| program(1.3) |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3605.1.1"}, {"coremlc-version", "3600.25.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] |
| { |
| func main<ios18>(tensor<int32, [1, 512]> attention_mask, tensor<int32, [1, 512]> input_ids) { |
| int32 var_5 = const()[name = string("op_5"), val = int32(1)]; |
| tensor<bool, [1, 512]> var_42 = not_equal(x = input_ids, y = var_5)[name = string("op_42")]; |
| string mask_dtype_0 = const()[name = string("mask_dtype_0"), val = string("int32")]; |
| bool var_44_exclusive_0 = const()[name = string("op_44_exclusive_0"), val = bool(false)]; |
| bool var_44_reverse_0 = const()[name = string("op_44_reverse_0"), val = bool(false)]; |
| tensor<int32, [1, 512]> mask = cast(dtype = mask_dtype_0, x = var_42)[name = string("cast_55")]; |
| tensor<int32, [1, 512]> var_44 = cumsum(axis = var_5, exclusive = var_44_exclusive_0, reverse = var_44_reverse_0, x = mask)[name = string("op_44")]; |
| tensor<int32, [1, 512]> incremental_indices = mul(x = var_44, y = mask)[name = string("incremental_indices")]; |
| int32 var_50 = const()[name = string("op_50"), val = int32(1)]; |
| tensor<int32, [1, 512]> input_3 = add(x = incremental_indices, y = var_50)[name = string("input_3")]; |
| 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, [40000, 768]> m_roberta_embeddings_word_embeddings_weight_to_fp16 = const()[name = string("m_roberta_embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [40000, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; |
| int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; |
| tensor<bool, [1, 512]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; |
| int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(40000)]; |
| tensor<int32, [1, 512]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")]; |
| tensor<int32, [1, 512]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")]; |
| int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)]; |
| tensor<fp16, [1, 512, 768]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0, validate_indices = inputs_embeds_validate_indices_0, x = m_roberta_embeddings_word_embeddings_weight_to_fp16)[name = string("inputs_embeds_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> token_type_embeddings_1_to_fp16 = const()[name = string("token_type_embeddings_1_to_fp16"), val = tensor<fp16, [1, 512, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61440128)))]; |
| tensor<fp16, [1, 512, 768]> embeddings_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = token_type_embeddings_1_to_fp16)[name = string("embeddings_1_cast_fp16")]; |
| int32 position_embeddings_1_axis_0 = const()[name = string("position_embeddings_1_axis_0"), val = int32(0)]; |
| int32 position_embeddings_1_batch_dims_0 = const()[name = string("position_embeddings_1_batch_dims_0"), val = int32(0)]; |
| bool position_embeddings_1_validate_indices_0 = const()[name = string("position_embeddings_1_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [514, 768]> m_roberta_embeddings_position_embeddings_weight_to_fp16 = const()[name = string("m_roberta_embeddings_position_embeddings_weight_to_fp16"), val = tensor<fp16, [514, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62226624)))]; |
| string input_3_to_uint16_dtype_0 = const()[name = string("input_3_to_uint16_dtype_0"), val = string("uint16")]; |
| tensor<uint16, [1, 512]> input_3_to_uint16 = cast(dtype = input_3_to_uint16_dtype_0, x = input_3)[name = string("cast_54")]; |
| tensor<fp16, [1, 512, 768]> position_embeddings_1_cast_fp16_cast_uint16 = gather(axis = position_embeddings_1_axis_0, batch_dims = position_embeddings_1_batch_dims_0, indices = input_3_to_uint16, validate_indices = position_embeddings_1_validate_indices_0, x = m_roberta_embeddings_position_embeddings_weight_to_fp16)[name = string("position_embeddings_1_cast_fp16_cast_uint16")]; |
| tensor<fp16, [1, 512, 768]> input_5_cast_fp16 = add(x = embeddings_1_cast_fp16, y = position_embeddings_1_cast_fp16_cast_uint16)[name = string("input_5_cast_fp16")]; |
| tensor<int32, [1]> input_7_axes_0 = const()[name = string("input_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_embeddings_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63016192)))]; |
| tensor<fp16, [768]> m_roberta_embeddings_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63017792)))]; |
| fp16 var_10_to_fp16 = const()[name = string("op_10_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 512, 768]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = m_roberta_embeddings_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_embeddings_LayerNorm_weight_to_fp16, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; |
| tensor<int32, [1]> var_68_axes_0 = const()[name = string("op_68_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [1, 1, 512]> var_68 = expand_dims(axes = var_68_axes_0, x = attention_mask)[name = string("op_68")]; |
| tensor<int32, [1]> var_69_axes_0 = const()[name = string("op_69_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [1, 1, 1, 512]> var_69 = expand_dims(axes = var_69_axes_0, x = var_68)[name = string("op_69")]; |
| tensor<int32, [4]> var_72_reps_0 = const()[name = string("op_72_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])]; |
| tensor<int32, [1, 1, 512, 512]> var_72 = tile(reps = var_72_reps_0, x = var_69)[name = string("op_72")]; |
| fp16 const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = fp16(0x1p+0)]; |
| string expanded_mask_to_fp16_dtype_0 = const()[name = string("expanded_mask_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1, 1, 512, 512]> var_72_to_fp16 = cast(dtype = expanded_mask_to_fp16_dtype_0, x = var_72)[name = string("cast_53")]; |
| tensor<fp16, [1, 1, 512, 512]> inverted_mask_cast_fp16 = sub(x = const_4_to_fp16, y = var_72_to_fp16)[name = string("inverted_mask_cast_fp16")]; |
| string var_77_dtype_0 = const()[name = string("op_77_dtype_0"), val = string("bool")]; |
| fp16 var_21_to_fp16 = const()[name = string("op_21_to_fp16"), val = fp16(-inf)]; |
| tensor<bool, [1, 1, 512, 512]> inverted_mask_cast_fp16_to_bool = cast(dtype = var_77_dtype_0, x = inverted_mask_cast_fp16)[name = string("cast_52")]; |
| tensor<fp16, [1, 1, 512, 512]> attention_mask_cast_fp16 = select(a = var_21_to_fp16, b = inverted_mask_cast_fp16, cond = inverted_mask_cast_fp16_to_bool)[name = string("attention_mask_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_0_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63019392)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64199104)))]; |
| tensor<fp16, [1, 512, 768]> linear_0_cast_fp16 = linear(bias = m_roberta_encoder_layer_0_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_0_attention_self_query_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; |
| tensor<int32, [4]> var_116 = const()[name = string("op_116"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_117_cast_fp16 = reshape(shape = var_116, x = linear_0_cast_fp16)[name = string("op_117_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_0_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64200704)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65380416)))]; |
| tensor<fp16, [1, 512, 768]> linear_1_cast_fp16 = linear(bias = m_roberta_encoder_layer_0_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_0_attention_self_key_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_1_cast_fp16")]; |
| tensor<int32, [4]> var_122 = const()[name = string("op_122"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_123_cast_fp16 = reshape(shape = var_122, x = linear_1_cast_fp16)[name = string("op_123_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_0_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65382016)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66561728)))]; |
| tensor<fp16, [1, 512, 768]> linear_2_cast_fp16 = linear(bias = m_roberta_encoder_layer_0_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_0_attention_self_value_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; |
| tensor<int32, [4]> var_128 = const()[name = string("op_128"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_129_cast_fp16 = reshape(shape = var_128, x = linear_2_cast_fp16)[name = string("op_129_cast_fp16")]; |
| tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_74 = transpose(perm = transpose_74_perm_0, x = var_129_cast_fp16)[name = string("transpose_153")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_123_cast_fp16)[name = string("transpose_154")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_72 = transpose(perm = transpose_72_perm_0, x = var_117_cast_fp16)[name = string("transpose_155")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_1_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_73, query = transpose_72, value = transpose_74)[name = string("attn_output_1_cast_fp16")]; |
| tensor<int32, [4]> attn_output_3_perm_0 = const()[name = string("attn_output_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_133 = const()[name = string("op_133"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_3_cast_fp16 = transpose(perm = attn_output_3_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_152")]; |
| tensor<fp16, [1, 512, 768]> input_9_cast_fp16 = reshape(shape = var_133, x = attn_output_3_cast_fp16)[name = string("input_9_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_0_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66563328)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67743040)))]; |
| tensor<fp16, [1, 512, 768]> linear_3_cast_fp16 = linear(bias = m_roberta_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_3_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_13_cast_fp16 = add(x = linear_3_cast_fp16, y = input_7_cast_fp16)[name = string("input_13_cast_fp16")]; |
| tensor<int32, [1]> input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67744640)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67746240)))]; |
| tensor<fp16, [1, 512, 768]> input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = m_roberta_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_0_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67747840)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_0_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72466496)))]; |
| tensor<fp16, [1, 512, 3072]> linear_4_cast_fp16 = linear(bias = m_roberta_encoder_layer_0_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_0_intermediate_dense_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_4_cast_fp16")]; |
| string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_4_cast_fp16)[name = string("input_19_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_0_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72472704)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77191360)))]; |
| tensor<fp16, [1, 512, 768]> linear_5_cast_fp16 = linear(bias = m_roberta_encoder_layer_0_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_0_output_dense_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_5_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_23_cast_fp16 = add(x = linear_5_cast_fp16, y = input_15_cast_fp16)[name = string("input_23_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_7_axes_0 = const()[name = string("hidden_states_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77192960)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_0_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_0_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77194560)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_7_cast_fp16 = layer_norm(axes = hidden_states_7_axes_0, beta = m_roberta_encoder_layer_0_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_0_output_LayerNorm_weight_to_fp16, x = input_23_cast_fp16)[name = string("hidden_states_7_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_1_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77196160)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78375872)))]; |
| tensor<fp16, [1, 512, 768]> linear_6_cast_fp16 = linear(bias = m_roberta_encoder_layer_1_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_1_attention_self_query_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = string("linear_6_cast_fp16")]; |
| tensor<int32, [4]> var_175 = const()[name = string("op_175"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_176_cast_fp16 = reshape(shape = var_175, x = linear_6_cast_fp16)[name = string("op_176_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_1_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78377472)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79557184)))]; |
| tensor<fp16, [1, 512, 768]> linear_7_cast_fp16 = linear(bias = m_roberta_encoder_layer_1_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_1_attention_self_key_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = string("linear_7_cast_fp16")]; |
| tensor<int32, [4]> var_181 = const()[name = string("op_181"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_182_cast_fp16 = reshape(shape = var_181, x = linear_7_cast_fp16)[name = string("op_182_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_1_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79558784)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80738496)))]; |
| tensor<fp16, [1, 512, 768]> linear_8_cast_fp16 = linear(bias = m_roberta_encoder_layer_1_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_1_attention_self_value_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = string("linear_8_cast_fp16")]; |
| tensor<int32, [4]> var_187 = const()[name = string("op_187"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_188_cast_fp16 = reshape(shape = var_187, x = linear_8_cast_fp16)[name = string("op_188_cast_fp16")]; |
| tensor<int32, [4]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_188_cast_fp16)[name = string("transpose_149")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_76 = transpose(perm = transpose_76_perm_0, x = var_182_cast_fp16)[name = string("transpose_150")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_176_cast_fp16)[name = string("transpose_151")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_5_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_76, query = transpose_75, value = transpose_77)[name = string("attn_output_5_cast_fp16")]; |
| tensor<int32, [4]> attn_output_7_perm_0 = const()[name = string("attn_output_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_192 = const()[name = string("op_192"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_7_cast_fp16 = transpose(perm = attn_output_7_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_148")]; |
| tensor<fp16, [1, 512, 768]> input_25_cast_fp16 = reshape(shape = var_192, x = attn_output_7_cast_fp16)[name = string("input_25_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_1_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80740096)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81919808)))]; |
| tensor<fp16, [1, 512, 768]> linear_9_cast_fp16 = linear(bias = m_roberta_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_9_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_29_cast_fp16 = add(x = linear_9_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("input_29_cast_fp16")]; |
| tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81921408)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81923008)))]; |
| tensor<fp16, [1, 512, 768]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = m_roberta_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_1_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81924608)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_1_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86643264)))]; |
| tensor<fp16, [1, 512, 3072]> linear_10_cast_fp16 = linear(bias = m_roberta_encoder_layer_1_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_1_intermediate_dense_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_10_cast_fp16")]; |
| string input_35_mode_0 = const()[name = string("input_35_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = string("input_35_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_1_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86649472)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91368128)))]; |
| tensor<fp16, [1, 512, 768]> linear_11_cast_fp16 = linear(bias = m_roberta_encoder_layer_1_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_1_output_dense_weight_to_fp16, x = input_35_cast_fp16)[name = string("linear_11_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_39_cast_fp16 = add(x = linear_11_cast_fp16, y = input_31_cast_fp16)[name = string("input_39_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_13_axes_0 = const()[name = string("hidden_states_13_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91369728)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_1_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_1_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91371328)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_13_cast_fp16 = layer_norm(axes = hidden_states_13_axes_0, beta = m_roberta_encoder_layer_1_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_1_output_LayerNorm_weight_to_fp16, x = input_39_cast_fp16)[name = string("hidden_states_13_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_2_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91372928)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92552640)))]; |
| tensor<fp16, [1, 512, 768]> linear_12_cast_fp16 = linear(bias = m_roberta_encoder_layer_2_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_2_attention_self_query_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = string("linear_12_cast_fp16")]; |
| tensor<int32, [4]> var_234 = const()[name = string("op_234"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_235_cast_fp16 = reshape(shape = var_234, x = linear_12_cast_fp16)[name = string("op_235_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_2_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92554240)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93733952)))]; |
| tensor<fp16, [1, 512, 768]> linear_13_cast_fp16 = linear(bias = m_roberta_encoder_layer_2_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_2_attention_self_key_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = string("linear_13_cast_fp16")]; |
| tensor<int32, [4]> var_240 = const()[name = string("op_240"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_241_cast_fp16 = reshape(shape = var_240, x = linear_13_cast_fp16)[name = string("op_241_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_2_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93735552)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94915264)))]; |
| tensor<fp16, [1, 512, 768]> linear_14_cast_fp16 = linear(bias = m_roberta_encoder_layer_2_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_2_attention_self_value_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = string("linear_14_cast_fp16")]; |
| tensor<int32, [4]> var_246 = const()[name = string("op_246"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_247_cast_fp16 = reshape(shape = var_246, x = linear_14_cast_fp16)[name = string("op_247_cast_fp16")]; |
| tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_80 = transpose(perm = transpose_80_perm_0, x = var_247_cast_fp16)[name = string("transpose_145")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_241_cast_fp16)[name = string("transpose_146")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_78 = transpose(perm = transpose_78_perm_0, x = var_235_cast_fp16)[name = string("transpose_147")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_9_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_79, query = transpose_78, value = transpose_80)[name = string("attn_output_9_cast_fp16")]; |
| tensor<int32, [4]> attn_output_11_perm_0 = const()[name = string("attn_output_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_251 = const()[name = string("op_251"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_11_cast_fp16 = transpose(perm = attn_output_11_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_144")]; |
| tensor<fp16, [1, 512, 768]> input_41_cast_fp16 = reshape(shape = var_251, x = attn_output_11_cast_fp16)[name = string("input_41_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_2_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94916864)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96096576)))]; |
| tensor<fp16, [1, 512, 768]> linear_15_cast_fp16 = linear(bias = m_roberta_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_15_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_45_cast_fp16 = add(x = linear_15_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("input_45_cast_fp16")]; |
| tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96098176)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96099776)))]; |
| tensor<fp16, [1, 512, 768]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = m_roberta_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_2_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96101376)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_2_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100820032)))]; |
| tensor<fp16, [1, 512, 3072]> linear_16_cast_fp16 = linear(bias = m_roberta_encoder_layer_2_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_2_intermediate_dense_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_16_cast_fp16")]; |
| string input_51_mode_0 = const()[name = string("input_51_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_16_cast_fp16)[name = string("input_51_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_2_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100826240)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105544896)))]; |
| tensor<fp16, [1, 512, 768]> linear_17_cast_fp16 = linear(bias = m_roberta_encoder_layer_2_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_2_output_dense_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_17_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_55_cast_fp16 = add(x = linear_17_cast_fp16, y = input_47_cast_fp16)[name = string("input_55_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_19_axes_0 = const()[name = string("hidden_states_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105546496)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_2_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_2_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105548096)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_19_cast_fp16 = layer_norm(axes = hidden_states_19_axes_0, beta = m_roberta_encoder_layer_2_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_2_output_LayerNorm_weight_to_fp16, x = input_55_cast_fp16)[name = string("hidden_states_19_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_3_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105549696)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106729408)))]; |
| tensor<fp16, [1, 512, 768]> linear_18_cast_fp16 = linear(bias = m_roberta_encoder_layer_3_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_3_attention_self_query_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = string("linear_18_cast_fp16")]; |
| tensor<int32, [4]> var_293 = const()[name = string("op_293"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_294_cast_fp16 = reshape(shape = var_293, x = linear_18_cast_fp16)[name = string("op_294_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_3_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106731008)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107910720)))]; |
| tensor<fp16, [1, 512, 768]> linear_19_cast_fp16 = linear(bias = m_roberta_encoder_layer_3_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_3_attention_self_key_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = string("linear_19_cast_fp16")]; |
| tensor<int32, [4]> var_299 = const()[name = string("op_299"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_300_cast_fp16 = reshape(shape = var_299, x = linear_19_cast_fp16)[name = string("op_300_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_3_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107912320)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109092032)))]; |
| tensor<fp16, [1, 512, 768]> linear_20_cast_fp16 = linear(bias = m_roberta_encoder_layer_3_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_3_attention_self_value_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = string("linear_20_cast_fp16")]; |
| tensor<int32, [4]> var_305 = const()[name = string("op_305"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_306_cast_fp16 = reshape(shape = var_305, x = linear_20_cast_fp16)[name = string("op_306_cast_fp16")]; |
| tensor<int32, [4]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_306_cast_fp16)[name = string("transpose_141")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_82 = transpose(perm = transpose_82_perm_0, x = var_300_cast_fp16)[name = string("transpose_142")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_294_cast_fp16)[name = string("transpose_143")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_13_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_82, query = transpose_81, value = transpose_83)[name = string("attn_output_13_cast_fp16")]; |
| tensor<int32, [4]> attn_output_15_perm_0 = const()[name = string("attn_output_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_310 = const()[name = string("op_310"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_15_cast_fp16 = transpose(perm = attn_output_15_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_140")]; |
| tensor<fp16, [1, 512, 768]> input_57_cast_fp16 = reshape(shape = var_310, x = attn_output_15_cast_fp16)[name = string("input_57_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_3_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109093632)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110273344)))]; |
| tensor<fp16, [1, 512, 768]> linear_21_cast_fp16 = linear(bias = m_roberta_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_21_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_61_cast_fp16 = add(x = linear_21_cast_fp16, y = hidden_states_19_cast_fp16)[name = string("input_61_cast_fp16")]; |
| tensor<int32, [1]> input_63_axes_0 = const()[name = string("input_63_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110274944)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110276544)))]; |
| tensor<fp16, [1, 512, 768]> input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = m_roberta_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_3_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110278144)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_3_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114996800)))]; |
| tensor<fp16, [1, 512, 3072]> linear_22_cast_fp16 = linear(bias = m_roberta_encoder_layer_3_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_3_intermediate_dense_weight_to_fp16, x = input_63_cast_fp16)[name = string("linear_22_cast_fp16")]; |
| string input_67_mode_0 = const()[name = string("input_67_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_67_cast_fp16 = gelu(mode = input_67_mode_0, x = linear_22_cast_fp16)[name = string("input_67_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_3_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115003008)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119721664)))]; |
| tensor<fp16, [1, 512, 768]> linear_23_cast_fp16 = linear(bias = m_roberta_encoder_layer_3_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_3_output_dense_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_23_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_71_cast_fp16 = add(x = linear_23_cast_fp16, y = input_63_cast_fp16)[name = string("input_71_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_25_axes_0 = const()[name = string("hidden_states_25_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119723264)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_3_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_3_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119724864)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_25_cast_fp16 = layer_norm(axes = hidden_states_25_axes_0, beta = m_roberta_encoder_layer_3_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_3_output_LayerNorm_weight_to_fp16, x = input_71_cast_fp16)[name = string("hidden_states_25_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_4_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119726464)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120906176)))]; |
| tensor<fp16, [1, 512, 768]> linear_24_cast_fp16 = linear(bias = m_roberta_encoder_layer_4_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_4_attention_self_query_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = string("linear_24_cast_fp16")]; |
| tensor<int32, [4]> var_352 = const()[name = string("op_352"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_353_cast_fp16 = reshape(shape = var_352, x = linear_24_cast_fp16)[name = string("op_353_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_4_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120907776)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122087488)))]; |
| tensor<fp16, [1, 512, 768]> linear_25_cast_fp16 = linear(bias = m_roberta_encoder_layer_4_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_4_attention_self_key_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = string("linear_25_cast_fp16")]; |
| tensor<int32, [4]> var_358 = const()[name = string("op_358"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_359_cast_fp16 = reshape(shape = var_358, x = linear_25_cast_fp16)[name = string("op_359_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_4_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122089088)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123268800)))]; |
| tensor<fp16, [1, 512, 768]> linear_26_cast_fp16 = linear(bias = m_roberta_encoder_layer_4_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_4_attention_self_value_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = string("linear_26_cast_fp16")]; |
| tensor<int32, [4]> var_364 = const()[name = string("op_364"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_365_cast_fp16 = reshape(shape = var_364, x = linear_26_cast_fp16)[name = string("op_365_cast_fp16")]; |
| tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_86 = transpose(perm = transpose_86_perm_0, x = var_365_cast_fp16)[name = string("transpose_137")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_85 = transpose(perm = transpose_85_perm_0, x = var_359_cast_fp16)[name = string("transpose_138")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_84 = transpose(perm = transpose_84_perm_0, x = var_353_cast_fp16)[name = string("transpose_139")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_17_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_85, query = transpose_84, value = transpose_86)[name = string("attn_output_17_cast_fp16")]; |
| tensor<int32, [4]> attn_output_19_perm_0 = const()[name = string("attn_output_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_369 = const()[name = string("op_369"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_19_cast_fp16 = transpose(perm = attn_output_19_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_136")]; |
| tensor<fp16, [1, 512, 768]> input_73_cast_fp16 = reshape(shape = var_369, x = attn_output_19_cast_fp16)[name = string("input_73_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_4_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123270400)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124450112)))]; |
| tensor<fp16, [1, 512, 768]> linear_27_cast_fp16 = linear(bias = m_roberta_encoder_layer_4_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_4_attention_output_dense_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_27_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_77_cast_fp16 = add(x = linear_27_cast_fp16, y = hidden_states_25_cast_fp16)[name = string("input_77_cast_fp16")]; |
| tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124451712)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124453312)))]; |
| tensor<fp16, [1, 512, 768]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = m_roberta_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_4_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124454912)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_4_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129173568)))]; |
| tensor<fp16, [1, 512, 3072]> linear_28_cast_fp16 = linear(bias = m_roberta_encoder_layer_4_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_4_intermediate_dense_weight_to_fp16, x = input_79_cast_fp16)[name = string("linear_28_cast_fp16")]; |
| string input_83_mode_0 = const()[name = string("input_83_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_28_cast_fp16)[name = string("input_83_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_4_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129179776)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133898432)))]; |
| tensor<fp16, [1, 512, 768]> linear_29_cast_fp16 = linear(bias = m_roberta_encoder_layer_4_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_4_output_dense_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_29_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_87_cast_fp16 = add(x = linear_29_cast_fp16, y = input_79_cast_fp16)[name = string("input_87_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_31_axes_0 = const()[name = string("hidden_states_31_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133900032)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_4_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_4_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133901632)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_31_cast_fp16 = layer_norm(axes = hidden_states_31_axes_0, beta = m_roberta_encoder_layer_4_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_4_output_LayerNorm_weight_to_fp16, x = input_87_cast_fp16)[name = string("hidden_states_31_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_5_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133903232)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135082944)))]; |
| tensor<fp16, [1, 512, 768]> linear_30_cast_fp16 = linear(bias = m_roberta_encoder_layer_5_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_5_attention_self_query_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = string("linear_30_cast_fp16")]; |
| tensor<int32, [4]> var_411 = const()[name = string("op_411"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_412_cast_fp16 = reshape(shape = var_411, x = linear_30_cast_fp16)[name = string("op_412_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_5_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135084544)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136264256)))]; |
| tensor<fp16, [1, 512, 768]> linear_31_cast_fp16 = linear(bias = m_roberta_encoder_layer_5_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_5_attention_self_key_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = string("linear_31_cast_fp16")]; |
| tensor<int32, [4]> var_417 = const()[name = string("op_417"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_418_cast_fp16 = reshape(shape = var_417, x = linear_31_cast_fp16)[name = string("op_418_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_5_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136265856)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137445568)))]; |
| tensor<fp16, [1, 512, 768]> linear_32_cast_fp16 = linear(bias = m_roberta_encoder_layer_5_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_5_attention_self_value_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = string("linear_32_cast_fp16")]; |
| tensor<int32, [4]> var_423 = const()[name = string("op_423"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_424_cast_fp16 = reshape(shape = var_423, x = linear_32_cast_fp16)[name = string("op_424_cast_fp16")]; |
| tensor<int32, [4]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_89 = transpose(perm = transpose_89_perm_0, x = var_424_cast_fp16)[name = string("transpose_133")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_88 = transpose(perm = transpose_88_perm_0, x = var_418_cast_fp16)[name = string("transpose_134")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_87 = transpose(perm = transpose_87_perm_0, x = var_412_cast_fp16)[name = string("transpose_135")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_21_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_88, query = transpose_87, value = transpose_89)[name = string("attn_output_21_cast_fp16")]; |
| tensor<int32, [4]> attn_output_23_perm_0 = const()[name = string("attn_output_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_428 = const()[name = string("op_428"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_23_cast_fp16 = transpose(perm = attn_output_23_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_132")]; |
| tensor<fp16, [1, 512, 768]> input_89_cast_fp16 = reshape(shape = var_428, x = attn_output_23_cast_fp16)[name = string("input_89_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_5_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137447168)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138626880)))]; |
| tensor<fp16, [1, 512, 768]> linear_33_cast_fp16 = linear(bias = m_roberta_encoder_layer_5_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_5_attention_output_dense_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_33_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_93_cast_fp16 = add(x = linear_33_cast_fp16, y = hidden_states_31_cast_fp16)[name = string("input_93_cast_fp16")]; |
| tensor<int32, [1]> input_95_axes_0 = const()[name = string("input_95_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138628480)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138630080)))]; |
| tensor<fp16, [1, 512, 768]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = m_roberta_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16, x = input_93_cast_fp16)[name = string("input_95_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_5_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138631680)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_5_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143350336)))]; |
| tensor<fp16, [1, 512, 3072]> linear_34_cast_fp16 = linear(bias = m_roberta_encoder_layer_5_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_5_intermediate_dense_weight_to_fp16, x = input_95_cast_fp16)[name = string("linear_34_cast_fp16")]; |
| string input_99_mode_0 = const()[name = string("input_99_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_34_cast_fp16)[name = string("input_99_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_5_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143356544)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148075200)))]; |
| tensor<fp16, [1, 512, 768]> linear_35_cast_fp16 = linear(bias = m_roberta_encoder_layer_5_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_5_output_dense_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_35_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_103_cast_fp16 = add(x = linear_35_cast_fp16, y = input_95_cast_fp16)[name = string("input_103_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_37_axes_0 = const()[name = string("hidden_states_37_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148076800)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_5_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_5_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148078400)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_37_cast_fp16 = layer_norm(axes = hidden_states_37_axes_0, beta = m_roberta_encoder_layer_5_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_5_output_LayerNorm_weight_to_fp16, x = input_103_cast_fp16)[name = string("hidden_states_37_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_6_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148080000)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149259712)))]; |
| tensor<fp16, [1, 512, 768]> linear_36_cast_fp16 = linear(bias = m_roberta_encoder_layer_6_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_6_attention_self_query_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = string("linear_36_cast_fp16")]; |
| tensor<int32, [4]> var_470 = const()[name = string("op_470"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_471_cast_fp16 = reshape(shape = var_470, x = linear_36_cast_fp16)[name = string("op_471_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_6_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149261312)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150441024)))]; |
| tensor<fp16, [1, 512, 768]> linear_37_cast_fp16 = linear(bias = m_roberta_encoder_layer_6_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_6_attention_self_key_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = string("linear_37_cast_fp16")]; |
| tensor<int32, [4]> var_476 = const()[name = string("op_476"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_477_cast_fp16 = reshape(shape = var_476, x = linear_37_cast_fp16)[name = string("op_477_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_6_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150442624)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151622336)))]; |
| tensor<fp16, [1, 512, 768]> linear_38_cast_fp16 = linear(bias = m_roberta_encoder_layer_6_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_6_attention_self_value_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = string("linear_38_cast_fp16")]; |
| tensor<int32, [4]> var_482 = const()[name = string("op_482"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_483_cast_fp16 = reshape(shape = var_482, x = linear_38_cast_fp16)[name = string("op_483_cast_fp16")]; |
| tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_92 = transpose(perm = transpose_92_perm_0, x = var_483_cast_fp16)[name = string("transpose_129")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_91 = transpose(perm = transpose_91_perm_0, x = var_477_cast_fp16)[name = string("transpose_130")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_90 = transpose(perm = transpose_90_perm_0, x = var_471_cast_fp16)[name = string("transpose_131")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_25_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_91, query = transpose_90, value = transpose_92)[name = string("attn_output_25_cast_fp16")]; |
| tensor<int32, [4]> attn_output_27_perm_0 = const()[name = string("attn_output_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_487 = const()[name = string("op_487"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_27_cast_fp16 = transpose(perm = attn_output_27_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_128")]; |
| tensor<fp16, [1, 512, 768]> input_105_cast_fp16 = reshape(shape = var_487, x = attn_output_27_cast_fp16)[name = string("input_105_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_6_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151623936)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152803648)))]; |
| tensor<fp16, [1, 512, 768]> linear_39_cast_fp16 = linear(bias = m_roberta_encoder_layer_6_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_6_attention_output_dense_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_39_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_109_cast_fp16 = add(x = linear_39_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("input_109_cast_fp16")]; |
| tensor<int32, [1]> input_111_axes_0 = const()[name = string("input_111_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152805248)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152806848)))]; |
| tensor<fp16, [1, 512, 768]> input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = m_roberta_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16, x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_6_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152808448)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_6_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157527104)))]; |
| tensor<fp16, [1, 512, 3072]> linear_40_cast_fp16 = linear(bias = m_roberta_encoder_layer_6_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_6_intermediate_dense_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_40_cast_fp16")]; |
| string input_115_mode_0 = const()[name = string("input_115_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_115_cast_fp16 = gelu(mode = input_115_mode_0, x = linear_40_cast_fp16)[name = string("input_115_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_6_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157533312)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162251968)))]; |
| tensor<fp16, [1, 512, 768]> linear_41_cast_fp16 = linear(bias = m_roberta_encoder_layer_6_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_6_output_dense_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_41_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_119_cast_fp16 = add(x = linear_41_cast_fp16, y = input_111_cast_fp16)[name = string("input_119_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_43_axes_0 = const()[name = string("hidden_states_43_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162253568)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_6_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_6_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162255168)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_43_cast_fp16 = layer_norm(axes = hidden_states_43_axes_0, beta = m_roberta_encoder_layer_6_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_6_output_LayerNorm_weight_to_fp16, x = input_119_cast_fp16)[name = string("hidden_states_43_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_7_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162256768)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163436480)))]; |
| tensor<fp16, [1, 512, 768]> linear_42_cast_fp16 = linear(bias = m_roberta_encoder_layer_7_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_7_attention_self_query_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = string("linear_42_cast_fp16")]; |
| tensor<int32, [4]> var_529 = const()[name = string("op_529"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_530_cast_fp16 = reshape(shape = var_529, x = linear_42_cast_fp16)[name = string("op_530_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_7_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163438080)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164617792)))]; |
| tensor<fp16, [1, 512, 768]> linear_43_cast_fp16 = linear(bias = m_roberta_encoder_layer_7_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_7_attention_self_key_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = string("linear_43_cast_fp16")]; |
| tensor<int32, [4]> var_535 = const()[name = string("op_535"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_536_cast_fp16 = reshape(shape = var_535, x = linear_43_cast_fp16)[name = string("op_536_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_7_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164619392)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165799104)))]; |
| tensor<fp16, [1, 512, 768]> linear_44_cast_fp16 = linear(bias = m_roberta_encoder_layer_7_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_7_attention_self_value_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = string("linear_44_cast_fp16")]; |
| tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_542_cast_fp16 = reshape(shape = var_541, x = linear_44_cast_fp16)[name = string("op_542_cast_fp16")]; |
| tensor<int32, [4]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_95 = transpose(perm = transpose_95_perm_0, x = var_542_cast_fp16)[name = string("transpose_125")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_94 = transpose(perm = transpose_94_perm_0, x = var_536_cast_fp16)[name = string("transpose_126")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_93 = transpose(perm = transpose_93_perm_0, x = var_530_cast_fp16)[name = string("transpose_127")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_29_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_94, query = transpose_93, value = transpose_95)[name = string("attn_output_29_cast_fp16")]; |
| tensor<int32, [4]> attn_output_31_perm_0 = const()[name = string("attn_output_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_31_cast_fp16 = transpose(perm = attn_output_31_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_124")]; |
| tensor<fp16, [1, 512, 768]> input_121_cast_fp16 = reshape(shape = var_546, x = attn_output_31_cast_fp16)[name = string("input_121_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_7_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165800704)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166980416)))]; |
| tensor<fp16, [1, 512, 768]> linear_45_cast_fp16 = linear(bias = m_roberta_encoder_layer_7_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_7_attention_output_dense_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_45_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_125_cast_fp16 = add(x = linear_45_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("input_125_cast_fp16")]; |
| tensor<int32, [1]> input_127_axes_0 = const()[name = string("input_127_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166982016)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166983616)))]; |
| tensor<fp16, [1, 512, 768]> input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = m_roberta_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16, x = input_125_cast_fp16)[name = string("input_127_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_7_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166985216)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_7_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171703872)))]; |
| tensor<fp16, [1, 512, 3072]> linear_46_cast_fp16 = linear(bias = m_roberta_encoder_layer_7_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_7_intermediate_dense_weight_to_fp16, x = input_127_cast_fp16)[name = string("linear_46_cast_fp16")]; |
| string input_131_mode_0 = const()[name = string("input_131_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_46_cast_fp16)[name = string("input_131_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_7_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171710080)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176428736)))]; |
| tensor<fp16, [1, 512, 768]> linear_47_cast_fp16 = linear(bias = m_roberta_encoder_layer_7_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_7_output_dense_weight_to_fp16, x = input_131_cast_fp16)[name = string("linear_47_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_135_cast_fp16 = add(x = linear_47_cast_fp16, y = input_127_cast_fp16)[name = string("input_135_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_49_axes_0 = const()[name = string("hidden_states_49_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176430336)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_7_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_7_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176431936)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_49_cast_fp16 = layer_norm(axes = hidden_states_49_axes_0, beta = m_roberta_encoder_layer_7_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_7_output_LayerNorm_weight_to_fp16, x = input_135_cast_fp16)[name = string("hidden_states_49_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_8_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176433536)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177613248)))]; |
| tensor<fp16, [1, 512, 768]> linear_48_cast_fp16 = linear(bias = m_roberta_encoder_layer_8_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_8_attention_self_query_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = string("linear_48_cast_fp16")]; |
| tensor<int32, [4]> var_588 = const()[name = string("op_588"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_589_cast_fp16 = reshape(shape = var_588, x = linear_48_cast_fp16)[name = string("op_589_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_8_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177614848)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178794560)))]; |
| tensor<fp16, [1, 512, 768]> linear_49_cast_fp16 = linear(bias = m_roberta_encoder_layer_8_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_8_attention_self_key_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = string("linear_49_cast_fp16")]; |
| tensor<int32, [4]> var_594 = const()[name = string("op_594"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_595_cast_fp16 = reshape(shape = var_594, x = linear_49_cast_fp16)[name = string("op_595_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_8_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178796160)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179975872)))]; |
| tensor<fp16, [1, 512, 768]> linear_50_cast_fp16 = linear(bias = m_roberta_encoder_layer_8_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_8_attention_self_value_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = string("linear_50_cast_fp16")]; |
| tensor<int32, [4]> var_600 = const()[name = string("op_600"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_601_cast_fp16 = reshape(shape = var_600, x = linear_50_cast_fp16)[name = string("op_601_cast_fp16")]; |
| tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_601_cast_fp16)[name = string("transpose_121")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_97 = transpose(perm = transpose_97_perm_0, x = var_595_cast_fp16)[name = string("transpose_122")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_589_cast_fp16)[name = string("transpose_123")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_33_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_97, query = transpose_96, value = transpose_98)[name = string("attn_output_33_cast_fp16")]; |
| tensor<int32, [4]> attn_output_35_perm_0 = const()[name = string("attn_output_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_605 = const()[name = string("op_605"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_35_cast_fp16 = transpose(perm = attn_output_35_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_120")]; |
| tensor<fp16, [1, 512, 768]> input_137_cast_fp16 = reshape(shape = var_605, x = attn_output_35_cast_fp16)[name = string("input_137_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_8_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179977472)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181157184)))]; |
| tensor<fp16, [1, 512, 768]> linear_51_cast_fp16 = linear(bias = m_roberta_encoder_layer_8_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_8_attention_output_dense_weight_to_fp16, x = input_137_cast_fp16)[name = string("linear_51_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_141_cast_fp16 = add(x = linear_51_cast_fp16, y = hidden_states_49_cast_fp16)[name = string("input_141_cast_fp16")]; |
| tensor<int32, [1]> input_143_axes_0 = const()[name = string("input_143_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181158784)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181160384)))]; |
| tensor<fp16, [1, 512, 768]> input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = m_roberta_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16, x = input_141_cast_fp16)[name = string("input_143_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_8_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181161984)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_8_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185880640)))]; |
| tensor<fp16, [1, 512, 3072]> linear_52_cast_fp16 = linear(bias = m_roberta_encoder_layer_8_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_8_intermediate_dense_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_52_cast_fp16")]; |
| string input_147_mode_0 = const()[name = string("input_147_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_147_cast_fp16 = gelu(mode = input_147_mode_0, x = linear_52_cast_fp16)[name = string("input_147_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_8_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185886848)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190605504)))]; |
| tensor<fp16, [1, 512, 768]> linear_53_cast_fp16 = linear(bias = m_roberta_encoder_layer_8_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_8_output_dense_weight_to_fp16, x = input_147_cast_fp16)[name = string("linear_53_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_151_cast_fp16 = add(x = linear_53_cast_fp16, y = input_143_cast_fp16)[name = string("input_151_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_55_axes_0 = const()[name = string("hidden_states_55_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190607104)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_8_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_8_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190608704)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_55_cast_fp16 = layer_norm(axes = hidden_states_55_axes_0, beta = m_roberta_encoder_layer_8_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_8_output_LayerNorm_weight_to_fp16, x = input_151_cast_fp16)[name = string("hidden_states_55_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_9_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190610304)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191790016)))]; |
| tensor<fp16, [1, 512, 768]> linear_54_cast_fp16 = linear(bias = m_roberta_encoder_layer_9_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_9_attention_self_query_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = string("linear_54_cast_fp16")]; |
| tensor<int32, [4]> var_647 = const()[name = string("op_647"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_648_cast_fp16 = reshape(shape = var_647, x = linear_54_cast_fp16)[name = string("op_648_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_9_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191791616)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192971328)))]; |
| tensor<fp16, [1, 512, 768]> linear_55_cast_fp16 = linear(bias = m_roberta_encoder_layer_9_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_9_attention_self_key_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = string("linear_55_cast_fp16")]; |
| tensor<int32, [4]> var_653 = const()[name = string("op_653"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_654_cast_fp16 = reshape(shape = var_653, x = linear_55_cast_fp16)[name = string("op_654_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_9_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192972928)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194152640)))]; |
| tensor<fp16, [1, 512, 768]> linear_56_cast_fp16 = linear(bias = m_roberta_encoder_layer_9_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_9_attention_self_value_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = string("linear_56_cast_fp16")]; |
| tensor<int32, [4]> var_659 = const()[name = string("op_659"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_660_cast_fp16 = reshape(shape = var_659, x = linear_56_cast_fp16)[name = string("op_660_cast_fp16")]; |
| tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_101 = transpose(perm = transpose_101_perm_0, x = var_660_cast_fp16)[name = string("transpose_117")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_654_cast_fp16)[name = string("transpose_118")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_99 = transpose(perm = transpose_99_perm_0, x = var_648_cast_fp16)[name = string("transpose_119")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_37_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_100, query = transpose_99, value = transpose_101)[name = string("attn_output_37_cast_fp16")]; |
| tensor<int32, [4]> attn_output_39_perm_0 = const()[name = string("attn_output_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_664 = const()[name = string("op_664"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_39_cast_fp16 = transpose(perm = attn_output_39_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_116")]; |
| tensor<fp16, [1, 512, 768]> input_153_cast_fp16 = reshape(shape = var_664, x = attn_output_39_cast_fp16)[name = string("input_153_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_9_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194154240)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195333952)))]; |
| tensor<fp16, [1, 512, 768]> linear_57_cast_fp16 = linear(bias = m_roberta_encoder_layer_9_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_9_attention_output_dense_weight_to_fp16, x = input_153_cast_fp16)[name = string("linear_57_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_157_cast_fp16 = add(x = linear_57_cast_fp16, y = hidden_states_55_cast_fp16)[name = string("input_157_cast_fp16")]; |
| tensor<int32, [1]> input_159_axes_0 = const()[name = string("input_159_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195335552)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195337152)))]; |
| tensor<fp16, [1, 512, 768]> input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = m_roberta_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16, x = input_157_cast_fp16)[name = string("input_159_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_9_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195338752)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_9_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200057408)))]; |
| tensor<fp16, [1, 512, 3072]> linear_58_cast_fp16 = linear(bias = m_roberta_encoder_layer_9_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_9_intermediate_dense_weight_to_fp16, x = input_159_cast_fp16)[name = string("linear_58_cast_fp16")]; |
| string input_163_mode_0 = const()[name = string("input_163_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_163_cast_fp16 = gelu(mode = input_163_mode_0, x = linear_58_cast_fp16)[name = string("input_163_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_9_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200063616)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204782272)))]; |
| tensor<fp16, [1, 512, 768]> linear_59_cast_fp16 = linear(bias = m_roberta_encoder_layer_9_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_9_output_dense_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_59_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_167_cast_fp16 = add(x = linear_59_cast_fp16, y = input_159_cast_fp16)[name = string("input_167_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_61_axes_0 = const()[name = string("hidden_states_61_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204783872)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_9_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_9_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204785472)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_61_cast_fp16 = layer_norm(axes = hidden_states_61_axes_0, beta = m_roberta_encoder_layer_9_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_9_output_LayerNorm_weight_to_fp16, x = input_167_cast_fp16)[name = string("hidden_states_61_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_10_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204787072)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205966784)))]; |
| tensor<fp16, [1, 512, 768]> linear_60_cast_fp16 = linear(bias = m_roberta_encoder_layer_10_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_10_attention_self_query_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = string("linear_60_cast_fp16")]; |
| tensor<int32, [4]> var_706 = const()[name = string("op_706"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_707_cast_fp16 = reshape(shape = var_706, x = linear_60_cast_fp16)[name = string("op_707_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_10_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205968384)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207148096)))]; |
| tensor<fp16, [1, 512, 768]> linear_61_cast_fp16 = linear(bias = m_roberta_encoder_layer_10_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_10_attention_self_key_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = string("linear_61_cast_fp16")]; |
| tensor<int32, [4]> var_712 = const()[name = string("op_712"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_713_cast_fp16 = reshape(shape = var_712, x = linear_61_cast_fp16)[name = string("op_713_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_10_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207149696)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208329408)))]; |
| tensor<fp16, [1, 512, 768]> linear_62_cast_fp16 = linear(bias = m_roberta_encoder_layer_10_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_10_attention_self_value_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = string("linear_62_cast_fp16")]; |
| tensor<int32, [4]> var_718 = const()[name = string("op_718"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_719_cast_fp16 = reshape(shape = var_718, x = linear_62_cast_fp16)[name = string("op_719_cast_fp16")]; |
| tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_719_cast_fp16)[name = string("transpose_113")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_103 = transpose(perm = transpose_103_perm_0, x = var_713_cast_fp16)[name = string("transpose_114")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_707_cast_fp16)[name = string("transpose_115")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_41_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_103, query = transpose_102, value = transpose_104)[name = string("attn_output_41_cast_fp16")]; |
| tensor<int32, [4]> attn_output_43_perm_0 = const()[name = string("attn_output_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_723 = const()[name = string("op_723"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_43_cast_fp16 = transpose(perm = attn_output_43_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_112")]; |
| tensor<fp16, [1, 512, 768]> input_169_cast_fp16 = reshape(shape = var_723, x = attn_output_43_cast_fp16)[name = string("input_169_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_10_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208331008)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209510720)))]; |
| tensor<fp16, [1, 512, 768]> linear_63_cast_fp16 = linear(bias = m_roberta_encoder_layer_10_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_10_attention_output_dense_weight_to_fp16, x = input_169_cast_fp16)[name = string("linear_63_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_173_cast_fp16 = add(x = linear_63_cast_fp16, y = hidden_states_61_cast_fp16)[name = string("input_173_cast_fp16")]; |
| tensor<int32, [1]> input_175_axes_0 = const()[name = string("input_175_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209512320)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209513920)))]; |
| tensor<fp16, [1, 512, 768]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, beta = m_roberta_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16, x = input_173_cast_fp16)[name = string("input_175_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_10_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209515520)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_10_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214234176)))]; |
| tensor<fp16, [1, 512, 3072]> linear_64_cast_fp16 = linear(bias = m_roberta_encoder_layer_10_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_10_intermediate_dense_weight_to_fp16, x = input_175_cast_fp16)[name = string("linear_64_cast_fp16")]; |
| string input_179_mode_0 = const()[name = string("input_179_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_64_cast_fp16)[name = string("input_179_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_10_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214240384)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218959040)))]; |
| tensor<fp16, [1, 512, 768]> linear_65_cast_fp16 = linear(bias = m_roberta_encoder_layer_10_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_10_output_dense_weight_to_fp16, x = input_179_cast_fp16)[name = string("linear_65_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_183_cast_fp16 = add(x = linear_65_cast_fp16, y = input_175_cast_fp16)[name = string("input_183_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_67_axes_0 = const()[name = string("hidden_states_67_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218960640)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_10_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_10_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218962240)))]; |
| tensor<fp16, [1, 512, 768]> hidden_states_67_cast_fp16 = layer_norm(axes = hidden_states_67_axes_0, beta = m_roberta_encoder_layer_10_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_10_output_LayerNorm_weight_to_fp16, x = input_183_cast_fp16)[name = string("hidden_states_67_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_11_attention_self_query_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_self_query_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218963840)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_attention_self_query_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_self_query_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220143552)))]; |
| tensor<fp16, [1, 512, 768]> linear_66_cast_fp16 = linear(bias = m_roberta_encoder_layer_11_attention_self_query_bias_to_fp16, weight = m_roberta_encoder_layer_11_attention_self_query_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = string("linear_66_cast_fp16")]; |
| tensor<int32, [4]> var_765 = const()[name = string("op_765"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_766_cast_fp16 = reshape(shape = var_765, x = linear_66_cast_fp16)[name = string("op_766_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_11_attention_self_key_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_self_key_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220145152)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_attention_self_key_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_self_key_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221324864)))]; |
| tensor<fp16, [1, 512, 768]> linear_67_cast_fp16 = linear(bias = m_roberta_encoder_layer_11_attention_self_key_bias_to_fp16, weight = m_roberta_encoder_layer_11_attention_self_key_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = string("linear_67_cast_fp16")]; |
| tensor<int32, [4]> var_771 = const()[name = string("op_771"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_772_cast_fp16 = reshape(shape = var_771, x = linear_67_cast_fp16)[name = string("op_772_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_11_attention_self_value_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_self_value_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221326464)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_attention_self_value_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_self_value_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222506176)))]; |
| tensor<fp16, [1, 512, 768]> linear_68_cast_fp16 = linear(bias = m_roberta_encoder_layer_11_attention_self_value_bias_to_fp16, weight = m_roberta_encoder_layer_11_attention_self_value_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = string("linear_68_cast_fp16")]; |
| tensor<int32, [4]> var_777 = const()[name = string("op_777"), val = tensor<int32, [4]>([1, -1, 12, 64])]; |
| tensor<fp16, [1, 512, 12, 64]> var_778_cast_fp16 = reshape(shape = var_777, x = linear_68_cast_fp16)[name = string("op_778_cast_fp16")]; |
| tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_107 = transpose(perm = transpose_107_perm_0, x = var_778_cast_fp16)[name = string("transpose_109")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_772_cast_fp16)[name = string("transpose_110")]; |
| tensor<fp16, [1, 12, 512, 64]> transpose_105 = transpose(perm = transpose_105_perm_0, x = var_766_cast_fp16)[name = string("transpose_111")]; |
| tensor<fp16, [1, 12, 512, 64]> attn_output_45_cast_fp16 = scaled_dot_product_attention(attn_mask = attention_mask_cast_fp16, key = transpose_106, query = transpose_105, value = transpose_107)[name = string("attn_output_45_cast_fp16")]; |
| tensor<int32, [4]> attn_output_perm_0 = const()[name = string("attn_output_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_782 = const()[name = string("op_782"), val = tensor<int32, [3]>([1, 512, 768])]; |
| tensor<fp16, [1, 512, 12, 64]> attn_output_cast_fp16 = transpose(perm = attn_output_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_108")]; |
| tensor<fp16, [1, 512, 768]> input_185_cast_fp16 = reshape(shape = var_782, x = attn_output_cast_fp16)[name = string("input_185_cast_fp16")]; |
| tensor<fp16, [768, 768]> m_roberta_encoder_layer_11_attention_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222507776)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_attention_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223687488)))]; |
| tensor<fp16, [1, 512, 768]> linear_69_cast_fp16 = linear(bias = m_roberta_encoder_layer_11_attention_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_11_attention_output_dense_weight_to_fp16, x = input_185_cast_fp16)[name = string("linear_69_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_189_cast_fp16 = add(x = linear_69_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("input_189_cast_fp16")]; |
| tensor<int32, [1]> input_191_axes_0 = const()[name = string("input_191_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223689088)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223690688)))]; |
| tensor<fp16, [1, 512, 768]> input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = m_roberta_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16, x = input_189_cast_fp16)[name = string("input_191_cast_fp16")]; |
| tensor<fp16, [3072, 768]> m_roberta_encoder_layer_11_intermediate_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223692288)))]; |
| tensor<fp16, [3072]> m_roberta_encoder_layer_11_intermediate_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228410944)))]; |
| tensor<fp16, [1, 512, 3072]> linear_70_cast_fp16 = linear(bias = m_roberta_encoder_layer_11_intermediate_dense_bias_to_fp16, weight = m_roberta_encoder_layer_11_intermediate_dense_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_70_cast_fp16")]; |
| string input_195_mode_0 = const()[name = string("input_195_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 512, 3072]> input_195_cast_fp16 = gelu(mode = input_195_mode_0, x = linear_70_cast_fp16)[name = string("input_195_cast_fp16")]; |
| tensor<fp16, [768, 3072]> m_roberta_encoder_layer_11_output_dense_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_output_dense_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228417152)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_output_dense_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_output_dense_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233135808)))]; |
| tensor<fp16, [1, 512, 768]> linear_71_cast_fp16 = linear(bias = m_roberta_encoder_layer_11_output_dense_bias_to_fp16, weight = m_roberta_encoder_layer_11_output_dense_weight_to_fp16, x = input_195_cast_fp16)[name = string("linear_71_cast_fp16")]; |
| tensor<fp16, [1, 512, 768]> input_199_cast_fp16 = add(x = linear_71_cast_fp16, y = input_191_cast_fp16)[name = string("input_199_cast_fp16")]; |
| tensor<int32, [1]> input_201_axes_0 = const()[name = string("input_201_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_output_LayerNorm_weight_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233137408)))]; |
| tensor<fp16, [768]> m_roberta_encoder_layer_11_output_LayerNorm_bias_to_fp16 = const()[name = string("m_roberta_encoder_layer_11_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233139008)))]; |
| tensor<fp16, [1, 512, 768]> input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = m_roberta_encoder_layer_11_output_LayerNorm_bias_to_fp16, epsilon = var_10_to_fp16, gamma = m_roberta_encoder_layer_11_output_LayerNorm_weight_to_fp16, x = input_199_cast_fp16)[name = string("input_201_cast_fp16")]; |
| tensor<fp16, [6, 768]> m_classifier_weight_to_fp16 = const()[name = string("m_classifier_weight_to_fp16"), val = tensor<fp16, [6, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233140608)))]; |
| tensor<fp16, [6]> m_classifier_bias_to_fp16 = const()[name = string("m_classifier_bias_to_fp16"), val = tensor<fp16, [6]>([0x1.ddcp-8, 0x1.264p-7, 0x1.59cp-7, -0x1.c5cp-5, -0x1.3p-6, -0x1.2fcp-5])]; |
| tensor<fp16, [1, 512, 6]> logits = linear(bias = m_classifier_bias_to_fp16, weight = m_classifier_weight_to_fp16, x = input_201_cast_fp16)[name = string("linear_72_cast_fp16")]; |
| } -> (logits); |
| } |
