fp16 = reshape(shape = var_1069, x = var_1065_cast_fp16)[name = string("value_layer_9_cast_fp16")]; |
| tensor<int32, [3]> var_1079_perm_0 = const()[name = string("op_1079_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1081_y_0_to_fp16 = const()[name = string("_inversed_1081_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_1079_cast_fp16 = transpose(perm = var_1079_perm_0, x = key_layer_9_cast_fp16)[name = string("transpose_52")]; |
| tensor<fp16, [12, 64, 512]> _inversed_1081_cast_fp16 = mul(x = var_1079_cast_fp16, y = _inversed_1081_y_0_to_fp16)[name = string("_inversed_1081_cast_fp16")]; |
| bool attention_scores_25_transpose_x_0 = const()[name = string("attention_scores_25_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_25_transpose_y_0 = const()[name = string("attention_scores_25_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_25_cast_fp16 = matmul(transpose_x = attention_scores_25_transpose_x_0, transpose_y = attention_scores_25_transpose_y_0, x = query_layer_27_cast_fp16, y = _inversed_1081_cast_fp16)[name = string("attention_scores_25_cast_fp16")]; |
| bool c2p_att_17_transpose_x_0 = const()[name = string("c2p_att_17_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_17_transpose_y_0 = const()[name = string("c2p_att_17_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_1137_to_fp16 = const()[name = string("op_1137_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275744832)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_17_cast_fp16 = matmul(transpose_x = c2p_att_17_transpose_x_0, transpose_y = c2p_att_17_transpose_y_0, x = query_layer_27_cast_fp16, y = var_1137_to_fp16)[name = string("c2p_att_17_cast_fp16")]; |
| bool c2p_att_19_validate_indices_0 = const()[name = string("c2p_att_19_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_19_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_19_validate_indices_0, x = c2p_att_17_cast_fp16)[name = string("c2p_att_19_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_1150_y_0_to_fp16 = const()[name = string("_inversed_1150_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_1150_cast_fp16 = mul(x = c2p_att_19_cast_fp16_cast_int16, y = _inversed_1150_y_0_to_fp16)[name = string("_inversed_1150_cast_fp16")]; |
| bool p2c_att_17_transpose_x_0 = const()[name = string("p2c_att_17_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_17_transpose_y_0 = const()[name = string("p2c_att_17_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_1165_to_fp16 = const()[name = string("op_1165_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276531328)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_17_cast_fp16 = matmul(transpose_x = p2c_att_17_transpose_x_0, transpose_y = p2c_att_17_transpose_y_0, x = key_layer_9_cast_fp16, y = var_1165_to_fp16)[name = string("p2c_att_17_cast_fp16")]; |
| bool var_1173_validate_indices_0 = const()[name = string("op_1173_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_1173_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_1173_validate_indices_0, x = p2c_att_17_cast_fp16)[name = string("op_1173_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_19_perm_0 = const()[name = string("p2c_att_19_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1176_y_0_to_fp16 = const()[name = string("_inversed_1176_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_19_cast_fp16 = transpose(perm = p2c_att_19_perm_0, x = var_1173_cast_fp16_cast_int16)[name = string("transpose_51")]; |
| tensor<fp16, [12, 512, 512]> _inversed_1176_cast_fp16 = mul(x = p2c_att_19_cast_fp16, y = _inversed_1176_y_0_to_fp16)[name = string("_inversed_1176_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_9_cast_fp16 = add(x = _inversed_1150_cast_fp16, y = _inversed_1176_cast_fp16)[name = string("rel_att_9_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_27_cast_fp16 = add(x = attention_scores_25_cast_fp16, y = rel_att_9_cast_fp16)[name = string("attention_scores_27_cast_fp16")]; |
| tensor<int32, [4]> var_1181 = const()[name = string("op_1181"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_29_cast_fp16 = reshape(shape = var_1181, x = attention_scores_27_cast_fp16)[name = string("attention_scores_29_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_105_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_29_cast_fp16, cond = var_308)[name = string("input_105_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_107_cast_fp16 = softmax(axis = var_32, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")]; |
| tensor<int32, [3]> var_1190 = const()[name = string("op_1190"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_1191_cast_fp16 = reshape(shape = var_1190, x = input_107_cast_fp16)[name = string("op_1191_cast_fp16")]; |
| bool context_layer_17_transpose_x_0 = const()[name = string("context_layer_17_transpose_x_0"), val = bool(false)]; |
| bool context_layer_17_transpose_y_0 = const()[name = string("context_layer_17_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_17_cast_fp16 = matmul(transpose_x = context_layer_17_transpose_x_0, transpose_y = context_layer_17_transpose_y_0, x = var_1191_cast_fp16, y = value_layer_9_cast_fp16)[name = string("context_layer_17_cast_fp16")]; |
| tensor<int32, [4]> var_1195 = const()[name = string("op_1195"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1196_cast_fp16 = reshape(shape = var_1195, x = context_layer_17_cast_fp16)[name = string("op_1196_cast_fp16")]; |
| tensor<int32, [4]> var_1197 = const()[name = string("op_1197"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1202 = const()[name = string("op_1202"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_1198_cast_fp16 = transpose(perm = var_1197, x = var_1196_cast_fp16)[name = string("transpose_50")]; |
| tensor<fp16, [1, 512, 768]> input_109_cast_fp16 = reshape(shape = var_1202, x = var_1198_cast_fp16)[name = string("input_109_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_4_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277317824)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278497536)))]; |
| tensor<fp16, [1, 512, 768]> linear_37_cast_fp16 = linear(bias = deberta_encoder_layer_4_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_4_attention_output_dense_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_37_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_113_cast_fp16 = add(x = linear_37_cast_fp16, y = input_101_cast_fp16)[name = string("input_113_cast_fp16")]; |
| tensor<int32, [1]> input_115_axes_0 = const()[name = string("input_115_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278499136)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278500736)))]; |
| tensor<fp16, [1, 512, 768]> input_115_cast_fp16 = layer_norm(axes = input_115_axes_0, beta = deberta_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16, x = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_4_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_4_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278502336)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_4_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_4_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283220992)))]; |
| tensor<fp16, [1, 512, 3072]> linear_38_cast_fp16 = linear(bias = deberta_encoder_layer_4_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_4_intermediate_dense_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_38_cast_fp16")]; |
| string input_119_mode_0 = const()[name = string("input_119_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = string("input_119_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_4_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_4_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283227200)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_4_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287945856)))]; |
| tensor<fp16, [1, 512, 768]> linear_39_cast_fp16 = linear(bias = deberta_encoder_layer_4_output_dense_bias_to_fp16, weight = deberta_encoder_layer_4_output_dense_weight_to_fp16, x = input_119_cast_fp16)[name = string("linear_39_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_123_cast_fp16 = add(x = linear_39_cast_fp16, y = input_115_cast_fp16)[name = string("input_123_cast_fp16")]; |
| tensor<int32, [1]> input_125_axes_0 = const()[name = string("input_125_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_4_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287947456)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_4_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_4_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287949056)))]; |
| tensor<fp16, [1, 512, 768]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = deberta_encoder_layer_4_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_4_output_LayerNorm_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_5_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287950656)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289130368)))]; |
| tensor<fp16, [1, 512, 768]> linear_40_cast_fp16 = linear(bias = deberta_encoder_layer_5_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_5_attention_self_query_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_40_cast_fp16")]; |
| tensor<int32, [4]> var_1253 = const()[name = string("op_1253"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_103_cast_fp16 = reshape(shape = var_1253, x = linear_40_cast_fp16)[name = string("x_103_cast_fp16")]; |
| tensor<int32, [4]> var_1255 = const()[name = string("op_1255"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1260 = const()[name = string("op_1260"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1256_cast_fp16 = transpose(perm = var_1255, x = x_103_cast_fp16)[name = string("transpose_49")]; |
| tensor<fp16, [12, 512, 64]> query_layer_33_cast_fp16 = reshape(shape = var_1260, x = var_1256_cast_fp16)[name = string("query_layer_33_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_5_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289131968)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290311680)))]; |
| tensor<fp16, [1, 512, 768]> linear_41_cast_fp16 = linear(bias = deberta_encoder_layer_5_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_5_attention_self_key_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_41_cast_fp16")]; |
| tensor<int32, [4]> var_1267 = const()[name = string("op_1267"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_107_cast_fp16 = reshape(shape = var_1267, x = linear_41_cast_fp16)[name = string("x_107_cast_fp16")]; |
| tensor<int32, [4]> var_1269 = const()[name = string("op_1269"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1274 = const()[name = string("op_1274"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1270_cast_fp16 = transpose(perm = var_1269, x = x_107_cast_fp16)[name = string("transpose_48")]; |
| tensor<fp16, [12, 512, 64]> key_layer_11_cast_fp16 = reshape(shape = var_1274, x = var_1270_cast_fp16)[name = string("key_layer_11_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_5_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290313280)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291492992)))]; |
| tensor<fp16, [1, 512, 768]> linear_42_cast_fp16 = linear(bias = deberta_encoder_layer_5_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_5_attention_self_value_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_42_cast_fp16")]; |
| tensor<int32, [4]> var_1281 = const()[name = string("op_1281"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_111_cast_fp16 = reshape(shape = var_1281, x = linear_42_cast_fp16)[name = string("x_111_cast_fp16")]; |
| tensor<int32, [4]> var_1283 = const()[name = string("op_1283"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1288 = const()[name = string("op_1288"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1284_cast_fp16 = transpose(perm = var_1283, x = x_111_cast_fp16)[name = string("transpose_47")]; |
| tensor<fp16, [12, 512, 64]> value_layer_11_cast_fp16 = reshape(shape = var_1288, x = var_1284_cast_fp16)[name = string("value_layer_11_cast_fp16")]; |
| tensor<int32, [3]> var_1298_perm_0 = const()[name = string("op_1298_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1300_y_0_to_fp16 = const()[name = string("_inversed_1300_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_1298_cast_fp16 = transpose(perm = var_1298_perm_0, x = key_layer_11_cast_fp16)[name = string("transpose_46")]; |
| tensor<fp16, [12, 64, 512]> _inversed_1300_cast_fp16 = mul(x = var_1298_cast_fp16, y = _inversed_1300_y_0_to_fp16)[name = string("_inversed_1300_cast_fp16")]; |
| bool attention_scores_31_transpose_x_0 = const()[name = string("attention_scores_31_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_31_transpose_y_0 = const()[name = string("attention_scores_31_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_31_cast_fp16 = matmul(transpose_x = attention_scores_31_transpose_x_0, transpose_y = attention_scores_31_transpose_y_0, x = query_layer_33_cast_fp16, y = _inversed_1300_cast_fp16)[name = string("attention_scores_31_cast_fp16")]; |
| bool c2p_att_21_transpose_x_0 = const()[name = string("c2p_att_21_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_21_transpose_y_0 = const()[name = string("c2p_att_21_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_1356_to_fp16 = const()[name = string("op_1356_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291494592)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_21_cast_fp16 = matmul(transpose_x = c2p_att_21_transpose_x_0, transpose_y = c2p_att_21_transpose_y_0, x = query_layer_33_cast_fp16, y = var_1356_to_fp16)[name = string("c2p_att_21_cast_fp16")]; |
| bool c2p_att_23_validate_indices_0 = const()[name = string("c2p_att_23_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_23_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_23_validate_indices_0, x = c2p_att_21_cast_fp16)[name = string("c2p_att_23_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_1369_y_0_to_fp16 = const()[name = string("_inversed_1369_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_1369_cast_fp16 = mul(x = c2p_att_23_cast_fp16_cast_int16, y = _inversed_1369_y_0_to_fp16)[name = string("_inversed_1369_cast_fp16")]; |
| bool p2c_att_21_transpose_x_0 = const()[name = string("p2c_att_21_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_21_transpose_y_0 = const()[name = string("p2c_att_21_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_1384_to_fp16 = const()[name = string("op_1384_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292281088)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_21_cast_fp16 = matmul(transpose_x = p2c_att_21_transpose_x_0, transpose_y = p2c_att_21_transpose_y_0, x = key_layer_11_cast_fp16, y = var_1384_to_fp16)[name = string("p2c_att_21_cast_fp16")]; |
| bool var_1392_validate_indices_0 = const()[name = string("op_1392_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_1392_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_1392_validate_indices_0, x = p2c_att_21_cast_fp16)[name = string("op_1392_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_23_perm_0 = const()[name = string("p2c_att_23_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1395_y_0_to_fp16 = const()[name = string("_inversed_1395_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_23_cast_fp16 = transpose(perm = p2c_att_23_perm_0, x = var_1392_cast_fp16_cast_int16)[name = string("transpose_45")]; |
| tensor<fp16, [12, 512, 512]> _inversed_1395_cast_fp16 = mul(x = p2c_att_23_cast_fp16, y = _inversed_1395_y_0_to_fp16)[name = string("_inversed_1395_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_11_cast_fp16 = add(x = _inversed_1369_cast_fp16, y = _inversed_1395_cast_fp16)[name = string("rel_att_11_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_33_cast_fp16 = add(x = attention_scores_31_cast_fp16, y = rel_att_11_cast_fp16)[name = string("attention_scores_33_cast_fp16")]; |
| tensor<int32, [4]> var_1400 = const()[name = string("op_1400"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_35_cast_fp16 = reshape(shape = var_1400, x = attention_scores_33_cast_fp16)[name = string("attention_scores_35_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_129_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_35_cast_fp16, cond = var_308)[name = string("input_129_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_131_cast_fp16 = softmax(axis = var_32, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; |
| tensor<int32, [3]> var_1409 = const()[name = string("op_1409"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_1410_cast_fp16 = reshape(shape = var_1409, x = input_131_cast_fp16)[name = string("op_1410_cast_fp16")]; |
| bool context_layer_21_transpose_x_0 = const()[name = string("context_layer_21_transpose_x_0"), val = bool(false)]; |
| bool context_layer_21_transpose_y_0 = const()[name = string("context_layer_21_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_21_cast_fp16 = matmul(transpose_x = context_layer_21_transpose_x_0, transpose_y = context_layer_21_transpose_y_0, x = var_1410_cast_fp16, y = value_layer_11_cast_fp16)[name = string("context_layer_21_cast_fp16")]; |
| tensor<int32, [4]> var_1414 = const()[name = string("op_1414"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1415_cast_fp16 = reshape(shape = var_1414, x = context_layer_21_cast_fp16)[name = string("op_1415_cast_fp16")]; |
| tensor<int32, [4]> var_1416 = const()[name = string("op_1416"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1421 = const()[name = string("op_1421"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_1417_cast_fp16 = transpose(perm = var_1416, x = var_1415_cast_fp16)[name = string("transpose_44")]; |
| tensor<fp16, [1, 512, 768]> input_133_cast_fp16 = reshape(shape = var_1421, x = var_1417_cast_fp16)[name = string("input_133_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_5_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293067584)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294247296)))]; |
| tensor<fp16, [1, 512, 768]> linear_45_cast_fp16 = linear(bias = deberta_encoder_layer_5_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_5_attention_output_dense_weight_to_fp16, x = input_133_cast_fp16)[name = string("linear_45_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_137_cast_fp16 = add(x = linear_45_cast_fp16, y = input_125_cast_fp16)[name = string("input_137_cast_fp16")]; |
| tensor<int32, [1]> input_139_axes_0 = const()[name = string("input_139_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294248896)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294250496)))]; |
| tensor<fp16, [1, 512, 768]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = deberta_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_5_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_5_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294252096)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_5_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_5_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298970752)))]; |
| tensor<fp16, [1, 512, 3072]> linear_46_cast_fp16 = linear(bias = deberta_encoder_layer_5_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_5_intermediate_dense_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_46_cast_fp16")]; |
| string input_143_mode_0 = const()[name = string("input_143_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = string("input_143_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_5_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_5_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298976960)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_5_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303695616)))]; |
| tensor<fp16, [1, 512, 768]> linear_47_cast_fp16 = linear(bias = deberta_encoder_layer_5_output_dense_bias_to_fp16, weight = deberta_encoder_layer_5_output_dense_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_47_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_147_cast_fp16 = add(x = linear_47_cast_fp16, y = input_139_cast_fp16)[name = string("input_147_cast_fp16")]; |
| tensor<int32, [1]> input_149_axes_0 = const()[name = string("input_149_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_5_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303697216)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_5_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_5_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303698816)))]; |
| tensor<fp16, [1, 512, 768]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = deberta_encoder_layer_5_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_5_output_LayerNorm_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_6_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303700416)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304880128)))]; |
| tensor<fp16, [1, 512, 768]> linear_48_cast_fp16 = linear(bias = deberta_encoder_layer_6_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_6_attention_self_query_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_48_cast_fp16")]; |
| tensor<int32, [4]> var_1472 = const()[name = string("op_1472"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_123_cast_fp16 = reshape(shape = var_1472, x = linear_48_cast_fp16)[name = string("x_123_cast_fp16")]; |
| tensor<int32, [4]> var_1474 = const()[name = string("op_1474"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1479 = const()[name = string("op_1479"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1475_cast_fp16 = transpose(perm = var_1474, x = x_123_cast_fp16)[name = string("transpose_43")]; |
| tensor<fp16, [12, 512, 64]> query_layer_39_cast_fp16 = reshape(shape = var_1479, x = var_1475_cast_fp16)[name = string("query_layer_39_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_6_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304881728)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306061440)))]; |
| tensor<fp16, [1, 512, 768]> linear_49_cast_fp16 = linear(bias = deberta_encoder_layer_6_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_6_attention_self_key_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_49_cast_fp16")]; |
| tensor<int32, [4]> var_1486 = const()[name = string("op_1486"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_127_cast_fp16 = reshape(shape = var_1486, x = linear_49_cast_fp16)[name = string("x_127_cast_fp16")]; |
| tensor<int32, [4]> var_1488 = const()[name = string("op_1488"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1493 = const()[name = string("op_1493"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1489_cast_fp16 = transpose(perm = var_1488, x = x_127_cast_fp16)[name = string("transpose_42")]; |
| tensor<fp16, [12, 512, 64]> key_layer_13_cast_fp16 = reshape(shape = var_1493, x = var_1489_cast_fp16)[name = string("key_layer_13_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_6_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306063040)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307242752)))]; |
| tensor<fp16, [1, 512, 768]> linear_50_cast_fp16 = linear(bias = deberta_encoder_layer_6_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_6_attention_self_value_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_50_cast_fp16")]; |
| tensor<int32, [4]> var_1500 = const()[name = string("op_1500"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_131_cast_fp16 = reshape(shape = var_1500, x = linear_50_cast_fp16)[name = string("x_131_cast_fp16")]; |
| tensor<int32, [4]> var_1502 = const()[name = string("op_1502"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1507 = const()[name = string("op_1507"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1503_cast_fp16 = transpose(perm = var_1502, x = x_131_cast_fp16)[name = string("transpose_41")]; |
| tensor<fp16, [12, 512, 64]> value_layer_13_cast_fp16 = reshape(shape = var_1507, x = var_1503_cast_fp16)[name = string("value_layer_13_cast_fp16")]; |
| tensor<int32, [3]> var_1517_perm_0 = const()[name = string("op_1517_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1519_y_0_to_fp16 = const()[name = string("_inversed_1519_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_1517_cast_fp16 = transpose(perm = var_1517_perm_0, x = key_layer_13_cast_fp16)[name = string("transpose_40")]; |
| tensor<fp16, [12, 64, 512]> _inversed_1519_cast_fp16 = mul(x = var_1517_cast_fp16, y = _inversed_1519_y_0_to_fp16)[name = string("_inversed_1519_cast_fp16")]; |
| bool attention_scores_37_transpose_x_0 = const()[name = string("attention_scores_37_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_37_transpose_y_0 = const()[name = string("attention_scores_37_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_37_cast_fp16 = matmul(transpose_x = attention_scores_37_transpose_x_0, transpose_y = attention_scores_37_transpose_y_0, x = query_layer_39_cast_fp16, y = _inversed_1519_cast_fp16)[name = string("attention_scores_37_cast_fp16")]; |
| bool c2p_att_25_transpose_x_0 = const()[name = string("c2p_att_25_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_25_transpose_y_0 = const()[name = string("c2p_att_25_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_1575_to_fp16 = const()[name = string("op_1575_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307244352)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_25_cast_fp16 = matmul(transpose_x = c2p_att_25_transpose_x_0, transpose_y = c2p_att_25_transpose_y_0, x = query_layer_39_cast_fp16, y = var_1575_to_fp16)[name = string("c2p_att_25_cast_fp16")]; |
| bool c2p_att_27_validate_indices_0 = const()[name = string("c2p_att_27_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_27_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_27_validate_indices_0, x = c2p_att_25_cast_fp16)[name = string("c2p_att_27_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_1588_y_0_to_fp16 = const()[name = string("_inversed_1588_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_1588_cast_fp16 = mul(x = c2p_att_27_cast_fp16_cast_int16, y = _inversed_1588_y_0_to_fp16)[name = string("_inversed_1588_cast_fp16")]; |
| bool p2c_att_25_transpose_x_0 = const()[name = string("p2c_att_25_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_25_transpose_y_0 = const()[name = string("p2c_att_25_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_1603_to_fp16 = const()[name = string("op_1603_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308030848)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_25_cast_fp16 = matmul(transpose_x = p2c_att_25_transpose_x_0, transpose_y = p2c_att_25_transpose_y_0, x = key_layer_13_cast_fp16, y = var_1603_to_fp16)[name = string("p2c_att_25_cast_fp16")]; |
| bool var_1611_validate_indices_0 = const()[name = string("op_1611_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_1611_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_1611_validate_indices_0, x = p2c_att_25_cast_fp16)[name = string("op_1611_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_27_perm_0 = const()[name = string("p2c_att_27_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1614_y_0_to_fp16 = const()[name = string("_inversed_1614_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_27_cast_fp16 = transpose(perm = p2c_att_27_perm_0, x = var_1611_cast_fp16_cast_int16)[name = string("transpose_39")]; |
| tensor<fp16, [12, 512, 512]> _inversed_1614_cast_fp16 = mul(x = p2c_att_27_cast_fp16, y = _inversed_1614_y_0_to_fp16)[name = string("_inversed_1614_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_13_cast_fp16 = add(x = _inversed_1588_cast_fp16, y = _inversed_1614_cast_fp16)[name = string("rel_att_13_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_39_cast_fp16 = add(x = attention_scores_37_cast_fp16, y = rel_att_13_cast_fp16)[name = string("attention_scores_39_cast_fp16")]; |
| tensor<int32, [4]> var_1619 = const()[name = string("op_1619"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_41_cast_fp16 = reshape(shape = var_1619, x = attention_scores_39_cast_fp16)[name = string("attention_scores_41_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_153_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_41_cast_fp16, cond = var_308)[name = string("input_153_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_155_cast_fp16 = softmax(axis = var_32, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; |
| tensor<int32, [3]> var_1628 = const()[name = string("op_1628"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_1629_cast_fp16 = reshape(shape = var_1628, x = input_155_cast_fp16)[name = string("op_1629_cast_fp16")]; |
| bool context_layer_25_transpose_x_0 = const()[name = string("context_layer_25_transpose_x_0"), val = bool(false)]; |
| bool context_layer_25_transpose_y_0 = const()[name = string("context_layer_25_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_25_cast_fp16 = matmul(transpose_x = context_layer_25_transpose_x_0, transpose_y = context_layer_25_transpose_y_0, x = var_1629_cast_fp16, y = value_layer_13_cast_fp16)[name = string("context_layer_25_cast_fp16")]; |
| tensor<int32, [4]> var_1633 = const()[name = string("op_1633"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1634_cast_fp16 = reshape(shape = var_1633, x = context_layer_25_cast_fp16)[name = string("op_1634_cast_fp16")]; |
| tensor<int32, [4]> var_1635 = const()[name = string("op_1635"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1640 = const()[name = string("op_1640"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_1636_cast_fp16 = transpose(perm = var_1635, x = var_1634_cast_fp16)[name = string("transpose_38")]; |
| tensor<fp16, [1, 512, 768]> input_157_cast_fp16 = reshape(shape = var_1640, x = var_1636_cast_fp16)[name = string("input_157_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_6_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308817344)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309997056)))]; |
| tensor<fp16, [1, 512, 768]> linear_53_cast_fp16 = linear(bias = deberta_encoder_layer_6_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_6_attention_output_dense_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_53_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_161_cast_fp16 = add(x = linear_53_cast_fp16, y = input_149_cast_fp16)[name = string("input_161_cast_fp16")]; |
| tensor<int32, [1]> input_163_axes_0 = const()[name = string("input_163_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309998656)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310000256)))]; |
| tensor<fp16, [1, 512, 768]> input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = deberta_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16, x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_6_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_6_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310001856)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_6_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_6_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314720512)))]; |
| tensor<fp16, [1, 512, 3072]> linear_54_cast_fp16 = linear(bias = deberta_encoder_layer_6_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_6_intermediate_dense_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_54_cast_fp16")]; |
| string input_167_mode_0 = const()[name = string("input_167_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = string("input_167_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_6_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_6_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314726720)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_6_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319445376)))]; |
| tensor<fp16, [1, 512, 768]> linear_55_cast_fp16 = linear(bias = deberta_encoder_layer_6_output_dense_bias_to_fp16, weight = deberta_encoder_layer_6_output_dense_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_55_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_171_cast_fp16 = add(x = linear_55_cast_fp16, y = input_163_cast_fp16)[name = string("input_171_cast_fp16")]; |
| tensor<int32, [1]> input_173_axes_0 = const()[name = string("input_173_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_6_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319446976)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_6_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_6_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319448576)))]; |
| tensor<fp16, [1, 512, 768]> input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = deberta_encoder_layer_6_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_6_output_LayerNorm_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_7_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319450176)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320629888)))]; |
| tensor<fp16, [1, 512, 768]> linear_56_cast_fp16 = linear(bias = deberta_encoder_layer_7_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_7_attention_self_query_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_56_cast_fp16")]; |
| tensor<int32, [4]> var_1691 = const()[name = string("op_1691"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_143_cast_fp16 = reshape(shape = var_1691, x = linear_56_cast_fp16)[name = string("x_143_cast_fp16")]; |
| tensor<int32, [4]> var_1693 = const()[name = string("op_1693"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1698 = const()[name = string("op_1698"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1694_cast_fp16 = transpose(perm = var_1693, x = x_143_cast_fp16)[name = string("transpose_37")]; |
| tensor<fp16, [12, 512, 64]> query_layer_45_cast_fp16 = reshape(shape = var_1698, x = var_1694_cast_fp16)[name = string("query_layer_45_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_7_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320631488)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321811200)))]; |
| tensor<fp16, [1, 512, 768]> linear_57_cast_fp16 = linear(bias = deberta_encoder_layer_7_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_7_attention_self_key_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_57_cast_fp16")]; |
| tensor<int32, [4]> var_1705 = const()[name = string("op_1705"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_147_cast_fp16 = reshape(shape = var_1705, x = linear_57_cast_fp16)[name = string("x_147_cast_fp16")]; |
| tensor<int32, [4]> var_1707 = const()[name = string("op_1707"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1712 = const()[name = string("op_1712"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1708_cast_fp16 = transpose(perm = var_1707, x = x_147_cast_fp16)[name = string("transpose_36")]; |
| tensor<fp16, [12, 512, 64]> key_layer_15_cast_fp16 = reshape(shape = var_1712, x = var_1708_cast_fp16)[name = string("key_layer_15_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_7_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321812800)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322992512)))]; |
| tensor<fp16, [1, 512, 768]> linear_58_cast_fp16 = linear(bias = deberta_encoder_layer_7_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_7_attention_self_value_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_58_cast_fp16")]; |
| tensor<int32, [4]> var_1719 = const()[name = string("op_1719"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_151_cast_fp16 = reshape(shape = var_1719, x = linear_58_cast_fp16)[name = string("x_151_cast_fp16")]; |
| tensor<int32, [4]> var_1721 = const()[name = string("op_1721"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1726 = const()[name = string("op_1726"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1722_cast_fp16 = transpose(perm = var_1721, x = x_151_cast_fp16)[name = string("transpose_35")]; |
| tensor<fp16, [12, 512, 64]> value_layer_15_cast_fp16 = reshape(shape = var_1726, x = var_1722_cast_fp16)[name = string("value_layer_15_cast_fp16")]; |
| tensor<int32, [3]> var_1736_perm_0 = const()[name = string("op_1736_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1738_y_0_to_fp16 = const()[name = string("_inversed_1738_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_1736_cast_fp16 = transpose(perm = var_1736_perm_0, x = key_layer_15_cast_fp16)[name = string("transpose_34")]; |
| tensor<fp16, [12, 64, 512]> _inversed_1738_cast_fp16 = mul(x = var_1736_cast_fp16, y = _inversed_1738_y_0_to_fp16)[name = string("_inversed_1738_cast_fp16")]; |
| bool attention_scores_43_transpose_x_0 = const()[name = string("attention_scores_43_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_43_transpose_y_0 = const()[name = string("attention_scores_43_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_43_cast_fp16 = matmul(transpose_x = attention_scores_43_transpose_x_0, transpose_y = attention_scores_43_transpose_y_0, x = query_layer_45_cast_fp16, y = _inversed_1738_cast_fp16)[name = string("attention_scores_43_cast_fp16")]; |
| bool c2p_att_29_transpose_x_0 = const()[name = string("c2p_att_29_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_29_transpose_y_0 = const()[name = string("c2p_att_29_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_1794_to_fp16 = const()[name = string("op_1794_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322994112)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_29_cast_fp16 = matmul(transpose_x = c2p_att_29_transpose_x_0, transpose_y = c2p_att_29_transpose_y_0, x = query_layer_45_cast_fp16, y = var_1794_to_fp16)[name = string("c2p_att_29_cast_fp16")]; |
| bool c2p_att_31_validate_indices_0 = const()[name = string("c2p_att_31_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_31_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_31_validate_indices_0, x = c2p_att_29_cast_fp16)[name = string("c2p_att_31_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_1807_y_0_to_fp16 = const()[name = string("_inversed_1807_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_1807_cast_fp16 = mul(x = c2p_att_31_cast_fp16_cast_int16, y = _inversed_1807_y_0_to_fp16)[name = string("_inversed_1807_cast_fp16")]; |
| bool p2c_att_29_transpose_x_0 = const()[name = string("p2c_att_29_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_29_transpose_y_0 = const()[name = string("p2c_att_29_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_1822_to_fp16 = const()[name = string("op_1822_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323780608)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_29_cast_fp16 = matmul(transpose_x = p2c_att_29_transpose_x_0, transpose_y = p2c_att_29_transpose_y_0, x = key_layer_15_cast_fp16, y = var_1822_to_fp16)[name = string("p2c_att_29_cast_fp16")]; |
| bool var_1830_validate_indices_0 = const()[name = string("op_1830_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_1830_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_1830_validate_indices_0, x = p2c_att_29_cast_fp16)[name = string("op_1830_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_31_perm_0 = const()[name = string("p2c_att_31_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1833_y_0_to_fp16 = const()[name = string("_inversed_1833_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_31_cast_fp16 = transpose(perm = p2c_att_31_perm_0, x = var_1830_cast_fp16_cast_int16)[name = string("transpose_33")]; |
| tensor<fp16, [12, 512, 512]> _inversed_1833_cast_fp16 = mul(x = p2c_att_31_cast_fp16, y = _inversed_1833_y_0_to_fp16)[name = string("_inversed_1833_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_15_cast_fp16 = add(x = _inversed_1807_cast_fp16, y = _inversed_1833_cast_fp16)[name = string("rel_att_15_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_45_cast_fp16 = add(x = attention_scores_43_cast_fp16, y = rel_att_15_cast_fp16)[name = string("attention_scores_45_cast_fp16")]; |
| tensor<int32, [4]> var_1838 = const()[name = string("op_1838"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_47_cast_fp16 = reshape(shape = var_1838, x = attention_scores_45_cast_fp16)[name = string("attention_scores_47_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_177_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_47_cast_fp16, cond = var_308)[name = string("input_177_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_179_cast_fp16 = softmax(axis = var_32, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")]; |
| tensor<int32, [3]> var_1847 = const()[name = string("op_1847"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_1848_cast_fp16 = reshape(shape = var_1847, x = input_179_cast_fp16)[name = string("op_1848_cast_fp16")]; |
| bool context_layer_29_transpose_x_0 = const()[name = string("context_layer_29_transpose_x_0"), val = bool(false)]; |
| bool context_layer_29_transpose_y_0 = const()[name = string("context_layer_29_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_29_cast_fp16 = matmul(transpose_x = context_layer_29_transpose_x_0, transpose_y = context_layer_29_transpose_y_0, x = var_1848_cast_fp16, y = value_layer_15_cast_fp16)[name = string("context_layer_29_cast_fp16")]; |
| tensor<int32, [4]> var_1852 = const()[name = string("op_1852"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1853_cast_fp16 = reshape(shape = var_1852, x = context_layer_29_cast_fp16)[name = string("op_1853_cast_fp16")]; |
| tensor<int32, [4]> var_1854 = const()[name = string("op_1854"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1859 = const()[name = string("op_1859"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_1855_cast_fp16 = transpose(perm = var_1854, x = var_1853_cast_fp16)[name = string("transpose_32")]; |
| tensor<fp16, [1, 512, 768]> input_181_cast_fp16 = reshape(shape = var_1859, x = var_1855_cast_fp16)[name = string("input_181_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_7_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324567104)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325746816)))]; |
| tensor<fp16, [1, 512, 768]> linear_61_cast_fp16 = linear(bias = deberta_encoder_layer_7_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_7_attention_output_dense_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_61_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_185_cast_fp16 = add(x = linear_61_cast_fp16, y = input_173_cast_fp16)[name = string("input_185_cast_fp16")]; |
| tensor<int32, [1]> input_187_axes_0 = const()[name = string("input_187_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325748416)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325750016)))]; |
| tensor<fp16, [1, 512, 768]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = deberta_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16, x = input_185_cast_fp16)[name = string("input_187_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_7_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_7_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325751616)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_7_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_7_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330470272)))]; |
| tensor<fp16, [1, 512, 3072]> linear_62_cast_fp16 = linear(bias = deberta_encoder_layer_7_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_7_intermediate_dense_weight_to_fp16, x = input_187_cast_fp16)[name = string("linear_62_cast_fp16")]; |
| string input_191_mode_0 = const()[name = string("input_191_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = string("input_191_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_7_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_7_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330476480)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_7_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335195136)))]; |
| tensor<fp16, [1, 512, 768]> linear_63_cast_fp16 = linear(bias = deberta_encoder_layer_7_output_dense_bias_to_fp16, weight = deberta_encoder_layer_7_output_dense_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_63_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_195_cast_fp16 = add(x = linear_63_cast_fp16, y = input_187_cast_fp16)[name = string("input_195_cast_fp16")]; |
| tensor<int32, [1]> input_197_axes_0 = const()[name = string("input_197_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_7_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335196736)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_7_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_7_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335198336)))]; |
| tensor<fp16, [1, 512, 768]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = deberta_encoder_layer_7_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_7_output_LayerNorm_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_8_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335199936)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336379648)))]; |
| tensor<fp16, [1, 512, 768]> linear_64_cast_fp16 = linear(bias = deberta_encoder_layer_8_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_8_attention_self_query_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_64_cast_fp16")]; |
| tensor<int32, [4]> var_1910 = const()[name = string("op_1910"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_163_cast_fp16 = reshape(shape = var_1910, x = linear_64_cast_fp16)[name = string("x_163_cast_fp16")]; |
| tensor<int32, [4]> var_1912 = const()[name = string("op_1912"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1917 = const()[name = string("op_1917"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1913_cast_fp16 = transpose(perm = var_1912, x = x_163_cast_fp16)[name = string("transpose_31")]; |
| tensor<fp16, [12, 512, 64]> query_layer_51_cast_fp16 = reshape(shape = var_1917, x = var_1913_cast_fp16)[name = string("query_layer_51_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_8_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336381248)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337560960)))]; |
| tensor<fp16, [1, 512, 768]> linear_65_cast_fp16 = linear(bias = deberta_encoder_layer_8_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_8_attention_self_key_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_65_cast_fp16")]; |
| tensor<int32, [4]> var_1924 = const()[name = string("op_1924"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_167_cast_fp16 = reshape(shape = var_1924, x = linear_65_cast_fp16)[name = string("x_167_cast_fp16")]; |
| tensor<int32, [4]> var_1926 = const()[name = string("op_1926"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1931 = const()[name = string("op_1931"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1927_cast_fp16 = transpose(perm = var_1926, x = x_167_cast_fp16)[name = string("transpose_30")]; |
| tensor<fp16, [12, 512, 64]> key_layer_17_cast_fp16 = reshape(shape = var_1931, x = var_1927_cast_fp16)[name = string("key_layer_17_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_8_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337562560)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338742272)))]; |
| tensor<fp16, [1, 512, 768]> linear_66_cast_fp16 = linear(bias = deberta_encoder_layer_8_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_8_attention_self_value_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_66_cast_fp16")]; |
| tensor<int32, [4]> var_1938 = const()[name = string("op_1938"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_171_cast_fp16 = reshape(shape = var_1938, x = linear_66_cast_fp16)[name = string("x_171_cast_fp16")]; |
| tensor<int32, [4]> var_1940 = const()[name = string("op_1940"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1945 = const()[name = string("op_1945"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_1941_cast_fp16 = transpose(perm = var_1940, x = x_171_cast_fp16)[name = string("transpose_29")]; |
| tensor<fp16, [12, 512, 64]> value_layer_17_cast_fp16 = reshape(shape = var_1945, x = var_1941_cast_fp16)[name = string("value_layer_17_cast_fp16")]; |
| tensor<int32, [3]> var_1955_perm_0 = const()[name = string("op_1955_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_1957_y_0_to_fp16 = const()[name = string("_inversed_1957_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_1955_cast_fp16 = transpose(perm = var_1955_perm_0, x = key_layer_17_cast_fp16)[name = string("transpose_28")]; |
| tensor<fp16, [12, 64, 512]> _inversed_1957_cast_fp16 = mul(x = var_1955_cast_fp16, y = _inversed_1957_y_0_to_fp16)[name = string("_inversed_1957_cast_fp16")]; |
| bool attention_scores_49_transpose_x_0 = const()[name = string("attention_scores_49_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_49_transpose_y_0 = const()[name = string("attention_scores_49_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_49_cast_fp16 = matmul(transpose_x = attention_scores_49_transpose_x_0, transpose_y = attention_scores_49_transpose_y_0, x = query_layer_51_cast_fp16, y = _inversed_1957_cast_fp16)[name = string("attention_scores_49_cast_fp16")]; |
| bool c2p_att_33_transpose_x_0 = const()[name = string("c2p_att_33_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_33_transpose_y_0 = const()[name = string("c2p_att_33_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_2013_to_fp16 = const()[name = string("op_2013_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338743872)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_33_cast_fp16 = matmul(transpose_x = c2p_att_33_transpose_x_0, transpose_y = c2p_att_33_transpose_y_0, x = query_layer_51_cast_fp16, y = var_2013_to_fp16)[name = string("c2p_att_33_cast_fp16")]; |
| bool c2p_att_35_validate_indices_0 = const()[name = string("c2p_att_35_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_35_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_35_validate_indices_0, x = c2p_att_33_cast_fp16)[name = string("c2p_att_35_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_2026_y_0_to_fp16 = const()[name = string("_inversed_2026_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_2026_cast_fp16 = mul(x = c2p_att_35_cast_fp16_cast_int16, y = _inversed_2026_y_0_to_fp16)[name = string("_inversed_2026_cast_fp16")]; |
| bool p2c_att_33_transpose_x_0 = const()[name = string("p2c_att_33_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_33_transpose_y_0 = const()[name = string("p2c_att_33_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_2041_to_fp16 = const()[name = string("op_2041_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339530368)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_33_cast_fp16 = matmul(transpose_x = p2c_att_33_transpose_x_0, transpose_y = p2c_att_33_transpose_y_0, x = key_layer_17_cast_fp16, y = var_2041_to_fp16)[name = string("p2c_att_33_cast_fp16")]; |
| bool var_2049_validate_indices_0 = const()[name = string("op_2049_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_2049_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_2049_validate_indices_0, x = p2c_att_33_cast_fp16)[name = string("op_2049_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_35_perm_0 = const()[name = string("p2c_att_35_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_2052_y_0_to_fp16 = const()[name = string("_inversed_2052_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_35_cast_fp16 = transpose(perm = p2c_att_35_perm_0, x = var_2049_cast_fp16_cast_int16)[name = string("transpose_27")]; |
| tensor<fp16, [12, 512, 512]> _inversed_2052_cast_fp16 = mul(x = p2c_att_35_cast_fp16, y = _inversed_2052_y_0_to_fp16)[name = string("_inversed_2052_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_17_cast_fp16 = add(x = _inversed_2026_cast_fp16, y = _inversed_2052_cast_fp16)[name = string("rel_att_17_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_51_cast_fp16 = add(x = attention_scores_49_cast_fp16, y = rel_att_17_cast_fp16)[name = string("attention_scores_51_cast_fp16")]; |
| tensor<int32, [4]> var_2057 = const()[name = string("op_2057"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_53_cast_fp16 = reshape(shape = var_2057, x = attention_scores_51_cast_fp16)[name = string("attention_scores_53_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_201_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_53_cast_fp16, cond = var_308)[name = string("input_201_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_203_cast_fp16 = softmax(axis = var_32, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")]; |
| tensor<int32, [3]> var_2066 = const()[name = string("op_2066"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_2067_cast_fp16 = reshape(shape = var_2066, x = input_203_cast_fp16)[name = string("op_2067_cast_fp16")]; |
| bool context_layer_33_transpose_x_0 = const()[name = string("context_layer_33_transpose_x_0"), val = bool(false)]; |
| bool context_layer_33_transpose_y_0 = const()[name = string("context_layer_33_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_33_cast_fp16 = matmul(transpose_x = context_layer_33_transpose_x_0, transpose_y = context_layer_33_transpose_y_0, x = var_2067_cast_fp16, y = value_layer_17_cast_fp16)[name = string("context_layer_33_cast_fp16")]; |
| tensor<int32, [4]> var_2071 = const()[name = string("op_2071"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2072_cast_fp16 = reshape(shape = var_2071, x = context_layer_33_cast_fp16)[name = string("op_2072_cast_fp16")]; |
| tensor<int32, [4]> var_2073 = const()[name = string("op_2073"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2078 = const()[name = string("op_2078"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_2074_cast_fp16 = transpose(perm = var_2073, x = var_2072_cast_fp16)[name = string("transpose_26")]; |
| tensor<fp16, [1, 512, 768]> input_205_cast_fp16 = reshape(shape = var_2078, x = var_2074_cast_fp16)[name = string("input_205_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_8_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340316864)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341496576)))]; |
| tensor<fp16, [1, 512, 768]> linear_69_cast_fp16 = linear(bias = deberta_encoder_layer_8_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_8_attention_output_dense_weight_to_fp16, x = input_205_cast_fp16)[name = string("linear_69_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_209_cast_fp16 = add(x = linear_69_cast_fp16, y = input_197_cast_fp16)[name = string("input_209_cast_fp16")]; |
| tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341498176)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341499776)))]; |
| tensor<fp16, [1, 512, 768]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, beta = deberta_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_8_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_8_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341501376)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_8_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_8_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346220032)))]; |
| tensor<fp16, [1, 512, 3072]> linear_70_cast_fp16 = linear(bias = deberta_encoder_layer_8_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_8_intermediate_dense_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_70_cast_fp16")]; |
| string input_215_mode_0 = const()[name = string("input_215_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = string("input_215_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_8_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_8_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346226240)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_8_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350944896)))]; |
| tensor<fp16, [1, 512, 768]> linear_71_cast_fp16 = linear(bias = deberta_encoder_layer_8_output_dense_bias_to_fp16, weight = deberta_encoder_layer_8_output_dense_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_71_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_219_cast_fp16 = add(x = linear_71_cast_fp16, y = input_211_cast_fp16)[name = string("input_219_cast_fp16")]; |
| tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_8_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350946496)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_8_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_8_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350948096)))]; |
| tensor<fp16, [1, 512, 768]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = deberta_encoder_layer_8_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_8_output_LayerNorm_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_9_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350949696)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352129408)))]; |
| tensor<fp16, [1, 512, 768]> linear_72_cast_fp16 = linear(bias = deberta_encoder_layer_9_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_9_attention_self_query_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_72_cast_fp16")]; |
| tensor<int32, [4]> var_2129 = const()[name = string("op_2129"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_183_cast_fp16 = reshape(shape = var_2129, x = linear_72_cast_fp16)[name = string("x_183_cast_fp16")]; |
| tensor<int32, [4]> var_2131 = const()[name = string("op_2131"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2136 = const()[name = string("op_2136"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2132_cast_fp16 = transpose(perm = var_2131, x = x_183_cast_fp16)[name = string("transpose_25")]; |
| tensor<fp16, [12, 512, 64]> query_layer_57_cast_fp16 = reshape(shape = var_2136, x = var_2132_cast_fp16)[name = string("query_layer_57_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_9_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352131008)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353310720)))]; |
| tensor<fp16, [1, 512, 768]> linear_73_cast_fp16 = linear(bias = deberta_encoder_layer_9_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_9_attention_self_key_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_73_cast_fp16")]; |
| tensor<int32, [4]> var_2143 = const()[name = string("op_2143"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_187_cast_fp16 = reshape(shape = var_2143, x = linear_73_cast_fp16)[name = string("x_187_cast_fp16")]; |
| tensor<int32, [4]> var_2145 = const()[name = string("op_2145"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2150 = const()[name = string("op_2150"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2146_cast_fp16 = transpose(perm = var_2145, x = x_187_cast_fp16)[name = string("transpose_24")]; |
| tensor<fp16, [12, 512, 64]> key_layer_19_cast_fp16 = reshape(shape = var_2150, x = var_2146_cast_fp16)[name = string("key_layer_19_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_9_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353312320)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354492032)))]; |
| tensor<fp16, [1, 512, 768]> linear_74_cast_fp16 = linear(bias = deberta_encoder_layer_9_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_9_attention_self_value_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_74_cast_fp16")]; |
| tensor<int32, [4]> var_2157 = const()[name = string("op_2157"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_191_cast_fp16 = reshape(shape = var_2157, x = linear_74_cast_fp16)[name = string("x_191_cast_fp16")]; |
| tensor<int32, [4]> var_2159 = const()[name = string("op_2159"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2164 = const()[name = string("op_2164"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2160_cast_fp16 = transpose(perm = var_2159, x = x_191_cast_fp16)[name = string("transpose_23")]; |
| tensor<fp16, [12, 512, 64]> value_layer_19_cast_fp16 = reshape(shape = var_2164, x = var_2160_cast_fp16)[name = string("value_layer_19_cast_fp16")]; |
| tensor<int32, [3]> var_2174_perm_0 = const()[name = string("op_2174_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_2176_y_0_to_fp16 = const()[name = string("_inversed_2176_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_2174_cast_fp16 = transpose(perm = var_2174_perm_0, x = key_layer_19_cast_fp16)[name = string("transpose_22")]; |
| tensor<fp16, [12, 64, 512]> _inversed_2176_cast_fp16 = mul(x = var_2174_cast_fp16, y = _inversed_2176_y_0_to_fp16)[name = string("_inversed_2176_cast_fp16")]; |
| bool attention_scores_55_transpose_x_0 = const()[name = string("attention_scores_55_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_55_transpose_y_0 = const()[name = string("attention_scores_55_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_55_cast_fp16 = matmul(transpose_x = attention_scores_55_transpose_x_0, transpose_y = attention_scores_55_transpose_y_0, x = query_layer_57_cast_fp16, y = _inversed_2176_cast_fp16)[name = string("attention_scores_55_cast_fp16")]; |
| bool c2p_att_37_transpose_x_0 = const()[name = string("c2p_att_37_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_37_transpose_y_0 = const()[name = string("c2p_att_37_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354493632)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_37_cast_fp16 = matmul(transpose_x = c2p_att_37_transpose_x_0, transpose_y = c2p_att_37_transpose_y_0, x = query_layer_57_cast_fp16, y = var_2232_to_fp16)[name = string("c2p_att_37_cast_fp16")]; |
| bool c2p_att_39_validate_indices_0 = const()[name = string("c2p_att_39_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_39_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_39_validate_indices_0, x = c2p_att_37_cast_fp16)[name = string("c2p_att_39_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_2245_y_0_to_fp16 = const()[name = string("_inversed_2245_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_2245_cast_fp16 = mul(x = c2p_att_39_cast_fp16_cast_int16, y = _inversed_2245_y_0_to_fp16)[name = string("_inversed_2245_cast_fp16")]; |
| bool p2c_att_37_transpose_x_0 = const()[name = string("p2c_att_37_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_37_transpose_y_0 = const()[name = string("p2c_att_37_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_2260_to_fp16 = const()[name = string("op_2260_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355280128)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_37_cast_fp16 = matmul(transpose_x = p2c_att_37_transpose_x_0, transpose_y = p2c_att_37_transpose_y_0, x = key_layer_19_cast_fp16, y = var_2260_to_fp16)[name = string("p2c_att_37_cast_fp16")]; |
| bool var_2268_validate_indices_0 = const()[name = string("op_2268_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_2268_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_2268_validate_indices_0, x = p2c_att_37_cast_fp16)[name = string("op_2268_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_39_perm_0 = const()[name = string("p2c_att_39_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_2271_y_0_to_fp16 = const()[name = string("_inversed_2271_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_39_cast_fp16 = transpose(perm = p2c_att_39_perm_0, x = var_2268_cast_fp16_cast_int16)[name = string("transpose_21")]; |
| tensor<fp16, [12, 512, 512]> _inversed_2271_cast_fp16 = mul(x = p2c_att_39_cast_fp16, y = _inversed_2271_y_0_to_fp16)[name = string("_inversed_2271_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_19_cast_fp16 = add(x = _inversed_2245_cast_fp16, y = _inversed_2271_cast_fp16)[name = string("rel_att_19_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_57_cast_fp16 = add(x = attention_scores_55_cast_fp16, y = rel_att_19_cast_fp16)[name = string("attention_scores_57_cast_fp16")]; |
| tensor<int32, [4]> var_2276 = const()[name = string("op_2276"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_59_cast_fp16 = reshape(shape = var_2276, x = attention_scores_57_cast_fp16)[name = string("attention_scores_59_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_225_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_59_cast_fp16, cond = var_308)[name = string("input_225_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_227_cast_fp16 = softmax(axis = var_32, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")]; |
| tensor<int32, [3]> var_2285 = const()[name = string("op_2285"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_2286_cast_fp16 = reshape(shape = var_2285, x = input_227_cast_fp16)[name = string("op_2286_cast_fp16")]; |
| bool context_layer_37_transpose_x_0 = const()[name = string("context_layer_37_transpose_x_0"), val = bool(false)]; |
| bool context_layer_37_transpose_y_0 = const()[name = string("context_layer_37_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_37_cast_fp16 = matmul(transpose_x = context_layer_37_transpose_x_0, transpose_y = context_layer_37_transpose_y_0, x = var_2286_cast_fp16, y = value_layer_19_cast_fp16)[name = string("context_layer_37_cast_fp16")]; |
| tensor<int32, [4]> var_2290 = const()[name = string("op_2290"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2291_cast_fp16 = reshape(shape = var_2290, x = context_layer_37_cast_fp16)[name = string("op_2291_cast_fp16")]; |
| tensor<int32, [4]> var_2292 = const()[name = string("op_2292"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2297 = const()[name = string("op_2297"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_2293_cast_fp16 = transpose(perm = var_2292, x = var_2291_cast_fp16)[name = string("transpose_20")]; |
| tensor<fp16, [1, 512, 768]> input_229_cast_fp16 = reshape(shape = var_2297, x = var_2293_cast_fp16)[name = string("input_229_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_9_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356066624)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357246336)))]; |
| tensor<fp16, [1, 512, 768]> linear_77_cast_fp16 = linear(bias = deberta_encoder_layer_9_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_9_attention_output_dense_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_77_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_233_cast_fp16 = add(x = linear_77_cast_fp16, y = input_221_cast_fp16)[name = string("input_233_cast_fp16")]; |
| tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357247936)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357249536)))]; |
| tensor<fp16, [1, 512, 768]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = deberta_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_9_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_9_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357251136)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_9_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_9_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361969792)))]; |
| tensor<fp16, [1, 512, 3072]> linear_78_cast_fp16 = linear(bias = deberta_encoder_layer_9_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_9_intermediate_dense_weight_to_fp16, x = input_235_cast_fp16)[name = string("linear_78_cast_fp16")]; |
| string input_239_mode_0 = const()[name = string("input_239_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = string("input_239_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_9_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_9_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361976000)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_9_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366694656)))]; |
| tensor<fp16, [1, 512, 768]> linear_79_cast_fp16 = linear(bias = deberta_encoder_layer_9_output_dense_bias_to_fp16, weight = deberta_encoder_layer_9_output_dense_weight_to_fp16, x = input_239_cast_fp16)[name = string("linear_79_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_243_cast_fp16 = add(x = linear_79_cast_fp16, y = input_235_cast_fp16)[name = string("input_243_cast_fp16")]; |
| tensor<int32, [1]> input_245_axes_0 = const()[name = string("input_245_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_9_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366696256)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_9_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_9_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366697856)))]; |
| tensor<fp16, [1, 512, 768]> input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = deberta_encoder_layer_9_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_9_output_LayerNorm_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_10_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366699456)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367879168)))]; |
| tensor<fp16, [1, 512, 768]> linear_80_cast_fp16 = linear(bias = deberta_encoder_layer_10_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_10_attention_self_query_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_80_cast_fp16")]; |
| tensor<int32, [4]> var_2348 = const()[name = string("op_2348"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_203_cast_fp16 = reshape(shape = var_2348, x = linear_80_cast_fp16)[name = string("x_203_cast_fp16")]; |
| tensor<int32, [4]> var_2350 = const()[name = string("op_2350"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2355 = const()[name = string("op_2355"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2351_cast_fp16 = transpose(perm = var_2350, x = x_203_cast_fp16)[name = string("transpose_19")]; |
| tensor<fp16, [12, 512, 64]> query_layer_63_cast_fp16 = reshape(shape = var_2355, x = var_2351_cast_fp16)[name = string("query_layer_63_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_10_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367880768)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369060480)))]; |
| tensor<fp16, [1, 512, 768]> linear_81_cast_fp16 = linear(bias = deberta_encoder_layer_10_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_10_attention_self_key_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_81_cast_fp16")]; |
| tensor<int32, [4]> var_2362 = const()[name = string("op_2362"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_207_cast_fp16 = reshape(shape = var_2362, x = linear_81_cast_fp16)[name = string("x_207_cast_fp16")]; |
| tensor<int32, [4]> var_2364 = const()[name = string("op_2364"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2369 = const()[name = string("op_2369"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2365_cast_fp16 = transpose(perm = var_2364, x = x_207_cast_fp16)[name = string("transpose_18")]; |
| tensor<fp16, [12, 512, 64]> key_layer_21_cast_fp16 = reshape(shape = var_2369, x = var_2365_cast_fp16)[name = string("key_layer_21_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_10_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369062080)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370241792)))]; |
| tensor<fp16, [1, 512, 768]> linear_82_cast_fp16 = linear(bias = deberta_encoder_layer_10_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_10_attention_self_value_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_82_cast_fp16")]; |
| tensor<int32, [4]> var_2376 = const()[name = string("op_2376"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_211_cast_fp16 = reshape(shape = var_2376, x = linear_82_cast_fp16)[name = string("x_211_cast_fp16")]; |
| tensor<int32, [4]> var_2378 = const()[name = string("op_2378"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2383 = const()[name = string("op_2383"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2379_cast_fp16 = transpose(perm = var_2378, x = x_211_cast_fp16)[name = string("transpose_17")]; |
| tensor<fp16, [12, 512, 64]> value_layer_21_cast_fp16 = reshape(shape = var_2383, x = var_2379_cast_fp16)[name = string("value_layer_21_cast_fp16")]; |
| tensor<int32, [3]> var_2393_perm_0 = const()[name = string("op_2393_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_2395_y_0_to_fp16 = const()[name = string("_inversed_2395_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_2393_cast_fp16 = transpose(perm = var_2393_perm_0, x = key_layer_21_cast_fp16)[name = string("transpose_16")]; |
| tensor<fp16, [12, 64, 512]> _inversed_2395_cast_fp16 = mul(x = var_2393_cast_fp16, y = _inversed_2395_y_0_to_fp16)[name = string("_inversed_2395_cast_fp16")]; |
| bool attention_scores_61_transpose_x_0 = const()[name = string("attention_scores_61_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_61_transpose_y_0 = const()[name = string("attention_scores_61_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_61_cast_fp16 = matmul(transpose_x = attention_scores_61_transpose_x_0, transpose_y = attention_scores_61_transpose_y_0, x = query_layer_63_cast_fp16, y = _inversed_2395_cast_fp16)[name = string("attention_scores_61_cast_fp16")]; |
| bool c2p_att_41_transpose_x_0 = const()[name = string("c2p_att_41_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_41_transpose_y_0 = const()[name = string("c2p_att_41_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_2451_to_fp16 = const()[name = string("op_2451_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370243392)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_41_cast_fp16 = matmul(transpose_x = c2p_att_41_transpose_x_0, transpose_y = c2p_att_41_transpose_y_0, x = query_layer_63_cast_fp16, y = var_2451_to_fp16)[name = string("c2p_att_41_cast_fp16")]; |
| bool c2p_att_43_validate_indices_0 = const()[name = string("c2p_att_43_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_43_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_43_validate_indices_0, x = c2p_att_41_cast_fp16)[name = string("c2p_att_43_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_2464_y_0_to_fp16 = const()[name = string("_inversed_2464_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_2464_cast_fp16 = mul(x = c2p_att_43_cast_fp16_cast_int16, y = _inversed_2464_y_0_to_fp16)[name = string("_inversed_2464_cast_fp16")]; |
| bool p2c_att_41_transpose_x_0 = const()[name = string("p2c_att_41_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_41_transpose_y_0 = const()[name = string("p2c_att_41_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_2479_to_fp16 = const()[name = string("op_2479_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371029888)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_41_cast_fp16 = matmul(transpose_x = p2c_att_41_transpose_x_0, transpose_y = p2c_att_41_transpose_y_0, x = key_layer_21_cast_fp16, y = var_2479_to_fp16)[name = string("p2c_att_41_cast_fp16")]; |
| bool var_2487_validate_indices_0 = const()[name = string("op_2487_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_2487_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_2487_validate_indices_0, x = p2c_att_41_cast_fp16)[name = string("op_2487_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_43_perm_0 = const()[name = string("p2c_att_43_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_2490_y_0_to_fp16 = const()[name = string("_inversed_2490_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_43_cast_fp16 = transpose(perm = p2c_att_43_perm_0, x = var_2487_cast_fp16_cast_int16)[name = string("transpose_15")]; |
| tensor<fp16, [12, 512, 512]> _inversed_2490_cast_fp16 = mul(x = p2c_att_43_cast_fp16, y = _inversed_2490_y_0_to_fp16)[name = string("_inversed_2490_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_21_cast_fp16 = add(x = _inversed_2464_cast_fp16, y = _inversed_2490_cast_fp16)[name = string("rel_att_21_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_63_cast_fp16 = add(x = attention_scores_61_cast_fp16, y = rel_att_21_cast_fp16)[name = string("attention_scores_63_cast_fp16")]; |
| tensor<int32, [4]> var_2495 = const()[name = string("op_2495"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_65_cast_fp16 = reshape(shape = var_2495, x = attention_scores_63_cast_fp16)[name = string("attention_scores_65_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_249_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_65_cast_fp16, cond = var_308)[name = string("input_249_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_251_cast_fp16 = softmax(axis = var_32, x = input_249_cast_fp16)[name = string("input_251_cast_fp16")]; |
| tensor<int32, [3]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_2505_cast_fp16 = reshape(shape = var_2504, x = input_251_cast_fp16)[name = string("op_2505_cast_fp16")]; |
| bool context_layer_41_transpose_x_0 = const()[name = string("context_layer_41_transpose_x_0"), val = bool(false)]; |
| bool context_layer_41_transpose_y_0 = const()[name = string("context_layer_41_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_41_cast_fp16 = matmul(transpose_x = context_layer_41_transpose_x_0, transpose_y = context_layer_41_transpose_y_0, x = var_2505_cast_fp16, y = value_layer_21_cast_fp16)[name = string("context_layer_41_cast_fp16")]; |
| tensor<int32, [4]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2510_cast_fp16 = reshape(shape = var_2509, x = context_layer_41_cast_fp16)[name = string("op_2510_cast_fp16")]; |
| tensor<int32, [4]> var_2511 = const()[name = string("op_2511"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2516 = const()[name = string("op_2516"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_2512_cast_fp16 = transpose(perm = var_2511, x = var_2510_cast_fp16)[name = string("transpose_14")]; |
| tensor<fp16, [1, 512, 768]> input_253_cast_fp16 = reshape(shape = var_2516, x = var_2512_cast_fp16)[name = string("input_253_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_10_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371816384)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372996096)))]; |
| tensor<fp16, [1, 512, 768]> linear_85_cast_fp16 = linear(bias = deberta_encoder_layer_10_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_10_attention_output_dense_weight_to_fp16, x = input_253_cast_fp16)[name = string("linear_85_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_257_cast_fp16 = add(x = linear_85_cast_fp16, y = input_245_cast_fp16)[name = string("input_257_cast_fp16")]; |
| tensor<int32, [1]> input_259_axes_0 = const()[name = string("input_259_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372997696)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372999296)))]; |
| tensor<fp16, [1, 512, 768]> input_259_cast_fp16 = layer_norm(axes = input_259_axes_0, beta = deberta_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_10_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_10_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373000896)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_10_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_10_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377719552)))]; |
| tensor<fp16, [1, 512, 3072]> linear_86_cast_fp16 = linear(bias = deberta_encoder_layer_10_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_10_intermediate_dense_weight_to_fp16, x = input_259_cast_fp16)[name = string("linear_86_cast_fp16")]; |
| string input_263_mode_0 = const()[name = string("input_263_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = string("input_263_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_10_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_10_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377725760)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_10_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382444416)))]; |
| tensor<fp16, [1, 512, 768]> linear_87_cast_fp16 = linear(bias = deberta_encoder_layer_10_output_dense_bias_to_fp16, weight = deberta_encoder_layer_10_output_dense_weight_to_fp16, x = input_263_cast_fp16)[name = string("linear_87_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_267_cast_fp16 = add(x = linear_87_cast_fp16, y = input_259_cast_fp16)[name = string("input_267_cast_fp16")]; |
| tensor<int32, [1]> input_269_axes_0 = const()[name = string("input_269_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_10_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382446016)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_10_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_10_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382447616)))]; |
| tensor<fp16, [1, 512, 768]> input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = deberta_encoder_layer_10_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_10_output_LayerNorm_weight_to_fp16, x = input_267_cast_fp16)[name = string("input_269_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_11_attention_self_query_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382449216)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_attention_self_query_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383628928)))]; |
| tensor<fp16, [1, 512, 768]> linear_88_cast_fp16 = linear(bias = deberta_encoder_layer_11_attention_self_query_proj_bias_to_fp16, weight = deberta_encoder_layer_11_attention_self_query_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_88_cast_fp16")]; |
| tensor<int32, [4]> var_2567 = const()[name = string("op_2567"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_223_cast_fp16 = reshape(shape = var_2567, x = linear_88_cast_fp16)[name = string("x_223_cast_fp16")]; |
| tensor<int32, [4]> var_2569 = const()[name = string("op_2569"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2570_cast_fp16 = transpose(perm = var_2569, x = x_223_cast_fp16)[name = string("transpose_13")]; |
| tensor<fp16, [12, 512, 64]> query_layer_69_cast_fp16 = reshape(shape = var_2574, x = var_2570_cast_fp16)[name = string("query_layer_69_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_11_attention_self_key_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383630528)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_attention_self_key_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384810240)))]; |
| tensor<fp16, [1, 512, 768]> linear_89_cast_fp16 = linear(bias = deberta_encoder_layer_11_attention_self_key_proj_bias_to_fp16, weight = deberta_encoder_layer_11_attention_self_key_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_89_cast_fp16")]; |
| tensor<int32, [4]> var_2581 = const()[name = string("op_2581"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_227_cast_fp16 = reshape(shape = var_2581, x = linear_89_cast_fp16)[name = string("x_227_cast_fp16")]; |
| tensor<int32, [4]> var_2583 = const()[name = string("op_2583"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2588 = const()[name = string("op_2588"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2584_cast_fp16 = transpose(perm = var_2583, x = x_227_cast_fp16)[name = string("transpose_12")]; |
| tensor<fp16, [12, 512, 64]> key_layer_cast_fp16 = reshape(shape = var_2588, x = var_2584_cast_fp16)[name = string("key_layer_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_11_attention_self_value_proj_weight_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384811840)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_attention_self_value_proj_bias_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385991552)))]; |
| tensor<fp16, [1, 512, 768]> linear_90_cast_fp16 = linear(bias = deberta_encoder_layer_11_attention_self_value_proj_bias_to_fp16, weight = deberta_encoder_layer_11_attention_self_value_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_90_cast_fp16")]; |
| tensor<int32, [4]> var_2595 = const()[name = string("op_2595"), val = tensor<int32, [4]>([1, 512, 12, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> x_231_cast_fp16 = reshape(shape = var_2595, x = linear_90_cast_fp16)[name = string("x_231_cast_fp16")]; |
| tensor<int32, [4]> var_2597 = const()[name = string("op_2597"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2602 = const()[name = string("op_2602"), val = tensor<int32, [3]>([-1, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2598_cast_fp16 = transpose(perm = var_2597, x = x_231_cast_fp16)[name = string("transpose_11")]; |
| tensor<fp16, [12, 512, 64]> value_layer_cast_fp16 = reshape(shape = var_2602, x = var_2598_cast_fp16)[name = string("value_layer_cast_fp16")]; |
| tensor<int32, [3]> var_2612_perm_0 = const()[name = string("op_2612_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_2614_y_0_to_fp16 = const()[name = string("_inversed_2614_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 64, 512]> var_2612_cast_fp16 = transpose(perm = var_2612_perm_0, x = key_layer_cast_fp16)[name = string("transpose_10")]; |
| tensor<fp16, [12, 64, 512]> _inversed_2614_cast_fp16 = mul(x = var_2612_cast_fp16, y = _inversed_2614_y_0_to_fp16)[name = string("_inversed_2614_cast_fp16")]; |
| bool attention_scores_67_transpose_x_0 = const()[name = string("attention_scores_67_transpose_x_0"), val = bool(false)]; |
| bool attention_scores_67_transpose_y_0 = const()[name = string("attention_scores_67_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> attention_scores_67_cast_fp16 = matmul(transpose_x = attention_scores_67_transpose_x_0, transpose_y = attention_scores_67_transpose_y_0, x = query_layer_69_cast_fp16, y = _inversed_2614_cast_fp16)[name = string("attention_scores_67_cast_fp16")]; |
| bool c2p_att_45_transpose_x_0 = const()[name = string("c2p_att_45_transpose_x_0"), val = bool(false)]; |
| bool c2p_att_45_transpose_y_0 = const()[name = string("c2p_att_45_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_2670_to_fp16 = const()[name = string("op_2670_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385993152)))]; |
| tensor<fp16, [12, 512, 512]> c2p_att_45_cast_fp16 = matmul(transpose_x = c2p_att_45_transpose_x_0, transpose_y = c2p_att_45_transpose_y_0, x = query_layer_69_cast_fp16, y = var_2670_to_fp16)[name = string("c2p_att_45_cast_fp16")]; |
| bool c2p_att_validate_indices_0 = const()[name = string("c2p_att_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> c2p_att_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_271_to_uint16, validate_indices = c2p_att_validate_indices_0, x = c2p_att_45_cast_fp16)[name = string("c2p_att_cast_fp16_cast_int16")]; |
| tensor<fp16, [1]> _inversed_2683_y_0_to_fp16 = const()[name = string("_inversed_2683_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> _inversed_2683_cast_fp16 = mul(x = c2p_att_cast_fp16_cast_int16, y = _inversed_2683_y_0_to_fp16)[name = string("_inversed_2683_cast_fp16")]; |
| bool p2c_att_45_transpose_x_0 = const()[name = string("p2c_att_45_transpose_x_0"), val = bool(false)]; |
| bool p2c_att_45_transpose_y_0 = const()[name = string("p2c_att_45_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 64, 512]> var_2698_to_fp16 = const()[name = string("op_2698_to_fp16"), val = tensor<fp16, [12, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386779648)))]; |
| tensor<fp16, [12, 512, 512]> p2c_att_45_cast_fp16 = matmul(transpose_x = p2c_att_45_transpose_x_0, transpose_y = p2c_att_45_transpose_y_0, x = key_layer_cast_fp16, y = var_2698_to_fp16)[name = string("p2c_att_45_cast_fp16")]; |
| bool var_2706_validate_indices_0 = const()[name = string("op_2706_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 512]> var_2706_cast_fp16_cast_int16 = gather_along_axis(axis = var_32, indices = var_296_to_uint16, validate_indices = var_2706_validate_indices_0, x = p2c_att_45_cast_fp16)[name = string("op_2706_cast_fp16_cast_int16")]; |
| tensor<int32, [3]> p2c_att_perm_0 = const()[name = string("p2c_att_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<fp16, [1]> _inversed_2709_y_0_to_fp16 = const()[name = string("_inversed_2709_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])]; |
| tensor<fp16, [12, 512, 512]> p2c_att_cast_fp16 = transpose(perm = p2c_att_perm_0, x = var_2706_cast_fp16_cast_int16)[name = string("transpose_9")]; |
| tensor<fp16, [12, 512, 512]> _inversed_2709_cast_fp16 = mul(x = p2c_att_cast_fp16, y = _inversed_2709_y_0_to_fp16)[name = string("_inversed_2709_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> rel_att_cast_fp16 = add(x = _inversed_2683_cast_fp16, y = _inversed_2709_cast_fp16)[name = string("rel_att_cast_fp16")]; |
| tensor<fp16, [12, 512, 512]> attention_scores_69_cast_fp16 = add(x = attention_scores_67_cast_fp16, y = rel_att_cast_fp16)[name = string("attention_scores_69_cast_fp16")]; |
| tensor<int32, [4]> var_2714 = const()[name = string("op_2714"), val = tensor<int32, [4]>([-1, 12, 512, 512])]; |
| tensor<fp16, [1, 12, 512, 512]> attention_scores_cast_fp16 = reshape(shape = var_2714, x = attention_scores_69_cast_fp16)[name = string("attention_scores_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_273_cast_fp16 = select(a = var_41_to_fp16, b = attention_scores_cast_fp16, cond = var_308)[name = string("input_273_cast_fp16")]; |
| tensor<fp16, [1, 12, 512, 512]> input_275_cast_fp16 = softmax(axis = var_32, x = input_273_cast_fp16)[name = string("input_275_cast_fp16")]; |
| tensor<int32, [3]> var_2723 = const()[name = string("op_2723"), val = tensor<int32, [3]>([-1, 512, 512])]; |
| tensor<fp16, [12, 512, 512]> var_2724_cast_fp16 = reshape(shape = var_2723, x = input_275_cast_fp16)[name = string("op_2724_cast_fp16")]; |
| bool context_layer_45_transpose_x_0 = const()[name = string("context_layer_45_transpose_x_0"), val = bool(false)]; |
| bool context_layer_45_transpose_y_0 = const()[name = string("context_layer_45_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [12, 512, 64]> context_layer_45_cast_fp16 = matmul(transpose_x = context_layer_45_transpose_x_0, transpose_y = context_layer_45_transpose_y_0, x = var_2724_cast_fp16, y = value_layer_cast_fp16)[name = string("context_layer_45_cast_fp16")]; |
| tensor<int32, [4]> var_2728 = const()[name = string("op_2728"), val = tensor<int32, [4]>([-1, 12, 512, 64])]; |
| tensor<fp16, [1, 12, 512, 64]> var_2729_cast_fp16 = reshape(shape = var_2728, x = context_layer_45_cast_fp16)[name = string("op_2729_cast_fp16")]; |
| tensor<int32, [4]> var_2730 = const()[name = string("op_2730"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2735 = const()[name = string("op_2735"), val = tensor<int32, [3]>([1, 512, -1])]; |
| tensor<fp16, [1, 512, 12, 64]> var_2731_cast_fp16 = transpose(perm = var_2730, x = var_2729_cast_fp16)[name = string("transpose_8")]; |
| tensor<fp16, [1, 512, 768]> input_277_cast_fp16 = reshape(shape = var_2735, x = var_2731_cast_fp16)[name = string("input_277_cast_fp16")]; |
| tensor<fp16, [768, 768]> deberta_encoder_layer_11_attention_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387566144)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_attention_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388745856)))]; |
| tensor<fp16, [1, 512, 768]> linear_93_cast_fp16 = linear(bias = deberta_encoder_layer_11_attention_output_dense_bias_to_fp16, weight = deberta_encoder_layer_11_attention_output_dense_weight_to_fp16, x = input_277_cast_fp16)[name = string("linear_93_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_281_cast_fp16 = add(x = linear_93_cast_fp16, y = input_269_cast_fp16)[name = string("input_281_cast_fp16")]; |
| tensor<int32, [1]> input_283_axes_0 = const()[name = string("input_283_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388747456)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388749056)))]; |
| tensor<fp16, [1, 512, 768]> input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, beta = deberta_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")]; |
| tensor<fp16, [3072, 768]> deberta_encoder_layer_11_intermediate_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_11_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388750656)))]; |
| tensor<fp16, [3072]> deberta_encoder_layer_11_intermediate_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_11_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393469312)))]; |
| tensor<fp16, [1, 512, 3072]> linear_94_cast_fp16 = linear(bias = deberta_encoder_layer_11_intermediate_dense_bias_to_fp16, weight = deberta_encoder_layer_11_intermediate_dense_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_94_cast_fp16")]; |
| string input_287_mode_0 = const()[name = string("input_287_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_287_cast_fp16 = gelu(mode = input_287_mode_0, x = linear_94_cast_fp16)[name = string("input_287_cast_fp16")]; |
| tensor<fp16, [768, 3072]> deberta_encoder_layer_11_output_dense_weight_to_fp16 = const()[name = string("deberta_encoder_layer_11_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393475520)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_output_dense_bias_to_fp16 = const()[name = string("deberta_encoder_layer_11_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398194176)))]; |
| tensor<fp16, [1, 512, 768]> linear_95_cast_fp16 = linear(bias = deberta_encoder_layer_11_output_dense_bias_to_fp16, weight = deberta_encoder_layer_11_output_dense_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_95_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_291_cast_fp16 = add(x = linear_95_cast_fp16, y = input_283_cast_fp16)[name = string("input_291_cast_fp16")]; |
| tensor<int32, [1]> token_embeds_axes_0 = const()[name = string("token_embeds_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_output_LayerNorm_weight_to_fp16 = const()[name = string("deberta_encoder_layer_11_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398195776)))]; |
| tensor<fp16, [768]> deberta_encoder_layer_11_output_LayerNorm_bias_to_fp16 = const()[name = string("deberta_encoder_layer_11_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398197376)))]; |
| tensor<fp16, [1, 512, 768]> token_embeds_cast_fp16 = layer_norm(axes = token_embeds_axes_0, beta = deberta_encoder_layer_11_output_LayerNorm_bias_to_fp16, epsilon = var_51_to_fp16, gamma = deberta_encoder_layer_11_output_LayerNorm_weight_to_fp16, x = input_291_cast_fp16)[name = string("token_embeds_cast_fp16")]; |
| int32 var_2769 = const()[name = string("op_2769"), val = int32(128001)]; |
| tensor<bool, [1, 512]> var_2770 = equal(x = input_ids, y = var_2769)[name = string("op_2770")]; |
| string cast_263_to_fp16_dtype_0 = const()[name = string("cast_263_to_fp16_dtype_0"), val = string("fp16")]; |
| string cast_262_to_fp16_dtype_0 = const()[name = string("cast_262_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1, 512]> attention_mask_to_fp16 = cast(dtype = cast_262_to_fp16_dtype_0, x = attention_mask)[name = string("cast_290")]; |
| tensor<fp16, [1, 512]> var_2770_to_fp16 = cast(dtype = cast_263_to_fp16_dtype_0, x = var_2770)[name = string("cast_291")]; |
| tensor<fp16, [1, 512]> ent_hits_cast_fp16 = mul(x = var_2770_to_fp16, y = attention_mask_to_fp16)[name = string("ent_hits_cast_fp16")]; |
| int32 var_2777 = const()[name = string("op_2777"), val = int32(1)]; |
| bool var_2779_exclusive_0 = const()[name = string("op_2779_exclusive_0"), val = bool(false)]; |
| bool var_2779_reverse_0 = const()[name = string("op_2779_reverse_0"), val = bool(false)]; |
| tensor<fp16, [1, 512]> var_2779_cast_fp16 = cumsum(axis = var_2777, exclusive = var_2779_exclusive_0, reverse = var_2779_reverse_0, x = ent_hits_cast_fp16)[name = string("op_2779_cast_fp16")]; |
| tensor<fp16, [1, 512]> ent_order_cast_fp16 = mul(x = var_2779_cast_fp16, y = ent_hits_cast_fp16)[name = string("ent_order_cast_fp16")]; |
| tensor<int32, [1]> var_2782_axes_0 = const()[name = string("op_2782_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1, 1, 512]> var_2782_cast_fp16 = expand_dims(axes = var_2782_axes_0, x = ent_order_cast_fp16)[name = string("op_2782_cast_fp16")]; |
| tensor<fp16, [1, 12, 1]> ent_slots_to_fp16 = const()[name = string("ent_slots_to_fp16"), val = tensor<fp16, [1, 12, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398198976)))]; |
| tensor<bool, [1, 12, 512]> var_2783_cast_fp16 = equal(x = var_2782_cast_fp16, y = ent_slots_to_fp16)[name = string("op_2783_cast_fp16")]; |
| bool input_297_transpose_x_0 = const()[name = string("input_297_transpose_x_0"), val = bool(false)]; |
| bool input_297_transpose_y_0 = const()[name = string("input_297_transpose_y_0"), val = bool(false)]; |
| string cast_264_to_fp16_dtype_0 = const()[name = string("cast_264_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1, 12, 512]> var_2783_cast_fp16_to_fp16 = cast(dtype = cast_264_to_fp16_dtype_0, x = var_2783_cast_fp16)[name = string("cast_289")]; |
| tensor<fp16, [1, 12, 768]> input_297_cast_fp16 = matmul(transpose_x = input_297_transpose_x_0, transpose_y = input_297_transpose_y_0, x = var_2783_cast_fp16_to_fp16, y = token_embeds_cast_fp16)[name = string("input_297_cast_fp16")]; |
| int32 var_2790 = const()[name = string("op_2790"), val = int32(128003)]; |
| tensor<bool, [1, 512]> var_2791 = equal(x = input_ids, y = var_2790)[name = string("op_2791")]; |
| string cast_265_to_fp16_dtype_0 = const()[name = string("cast_265_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1, 512]> var_2791_to_fp16 = cast(dtype = cast_265_to_fp16_dtype_0, x = var_2791)[name = string("cast_288")]; |
| tensor<fp16, [1, 512]> rel_hits_cast_fp16 = mul(x = var_2791_to_fp16, y = attention_mask_to_fp16)[name = string("rel_hits_cast_fp16")]; |
| int32 var_2798 = const()[name = string("op_2798"), val = int32(1)]; |
| bool var_2800_exclusive_0 = const()[name = string("op_2800_exclusive_0"), val = bool(false)]; |
| bool var_2800_reverse_0 = const()[name = string("op_2800_reverse_0"), val = bool(false)]; |
| tensor<fp16, [1, 512]> var_2800_cast_fp16 = cumsum(axis = var_2798, exclusive = var_2800_exclusive_0, reverse = var_2800_reverse_0, x = rel_hits_cast_fp16)[name = string("op_2800_cast_fp16")]; |
| tensor<fp16, [1, 512]> rel_order_cast_fp16 = mul(x = var_2800_cast_fp16, y = rel_hits_cast_fp16)[name = string("rel_order_cast_fp16")]; |
| tensor<int32, [1]> var_2803_axes_0 = const()[name = string("op_2803_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1, 1, 512]> var_2803_cast_fp16 = expand_dims(axes = var_2803_axes_0, x = rel_order_cast_fp16)[name = string("op_2803_cast_fp16")]; |
| tensor<fp16, [1, 8, 1]> rel_slots_to_fp16 = const()[name = string("rel_slots_to_fp16"), val = tensor<fp16, [1, 8, 1]>([[[0x1p+0], [0x1p+1], [0x1.8p+1], [0x1p+2], [0x1.4p+2], [0x1.8p+2], [0x1.cp+2], [0x1p+3]]])]; |
| tensor<bool, [1, 8, 512]> var_2804_cast_fp16 = equal(x = var_2803_cast_fp16, y = rel_slots_to_fp16)[name = string("op_2804_cast_fp16")]; |
| bool var_2810_transpose_x_0 = const()[name = string("op_2810_transpose_x_0"), val = bool(false)]; |
| bool var_2810_transpose_y_0 = const()[name = string("op_2810_transpose_y_0"), val = bool(false)]; |
| string cast_266_to_fp16_dtype_0 = const()[name = string("cast_266_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1, 8, 512]> var_2804_cast_fp16_to_fp16 = cast(dtype = cast_266_to_fp16_dtype_0, x = var_2804_cast_fp16)[name = string("cast_287")]; |
| tensor<fp16, [1, 8, 768]> var_2810_cast_fp16 = matmul(transpose_x = var_2810_transpose_x_0, transpose_y = var_2810_transpose_y_0, x = var_2804_cast_fp16_to_fp16, y = token_embeds_cast_fp16)[name = string("op_2810_cast_fp16")]; |
| tensor<int32, [1]> var_2817_axes_0 = const()[name = string("op_2817_axes_0"), val = tensor<int32, [1]>([1])]; |
| string cast_267_to_fp16_dtype_0 = const()[name = string("cast_267_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1, 512]> words_mask_to_fp16 = cast(dtype = cast_267_to_fp16_dtype_0, x = words_mask)[name = string("cast_286")]; |
| tensor<fp16, [1, 1, 512]> var_2817_cast_fp16 = expand_dims(axes = var_2817_axes_0, x = words_mask_to_fp16)[name = string("op_2817_cast_fp16")]; |
| tensor<fp16, [1, 320, 1]> word_slots_to_fp16 = const()[name = string("word_slots_to_fp16"), val = tensor<fp16, [1, 320, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398199104)))]; |
| tensor<bool, [1, 320, 512]> word_sel_cast_fp16 = equal(x = var_2817_cast_fp16, y = word_slots_to_fp16)[name = string("word_sel_cast_fp16")]; |
| bool input_293_transpose_x_0 = const()[name = string("input_293_transpose_x_0"), val = bool(false)]; |
| bool input_293_transpose_y_0 = const()[name = string("input_293_transpose_y_0"), val = bool(false)]; |
| string cast_268_to_fp16_dtype_0 = const()[name = string("cast_268_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1, 320, 512]> word_sel_cast_fp16_to_fp16 = cast(dtype = cast_268_to_fp16_dtype_0, x = word_sel_cast_fp16)[name = string("cast_285")]; |
| tensor<fp16, [1, 320, 768]> input_293_cast_fp16 = matmul(transpose_x = input_293_transpose_x_0, transpose_y = input_293_transpose_y_0, x = word_sel_cast_fp16_to_fp16, y = token_embeds_cast_fp16)[name = string("input_293_cast_fp16")]; |
| tensor<int32, [3]> input_293_batch_first_transpose_perm_0 = const()[name = string("input_293_batch_first_transpose_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| string fwd_out_batch_first_direction_0 = const()[name = string("fwd_out_batch_first_direction_0"), val = string("forward")]; |
| bool fwd_out_batch_first_output_sequence_0 = const()[name = string("fwd_out_batch_first_output_sequence_0"), val = bool(true)]; |
| string fwd_out_batch_first_recurrent_activation_0 = const()[name = string("fwd_out_batch_first_recurrent_activation_0"), val = string("sigmoid")]; |
| string fwd_out_batch_first_cell_activation_0 = const()[name = string("fwd_out_batch_first_cell_activation_0"), val = string("tanh")]; |
| string fwd_out_batch_first_activation_0 = const()[name = string("fwd_out_batch_first_activation_0"), val = string("tanh")]; |
| tensor<fp16, [1, 384]> fwd_out_batch_first_lstm_h0_squeeze_to_fp16 = const()[name = string("fwd_out_batch_first_lstm_h0_squeeze_to_fp16"), val = tensor<fp16, [1, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398199808)))]; |
| tensor<fp16, [1536, 768]> concat_27_to_fp16 = const()[name = string("concat_27_to_fp16"), val = tensor<fp16, [1536, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398200640)))]; |
| tensor<fp16, [1536, 384]> concat_28_to_fp16 = const()[name = string("concat_28_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400560000)))]; |
| tensor<fp16, [1536]> concat_26_to_fp16 = const()[name = string("concat_26_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401739712)))]; |
| tensor<fp16, [320, 1, 768]> input_293_batch_first_transpose_cast_fp16 = transpose(perm = input_293_batch_first_transpose_perm_0, x = input_293_cast_fp16)[name = string("transpose_7")]; |
| tensor<fp16, [320, 1, 384]> fwd_out_batch_first_cast_fp16_0, tensor<fp16, [1, 384]> fwd_out_batch_first_cast_fp16_1, tensor<fp16, [1, 384]> fwd_out_batch_first_cast_fp16_2 = lstm(activation = fwd_out_batch_first_activation_0, bias = concat_26_to_fp16, cell_activation = fwd_out_batch_first_cell_activation_0, direction = fwd_out_batch_first_direction_0, initial_c = fwd_out_batch_first_lstm_h0_squeeze_to_fp16, initial_h = fwd_out_batch_first_lstm_h0_squeeze_to_fp16, output_sequence = fwd_out_batch_first_output_sequence_0, recurrent_activation = fwd_out_batch_first_recurrent_activation_0, weight_hh = concat_28_to_fp16, weight_ih = concat_27_to_fp16, x = input_293_batch_first_transpose_cast_fp16)[name = string("fwd_out_batch_first_cast_fp16")]; |
| tensor<int32, [3]> fwd_out_perm_0 = const()[name = string("fwd_out_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| int32 var_2858 = const()[name = string("op_2858"), val = int32(1)]; |
| tensor<int32, [1, 1]> var_2859 = sub(x = text_lengths, y = var_2858)[name = string("op_2859")]; |
| tensor<int32, [1, 320]> var_2861 = sub(x = var_2859, y = word_positions)[name = string("op_2861")]; |
| string var_2861_promoted_to_fp16_dtype_0 = const()[name = string("op_2861_promoted_to_fp16_dtype_0"), val = string("fp16")]; |
| fp16 var_2862_promoted_to_fp16 = const()[name = string("op_2862_promoted_to_fp16"), val = fp16(0x0p+0)]; |
| fp16 const_574_to_fp16 = const()[name = string("const_574_to_fp16"), val = fp16(inf)]; |
| tensor<fp16, [1, 320]> var_2861_to_fp16 = cast(dtype = var_2861_promoted_to_fp16_dtype_0, x = var_2861)[name = string("cast_284")]; |
| tensor<fp16, [1, 320]> clip_0_cast_fp16 = clip(alpha = var_2862_promoted_to_fp16, beta = const_574_to_fp16, x = var_2861_to_fp16)[name = string("clip_0_cast_fp16")]; |
| string cast_271_dtype_0 = const()[name = string("cast_271_dtype_0"), val = string("int32")]; |
| tensor<int32, [1]> var_2876_axes_0 = const()[name = string("op_2876_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<int32, [1, 320]> clip_0_cast_fp16_to_int32 = cast(dtype = cast_271_dtype_0, x = clip_0_cast_fp16)[name = string("cast_283")]; |
| tensor<int32, [1, 320, 1]> var_2876 = expand_dims(axes = var_2876_axes_0, x = clip_0_cast_fp16_to_int32)[name = string("op_2876")]; |
| tensor<int32, [3]> rev_gather_in_reps_0 = const()[name = string("rev_gather_in_reps_0"), val = tensor<int32, [3]>([1, 1, 768])]; |
| tensor<int32, [1, 320, 768]> rev_gather_in = tile(reps = rev_gather_in_reps_0, x = var_2876)[name = string("rev_gather_in")]; |
| tensor<int32, [3]> rev_gather_out_reps_0 = const()[name = string("rev_gather_out_reps_0"), val = tensor<int32, [3]>([1, 1, 384])]; |
| tensor<int32, [1, 320, 384]> rev_gather_out = tile(reps = rev_gather_out_reps_0, x = var_2876)[name = string("rev_gather_out")]; |
| int32 var_2891 = const()[name = string("op_2891"), val = int32(1)]; |
| bool input_295_validate_indices_0 = const()[name = string("input_295_validate_indices_0"), val = bool(false)]; |
| string rev_gather_in_to_int16_dtype_0 = const()[name = string("rev_gather_in_to_int16_dtype_0"), val = string("int16")]; |
| tensor<int16, [1, 320, 768]> rev_gather_in_to_int16 = cast(dtype = rev_gather_in_to_int16_dtype_0, x = rev_gather_in)[name = string("cast_282")]; |
| tensor<fp16, [1, 320, 768]> input_295_cast_fp16_cast_uint16 = gather_along_axis(axis = var_2891, indices = rev_gather_in_to_int16, validate_indices = input_295_validate_indices_0, x = input_293_cast_fp16)[name = string("input_295_cast_fp16_cast_uint16")]; |
| tensor<int32, [3]> input_295_batch_first_transpose_perm_0 = const()[name = string("input_295_batch_first_transpose_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| string bwd_rev_batch_first_direction_0 = const()[name = string("bwd_rev_batch_first_direction_0"), val = string("forward")]; |
| bool bwd_rev_batch_first_output_sequence_0 = const()[name = string("bwd_rev_batch_first_output_sequence_0"), val = bool(true)]; |
| string bwd_rev_batch_first_recurrent_activation_0 = const()[name = string("bwd_rev_batch_first_recurrent_activation_0"), val = string("sigmoid")]; |
| string bwd_rev_batch_first_cell_activation_0 = const()[name = string("bwd_rev_batch_first_cell_activation_0"), val = string("tanh")]; |
| string bwd_rev_batch_first_activation_0 = const()[name = string("bwd_rev_batch_first_activation_0"), val = string("tanh")]; |
| tensor<fp16, [1536, 768]> concat_32_to_fp16 = const()[name = string("concat_32_to_fp16"), val = tensor<fp16, [1536, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401742848)))]; |
| tensor<fp16, [1536, 384]> concat_33_to_fp16 = const()[name = string("concat_33_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404102208)))]; |
| tensor<fp16, [1536]> concat_31_to_fp16 = const()[name = string("concat_31_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405281920)))]; |
| tensor<fp16, [320, 1, 768]> input_295_batch_first_transpose_cast_fp16 = transpose(perm = input_295_batch_first_transpose_perm_0, x = input_295_cast_fp16_cast_uint16)[name = string("transpose_5")]; |
| tensor<fp16, [320, 1, 384]> bwd_rev_batch_first_cast_fp16_0, tensor<fp16, [1, 384]> bwd_rev_batch_first_cast_fp16_1, tensor<fp16, [1, 384]> bwd_rev_batch_first_cast_fp16_2 = lstm(activation = bwd_rev_batch_first_activation_0, bias = concat_31_to_fp16, cell_activation = bwd_rev_batch_first_cell_activation_0, direction = bwd_rev_batch_first_direction_0, initial_c = fwd_out_batch_first_lstm_h0_squeeze_to_fp16, initial_h = fwd_out_batch_first_lstm_h0_squeeze_to_fp16, output_sequence = bwd_rev_batch_first_output_sequence_0, recurrent_activation = bwd_rev_batch_first_recurrent_activation_0, weight_hh = concat_33_to_fp16, weight_ih = concat_32_to_fp16, x = input_295_batch_first_transpose_cast_fp16)[name = string("bwd_rev_batch_first_cast_fp16")]; |
| tensor<int32, [3]> bwd_rev_perm_0 = const()[name = string("bwd_rev_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| int32 var_2917 = const()[name = string("op_2917"), val = int32(1)]; |
| bool bwd_out_validate_indices_0 = const()[name = string("bwd_out_validate_indices_0"), val = bool(false)]; |
| string rev_gather_out_to_uint16_dtype_0 = const()[name = string("rev_gather_out_to_uint16_dtype_0"), val = string("uint16")]; |
| tensor<uint16, [1, 320, 384]> rev_gather_out_to_uint16 = cast(dtype = rev_gather_out_to_uint16_dtype_0, x = rev_gather_out)[name = string("cast_281")]; |
| tensor<fp16, [1, 320, 384]> bwd_rev_cast_fp16 = transpose(perm = bwd_rev_perm_0, x = bwd_rev_batch_first_cast_fp16_0)[name = string("transpose_4")]; |
| tensor<fp16, [1, 320, 384]> bwd_out_cast_fp16_cast_uint16 = gather_along_axis(axis = var_2917, indices = rev_gather_out_to_uint16, validate_indices = bwd_out_validate_indices_0, x = bwd_rev_cast_fp16)[name = string("bwd_out_cast_fp16_cast_uint16")]; |
| int32 var_2921 = const()[name = string("op_2921"), val = int32(-1)]; |
| bool token_rep_1_interleave_0 = const()[name = string("token_rep_1_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 320, 384]> fwd_out_cast_fp16 = transpose(perm = fwd_out_perm_0, x = fwd_out_batch_first_cast_fp16_0)[name = string("transpose_6")]; |
| tensor<fp16, [1, 320, 768]> token_rep_1_cast_fp16 = concat(axis = var_2921, interleave = token_rep_1_interleave_0, values = (fwd_out_cast_fp16, bwd_out_cast_fp16_cast_uint16))[name = string("token_rep_1_cast_fp16")]; |
| tensor<fp16, [3072, 768]> prompt_rep_layer_0_weight_to_fp16 = const()[name = string("prompt_rep_layer_0_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405285056)))]; |
| tensor<fp16, [3072]> prompt_rep_layer_0_bias_to_fp16 = const()[name = string("prompt_rep_layer_0_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410003712)))]; |
| tensor<fp16, [1, 12, 3072]> linear_96_cast_fp16 = linear(bias = prompt_rep_layer_0_bias_to_fp16, weight = prompt_rep_layer_0_weight_to_fp16, x = input_297_cast_fp16)[name = string("linear_96_cast_fp16")]; |
| tensor<fp16, [1, 12, 3072]> input_301_cast_fp16 = relu(x = linear_96_cast_fp16)[name = string("input_301_cast_fp16")]; |
| tensor<fp16, [768, 3072]> prompt_rep_layer_3_weight_to_fp16 = const()[name = string("prompt_rep_layer_3_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410009920)))]; |
| tensor<fp16, [768]> prompt_rep_layer_3_bias_to_fp16 = const()[name = string("prompt_rep_layer_3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414728576)))]; |
| tensor<fp16, [1, 12, 768]> linear_97_cast_fp16 = linear(bias = prompt_rep_layer_3_bias_to_fp16, weight = prompt_rep_layer_3_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_97_cast_fp16")]; |
| int32 var_2939 = const()[name = string("op_2939"), val = int32(-1)]; |
| tensor<fp16, [1536, 768]> scorer_proj_token_weight_to_fp16 = const()[name = string("scorer_proj_token_weight_to_fp16"), val = tensor<fp16, [1536, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414730176)))]; |
| tensor<fp16, [1536]> scorer_proj_token_bias_to_fp16 = const()[name = string("scorer_proj_token_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417089536)))]; |
| tensor<fp16, [1, 320, 1536]> linear_98_cast_fp16 = linear(bias = scorer_proj_token_bias_to_fp16, weight = scorer_proj_token_weight_to_fp16, x = token_rep_1_cast_fp16)[name = string("linear_98_cast_fp16")]; |
| tensor<int32, [5]> var_2953 = const()[name = string("op_2953"), val = tensor<int32, [5]>([1, 320, 1, 2, 768])]; |
| tensor<fp16, [1, 320, 1, 2, 768]> token_rep_3_cast_fp16 = reshape(shape = var_2953, x = linear_98_cast_fp16)[name = string("token_rep_3_cast_fp16")]; |
| tensor<fp16, [1536, 768]> scorer_proj_label_weight_to_fp16 = const()[name = string("scorer_proj_label_weight_to_fp16"), val = tensor<fp16, [1536, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417092672)))]; |
| tensor<fp16, [1536]> scorer_proj_label_bias_to_fp16 = const()[name = string("scorer_proj_label_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419452032)))]; |
| tensor<fp16, [1, 12, 1536]> linear_99_cast_fp16 = linear(bias = scorer_proj_label_bias_to_fp16, weight = scorer_proj_label_weight_to_fp16, x = linear_97_cast_fp16)[name = string("linear_99_cast_fp16")]; |
| tensor<int32, [5]> var_2958 = const()[name = string("op_2958"), val = tensor<int32, [5]>([1, 1, 12, 2, 768])]; |
| tensor<fp16, [1, 1, 12, 2, 768]> label_rep_3_cast_fp16 = reshape(shape = var_2958, x = linear_99_cast_fp16)[name = string("label_rep_3_cast_fp16")]; |
| tensor<int32, [5]> var_2961_reps_0 = const()[name = string("op_2961_reps_0"), val = tensor<int32, [5]>([1, 1, 12, 1, 1])]; |
| tensor<fp16, [1, 320, 12, 2, 768]> var_2961_cast_fp16 = tile(reps = var_2961_reps_0, x = token_rep_3_cast_fp16)[name = string("op_2961_cast_fp16")]; |
| tensor<int32, [5]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [5]>([3, 0, 1, 2, 4])]; |
| tensor<int32, [5]> var_2965_reps_0 = const()[name = string("op_2965_reps_0"), val = tensor<int32, [5]>([1, 320, 1, 1, 1])]; |
| tensor<fp16, [1, 320, 12, 2, 768]> var_2965_cast_fp16 = tile(reps = var_2965_reps_0, x = label_rep_3_cast_fp16)[name = string("op_2965_cast_fp16")]; |
| tensor<int32, [5]> var_2966 = const()[name = string("op_2966"), val = tensor<int32, [5]>([3, 0, 1, 2, 4])]; |
| tensor<int32, [5]> var_2968_begin_0 = const()[name = string("op_2968_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> var_2968_end_0 = const()[name = string("op_2968_end_0"), val = tensor<int32, [5]>([1, 1, 320, 12, 768])]; |
| tensor<bool, [5]> var_2968_end_mask_0 = const()[name = string("op_2968_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> var_2968_squeeze_mask_0 = const()[name = string("op_2968_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp16, [2, 1, 320, 12, 768]> token_rep_cast_fp16 = transpose(perm = var_2962, x = var_2961_cast_fp16)[name = string("transpose_3")]; |
| tensor<fp16, [1, 320, 12, 768]> var_2968_cast_fp16 = slice_by_index(begin = var_2968_begin_0, end = var_2968_end_0, end_mask = var_2968_end_mask_0, squeeze_mask = var_2968_squeeze_mask_0, x = token_rep_cast_fp16)[name = string("op_2968_cast_fp16")]; |
| tensor<int32, [5]> var_2969_begin_0 = const()[name = string("op_2969_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> var_2969_end_0 = const()[name = string("op_2969_end_0"), val = tensor<int32, [5]>([1, 1, 320, 12, 768])]; |
| tensor<bool, [5]> var_2969_end_mask_0 = const()[name = string("op_2969_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> var_2969_squeeze_mask_0 = const()[name = string("op_2969_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp16, [2, 1, 320, 12, 768]> label_rep_cast_fp16 = transpose(perm = var_2966, x = var_2965_cast_fp16)[name = string("transpose_2")]; |
| tensor<fp16, [1, 320, 12, 768]> var_2969_cast_fp16 = slice_by_index(begin = var_2969_begin_0, end = var_2969_end_0, end_mask = var_2969_end_mask_0, squeeze_mask = var_2969_squeeze_mask_0, x = label_rep_cast_fp16)[name = string("op_2969_cast_fp16")]; |
| tensor<int32, [5]> var_2970_begin_0 = const()[name = string("op_2970_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> var_2970_end_0 = const()[name = string("op_2970_end_0"), val = tensor<int32, [5]>([2, 1, 320, 12, 768])]; |
| tensor<bool, [5]> var_2970_end_mask_0 = const()[name = string("op_2970_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> var_2970_squeeze_mask_0 = const()[name = string("op_2970_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp16, [1, 320, 12, 768]> var_2970_cast_fp16 = slice_by_index(begin = var_2970_begin_0, end = var_2970_end_0, end_mask = var_2970_end_mask_0, squeeze_mask = var_2970_squeeze_mask_0, x = token_rep_cast_fp16)[name = string("op_2970_cast_fp16")]; |
| tensor<int32, [5]> var_2971_begin_0 = const()[name = string("op_2971_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> var_2971_end_0 = const()[name = string("op_2971_end_0"), val = tensor<int32, [5]>([2, 1, 320, 12, 768])]; |
| tensor<bool, [5]> var_2971_end_mask_0 = const()[name = string("op_2971_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> var_2971_squeeze_mask_0 = const()[name = string("op_2971_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp16, [1, 320, 12, 768]> var_2971_cast_fp16 = slice_by_index(begin = var_2971_begin_0, end = var_2971_end_0, end_mask = var_2971_end_mask_0, squeeze_mask = var_2971_squeeze_mask_0, x = label_rep_cast_fp16)[name = string("op_2971_cast_fp16")]; |
| tensor<fp16, [1, 320, 12, 768]> var_2972_cast_fp16 = mul(x = var_2970_cast_fp16, y = var_2971_cast_fp16)[name = string("op_2972_cast_fp16")]; |
| bool input_305_interleave_0 = const()[name = string("input_305_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 320, 12, 2304]> input_305_cast_fp16 = concat(axis = var_2939, interleave = input_305_interleave_0, values = (var_2968_cast_fp16, var_2969_cast_fp16, var_2972_cast_fp16))[name = string("input_305_cast_fp16")]; |
| tensor<fp16, [3072, 2304]> scorer_out_mlp_0_weight_to_fp16 = const()[name = string("scorer_out_mlp_0_weight_to_fp16"), val = tensor<fp16, [3072, 2304]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419455168)))]; |
| tensor<fp16, [3072]> scorer_out_mlp_0_bias_to_fp16 = const()[name = string("scorer_out_mlp_0_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433611008)))]; |
| tensor<fp16, [1, 320, 12, 3072]> linear_100_cast_fp16 = linear(bias = scorer_out_mlp_0_bias_to_fp16, weight = scorer_out_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = string("linear_100_cast_fp16")]; |
| tensor<fp16, [1, 320, 12, 3072]> input_311_cast_fp16 = relu(x = linear_100_cast_fp16)[name = string("input_311_cast_fp16")]; |
| tensor<fp16, [3, 3072]> scorer_out_mlp_3_weight_to_fp16 = const()[name = string("scorer_out_mlp_3_weight_to_fp16"), val = tensor<fp16, [3, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433617216)))]; |
| tensor<fp16, [3]> scorer_out_mlp_3_bias_to_fp16 = const()[name = string("scorer_out_mlp_3_bias_to_fp16"), val = tensor<fp16, [3]>([-0x1.984p-8, -0x1.778p-7, 0x1.34p-9])]; |
| tensor<fp16, [1, 320, 12, 3]> linear_101_cast_fp16 = linear(bias = scorer_out_mlp_3_bias_to_fp16, weight = scorer_out_mlp_3_weight_to_fp16, x = input_311_cast_fp16)[name = string("linear_101_cast_fp16")]; |
| string linear_101_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_101_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; |
| tensor<int32, [1]> var_2996_axes_0 = const()[name = string("op_2996_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<int32, [1, 16, 1]> var_2996 = expand_dims(axes = var_2996_axes_0, x = span_mask)[name = string("op_2996")]; |
| tensor<int32, [1, 16, 2]> span_idx_1 = mul(x = span_idx, y = var_2996)[name = string("span_idx")]; |
| int32 var_2998 = const()[name = string("op_2998"), val = int32(-1)]; |
| int32 var_3000 = const()[name = string("op_3000"), val = int32(1)]; |
| tensor<fp16, [3072, 768]> token_marker_project_start_0_weight_to_fp16 = const()[name = string("token_marker_project_start_0_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433635712)))]; |
| tensor<fp16, [3072]> token_marker_project_start_0_bias_to_fp16 = const()[name = string("token_marker_project_start_0_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438354368)))]; |
| tensor<fp16, [1, 320, 3072]> linear_102_cast_fp16 = linear(bias = token_marker_project_start_0_bias_to_fp16, weight = token_marker_project_start_0_weight_to_fp16, x = token_rep_1_cast_fp16)[name = string("linear_102_cast_fp16")]; |
| tensor<fp16, [1, 320, 3072]> input_315_cast_fp16 = relu(x = linear_102_cast_fp16)[name = string("input_315_cast_fp16")]; |
| tensor<fp16, [768, 3072]> token_marker_project_start_3_weight_to_fp16 = const()[name = string("token_marker_project_start_3_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438360576)))]; |
| tensor<fp16, [768]> token_marker_project_start_3_bias_to_fp16 = const()[name = string("token_marker_project_start_3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443079232)))]; |
| tensor<fp16, [1, 320, 768]> linear_103_cast_fp16 = linear(bias = token_marker_project_start_3_bias_to_fp16, weight = token_marker_project_start_3_weight_to_fp16, x = input_315_cast_fp16)[name = string("linear_103_cast_fp16")]; |
| tensor<fp16, [3072, 768]> token_marker_project_end_0_weight_to_fp16 = const()[name = string("token_marker_project_end_0_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443080832)))]; |
| tensor<fp16, [3072]> token_marker_project_end_0_bias_to_fp16 = const()[name = string("token_marker_project_end_0_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447799488)))]; |
| tensor<fp16, [1, 320, 3072]> linear_104_cast_fp16 = linear(bias = token_marker_project_end_0_bias_to_fp16, weight = token_marker_project_end_0_weight_to_fp16, x = token_rep_1_cast_fp16)[name = string("linear_104_cast_fp16")]; |
| tensor<fp16, [1, 320, 3072]> input_321_cast_fp16 = relu(x = linear_104_cast_fp16)[name = string("input_321_cast_fp16")]; |
| tensor<fp16, [768, 3072]> token_marker_project_end_3_weight_to_fp16 = const()[name = string("token_marker_project_end_3_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447805696)))]; |
| tensor<fp16, [768]> token_marker_project_end_3_bias_to_fp16 = const()[name = string("token_marker_project_end_3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452524352)))]; |
| tensor<fp16, [1, 320, 768]> linear_105_cast_fp16 = linear(bias = token_marker_project_end_3_bias_to_fp16, weight = token_marker_project_end_3_weight_to_fp16, x = input_321_cast_fp16)[name = string("linear_105_cast_fp16")]; |
| tensor<int32, [3]> indices_1_begin_0 = const()[name = string("indices_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> indices_1_end_0 = const()[name = string("indices_1_end_0"), val = tensor<int32, [3]>([1, 16, 1])]; |
| tensor<bool, [3]> indices_1_end_mask_0 = const()[name = string("indices_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])]; |
| tensor<bool, [3]> indices_1_squeeze_mask_0 = const()[name = string("indices_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])]; |
| tensor<int32, [1, 16]> indices_1 = slice_by_index(begin = indices_1_begin_0, end = indices_1_end_0, end_mask = indices_1_end_mask_0, squeeze_mask = indices_1_squeeze_mask_0, x = span_idx_1)[name = string("indices_1")]; |
| tensor<int32, [1]> var_3032_axes_0 = const()[name = string("op_3032_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [1, 16, 1]> var_3032 = expand_dims(axes = var_3032_axes_0, x = indices_1)[name = string("op_3032")]; |
| tensor<int32, [3]> expanded_indices_1_reps_0 = const()[name = string("expanded_indices_1_reps_0"), val = tensor<int32, [3]>([1, 1, 768])]; |
| tensor<int32, [1, 16, 768]> expanded_indices_1 = tile(reps = expanded_indices_1_reps_0, x = var_3032)[name = string("expanded_indices_1")]; |
| bool start_span_rep_validate_indices_0 = const()[name = string("start_span_rep_validate_indices_0"), val = bool(false)]; |
| string expanded_indices_1_to_uint16_dtype_0 = const()[name = string("expanded_indices_1_to_uint16_dtype_0"), val = string("uint16")]; |
| tensor<uint16, [1, 16, 768]> expanded_indices_1_to_uint16 = cast(dtype = expanded_indices_1_to_uint16_dtype_0, x = expanded_indices_1)[name = string("cast_279")]; |
| tensor<fp16, [1, 16, 768]> start_span_rep_cast_fp16_cast_uint16 = gather_along_axis(axis = var_3000, indices = expanded_indices_1_to_uint16, validate_indices = start_span_rep_validate_indices_0, x = linear_103_cast_fp16)[name = string("start_span_rep_cast_fp16_cast_uint16")]; |
| tensor<int32, [3]> indices_begin_0 = const()[name = string("indices_begin_0"), val = tensor<int32, [3]>([0, 0, 1])]; |
| tensor<int32, [3]> indices_end_0 = const()[name = string("indices_end_0"), val = tensor<int32, [3]>([1, 16, 2])]; |
| tensor<bool, [3]> indices_end_mask_0 = const()[name = string("indices_end_mask_0"), val = tensor<bool, [3]>([true, true, false])]; |
| tensor<bool, [3]> indices_squeeze_mask_0 = const()[name = string("indices_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])]; |
| tensor<int32, [1, 16]> indices = slice_by_index(begin = indices_begin_0, end = indices_end_0, end_mask = indices_end_mask_0, squeeze_mask = indices_squeeze_mask_0, x = span_idx_1)[name = string("indices")]; |
| tensor<int32, [1]> var_3040_axes_0 = const()[name = string("op_3040_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [1, 16, 1]> var_3040 = expand_dims(axes = var_3040_axes_0, x = indices)[name = string("op_3040")]; |
| tensor<int32, [3]> expanded_indices_reps_0 = const()[name = string("expanded_indices_reps_0"), val = tensor<int32, [3]>([1, 1, 768])]; |
| tensor<int32, [1, 16, 768]> expanded_indices = tile(reps = expanded_indices_reps_0, x = var_3040)[name = string("expanded_indices")]; |
| bool end_span_rep_validate_indices_0 = const()[name = string("end_span_rep_validate_indices_0"), val = bool(false)]; |
| string expanded_indices_to_uint16_dtype_0 = const()[name = string("expanded_indices_to_uint16_dtype_0"), val = string("uint16")]; |
| tensor<uint16, [1, 16, 768]> expanded_indices_to_uint16 = cast(dtype = expanded_indices_to_uint16_dtype_0, x = expanded_indices)[name = string("cast_278")]; |
| tensor<fp16, [1, 16, 768]> end_span_rep_cast_fp16_cast_uint16 = gather_along_axis(axis = var_3000, indices = expanded_indices_to_uint16, validate_indices = end_span_rep_validate_indices_0, x = linear_105_cast_fp16)[name = string("end_span_rep_cast_fp16_cast_uint16")]; |
| bool var_3045_interleave_0 = const()[name = string("op_3045_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 1536]> var_3045_cast_fp16 = concat(axis = var_2998, interleave = var_3045_interleave_0, values = (start_span_rep_cast_fp16_cast_uint16, end_span_rep_cast_fp16_cast_uint16))[name = string("op_3045_cast_fp16")]; |
| tensor<fp16, [1, 16, 1536]> input_325_cast_fp16 = relu(x = var_3045_cast_fp16)[name = string("input_325_cast_fp16")]; |
| tensor<fp16, [3072, 1536]> token_marker_out_project_0_weight_to_fp16 = const()[name = string("token_marker_out_project_0_weight_to_fp16"), val = tensor<fp16, [3072, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452525952)))]; |
| tensor<fp16, [3072]> token_marker_out_project_0_bias_to_fp16 = const()[name = string("token_marker_out_project_0_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461963200)))]; |
| tensor<fp16, [1, 16, 3072]> linear_106_cast_fp16 = linear(bias = token_marker_out_project_0_bias_to_fp16, weight = token_marker_out_project_0_weight_to_fp16, x = input_325_cast_fp16)[name = string("linear_106_cast_fp16")]; |
| tensor<fp16, [1, 16, 3072]> input_329_cast_fp16 = relu(x = linear_106_cast_fp16)[name = string("input_329_cast_fp16")]; |
| tensor<fp16, [768, 3072]> token_marker_out_project_3_weight_to_fp16 = const()[name = string("token_marker_out_project_3_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461969408)))]; |
| tensor<fp16, [768]> token_marker_out_project_3_bias_to_fp16 = const()[name = string("token_marker_out_project_3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466688064)))]; |
| tensor<fp16, [1, 16, 768]> linear_107_cast_fp16 = linear(bias = token_marker_out_project_3_bias_to_fp16, weight = token_marker_out_project_3_weight_to_fp16, x = input_329_cast_fp16)[name = string("linear_107_cast_fp16")]; |
| int32 var_3057 = const()[name = string("op_3057"), val = int32(1)]; |
| int32 head_batch_dims_0 = const()[name = string("head_batch_dims_0"), val = int32(0)]; |
| bool head_validate_indices_0 = const()[name = string("head_validate_indices_0"), val = bool(false)]; |
| tensor<uint16, [240]> pair_head_to_uint16 = const()[name = string("pair_head_to_uint16"), val = tensor<uint16, [240]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466689664)))]; |
| tensor<fp16, [1, 240, 768]> head_cast_fp16_cast_uint16 = gather(axis = var_3057, batch_dims = head_batch_dims_0, indices = pair_head_to_uint16, validate_indices = head_validate_indices_0, x = linear_107_cast_fp16)[name = string("head_cast_fp16_cast_uint16")]; |
| int32 var_3059 = const()[name = string("op_3059"), val = int32(1)]; |
| int32 tail_batch_dims_0 = const()[name = string("tail_batch_dims_0"), val = int32(0)]; |
| bool tail_validate_indices_0 = const()[name = string("tail_validate_indices_0"), val = bool(false)]; |
| tensor<uint16, [240]> pair_tail_to_uint16 = const()[name = string("pair_tail_to_uint16"), val = tensor<uint16, [240]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466690240)))]; |
| tensor<fp16, [1, 240, 768]> tail_cast_fp16_cast_uint16 = gather(axis = var_3059, batch_dims = tail_batch_dims_0, indices = pair_tail_to_uint16, validate_indices = tail_validate_indices_0, x = linear_107_cast_fp16)[name = string("tail_cast_fp16_cast_uint16")]; |
| int32 var_3062 = const()[name = string("op_3062"), val = int32(-1)]; |
| bool input_333_interleave_0 = const()[name = string("input_333_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 240, 1536]> input_333_cast_fp16 = concat(axis = var_3062, interleave = input_333_interleave_0, values = (head_cast_fp16_cast_uint16, tail_cast_fp16_cast_uint16))[name = string("input_333_cast_fp16")]; |
| tensor<fp16, [3072, 1536]> pair_rep_layer_0_weight_to_fp16 = const()[name = string("pair_rep_layer_0_weight_to_fp16"), val = tensor<fp16, [3072, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466690816)))]; |
| tensor<fp16, [3072]> pair_rep_layer_0_bias_to_fp16 = const()[name = string("pair_rep_layer_0_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476128064)))]; |
| tensor<fp16, [1, 240, 3072]> linear_108_cast_fp16 = linear(bias = pair_rep_layer_0_bias_to_fp16, weight = pair_rep_layer_0_weight_to_fp16, x = input_333_cast_fp16)[name = string("linear_108_cast_fp16")]; |
| tensor<fp16, [1, 240, 3072]> input_337_cast_fp16 = relu(x = linear_108_cast_fp16)[name = string("input_337_cast_fp16")]; |
| tensor<fp16, [768, 3072]> pair_rep_layer_3_weight_to_fp16 = const()[name = string("pair_rep_layer_3_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476134272)))]; |
| tensor<fp16, [768]> pair_rep_layer_3_bias_to_fp16 = const()[name = string("pair_rep_layer_3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480852928)))]; |
| tensor<fp16, [1, 240, 768]> linear_109_cast_fp16 = linear(bias = pair_rep_layer_3_bias_to_fp16, weight = pair_rep_layer_3_weight_to_fp16, x = input_337_cast_fp16)[name = string("linear_109_cast_fp16")]; |
| bool matmul_0_transpose_x_1 = const()[name = string("matmul_0_transpose_x_1"), val = bool(false)]; |
| bool matmul_0_transpose_y_1 = const()[name = string("matmul_0_transpose_y_1"), val = bool(true)]; |
| tensor<fp16, [1, 240, 8]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_1, transpose_y = matmul_0_transpose_y_1, x = linear_109_cast_fp16, y = var_2810_cast_fp16)[name = string("matmul_0_cast_fp16")]; |
| string var_3079_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3079_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; |
| tensor<fp32, [1, 240, 8]> rel_logits = cast(dtype = var_3079_cast_fp16_to_fp32_dtype_0, x = matmul_0_cast_fp16)[name = string("cast_277")]; |
| tensor<fp32, [1, 320, 12, 3]> logits = cast(dtype = linear_101_cast_fp16_to_fp32_dtype_0, x = linear_101_cast_fp16)[name = string("cast_280")]; |
| } -> (logits, rel_logits); |
| } |
