| program(1.0) |
| [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3400.43.1"}, {"coremlc-version", "3400.58.2"}, {"coremltools-component-torch", "2.4.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})] |
| { |
| func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, tensor<fp16, [1, 768, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 9216, 1, 448]> key_cache, tensor<fp16, [1, 448]> kv_cache_update_mask, tensor<fp16, [1, 9216, 1, 448]> value_cache) { |
| tensor<int32, []> var_40_axis_0 = const()[name = tensor<string, []>("op_40_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> var_40_batch_dims_0 = const()[name = tensor<string, []>("op_40_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [51865, 768]> embed_tokens_weight_to_fp16 = const()[name = tensor<string, []>("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51865, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))]; |
| tensor<fp16, [1, 768]> var_40_cast_fp16 = gather(axis = var_40_axis_0, batch_dims = var_40_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_40_cast_fp16")]; |
| tensor<int32, []> var_44_axis_0 = const()[name = tensor<string, []>("op_44_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> var_44_batch_dims_0 = const()[name = tensor<string, []>("op_44_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [448, 768]> embed_positions_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79664768)))]; |
| tensor<fp16, [1, 768]> var_44_cast_fp16 = gather(axis = var_44_axis_0, batch_dims = var_44_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_44_cast_fp16")]; |
| tensor<fp16, [1, 768]> hidden_states_1_cast_fp16 = add(x = var_40_cast_fp16, y = var_44_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")]; |
| tensor<int32, [1]> var_58_axes_0 = const()[name = tensor<string, []>("op_58_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 768, 1]> var_58_cast_fp16 = expand_dims(axes = var_58_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_58_cast_fp16")]; |
| tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_58_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")]; |
| tensor<int32, [12]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [12]>([768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768])]; |
| tensor<int32, []> var_63_axis_0 = const()[name = tensor<string, []>("op_63_axis_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_0, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_1, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_2, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_3, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_4, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_5, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_6, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_7, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_8, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_9, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_10, tensor<fp16, [1, 768, 1, 448]> var_63_cast_fp16_11 = split(axis = var_63_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_63_cast_fp16")]; |
| tensor<int32, [12]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [12]>([768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768])]; |
| tensor<int32, []> var_78_axis_0 = const()[name = tensor<string, []>("op_78_axis_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_0, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_1, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_2, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_3, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_4, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_5, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_6, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_7, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_8, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_9, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_10, tensor<fp16, [1, 768, 1, 448]> var_78_cast_fp16_11 = split(axis = var_78_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_78_cast_fp16")]; |
| tensor<int32, []> var_96 = const()[name = tensor<string, []>("op_96"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_122_to_fp16 = const()[name = tensor<string, []>("op_122_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_122_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")]; |
| tensor<fp16, [768]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80352960)))]; |
| tensor<fp16, [768]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80354560)))]; |
| tensor<fp16, [768]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80356160)))]; |
| tensor<fp16, [768]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80357760)))]; |
| tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")]; |
| tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_1_strides_0 = const()[name = tensor<string, []>("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_1_dilations_0 = const()[name = tensor<string, []>("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_1_groups_0 = const()[name = tensor<string, []>("query_1_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80359360)))]; |
| tensor<fp16, [768]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81539072)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")]; |
| tensor<string, []> current_key_1_pad_type_0 = const()[name = tensor<string, []>("current_key_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_1_strides_0 = const()[name = tensor<string, []>("current_key_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_1_pad_0 = const()[name = tensor<string, []>("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_1_dilations_0 = const()[name = tensor<string, []>("current_key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_1_groups_0 = const()[name = tensor<string, []>("current_key_1_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81540672)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")]; |
| tensor<string, []> current_value_1_pad_type_0 = const()[name = tensor<string, []>("current_value_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_1_strides_0 = const()[name = tensor<string, []>("current_value_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_1_pad_0 = const()[name = tensor<string, []>("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_1_dilations_0 = const()[name = tensor<string, []>("current_value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_1_groups_0 = const()[name = tensor<string, []>("current_value_1_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82720384)))]; |
| tensor<fp16, [768]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83900096)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")]; |
| tensor<int32, [1]> var_157_axes_0 = const()[name = tensor<string, []>("op_157_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1, 1, 448]> var_157_cast_fp16 = expand_dims(axes = var_157_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_157_cast_fp16")]; |
| tensor<int32, [1]> var_158_axes_0 = const()[name = tensor<string, []>("op_158_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1, 1, 448]> var_158_cast_fp16 = expand_dims(axes = var_158_axes_0, x = var_157_cast_fp16)[name = tensor<string, []>("op_158_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_160_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_160_cast_fp16")]; |
| tensor<fp16, []> var_97_to_fp16 = const()[name = tensor<string, []>("op_97_to_fp16"), val = tensor<fp16, []>(0x1p+0)]; |
| tensor<fp16, [1, 1, 1, 448]> var_161_cast_fp16 = sub(x = var_97_to_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_161_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_162_cast_fp16 = mul(x = var_63_cast_fp16_0, y = var_161_cast_fp16)[name = tensor<string, []>("op_162_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_1_cast_fp16 = add(x = var_160_cast_fp16, y = var_162_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_164_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_164_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_166_cast_fp16 = mul(x = var_78_cast_fp16_0, y = var_161_cast_fp16)[name = tensor<string, []>("op_166_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_1_cast_fp16 = add(x = var_164_cast_fp16, y = var_166_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")]; |
| tensor<int32, [4]> var_169 = const()[name = tensor<string, []>("op_169"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_169, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")]; |
| tensor<fp16, []> var_171_to_fp16 = const()[name = tensor<string, []>("op_171_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_172_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_171_to_fp16)[name = tensor<string, []>("op_172_cast_fp16")]; |
| tensor<int32, [4]> var_173 = const()[name = tensor<string, []>("op_173"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_174_cast_fp16 = reshape(shape = var_173, x = key_1_cast_fp16)[name = tensor<string, []>("op_174_cast_fp16")]; |
| tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_172_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")]; |
| tensor<int32, [1]> var_178_axes_0 = const()[name = tensor<string, []>("op_178_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1, 1, 448]> var_178_cast_fp16 = expand_dims(axes = var_178_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_178_cast_fp16")]; |
| tensor<int32, [1]> var_179_axes_0 = const()[name = tensor<string, []>("op_179_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1, 1, 448]> var_179_cast_fp16 = expand_dims(axes = var_179_axes_0, x = var_178_cast_fp16)[name = tensor<string, []>("op_179_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_182_cast_fp16 = softmax(axis = var_96, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_182_cast_fp16")]; |
| tensor<int32, [4]> var_183 = const()[name = tensor<string, []>("op_183"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_184_cast_fp16 = reshape(shape = var_183, x = value_1_cast_fp16)[name = tensor<string, []>("op_184_cast_fp16")]; |
| tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_184_cast_fp16, y = var_182_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")]; |
| tensor<int32, [4]> var_187 = const()[name = tensor<string, []>("op_187"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_1_cast_fp16 = reshape(shape = var_187, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")]; |
| tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_7_strides_0 = const()[name = tensor<string, []>("obj_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_7_dilations_0 = const()[name = tensor<string, []>("obj_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_7_groups_0 = const()[name = tensor<string, []>("obj_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83901696)))]; |
| tensor<fp16, [768]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85081408)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")]; |
| tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_209_to_fp16 = const()[name = tensor<string, []>("op_209_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_209_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")]; |
| tensor<fp16, [768]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85083008)))]; |
| tensor<fp16, [768]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85084608)))]; |
| tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")]; |
| tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_3_strides_0 = const()[name = tensor<string, []>("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_3_dilations_0 = const()[name = tensor<string, []>("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_3_groups_0 = const()[name = tensor<string, []>("query_3_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85086208)))]; |
| tensor<fp16, [768]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86265920)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")]; |
| tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_3_strides_0 = const()[name = tensor<string, []>("key_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_3_dilations_0 = const()[name = tensor<string, []>("key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_3_groups_0 = const()[name = tensor<string, []>("key_3_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86267520)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_3_cast_fp16")]; |
| tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_3_strides_0 = const()[name = tensor<string, []>("value_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_3_dilations_0 = const()[name = tensor<string, []>("value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_3_groups_0 = const()[name = tensor<string, []>("value_3_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87447232)))]; |
| tensor<fp16, [768]> layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88626944)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_3_cast_fp16")]; |
| tensor<int32, [4]> var_244 = const()[name = tensor<string, []>("op_244"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_244, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")]; |
| tensor<fp16, []> var_246_to_fp16 = const()[name = tensor<string, []>("op_246_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_247_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_246_to_fp16)[name = tensor<string, []>("op_247_cast_fp16")]; |
| tensor<int32, [4]> var_248 = const()[name = tensor<string, []>("op_248"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_249_cast_fp16 = reshape(shape = var_248, x = key_3_cast_fp16)[name = tensor<string, []>("op_249_cast_fp16")]; |
| tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_247_cast_fp16, y = var_249_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_13_cast_fp16 = softmax(axis = var_96, x = mh_w_5_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")]; |
| tensor<int32, [4]> var_253 = const()[name = tensor<string, []>("op_253"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_254_cast_fp16 = reshape(shape = var_253, x = value_3_cast_fp16)[name = tensor<string, []>("op_254_cast_fp16")]; |
| tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_254_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")]; |
| tensor<int32, [4]> var_257 = const()[name = tensor<string, []>("op_257"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_3_cast_fp16 = reshape(shape = var_257, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")]; |
| tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_11_strides_0 = const()[name = tensor<string, []>("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_11_dilations_0 = const()[name = tensor<string, []>("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_11_groups_0 = const()[name = tensor<string, []>("obj_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88628544)))]; |
| tensor<fp16, [768]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89808256)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")]; |
| tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_275_to_fp16 = const()[name = tensor<string, []>("op_275_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_275_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")]; |
| tensor<fp16, [768]> input_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_5_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89809856)))]; |
| tensor<fp16, [768]> input_5_beta_0_to_fp16 = const()[name = tensor<string, []>("input_5_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89811456)))]; |
| tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")]; |
| tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_7_strides_0 = const()[name = tensor<string, []>("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_7_dilations_0 = const()[name = tensor<string, []>("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_7_groups_0 = const()[name = tensor<string, []>("input_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89813056)))]; |
| tensor<fp16, [3072]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94531712)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")]; |
| tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")]; |
| tensor<string, []> hidden_states_3_pad_type_0 = const()[name = tensor<string, []>("hidden_states_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = tensor<string, []>("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = tensor<string, []>("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = tensor<string, []>("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_3_groups_0 = const()[name = tensor<string, []>("hidden_states_3_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94537920)))]; |
| tensor<fp16, [768]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99256576)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")]; |
| tensor<int32, []> var_310 = const()[name = tensor<string, []>("op_310"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_336_to_fp16 = const()[name = tensor<string, []>("op_336_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_336_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")]; |
| tensor<fp16, [768]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99258176)))]; |
| tensor<fp16, [768]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99259776)))]; |
| tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")]; |
| tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_5_strides_0 = const()[name = tensor<string, []>("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_5_dilations_0 = const()[name = tensor<string, []>("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_5_groups_0 = const()[name = tensor<string, []>("query_5_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99261376)))]; |
| tensor<fp16, [768]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100441088)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")]; |
| tensor<string, []> current_key_3_pad_type_0 = const()[name = tensor<string, []>("current_key_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_3_strides_0 = const()[name = tensor<string, []>("current_key_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_3_pad_0 = const()[name = tensor<string, []>("current_key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_3_dilations_0 = const()[name = tensor<string, []>("current_key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_3_groups_0 = const()[name = tensor<string, []>("current_key_3_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100442688)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_3_cast_fp16 = conv(dilations = current_key_3_dilations_0, groups = current_key_3_groups_0, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = current_key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")]; |
| tensor<string, []> current_value_3_pad_type_0 = const()[name = tensor<string, []>("current_value_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_3_strides_0 = const()[name = tensor<string, []>("current_value_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_3_pad_0 = const()[name = tensor<string, []>("current_value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_3_dilations_0 = const()[name = tensor<string, []>("current_value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_3_groups_0 = const()[name = tensor<string, []>("current_value_3_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101622400)))]; |
| tensor<fp16, [768]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102802112)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_3_dilations_0, groups = current_value_3_groups_0, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = current_value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_374_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_374_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_376_cast_fp16 = mul(x = var_63_cast_fp16_1, y = var_161_cast_fp16)[name = tensor<string, []>("op_376_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_5_cast_fp16 = add(x = var_374_cast_fp16, y = var_376_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_378_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_378_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_380_cast_fp16 = mul(x = var_78_cast_fp16_1, y = var_161_cast_fp16)[name = tensor<string, []>("op_380_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_5_cast_fp16 = add(x = var_378_cast_fp16, y = var_380_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")]; |
| tensor<int32, [4]> var_383 = const()[name = tensor<string, []>("op_383"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_383, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")]; |
| tensor<fp16, []> var_385_to_fp16 = const()[name = tensor<string, []>("op_385_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_386_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_385_to_fp16)[name = tensor<string, []>("op_386_cast_fp16")]; |
| tensor<int32, [4]> var_387 = const()[name = tensor<string, []>("op_387"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_388_cast_fp16 = reshape(shape = var_387, x = key_5_cast_fp16)[name = tensor<string, []>("op_388_cast_fp16")]; |
| tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_386_cast_fp16, y = var_388_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_396_cast_fp16 = softmax(axis = var_310, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_396_cast_fp16")]; |
| tensor<int32, [4]> var_397 = const()[name = tensor<string, []>("op_397"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_398_cast_fp16 = reshape(shape = var_397, x = value_5_cast_fp16)[name = tensor<string, []>("op_398_cast_fp16")]; |
| tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_398_cast_fp16, y = var_396_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")]; |
| tensor<int32, [4]> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_11_cast_fp16 = reshape(shape = var_401, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")]; |
| tensor<string, []> obj_21_pad_type_0 = const()[name = tensor<string, []>("obj_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_21_strides_0 = const()[name = tensor<string, []>("obj_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_21_pad_0 = const()[name = tensor<string, []>("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_21_dilations_0 = const()[name = tensor<string, []>("obj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_21_groups_0 = const()[name = tensor<string, []>("obj_21_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102803712)))]; |
| tensor<fp16, [768]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103983424)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")]; |
| tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_423_to_fp16 = const()[name = tensor<string, []>("op_423_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_423_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")]; |
| tensor<fp16, [768]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103985024)))]; |
| tensor<fp16, [768]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103986624)))]; |
| tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")]; |
| tensor<string, []> query_7_pad_type_0 = const()[name = tensor<string, []>("query_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_7_strides_0 = const()[name = tensor<string, []>("query_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_7_pad_0 = const()[name = tensor<string, []>("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_7_dilations_0 = const()[name = tensor<string, []>("query_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_7_groups_0 = const()[name = tensor<string, []>("query_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103988224)))]; |
| tensor<fp16, [768]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105167936)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")]; |
| tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_7_strides_0 = const()[name = tensor<string, []>("key_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_7_pad_0 = const()[name = tensor<string, []>("key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_7_dilations_0 = const()[name = tensor<string, []>("key_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_7_groups_0 = const()[name = tensor<string, []>("key_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105169536)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_7_cast_fp16")]; |
| tensor<string, []> value_7_pad_type_0 = const()[name = tensor<string, []>("value_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_7_strides_0 = const()[name = tensor<string, []>("value_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_7_pad_0 = const()[name = tensor<string, []>("value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_7_dilations_0 = const()[name = tensor<string, []>("value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_7_groups_0 = const()[name = tensor<string, []>("value_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106349248)))]; |
| tensor<fp16, [768]> layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107528960)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_7_cast_fp16")]; |
| tensor<int32, [4]> var_458 = const()[name = tensor<string, []>("op_458"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_7_cast_fp16 = reshape(shape = var_458, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")]; |
| tensor<fp16, []> var_460_to_fp16 = const()[name = tensor<string, []>("op_460_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_461_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_460_to_fp16)[name = tensor<string, []>("op_461_cast_fp16")]; |
| tensor<int32, [4]> var_462 = const()[name = tensor<string, []>("op_462"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_463_cast_fp16 = reshape(shape = var_462, x = key_7_cast_fp16)[name = tensor<string, []>("op_463_cast_fp16")]; |
| tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_461_cast_fp16, y = var_463_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_27_cast_fp16 = softmax(axis = var_310, x = mh_w_11_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")]; |
| tensor<int32, [4]> var_467 = const()[name = tensor<string, []>("op_467"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_468_cast_fp16 = reshape(shape = var_467, x = value_7_cast_fp16)[name = tensor<string, []>("op_468_cast_fp16")]; |
| tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_468_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")]; |
| tensor<int32, [4]> var_471 = const()[name = tensor<string, []>("op_471"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_13_cast_fp16 = reshape(shape = var_471, x = attn_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")]; |
| tensor<string, []> obj_25_pad_type_0 = const()[name = tensor<string, []>("obj_25_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_25_strides_0 = const()[name = tensor<string, []>("obj_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_25_pad_0 = const()[name = tensor<string, []>("obj_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_25_dilations_0 = const()[name = tensor<string, []>("obj_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_25_groups_0 = const()[name = tensor<string, []>("obj_25_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107530560)))]; |
| tensor<fp16, [768]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108710272)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")]; |
| tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_489_to_fp16 = const()[name = tensor<string, []>("op_489_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_489_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")]; |
| tensor<fp16, [768]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108711872)))]; |
| tensor<fp16, [768]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108713472)))]; |
| tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")]; |
| tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_17_strides_0 = const()[name = tensor<string, []>("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_17_dilations_0 = const()[name = tensor<string, []>("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_17_groups_0 = const()[name = tensor<string, []>("input_17_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108715072)))]; |
| tensor<fp16, [3072]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113433728)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")]; |
| tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")]; |
| tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = tensor<string, []>("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = tensor<string, []>("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_5_groups_0 = const()[name = tensor<string, []>("hidden_states_5_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113439936)))]; |
| tensor<fp16, [768]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118158592)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")]; |
| tensor<int32, []> var_524 = const()[name = tensor<string, []>("op_524"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_550_to_fp16 = const()[name = tensor<string, []>("op_550_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_550_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")]; |
| tensor<fp16, [768]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118160192)))]; |
| tensor<fp16, [768]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118161792)))]; |
| tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")]; |
| tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_9_strides_0 = const()[name = tensor<string, []>("query_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_9_dilations_0 = const()[name = tensor<string, []>("query_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_9_groups_0 = const()[name = tensor<string, []>("query_9_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118163392)))]; |
| tensor<fp16, [768]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119343104)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")]; |
| tensor<string, []> current_key_5_pad_type_0 = const()[name = tensor<string, []>("current_key_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_5_strides_0 = const()[name = tensor<string, []>("current_key_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_5_pad_0 = const()[name = tensor<string, []>("current_key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_5_dilations_0 = const()[name = tensor<string, []>("current_key_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_5_groups_0 = const()[name = tensor<string, []>("current_key_5_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119344704)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_5_cast_fp16 = conv(dilations = current_key_5_dilations_0, groups = current_key_5_groups_0, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = current_key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")]; |
| tensor<string, []> current_value_5_pad_type_0 = const()[name = tensor<string, []>("current_value_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_5_strides_0 = const()[name = tensor<string, []>("current_value_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_5_pad_0 = const()[name = tensor<string, []>("current_value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_5_dilations_0 = const()[name = tensor<string, []>("current_value_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_5_groups_0 = const()[name = tensor<string, []>("current_value_5_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120524416)))]; |
| tensor<fp16, [768]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121704128)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = current_value_5_dilations_0, groups = current_value_5_groups_0, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = current_value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_588_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_588_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_590_cast_fp16 = mul(x = var_63_cast_fp16_2, y = var_161_cast_fp16)[name = tensor<string, []>("op_590_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_9_cast_fp16 = add(x = var_588_cast_fp16, y = var_590_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_592_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_592_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_594_cast_fp16 = mul(x = var_78_cast_fp16_2, y = var_161_cast_fp16)[name = tensor<string, []>("op_594_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_9_cast_fp16 = add(x = var_592_cast_fp16, y = var_594_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")]; |
| tensor<int32, [4]> var_597 = const()[name = tensor<string, []>("op_597"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_9_cast_fp16 = reshape(shape = var_597, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")]; |
| tensor<fp16, []> var_599_to_fp16 = const()[name = tensor<string, []>("op_599_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_600_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_599_to_fp16)[name = tensor<string, []>("op_600_cast_fp16")]; |
| tensor<int32, [4]> var_601 = const()[name = tensor<string, []>("op_601"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_602_cast_fp16 = reshape(shape = var_601, x = key_9_cast_fp16)[name = tensor<string, []>("op_602_cast_fp16")]; |
| tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_600_cast_fp16, y = var_602_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_610_cast_fp16 = softmax(axis = var_524, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_610_cast_fp16")]; |
| tensor<int32, [4]> var_611 = const()[name = tensor<string, []>("op_611"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_612_cast_fp16 = reshape(shape = var_611, x = value_9_cast_fp16)[name = tensor<string, []>("op_612_cast_fp16")]; |
| tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_612_cast_fp16, y = var_610_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")]; |
| tensor<int32, [4]> var_615 = const()[name = tensor<string, []>("op_615"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_21_cast_fp16 = reshape(shape = var_615, x = attn_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")]; |
| tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_35_strides_0 = const()[name = tensor<string, []>("obj_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_35_dilations_0 = const()[name = tensor<string, []>("obj_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_35_groups_0 = const()[name = tensor<string, []>("obj_35_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121705728)))]; |
| tensor<fp16, [768]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122885440)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")]; |
| tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_637_to_fp16 = const()[name = tensor<string, []>("op_637_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_637_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")]; |
| tensor<fp16, [768]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122887040)))]; |
| tensor<fp16, [768]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122888640)))]; |
| tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")]; |
| tensor<string, []> query_11_pad_type_0 = const()[name = tensor<string, []>("query_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_11_strides_0 = const()[name = tensor<string, []>("query_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_11_pad_0 = const()[name = tensor<string, []>("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_11_dilations_0 = const()[name = tensor<string, []>("query_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_11_groups_0 = const()[name = tensor<string, []>("query_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122890240)))]; |
| tensor<fp16, [768]> layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124069952)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")]; |
| tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_11_strides_0 = const()[name = tensor<string, []>("key_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_11_dilations_0 = const()[name = tensor<string, []>("key_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_11_groups_0 = const()[name = tensor<string, []>("key_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124071552)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_11_cast_fp16")]; |
| tensor<string, []> value_11_pad_type_0 = const()[name = tensor<string, []>("value_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_11_strides_0 = const()[name = tensor<string, []>("value_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_11_pad_0 = const()[name = tensor<string, []>("value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_11_dilations_0 = const()[name = tensor<string, []>("value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_11_groups_0 = const()[name = tensor<string, []>("value_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125251264)))]; |
| tensor<fp16, [768]> layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126430976)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_11_cast_fp16")]; |
| tensor<int32, [4]> var_672 = const()[name = tensor<string, []>("op_672"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_11_cast_fp16 = reshape(shape = var_672, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")]; |
| tensor<fp16, []> var_674_to_fp16 = const()[name = tensor<string, []>("op_674_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_675_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_674_to_fp16)[name = tensor<string, []>("op_675_cast_fp16")]; |
| tensor<int32, [4]> var_676 = const()[name = tensor<string, []>("op_676"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_677_cast_fp16 = reshape(shape = var_676, x = key_11_cast_fp16)[name = tensor<string, []>("op_677_cast_fp16")]; |
| tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_675_cast_fp16, y = var_677_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_41_cast_fp16 = softmax(axis = var_524, x = mh_w_17_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")]; |
| tensor<int32, [4]> var_681 = const()[name = tensor<string, []>("op_681"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_682_cast_fp16 = reshape(shape = var_681, x = value_11_cast_fp16)[name = tensor<string, []>("op_682_cast_fp16")]; |
| tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_682_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")]; |
| tensor<int32, [4]> var_685 = const()[name = tensor<string, []>("op_685"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_23_cast_fp16 = reshape(shape = var_685, x = attn_11_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")]; |
| tensor<string, []> obj_39_pad_type_0 = const()[name = tensor<string, []>("obj_39_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_39_strides_0 = const()[name = tensor<string, []>("obj_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_39_pad_0 = const()[name = tensor<string, []>("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_39_dilations_0 = const()[name = tensor<string, []>("obj_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_39_groups_0 = const()[name = tensor<string, []>("obj_39_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126432576)))]; |
| tensor<fp16, [768]> layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127612288)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")]; |
| tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_703_to_fp16 = const()[name = tensor<string, []>("op_703_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_703_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")]; |
| tensor<fp16, [768]> input_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_25_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127613888)))]; |
| tensor<fp16, [768]> input_25_beta_0_to_fp16 = const()[name = tensor<string, []>("input_25_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127615488)))]; |
| tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")]; |
| tensor<string, []> input_27_pad_type_0 = const()[name = tensor<string, []>("input_27_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_27_strides_0 = const()[name = tensor<string, []>("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_27_pad_0 = const()[name = tensor<string, []>("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_27_dilations_0 = const()[name = tensor<string, []>("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_27_groups_0 = const()[name = tensor<string, []>("input_27_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127617088)))]; |
| tensor<fp16, [3072]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132335744)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")]; |
| tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")]; |
| tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = tensor<string, []>("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = tensor<string, []>("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_7_groups_0 = const()[name = tensor<string, []>("hidden_states_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_2_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132341952)))]; |
| tensor<fp16, [768]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137060608)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")]; |
| tensor<int32, []> var_738 = const()[name = tensor<string, []>("op_738"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_764_to_fp16 = const()[name = tensor<string, []>("op_764_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_764_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")]; |
| tensor<fp16, [768]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137062208)))]; |
| tensor<fp16, [768]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137063808)))]; |
| tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")]; |
| tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_13_strides_0 = const()[name = tensor<string, []>("query_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_13_dilations_0 = const()[name = tensor<string, []>("query_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_13_groups_0 = const()[name = tensor<string, []>("query_13_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137065408)))]; |
| tensor<fp16, [768]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138245120)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")]; |
| tensor<string, []> current_key_7_pad_type_0 = const()[name = tensor<string, []>("current_key_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_7_strides_0 = const()[name = tensor<string, []>("current_key_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_7_pad_0 = const()[name = tensor<string, []>("current_key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_7_dilations_0 = const()[name = tensor<string, []>("current_key_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_7_groups_0 = const()[name = tensor<string, []>("current_key_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138246720)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_7_cast_fp16 = conv(dilations = current_key_7_dilations_0, groups = current_key_7_groups_0, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = current_key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("current_key_7_cast_fp16")]; |
| tensor<string, []> current_value_7_pad_type_0 = const()[name = tensor<string, []>("current_value_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_7_strides_0 = const()[name = tensor<string, []>("current_value_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_7_pad_0 = const()[name = tensor<string, []>("current_value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_7_dilations_0 = const()[name = tensor<string, []>("current_value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_7_groups_0 = const()[name = tensor<string, []>("current_value_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139426432)))]; |
| tensor<fp16, [768]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140606144)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = current_value_7_dilations_0, groups = current_value_7_groups_0, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = current_value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("current_value_7_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_802_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_802_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_804_cast_fp16 = mul(x = var_63_cast_fp16_3, y = var_161_cast_fp16)[name = tensor<string, []>("op_804_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_13_cast_fp16 = add(x = var_802_cast_fp16, y = var_804_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_806_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_806_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_808_cast_fp16 = mul(x = var_78_cast_fp16_3, y = var_161_cast_fp16)[name = tensor<string, []>("op_808_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_13_cast_fp16 = add(x = var_806_cast_fp16, y = var_808_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")]; |
| tensor<int32, [4]> var_811 = const()[name = tensor<string, []>("op_811"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_13_cast_fp16 = reshape(shape = var_811, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")]; |
| tensor<fp16, []> var_813_to_fp16 = const()[name = tensor<string, []>("op_813_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_814_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_813_to_fp16)[name = tensor<string, []>("op_814_cast_fp16")]; |
| tensor<int32, [4]> var_815 = const()[name = tensor<string, []>("op_815"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_816_cast_fp16 = reshape(shape = var_815, x = key_13_cast_fp16)[name = tensor<string, []>("op_816_cast_fp16")]; |
| tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_814_cast_fp16, y = var_816_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_824_cast_fp16 = softmax(axis = var_738, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_824_cast_fp16")]; |
| tensor<int32, [4]> var_825 = const()[name = tensor<string, []>("op_825"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_826_cast_fp16 = reshape(shape = var_825, x = value_13_cast_fp16)[name = tensor<string, []>("op_826_cast_fp16")]; |
| tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_826_cast_fp16, y = var_824_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")]; |
| tensor<int32, [4]> var_829 = const()[name = tensor<string, []>("op_829"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_31_cast_fp16 = reshape(shape = var_829, x = attn_13_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")]; |
| tensor<string, []> obj_49_pad_type_0 = const()[name = tensor<string, []>("obj_49_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_49_strides_0 = const()[name = tensor<string, []>("obj_49_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_49_pad_0 = const()[name = tensor<string, []>("obj_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_49_dilations_0 = const()[name = tensor<string, []>("obj_49_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_49_groups_0 = const()[name = tensor<string, []>("obj_49_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140607744)))]; |
| tensor<fp16, [768]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141787456)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")]; |
| tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_851_to_fp16 = const()[name = tensor<string, []>("op_851_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_851_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")]; |
| tensor<fp16, [768]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141789056)))]; |
| tensor<fp16, [768]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141790656)))]; |
| tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")]; |
| tensor<string, []> query_15_pad_type_0 = const()[name = tensor<string, []>("query_15_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_15_strides_0 = const()[name = tensor<string, []>("query_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_15_pad_0 = const()[name = tensor<string, []>("query_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_15_dilations_0 = const()[name = tensor<string, []>("query_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_15_groups_0 = const()[name = tensor<string, []>("query_15_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141792256)))]; |
| tensor<fp16, [768]> layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142971968)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("query_15_cast_fp16")]; |
| tensor<string, []> key_15_pad_type_0 = const()[name = tensor<string, []>("key_15_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_15_strides_0 = const()[name = tensor<string, []>("key_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_15_pad_0 = const()[name = tensor<string, []>("key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_15_dilations_0 = const()[name = tensor<string, []>("key_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_15_groups_0 = const()[name = tensor<string, []>("key_15_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142973568)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_15_cast_fp16")]; |
| tensor<string, []> value_15_pad_type_0 = const()[name = tensor<string, []>("value_15_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_15_strides_0 = const()[name = tensor<string, []>("value_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_15_pad_0 = const()[name = tensor<string, []>("value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_15_dilations_0 = const()[name = tensor<string, []>("value_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_15_groups_0 = const()[name = tensor<string, []>("value_15_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144153280)))]; |
| tensor<fp16, [768]> layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145332992)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_15_cast_fp16")]; |
| tensor<int32, [4]> var_886 = const()[name = tensor<string, []>("op_886"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_15_cast_fp16 = reshape(shape = var_886, x = query_15_cast_fp16)[name = tensor<string, []>("mh_q_15_cast_fp16")]; |
| tensor<fp16, []> var_888_to_fp16 = const()[name = tensor<string, []>("op_888_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_889_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_888_to_fp16)[name = tensor<string, []>("op_889_cast_fp16")]; |
| tensor<int32, [4]> var_890 = const()[name = tensor<string, []>("op_890"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_891_cast_fp16 = reshape(shape = var_890, x = key_15_cast_fp16)[name = tensor<string, []>("op_891_cast_fp16")]; |
| tensor<bool, []> mh_w_23_transpose_x_0 = const()[name = tensor<string, []>("mh_w_23_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_23_transpose_y_0 = const()[name = tensor<string, []>("mh_w_23_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_889_cast_fp16, y = var_891_cast_fp16)[name = tensor<string, []>("mh_w_23_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_55_cast_fp16 = softmax(axis = var_738, x = mh_w_23_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")]; |
| tensor<int32, [4]> var_895 = const()[name = tensor<string, []>("op_895"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_896_cast_fp16 = reshape(shape = var_895, x = value_15_cast_fp16)[name = tensor<string, []>("op_896_cast_fp16")]; |
| tensor<bool, []> attn_15_transpose_x_0 = const()[name = tensor<string, []>("attn_15_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_15_transpose_y_0 = const()[name = tensor<string, []>("attn_15_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_896_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")]; |
| tensor<int32, [4]> var_899 = const()[name = tensor<string, []>("op_899"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_33_cast_fp16 = reshape(shape = var_899, x = attn_15_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")]; |
| tensor<string, []> obj_53_pad_type_0 = const()[name = tensor<string, []>("obj_53_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_53_strides_0 = const()[name = tensor<string, []>("obj_53_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_53_pad_0 = const()[name = tensor<string, []>("obj_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_53_dilations_0 = const()[name = tensor<string, []>("obj_53_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_53_groups_0 = const()[name = tensor<string, []>("obj_53_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145334592)))]; |
| tensor<fp16, [768]> layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146514304)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")]; |
| tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_917_to_fp16 = const()[name = tensor<string, []>("op_917_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_917_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")]; |
| tensor<fp16, [768]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146515904)))]; |
| tensor<fp16, [768]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146517504)))]; |
| tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")]; |
| tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_37_strides_0 = const()[name = tensor<string, []>("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_37_dilations_0 = const()[name = tensor<string, []>("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_37_groups_0 = const()[name = tensor<string, []>("input_37_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_3_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146519104)))]; |
| tensor<fp16, [3072]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151237760)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")]; |
| tensor<string, []> input_39_mode_0 = const()[name = tensor<string, []>("input_39_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")]; |
| tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = tensor<string, []>("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = tensor<string, []>("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_9_groups_0 = const()[name = tensor<string, []>("hidden_states_9_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151243968)))]; |
| tensor<fp16, [768]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155962624)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_25_cast_fp16")]; |
| tensor<int32, []> var_952 = const()[name = tensor<string, []>("op_952"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_25_axes_0 = const()[name = tensor<string, []>("out_25_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_978_to_fp16 = const()[name = tensor<string, []>("op_978_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_978_to_fp16, x = inputs_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")]; |
| tensor<fp16, [768]> obj_57_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_57_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155964224)))]; |
| tensor<fp16, [768]> obj_57_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_57_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155965824)))]; |
| tensor<fp16, []> obj_57_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_57_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor<string, []>("obj_57_cast_fp16")]; |
| tensor<string, []> query_17_pad_type_0 = const()[name = tensor<string, []>("query_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_17_strides_0 = const()[name = tensor<string, []>("query_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_17_pad_0 = const()[name = tensor<string, []>("query_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_17_dilations_0 = const()[name = tensor<string, []>("query_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_17_groups_0 = const()[name = tensor<string, []>("query_17_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155967424)))]; |
| tensor<fp16, [768]> layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157147136)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")]; |
| tensor<string, []> current_key_9_pad_type_0 = const()[name = tensor<string, []>("current_key_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_9_strides_0 = const()[name = tensor<string, []>("current_key_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_9_pad_0 = const()[name = tensor<string, []>("current_key_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_9_dilations_0 = const()[name = tensor<string, []>("current_key_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_9_groups_0 = const()[name = tensor<string, []>("current_key_9_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157148736)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_9_cast_fp16 = conv(dilations = current_key_9_dilations_0, groups = current_key_9_groups_0, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = current_key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("current_key_9_cast_fp16")]; |
| tensor<string, []> current_value_9_pad_type_0 = const()[name = tensor<string, []>("current_value_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_9_strides_0 = const()[name = tensor<string, []>("current_value_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_9_pad_0 = const()[name = tensor<string, []>("current_value_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_9_dilations_0 = const()[name = tensor<string, []>("current_value_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_9_groups_0 = const()[name = tensor<string, []>("current_value_9_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158328448)))]; |
| tensor<fp16, [768]> layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159508160)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = current_value_9_dilations_0, groups = current_value_9_groups_0, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = current_value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("current_value_9_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1016_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1016_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1018_cast_fp16 = mul(x = var_63_cast_fp16_4, y = var_161_cast_fp16)[name = tensor<string, []>("op_1018_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_17_cast_fp16 = add(x = var_1016_cast_fp16, y = var_1018_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1020_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1020_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1022_cast_fp16 = mul(x = var_78_cast_fp16_4, y = var_161_cast_fp16)[name = tensor<string, []>("op_1022_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_17_cast_fp16 = add(x = var_1020_cast_fp16, y = var_1022_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")]; |
| tensor<int32, [4]> var_1025 = const()[name = tensor<string, []>("op_1025"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_17_cast_fp16 = reshape(shape = var_1025, x = query_17_cast_fp16)[name = tensor<string, []>("mh_q_17_cast_fp16")]; |
| tensor<fp16, []> var_1027_to_fp16 = const()[name = tensor<string, []>("op_1027_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1028_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1027_to_fp16)[name = tensor<string, []>("op_1028_cast_fp16")]; |
| tensor<int32, [4]> var_1029 = const()[name = tensor<string, []>("op_1029"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1030_cast_fp16 = reshape(shape = var_1029, x = key_17_cast_fp16)[name = tensor<string, []>("op_1030_cast_fp16")]; |
| tensor<bool, []> mh_w_25_transpose_x_0 = const()[name = tensor<string, []>("mh_w_25_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_25_transpose_y_0 = const()[name = tensor<string, []>("mh_w_25_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1028_cast_fp16, y = var_1030_cast_fp16)[name = tensor<string, []>("mh_w_25_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_27_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_1038_cast_fp16 = softmax(axis = var_952, x = mh_w_27_cast_fp16)[name = tensor<string, []>("op_1038_cast_fp16")]; |
| tensor<int32, [4]> var_1039 = const()[name = tensor<string, []>("op_1039"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1040_cast_fp16 = reshape(shape = var_1039, x = value_17_cast_fp16)[name = tensor<string, []>("op_1040_cast_fp16")]; |
| tensor<bool, []> attn_17_transpose_x_0 = const()[name = tensor<string, []>("attn_17_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_17_transpose_y_0 = const()[name = tensor<string, []>("attn_17_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1040_cast_fp16, y = var_1038_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")]; |
| tensor<int32, [4]> var_1043 = const()[name = tensor<string, []>("op_1043"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_41_cast_fp16 = reshape(shape = var_1043, x = attn_17_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")]; |
| tensor<string, []> obj_63_pad_type_0 = const()[name = tensor<string, []>("obj_63_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_63_strides_0 = const()[name = tensor<string, []>("obj_63_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_63_pad_0 = const()[name = tensor<string, []>("obj_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_63_dilations_0 = const()[name = tensor<string, []>("obj_63_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_63_groups_0 = const()[name = tensor<string, []>("obj_63_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159509760)))]; |
| tensor<fp16, [768]> layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160689472)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_63_dilations_0, groups = obj_63_groups_0, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = obj_63_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("obj_63_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor<string, []>("inputs_27_cast_fp16")]; |
| tensor<int32, [1]> out_27_axes_0 = const()[name = tensor<string, []>("out_27_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1065_to_fp16 = const()[name = tensor<string, []>("op_1065_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1065_to_fp16, x = inputs_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")]; |
| tensor<fp16, [768]> obj_65_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_65_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160691072)))]; |
| tensor<fp16, [768]> obj_65_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_65_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160692672)))]; |
| tensor<fp16, []> obj_65_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_65_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor<string, []>("obj_65_cast_fp16")]; |
| tensor<string, []> query_19_pad_type_0 = const()[name = tensor<string, []>("query_19_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_19_strides_0 = const()[name = tensor<string, []>("query_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_19_pad_0 = const()[name = tensor<string, []>("query_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_19_dilations_0 = const()[name = tensor<string, []>("query_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_19_groups_0 = const()[name = tensor<string, []>("query_19_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_4_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160694272)))]; |
| tensor<fp16, [768]> layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161873984)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor<string, []>("query_19_cast_fp16")]; |
| tensor<string, []> key_19_pad_type_0 = const()[name = tensor<string, []>("key_19_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_19_strides_0 = const()[name = tensor<string, []>("key_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_19_pad_0 = const()[name = tensor<string, []>("key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_19_dilations_0 = const()[name = tensor<string, []>("key_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_19_groups_0 = const()[name = tensor<string, []>("key_19_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_4_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161875584)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_19_cast_fp16")]; |
| tensor<string, []> value_19_pad_type_0 = const()[name = tensor<string, []>("value_19_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_19_strides_0 = const()[name = tensor<string, []>("value_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_19_pad_0 = const()[name = tensor<string, []>("value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_19_dilations_0 = const()[name = tensor<string, []>("value_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_19_groups_0 = const()[name = tensor<string, []>("value_19_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_4_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163055296)))]; |
| tensor<fp16, [768]> layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164235008)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_19_cast_fp16")]; |
| tensor<int32, [4]> var_1100 = const()[name = tensor<string, []>("op_1100"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_19_cast_fp16 = reshape(shape = var_1100, x = query_19_cast_fp16)[name = tensor<string, []>("mh_q_19_cast_fp16")]; |
| tensor<fp16, []> var_1102_to_fp16 = const()[name = tensor<string, []>("op_1102_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1103_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1102_to_fp16)[name = tensor<string, []>("op_1103_cast_fp16")]; |
| tensor<int32, [4]> var_1104 = const()[name = tensor<string, []>("op_1104"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1105_cast_fp16 = reshape(shape = var_1104, x = key_19_cast_fp16)[name = tensor<string, []>("op_1105_cast_fp16")]; |
| tensor<bool, []> mh_w_29_transpose_x_0 = const()[name = tensor<string, []>("mh_w_29_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_29_transpose_y_0 = const()[name = tensor<string, []>("mh_w_29_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1103_cast_fp16, y = var_1105_cast_fp16)[name = tensor<string, []>("mh_w_29_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_69_cast_fp16 = softmax(axis = var_952, x = mh_w_29_cast_fp16)[name = tensor<string, []>("obj_69_cast_fp16")]; |
| tensor<int32, [4]> var_1109 = const()[name = tensor<string, []>("op_1109"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1110_cast_fp16 = reshape(shape = var_1109, x = value_19_cast_fp16)[name = tensor<string, []>("op_1110_cast_fp16")]; |
| tensor<bool, []> attn_19_transpose_x_0 = const()[name = tensor<string, []>("attn_19_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_19_transpose_y_0 = const()[name = tensor<string, []>("attn_19_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1110_cast_fp16, y = obj_69_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")]; |
| tensor<int32, [4]> var_1113 = const()[name = tensor<string, []>("op_1113"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_43_cast_fp16 = reshape(shape = var_1113, x = attn_19_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")]; |
| tensor<string, []> obj_67_pad_type_0 = const()[name = tensor<string, []>("obj_67_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_67_strides_0 = const()[name = tensor<string, []>("obj_67_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_67_pad_0 = const()[name = tensor<string, []>("obj_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_67_dilations_0 = const()[name = tensor<string, []>("obj_67_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_67_groups_0 = const()[name = tensor<string, []>("obj_67_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_4_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164236608)))]; |
| tensor<fp16, [768]> layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165416320)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("obj_67_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor<string, []>("inputs_29_cast_fp16")]; |
| tensor<int32, [1]> out_29_axes_0 = const()[name = tensor<string, []>("out_29_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1131_to_fp16 = const()[name = tensor<string, []>("op_1131_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1131_to_fp16, x = inputs_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")]; |
| tensor<fp16, [768]> input_45_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_45_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165417920)))]; |
| tensor<fp16, [768]> input_45_beta_0_to_fp16 = const()[name = tensor<string, []>("input_45_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165419520)))]; |
| tensor<fp16, []> input_45_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_45_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")]; |
| tensor<string, []> input_47_pad_type_0 = const()[name = tensor<string, []>("input_47_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_47_strides_0 = const()[name = tensor<string, []>("input_47_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_47_pad_0 = const()[name = tensor<string, []>("input_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_47_dilations_0 = const()[name = tensor<string, []>("input_47_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_47_groups_0 = const()[name = tensor<string, []>("input_47_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_4_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165421120)))]; |
| tensor<fp16, [3072]> layers_4_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170139776)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")]; |
| tensor<string, []> input_49_mode_0 = const()[name = tensor<string, []>("input_49_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")]; |
| tensor<string, []> hidden_states_11_pad_type_0 = const()[name = tensor<string, []>("hidden_states_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_11_strides_0 = const()[name = tensor<string, []>("hidden_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = tensor<string, []>("hidden_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_11_dilations_0 = const()[name = tensor<string, []>("hidden_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_11_groups_0 = const()[name = tensor<string, []>("hidden_states_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_4_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170145984)))]; |
| tensor<fp16, [768]> layers_4_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174864640)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("hidden_states_11_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor<string, []>("inputs_31_cast_fp16")]; |
| tensor<int32, []> var_1166 = const()[name = tensor<string, []>("op_1166"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_31_axes_0 = const()[name = tensor<string, []>("out_31_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1192_to_fp16 = const()[name = tensor<string, []>("op_1192_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1192_to_fp16, x = inputs_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")]; |
| tensor<fp16, [768]> obj_71_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_71_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174866240)))]; |
| tensor<fp16, [768]> obj_71_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_71_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174867840)))]; |
| tensor<fp16, []> obj_71_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_71_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor<string, []>("obj_71_cast_fp16")]; |
| tensor<string, []> query_21_pad_type_0 = const()[name = tensor<string, []>("query_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_21_strides_0 = const()[name = tensor<string, []>("query_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_21_pad_0 = const()[name = tensor<string, []>("query_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_21_dilations_0 = const()[name = tensor<string, []>("query_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_21_groups_0 = const()[name = tensor<string, []>("query_21_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174869440)))]; |
| tensor<fp16, [768]> layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176049152)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")]; |
| tensor<string, []> current_key_11_pad_type_0 = const()[name = tensor<string, []>("current_key_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_11_strides_0 = const()[name = tensor<string, []>("current_key_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_11_pad_0 = const()[name = tensor<string, []>("current_key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_11_dilations_0 = const()[name = tensor<string, []>("current_key_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_11_groups_0 = const()[name = tensor<string, []>("current_key_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176050752)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_11_cast_fp16 = conv(dilations = current_key_11_dilations_0, groups = current_key_11_groups_0, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = current_key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("current_key_11_cast_fp16")]; |
| tensor<string, []> current_value_11_pad_type_0 = const()[name = tensor<string, []>("current_value_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_11_strides_0 = const()[name = tensor<string, []>("current_value_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_11_pad_0 = const()[name = tensor<string, []>("current_value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_11_dilations_0 = const()[name = tensor<string, []>("current_value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_11_groups_0 = const()[name = tensor<string, []>("current_value_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177230464)))]; |
| tensor<fp16, [768]> layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178410176)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = current_value_11_dilations_0, groups = current_value_11_groups_0, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = current_value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("current_value_11_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1230_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1230_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1232_cast_fp16 = mul(x = var_63_cast_fp16_5, y = var_161_cast_fp16)[name = tensor<string, []>("op_1232_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_21_cast_fp16 = add(x = var_1230_cast_fp16, y = var_1232_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1234_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1234_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1236_cast_fp16 = mul(x = var_78_cast_fp16_5, y = var_161_cast_fp16)[name = tensor<string, []>("op_1236_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_21_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1236_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")]; |
| tensor<int32, [4]> var_1239 = const()[name = tensor<string, []>("op_1239"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_21_cast_fp16 = reshape(shape = var_1239, x = query_21_cast_fp16)[name = tensor<string, []>("mh_q_21_cast_fp16")]; |
| tensor<fp16, []> var_1241_to_fp16 = const()[name = tensor<string, []>("op_1241_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1242_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1241_to_fp16)[name = tensor<string, []>("op_1242_cast_fp16")]; |
| tensor<int32, [4]> var_1243 = const()[name = tensor<string, []>("op_1243"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1244_cast_fp16 = reshape(shape = var_1243, x = key_21_cast_fp16)[name = tensor<string, []>("op_1244_cast_fp16")]; |
| tensor<bool, []> mh_w_31_transpose_x_0 = const()[name = tensor<string, []>("mh_w_31_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_31_transpose_y_0 = const()[name = tensor<string, []>("mh_w_31_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_1242_cast_fp16, y = var_1244_cast_fp16)[name = tensor<string, []>("mh_w_31_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_33_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_1252_cast_fp16 = softmax(axis = var_1166, x = mh_w_33_cast_fp16)[name = tensor<string, []>("op_1252_cast_fp16")]; |
| tensor<int32, [4]> var_1253 = const()[name = tensor<string, []>("op_1253"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1254_cast_fp16 = reshape(shape = var_1253, x = value_21_cast_fp16)[name = tensor<string, []>("op_1254_cast_fp16")]; |
| tensor<bool, []> attn_21_transpose_x_0 = const()[name = tensor<string, []>("attn_21_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_21_transpose_y_0 = const()[name = tensor<string, []>("attn_21_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1254_cast_fp16, y = var_1252_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")]; |
| tensor<int32, [4]> var_1257 = const()[name = tensor<string, []>("op_1257"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_51_cast_fp16 = reshape(shape = var_1257, x = attn_21_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")]; |
| tensor<string, []> obj_77_pad_type_0 = const()[name = tensor<string, []>("obj_77_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_77_strides_0 = const()[name = tensor<string, []>("obj_77_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_77_pad_0 = const()[name = tensor<string, []>("obj_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_77_dilations_0 = const()[name = tensor<string, []>("obj_77_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_77_groups_0 = const()[name = tensor<string, []>("obj_77_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178411776)))]; |
| tensor<fp16, [768]> layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179591488)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_77_dilations_0, groups = obj_77_groups_0, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = obj_77_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("obj_77_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")]; |
| tensor<int32, [1]> out_33_axes_0 = const()[name = tensor<string, []>("out_33_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1279_to_fp16 = const()[name = tensor<string, []>("op_1279_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1279_to_fp16, x = inputs_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")]; |
| tensor<fp16, [768]> obj_79_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_79_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179593088)))]; |
| tensor<fp16, [768]> obj_79_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_79_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179594688)))]; |
| tensor<fp16, []> obj_79_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_79_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor<string, []>("obj_79_cast_fp16")]; |
| tensor<string, []> query_23_pad_type_0 = const()[name = tensor<string, []>("query_23_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_23_strides_0 = const()[name = tensor<string, []>("query_23_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_23_pad_0 = const()[name = tensor<string, []>("query_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_23_dilations_0 = const()[name = tensor<string, []>("query_23_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_23_groups_0 = const()[name = tensor<string, []>("query_23_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179596288)))]; |
| tensor<fp16, [768]> layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180776000)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor<string, []>("query_23_cast_fp16")]; |
| tensor<string, []> key_23_pad_type_0 = const()[name = tensor<string, []>("key_23_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_23_strides_0 = const()[name = tensor<string, []>("key_23_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_23_pad_0 = const()[name = tensor<string, []>("key_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_23_dilations_0 = const()[name = tensor<string, []>("key_23_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_23_groups_0 = const()[name = tensor<string, []>("key_23_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_5_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180777600)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_23_cast_fp16")]; |
| tensor<string, []> value_23_pad_type_0 = const()[name = tensor<string, []>("value_23_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_23_strides_0 = const()[name = tensor<string, []>("value_23_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_23_pad_0 = const()[name = tensor<string, []>("value_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_23_dilations_0 = const()[name = tensor<string, []>("value_23_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_23_groups_0 = const()[name = tensor<string, []>("value_23_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_5_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181957312)))]; |
| tensor<fp16, [768]> layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183137024)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_23_cast_fp16")]; |
| tensor<int32, [4]> var_1314 = const()[name = tensor<string, []>("op_1314"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_23_cast_fp16 = reshape(shape = var_1314, x = query_23_cast_fp16)[name = tensor<string, []>("mh_q_23_cast_fp16")]; |
| tensor<fp16, []> var_1316_to_fp16 = const()[name = tensor<string, []>("op_1316_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1317_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1316_to_fp16)[name = tensor<string, []>("op_1317_cast_fp16")]; |
| tensor<int32, [4]> var_1318 = const()[name = tensor<string, []>("op_1318"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1319_cast_fp16 = reshape(shape = var_1318, x = key_23_cast_fp16)[name = tensor<string, []>("op_1319_cast_fp16")]; |
| tensor<bool, []> mh_w_35_transpose_x_0 = const()[name = tensor<string, []>("mh_w_35_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_35_transpose_y_0 = const()[name = tensor<string, []>("mh_w_35_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_1317_cast_fp16, y = var_1319_cast_fp16)[name = tensor<string, []>("mh_w_35_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_83_cast_fp16 = softmax(axis = var_1166, x = mh_w_35_cast_fp16)[name = tensor<string, []>("obj_83_cast_fp16")]; |
| tensor<int32, [4]> var_1323 = const()[name = tensor<string, []>("op_1323"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1324_cast_fp16 = reshape(shape = var_1323, x = value_23_cast_fp16)[name = tensor<string, []>("op_1324_cast_fp16")]; |
| tensor<bool, []> attn_23_transpose_x_0 = const()[name = tensor<string, []>("attn_23_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_23_transpose_y_0 = const()[name = tensor<string, []>("attn_23_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1324_cast_fp16, y = obj_83_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")]; |
| tensor<int32, [4]> var_1327 = const()[name = tensor<string, []>("op_1327"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_53_cast_fp16 = reshape(shape = var_1327, x = attn_23_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")]; |
| tensor<string, []> obj_81_pad_type_0 = const()[name = tensor<string, []>("obj_81_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_81_strides_0 = const()[name = tensor<string, []>("obj_81_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_81_pad_0 = const()[name = tensor<string, []>("obj_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_81_dilations_0 = const()[name = tensor<string, []>("obj_81_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_81_groups_0 = const()[name = tensor<string, []>("obj_81_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183138624)))]; |
| tensor<fp16, [768]> layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184318336)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = obj_81_dilations_0, groups = obj_81_groups_0, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = obj_81_strides_0, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("obj_81_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor<string, []>("inputs_35_cast_fp16")]; |
| tensor<int32, [1]> out_35_axes_0 = const()[name = tensor<string, []>("out_35_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1348_to_fp16 = const()[name = tensor<string, []>("op_1348_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1348_to_fp16, x = inputs_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")]; |
| tensor<fp16, [768]> input_55_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_55_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184319936)))]; |
| tensor<fp16, [768]> input_55_beta_0_to_fp16 = const()[name = tensor<string, []>("input_55_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184321536)))]; |
| tensor<fp16, []> input_55_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_55_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_55_cast_fp16 = batch_norm(beta = input_55_beta_0_to_fp16, epsilon = input_55_epsilon_0_to_fp16, gamma = input_55_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")]; |
| tensor<string, []> input_57_pad_type_0 = const()[name = tensor<string, []>("input_57_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_57_strides_0 = const()[name = tensor<string, []>("input_57_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_57_pad_0 = const()[name = tensor<string, []>("input_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_57_dilations_0 = const()[name = tensor<string, []>("input_57_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_57_groups_0 = const()[name = tensor<string, []>("input_57_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_5_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184323136)))]; |
| tensor<fp16, [3072]> layers_5_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189041792)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")]; |
| tensor<string, []> input_59_mode_0 = const()[name = tensor<string, []>("input_59_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")]; |
| tensor<string, []> hidden_states_13_pad_type_0 = const()[name = tensor<string, []>("hidden_states_13_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = tensor<string, []>("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = tensor<string, []>("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = tensor<string, []>("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_13_groups_0 = const()[name = tensor<string, []>("hidden_states_13_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_5_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189048000)))]; |
| tensor<fp16, [768]> layers_5_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193766656)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor<string, []>("inputs_37_cast_fp16")]; |
| tensor<int32, []> var_1384 = const()[name = tensor<string, []>("op_1384"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_37_axes_0 = const()[name = tensor<string, []>("out_37_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1410_to_fp16 = const()[name = tensor<string, []>("op_1410_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1410_to_fp16, x = inputs_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")]; |
| tensor<fp16, [768]> obj_85_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_85_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193768256)))]; |
| tensor<fp16, [768]> obj_85_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_85_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193769856)))]; |
| tensor<fp16, []> obj_85_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_85_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor<string, []>("obj_85_cast_fp16")]; |
| tensor<string, []> query_25_pad_type_0 = const()[name = tensor<string, []>("query_25_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_25_strides_0 = const()[name = tensor<string, []>("query_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_25_pad_0 = const()[name = tensor<string, []>("query_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_25_dilations_0 = const()[name = tensor<string, []>("query_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_25_groups_0 = const()[name = tensor<string, []>("query_25_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193771456)))]; |
| tensor<fp16, [768]> layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194951168)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")]; |
| tensor<string, []> current_key_13_pad_type_0 = const()[name = tensor<string, []>("current_key_13_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_13_strides_0 = const()[name = tensor<string, []>("current_key_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_13_pad_0 = const()[name = tensor<string, []>("current_key_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_13_dilations_0 = const()[name = tensor<string, []>("current_key_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_13_groups_0 = const()[name = tensor<string, []>("current_key_13_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194952768)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_13_cast_fp16 = conv(dilations = current_key_13_dilations_0, groups = current_key_13_groups_0, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = current_key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("current_key_13_cast_fp16")]; |
| tensor<string, []> current_value_13_pad_type_0 = const()[name = tensor<string, []>("current_value_13_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_13_strides_0 = const()[name = tensor<string, []>("current_value_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_13_pad_0 = const()[name = tensor<string, []>("current_value_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_13_dilations_0 = const()[name = tensor<string, []>("current_value_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_13_groups_0 = const()[name = tensor<string, []>("current_value_13_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196132480)))]; |
| tensor<fp16, [768]> layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197312192)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = current_value_13_dilations_0, groups = current_value_13_groups_0, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = current_value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("current_value_13_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1448_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1448_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1450_cast_fp16 = mul(x = var_63_cast_fp16_6, y = var_161_cast_fp16)[name = tensor<string, []>("op_1450_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_25_cast_fp16 = add(x = var_1448_cast_fp16, y = var_1450_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1452_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1452_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1454_cast_fp16 = mul(x = var_78_cast_fp16_6, y = var_161_cast_fp16)[name = tensor<string, []>("op_1454_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_25_cast_fp16 = add(x = var_1452_cast_fp16, y = var_1454_cast_fp16)[name = tensor<string, []>("value_25_cast_fp16")]; |
| tensor<int32, [4]> var_1457 = const()[name = tensor<string, []>("op_1457"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_25_cast_fp16 = reshape(shape = var_1457, x = query_25_cast_fp16)[name = tensor<string, []>("mh_q_25_cast_fp16")]; |
| tensor<fp16, []> var_1459_to_fp16 = const()[name = tensor<string, []>("op_1459_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1460_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1459_to_fp16)[name = tensor<string, []>("op_1460_cast_fp16")]; |
| tensor<int32, [4]> var_1461 = const()[name = tensor<string, []>("op_1461"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1462_cast_fp16 = reshape(shape = var_1461, x = key_25_cast_fp16)[name = tensor<string, []>("op_1462_cast_fp16")]; |
| tensor<bool, []> mh_w_37_transpose_x_0 = const()[name = tensor<string, []>("mh_w_37_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_37_transpose_y_0 = const()[name = tensor<string, []>("mh_w_37_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1460_cast_fp16, y = var_1462_cast_fp16)[name = tensor<string, []>("mh_w_37_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_39_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_1470_cast_fp16 = softmax(axis = var_1384, x = mh_w_39_cast_fp16)[name = tensor<string, []>("op_1470_cast_fp16")]; |
| tensor<int32, [4]> var_1471 = const()[name = tensor<string, []>("op_1471"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1472_cast_fp16 = reshape(shape = var_1471, x = value_25_cast_fp16)[name = tensor<string, []>("op_1472_cast_fp16")]; |
| tensor<bool, []> attn_25_transpose_x_0 = const()[name = tensor<string, []>("attn_25_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_25_transpose_y_0 = const()[name = tensor<string, []>("attn_25_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1472_cast_fp16, y = var_1470_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")]; |
| tensor<int32, [4]> var_1475 = const()[name = tensor<string, []>("op_1475"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_61_cast_fp16 = reshape(shape = var_1475, x = attn_25_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")]; |
| tensor<string, []> obj_91_pad_type_0 = const()[name = tensor<string, []>("obj_91_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_91_strides_0 = const()[name = tensor<string, []>("obj_91_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_91_pad_0 = const()[name = tensor<string, []>("obj_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_91_dilations_0 = const()[name = tensor<string, []>("obj_91_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_91_groups_0 = const()[name = tensor<string, []>("obj_91_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197313792)))]; |
| tensor<fp16, [768]> layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198493504)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_91_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("obj_91_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor<string, []>("inputs_39_cast_fp16")]; |
| tensor<int32, [1]> out_39_axes_0 = const()[name = tensor<string, []>("out_39_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1497_to_fp16 = const()[name = tensor<string, []>("op_1497_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1497_to_fp16, x = inputs_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")]; |
| tensor<fp16, [768]> obj_93_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_93_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198495104)))]; |
| tensor<fp16, [768]> obj_93_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_93_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198496704)))]; |
| tensor<fp16, []> obj_93_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_93_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor<string, []>("obj_93_cast_fp16")]; |
| tensor<string, []> query_27_pad_type_0 = const()[name = tensor<string, []>("query_27_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_27_strides_0 = const()[name = tensor<string, []>("query_27_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_27_pad_0 = const()[name = tensor<string, []>("query_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_27_dilations_0 = const()[name = tensor<string, []>("query_27_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_27_groups_0 = const()[name = tensor<string, []>("query_27_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_6_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198498304)))]; |
| tensor<fp16, [768]> layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199678016)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor<string, []>("query_27_cast_fp16")]; |
| tensor<string, []> key_27_pad_type_0 = const()[name = tensor<string, []>("key_27_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_27_strides_0 = const()[name = tensor<string, []>("key_27_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_27_pad_0 = const()[name = tensor<string, []>("key_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_27_dilations_0 = const()[name = tensor<string, []>("key_27_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_27_groups_0 = const()[name = tensor<string, []>("key_27_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_6_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199679616)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_27_cast_fp16")]; |
| tensor<string, []> value_27_pad_type_0 = const()[name = tensor<string, []>("value_27_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_27_strides_0 = const()[name = tensor<string, []>("value_27_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_27_pad_0 = const()[name = tensor<string, []>("value_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_27_dilations_0 = const()[name = tensor<string, []>("value_27_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_27_groups_0 = const()[name = tensor<string, []>("value_27_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_6_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200859328)))]; |
| tensor<fp16, [768]> layers_6_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202039040)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_27_cast_fp16")]; |
| tensor<int32, [4]> var_1532 = const()[name = tensor<string, []>("op_1532"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_27_cast_fp16 = reshape(shape = var_1532, x = query_27_cast_fp16)[name = tensor<string, []>("mh_q_27_cast_fp16")]; |
| tensor<fp16, []> var_1534_to_fp16 = const()[name = tensor<string, []>("op_1534_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1535_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1534_to_fp16)[name = tensor<string, []>("op_1535_cast_fp16")]; |
| tensor<int32, [4]> var_1536 = const()[name = tensor<string, []>("op_1536"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1537_cast_fp16 = reshape(shape = var_1536, x = key_27_cast_fp16)[name = tensor<string, []>("op_1537_cast_fp16")]; |
| tensor<bool, []> mh_w_41_transpose_x_0 = const()[name = tensor<string, []>("mh_w_41_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_41_transpose_y_0 = const()[name = tensor<string, []>("mh_w_41_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1535_cast_fp16, y = var_1537_cast_fp16)[name = tensor<string, []>("mh_w_41_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_97_cast_fp16 = softmax(axis = var_1384, x = mh_w_41_cast_fp16)[name = tensor<string, []>("obj_97_cast_fp16")]; |
| tensor<int32, [4]> var_1541 = const()[name = tensor<string, []>("op_1541"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1542_cast_fp16 = reshape(shape = var_1541, x = value_27_cast_fp16)[name = tensor<string, []>("op_1542_cast_fp16")]; |
| tensor<bool, []> attn_27_transpose_x_0 = const()[name = tensor<string, []>("attn_27_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_27_transpose_y_0 = const()[name = tensor<string, []>("attn_27_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1542_cast_fp16, y = obj_97_cast_fp16)[name = tensor<string, []>("attn_27_cast_fp16")]; |
| tensor<int32, [4]> var_1545 = const()[name = tensor<string, []>("op_1545"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_63_cast_fp16 = reshape(shape = var_1545, x = attn_27_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")]; |
| tensor<string, []> obj_95_pad_type_0 = const()[name = tensor<string, []>("obj_95_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_95_strides_0 = const()[name = tensor<string, []>("obj_95_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_95_pad_0 = const()[name = tensor<string, []>("obj_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_95_dilations_0 = const()[name = tensor<string, []>("obj_95_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_95_groups_0 = const()[name = tensor<string, []>("obj_95_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_6_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202040640)))]; |
| tensor<fp16, [768]> layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203220352)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_95_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("obj_95_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor<string, []>("inputs_41_cast_fp16")]; |
| tensor<int32, [1]> out_41_axes_0 = const()[name = tensor<string, []>("out_41_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1563_to_fp16 = const()[name = tensor<string, []>("op_1563_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1563_to_fp16, x = inputs_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")]; |
| tensor<fp16, [768]> input_65_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_65_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203221952)))]; |
| tensor<fp16, [768]> input_65_beta_0_to_fp16 = const()[name = tensor<string, []>("input_65_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203223552)))]; |
| tensor<fp16, []> input_65_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_65_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_65_cast_fp16 = batch_norm(beta = input_65_beta_0_to_fp16, epsilon = input_65_epsilon_0_to_fp16, gamma = input_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")]; |
| tensor<string, []> input_67_pad_type_0 = const()[name = tensor<string, []>("input_67_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_67_strides_0 = const()[name = tensor<string, []>("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_67_pad_0 = const()[name = tensor<string, []>("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_67_dilations_0 = const()[name = tensor<string, []>("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_67_groups_0 = const()[name = tensor<string, []>("input_67_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_6_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203225152)))]; |
| tensor<fp16, [3072]> layers_6_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207943808)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")]; |
| tensor<string, []> input_69_mode_0 = const()[name = tensor<string, []>("input_69_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")]; |
| tensor<string, []> hidden_states_15_pad_type_0 = const()[name = tensor<string, []>("hidden_states_15_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_15_strides_0 = const()[name = tensor<string, []>("hidden_states_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = tensor<string, []>("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_15_dilations_0 = const()[name = tensor<string, []>("hidden_states_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_15_groups_0 = const()[name = tensor<string, []>("hidden_states_15_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_6_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207950016)))]; |
| tensor<fp16, [768]> layers_6_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212668672)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("hidden_states_15_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")]; |
| tensor<int32, []> var_1598 = const()[name = tensor<string, []>("op_1598"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_43_axes_0 = const()[name = tensor<string, []>("out_43_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1624_to_fp16 = const()[name = tensor<string, []>("op_1624_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1624_to_fp16, x = inputs_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")]; |
| tensor<fp16, [768]> obj_99_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_99_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212670272)))]; |
| tensor<fp16, [768]> obj_99_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_99_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212671872)))]; |
| tensor<fp16, []> obj_99_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_99_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor<string, []>("obj_99_cast_fp16")]; |
| tensor<string, []> query_29_pad_type_0 = const()[name = tensor<string, []>("query_29_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_29_strides_0 = const()[name = tensor<string, []>("query_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_29_pad_0 = const()[name = tensor<string, []>("query_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_29_dilations_0 = const()[name = tensor<string, []>("query_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_29_groups_0 = const()[name = tensor<string, []>("query_29_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212673472)))]; |
| tensor<fp16, [768]> layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213853184)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")]; |
| tensor<string, []> current_key_15_pad_type_0 = const()[name = tensor<string, []>("current_key_15_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_15_strides_0 = const()[name = tensor<string, []>("current_key_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_15_pad_0 = const()[name = tensor<string, []>("current_key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_15_dilations_0 = const()[name = tensor<string, []>("current_key_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_15_groups_0 = const()[name = tensor<string, []>("current_key_15_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213854784)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_15_cast_fp16 = conv(dilations = current_key_15_dilations_0, groups = current_key_15_groups_0, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = current_key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor<string, []>("current_key_15_cast_fp16")]; |
| tensor<string, []> current_value_15_pad_type_0 = const()[name = tensor<string, []>("current_value_15_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_15_strides_0 = const()[name = tensor<string, []>("current_value_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_15_pad_0 = const()[name = tensor<string, []>("current_value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_15_dilations_0 = const()[name = tensor<string, []>("current_value_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_15_groups_0 = const()[name = tensor<string, []>("current_value_15_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215034496)))]; |
| tensor<fp16, [768]> layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216214208)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = current_value_15_dilations_0, groups = current_value_15_groups_0, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = current_value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor<string, []>("current_value_15_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1662_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1662_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1664_cast_fp16 = mul(x = var_63_cast_fp16_7, y = var_161_cast_fp16)[name = tensor<string, []>("op_1664_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_29_cast_fp16 = add(x = var_1662_cast_fp16, y = var_1664_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1666_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1666_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1668_cast_fp16 = mul(x = var_78_cast_fp16_7, y = var_161_cast_fp16)[name = tensor<string, []>("op_1668_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_29_cast_fp16 = add(x = var_1666_cast_fp16, y = var_1668_cast_fp16)[name = tensor<string, []>("value_29_cast_fp16")]; |
| tensor<int32, [4]> var_1671 = const()[name = tensor<string, []>("op_1671"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_29_cast_fp16 = reshape(shape = var_1671, x = query_29_cast_fp16)[name = tensor<string, []>("mh_q_29_cast_fp16")]; |
| tensor<fp16, []> var_1673_to_fp16 = const()[name = tensor<string, []>("op_1673_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1674_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1673_to_fp16)[name = tensor<string, []>("op_1674_cast_fp16")]; |
| tensor<int32, [4]> var_1675 = const()[name = tensor<string, []>("op_1675"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1676_cast_fp16 = reshape(shape = var_1675, x = key_29_cast_fp16)[name = tensor<string, []>("op_1676_cast_fp16")]; |
| tensor<bool, []> mh_w_43_transpose_x_0 = const()[name = tensor<string, []>("mh_w_43_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_43_transpose_y_0 = const()[name = tensor<string, []>("mh_w_43_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_1674_cast_fp16, y = var_1676_cast_fp16)[name = tensor<string, []>("mh_w_43_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_45_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_1684_cast_fp16 = softmax(axis = var_1598, x = mh_w_45_cast_fp16)[name = tensor<string, []>("op_1684_cast_fp16")]; |
| tensor<int32, [4]> var_1685 = const()[name = tensor<string, []>("op_1685"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1686_cast_fp16 = reshape(shape = var_1685, x = value_29_cast_fp16)[name = tensor<string, []>("op_1686_cast_fp16")]; |
| tensor<bool, []> attn_29_transpose_x_0 = const()[name = tensor<string, []>("attn_29_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_29_transpose_y_0 = const()[name = tensor<string, []>("attn_29_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1686_cast_fp16, y = var_1684_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")]; |
| tensor<int32, [4]> var_1689 = const()[name = tensor<string, []>("op_1689"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_71_cast_fp16 = reshape(shape = var_1689, x = attn_29_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")]; |
| tensor<string, []> obj_105_pad_type_0 = const()[name = tensor<string, []>("obj_105_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_105_strides_0 = const()[name = tensor<string, []>("obj_105_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_105_pad_0 = const()[name = tensor<string, []>("obj_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_105_dilations_0 = const()[name = tensor<string, []>("obj_105_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_105_groups_0 = const()[name = tensor<string, []>("obj_105_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216215808)))]; |
| tensor<fp16, [768]> layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217395520)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_105_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_105_dilations_0, groups = obj_105_groups_0, pad = obj_105_pad_0, pad_type = obj_105_pad_type_0, strides = obj_105_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("obj_105_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor<string, []>("inputs_45_cast_fp16")]; |
| tensor<int32, [1]> out_45_axes_0 = const()[name = tensor<string, []>("out_45_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1711_to_fp16 = const()[name = tensor<string, []>("op_1711_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1711_to_fp16, x = inputs_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")]; |
| tensor<fp16, [768]> obj_107_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_107_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217397120)))]; |
| tensor<fp16, [768]> obj_107_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_107_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217398720)))]; |
| tensor<fp16, []> obj_107_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_107_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor<string, []>("obj_107_cast_fp16")]; |
| tensor<string, []> query_31_pad_type_0 = const()[name = tensor<string, []>("query_31_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_31_strides_0 = const()[name = tensor<string, []>("query_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_31_pad_0 = const()[name = tensor<string, []>("query_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_31_dilations_0 = const()[name = tensor<string, []>("query_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_31_groups_0 = const()[name = tensor<string, []>("query_31_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_7_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217400320)))]; |
| tensor<fp16, [768]> layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218580032)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor<string, []>("query_31_cast_fp16")]; |
| tensor<string, []> key_31_pad_type_0 = const()[name = tensor<string, []>("key_31_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_31_strides_0 = const()[name = tensor<string, []>("key_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_31_pad_0 = const()[name = tensor<string, []>("key_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_31_dilations_0 = const()[name = tensor<string, []>("key_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_31_groups_0 = const()[name = tensor<string, []>("key_31_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_7_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218581632)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_31_cast_fp16")]; |
| tensor<string, []> value_31_pad_type_0 = const()[name = tensor<string, []>("value_31_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_31_strides_0 = const()[name = tensor<string, []>("value_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_31_pad_0 = const()[name = tensor<string, []>("value_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_31_dilations_0 = const()[name = tensor<string, []>("value_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_31_groups_0 = const()[name = tensor<string, []>("value_31_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_7_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219761344)))]; |
| tensor<fp16, [768]> layers_7_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220941056)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_31_cast_fp16")]; |
| tensor<int32, [4]> var_1746 = const()[name = tensor<string, []>("op_1746"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_31_cast_fp16 = reshape(shape = var_1746, x = query_31_cast_fp16)[name = tensor<string, []>("mh_q_31_cast_fp16")]; |
| tensor<fp16, []> var_1748_to_fp16 = const()[name = tensor<string, []>("op_1748_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1749_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_1748_to_fp16)[name = tensor<string, []>("op_1749_cast_fp16")]; |
| tensor<int32, [4]> var_1750 = const()[name = tensor<string, []>("op_1750"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1751_cast_fp16 = reshape(shape = var_1750, x = key_31_cast_fp16)[name = tensor<string, []>("op_1751_cast_fp16")]; |
| tensor<bool, []> mh_w_47_transpose_x_0 = const()[name = tensor<string, []>("mh_w_47_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_47_transpose_y_0 = const()[name = tensor<string, []>("mh_w_47_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_1749_cast_fp16, y = var_1751_cast_fp16)[name = tensor<string, []>("mh_w_47_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_111_cast_fp16 = softmax(axis = var_1598, x = mh_w_47_cast_fp16)[name = tensor<string, []>("obj_111_cast_fp16")]; |
| tensor<int32, [4]> var_1755 = const()[name = tensor<string, []>("op_1755"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1756_cast_fp16 = reshape(shape = var_1755, x = value_31_cast_fp16)[name = tensor<string, []>("op_1756_cast_fp16")]; |
| tensor<bool, []> attn_31_transpose_x_0 = const()[name = tensor<string, []>("attn_31_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_31_transpose_y_0 = const()[name = tensor<string, []>("attn_31_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_1756_cast_fp16, y = obj_111_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")]; |
| tensor<int32, [4]> var_1759 = const()[name = tensor<string, []>("op_1759"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_73_cast_fp16 = reshape(shape = var_1759, x = attn_31_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")]; |
| tensor<string, []> obj_109_pad_type_0 = const()[name = tensor<string, []>("obj_109_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_109_strides_0 = const()[name = tensor<string, []>("obj_109_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_109_pad_0 = const()[name = tensor<string, []>("obj_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_109_dilations_0 = const()[name = tensor<string, []>("obj_109_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_109_groups_0 = const()[name = tensor<string, []>("obj_109_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_7_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220942656)))]; |
| tensor<fp16, [768]> layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222122368)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_109_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = obj_109_dilations_0, groups = obj_109_groups_0, pad = obj_109_pad_0, pad_type = obj_109_pad_type_0, strides = obj_109_strides_0, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("obj_109_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor<string, []>("inputs_47_cast_fp16")]; |
| tensor<int32, [1]> out_47_axes_0 = const()[name = tensor<string, []>("out_47_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1777_to_fp16 = const()[name = tensor<string, []>("op_1777_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1777_to_fp16, x = inputs_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")]; |
| tensor<fp16, [768]> input_75_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_75_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222123968)))]; |
| tensor<fp16, [768]> input_75_beta_0_to_fp16 = const()[name = tensor<string, []>("input_75_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222125568)))]; |
| tensor<fp16, []> input_75_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_75_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")]; |
| tensor<string, []> input_77_pad_type_0 = const()[name = tensor<string, []>("input_77_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_77_strides_0 = const()[name = tensor<string, []>("input_77_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_77_pad_0 = const()[name = tensor<string, []>("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_77_dilations_0 = const()[name = tensor<string, []>("input_77_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_77_groups_0 = const()[name = tensor<string, []>("input_77_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_7_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222127168)))]; |
| tensor<fp16, [3072]> layers_7_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226845824)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")]; |
| tensor<string, []> input_79_mode_0 = const()[name = tensor<string, []>("input_79_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")]; |
| tensor<string, []> hidden_states_17_pad_type_0 = const()[name = tensor<string, []>("hidden_states_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = tensor<string, []>("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = tensor<string, []>("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = tensor<string, []>("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_17_groups_0 = const()[name = tensor<string, []>("hidden_states_17_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_7_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226852032)))]; |
| tensor<fp16, [768]> layers_7_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231570688)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("hidden_states_17_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor<string, []>("inputs_49_cast_fp16")]; |
| tensor<int32, []> var_1812 = const()[name = tensor<string, []>("op_1812"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_49_axes_0 = const()[name = tensor<string, []>("out_49_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1838_to_fp16 = const()[name = tensor<string, []>("op_1838_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1838_to_fp16, x = inputs_49_cast_fp16)[name = tensor<string, []>("out_49_cast_fp16")]; |
| tensor<fp16, [768]> obj_113_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_113_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231572288)))]; |
| tensor<fp16, [768]> obj_113_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_113_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231573888)))]; |
| tensor<fp16, []> obj_113_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_113_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor<string, []>("obj_113_cast_fp16")]; |
| tensor<string, []> query_33_pad_type_0 = const()[name = tensor<string, []>("query_33_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_33_strides_0 = const()[name = tensor<string, []>("query_33_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_33_pad_0 = const()[name = tensor<string, []>("query_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_33_dilations_0 = const()[name = tensor<string, []>("query_33_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_33_groups_0 = const()[name = tensor<string, []>("query_33_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231575488)))]; |
| tensor<fp16, [768]> layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232755200)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")]; |
| tensor<string, []> current_key_17_pad_type_0 = const()[name = tensor<string, []>("current_key_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_17_strides_0 = const()[name = tensor<string, []>("current_key_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_17_pad_0 = const()[name = tensor<string, []>("current_key_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_17_dilations_0 = const()[name = tensor<string, []>("current_key_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_17_groups_0 = const()[name = tensor<string, []>("current_key_17_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232756800)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_17_cast_fp16 = conv(dilations = current_key_17_dilations_0, groups = current_key_17_groups_0, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = current_key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("current_key_17_cast_fp16")]; |
| tensor<string, []> current_value_17_pad_type_0 = const()[name = tensor<string, []>("current_value_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_17_strides_0 = const()[name = tensor<string, []>("current_value_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_17_pad_0 = const()[name = tensor<string, []>("current_value_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_17_dilations_0 = const()[name = tensor<string, []>("current_value_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_17_groups_0 = const()[name = tensor<string, []>("current_value_17_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233936512)))]; |
| tensor<fp16, [768]> layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235116224)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = current_value_17_dilations_0, groups = current_value_17_groups_0, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = current_value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("current_value_17_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1876_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1876_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1878_cast_fp16 = mul(x = var_63_cast_fp16_8, y = var_161_cast_fp16)[name = tensor<string, []>("op_1878_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_33_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1878_cast_fp16)[name = tensor<string, []>("key_33_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1880_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_1880_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_1882_cast_fp16 = mul(x = var_78_cast_fp16_8, y = var_161_cast_fp16)[name = tensor<string, []>("op_1882_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_33_cast_fp16 = add(x = var_1880_cast_fp16, y = var_1882_cast_fp16)[name = tensor<string, []>("value_33_cast_fp16")]; |
| tensor<int32, [4]> var_1885 = const()[name = tensor<string, []>("op_1885"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_33_cast_fp16 = reshape(shape = var_1885, x = query_33_cast_fp16)[name = tensor<string, []>("mh_q_33_cast_fp16")]; |
| tensor<fp16, []> var_1887_to_fp16 = const()[name = tensor<string, []>("op_1887_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1888_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_1887_to_fp16)[name = tensor<string, []>("op_1888_cast_fp16")]; |
| tensor<int32, [4]> var_1889 = const()[name = tensor<string, []>("op_1889"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1890_cast_fp16 = reshape(shape = var_1889, x = key_33_cast_fp16)[name = tensor<string, []>("op_1890_cast_fp16")]; |
| tensor<bool, []> mh_w_49_transpose_x_0 = const()[name = tensor<string, []>("mh_w_49_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_49_transpose_y_0 = const()[name = tensor<string, []>("mh_w_49_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_1888_cast_fp16, y = var_1890_cast_fp16)[name = tensor<string, []>("mh_w_49_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_51_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_1898_cast_fp16 = softmax(axis = var_1812, x = mh_w_51_cast_fp16)[name = tensor<string, []>("op_1898_cast_fp16")]; |
| tensor<int32, [4]> var_1899 = const()[name = tensor<string, []>("op_1899"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_1900_cast_fp16 = reshape(shape = var_1899, x = value_33_cast_fp16)[name = tensor<string, []>("op_1900_cast_fp16")]; |
| tensor<bool, []> attn_33_transpose_x_0 = const()[name = tensor<string, []>("attn_33_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_33_transpose_y_0 = const()[name = tensor<string, []>("attn_33_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_1900_cast_fp16, y = var_1898_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")]; |
| tensor<int32, [4]> var_1903 = const()[name = tensor<string, []>("op_1903"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_81_cast_fp16 = reshape(shape = var_1903, x = attn_33_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")]; |
| tensor<string, []> obj_119_pad_type_0 = const()[name = tensor<string, []>("obj_119_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_119_strides_0 = const()[name = tensor<string, []>("obj_119_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_119_pad_0 = const()[name = tensor<string, []>("obj_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_119_dilations_0 = const()[name = tensor<string, []>("obj_119_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_119_groups_0 = const()[name = tensor<string, []>("obj_119_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235117824)))]; |
| tensor<fp16, [768]> layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236297536)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_119_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("obj_119_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor<string, []>("inputs_51_cast_fp16")]; |
| tensor<int32, [1]> out_51_axes_0 = const()[name = tensor<string, []>("out_51_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1925_to_fp16 = const()[name = tensor<string, []>("op_1925_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1925_to_fp16, x = inputs_51_cast_fp16)[name = tensor<string, []>("out_51_cast_fp16")]; |
| tensor<fp16, [768]> obj_121_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_121_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236299136)))]; |
| tensor<fp16, [768]> obj_121_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_121_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236300736)))]; |
| tensor<fp16, []> obj_121_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_121_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor<string, []>("obj_121_cast_fp16")]; |
| tensor<string, []> query_35_pad_type_0 = const()[name = tensor<string, []>("query_35_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_35_strides_0 = const()[name = tensor<string, []>("query_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_35_pad_0 = const()[name = tensor<string, []>("query_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_35_dilations_0 = const()[name = tensor<string, []>("query_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_35_groups_0 = const()[name = tensor<string, []>("query_35_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_8_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236302336)))]; |
| tensor<fp16, [768]> layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237482048)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor<string, []>("query_35_cast_fp16")]; |
| tensor<string, []> key_35_pad_type_0 = const()[name = tensor<string, []>("key_35_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_35_strides_0 = const()[name = tensor<string, []>("key_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_35_pad_0 = const()[name = tensor<string, []>("key_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_35_dilations_0 = const()[name = tensor<string, []>("key_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_35_groups_0 = const()[name = tensor<string, []>("key_35_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_8_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237483648)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_35_cast_fp16")]; |
| tensor<string, []> value_35_pad_type_0 = const()[name = tensor<string, []>("value_35_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_35_strides_0 = const()[name = tensor<string, []>("value_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_35_pad_0 = const()[name = tensor<string, []>("value_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_35_dilations_0 = const()[name = tensor<string, []>("value_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_35_groups_0 = const()[name = tensor<string, []>("value_35_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_8_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238663360)))]; |
| tensor<fp16, [768]> layers_8_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239843072)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_35_cast_fp16")]; |
| tensor<int32, [4]> var_1960 = const()[name = tensor<string, []>("op_1960"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_35_cast_fp16 = reshape(shape = var_1960, x = query_35_cast_fp16)[name = tensor<string, []>("mh_q_35_cast_fp16")]; |
| tensor<fp16, []> var_1962_to_fp16 = const()[name = tensor<string, []>("op_1962_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_1963_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_1962_to_fp16)[name = tensor<string, []>("op_1963_cast_fp16")]; |
| tensor<int32, [4]> var_1964 = const()[name = tensor<string, []>("op_1964"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1965_cast_fp16 = reshape(shape = var_1964, x = key_35_cast_fp16)[name = tensor<string, []>("op_1965_cast_fp16")]; |
| tensor<bool, []> mh_w_53_transpose_x_0 = const()[name = tensor<string, []>("mh_w_53_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_53_transpose_y_0 = const()[name = tensor<string, []>("mh_w_53_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_1963_cast_fp16, y = var_1965_cast_fp16)[name = tensor<string, []>("mh_w_53_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_125_cast_fp16 = softmax(axis = var_1812, x = mh_w_53_cast_fp16)[name = tensor<string, []>("obj_125_cast_fp16")]; |
| tensor<int32, [4]> var_1969 = const()[name = tensor<string, []>("op_1969"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_1970_cast_fp16 = reshape(shape = var_1969, x = value_35_cast_fp16)[name = tensor<string, []>("op_1970_cast_fp16")]; |
| tensor<bool, []> attn_35_transpose_x_0 = const()[name = tensor<string, []>("attn_35_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_35_transpose_y_0 = const()[name = tensor<string, []>("attn_35_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_1970_cast_fp16, y = obj_125_cast_fp16)[name = tensor<string, []>("attn_35_cast_fp16")]; |
| tensor<int32, [4]> var_1973 = const()[name = tensor<string, []>("op_1973"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_83_cast_fp16 = reshape(shape = var_1973, x = attn_35_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")]; |
| tensor<string, []> obj_123_pad_type_0 = const()[name = tensor<string, []>("obj_123_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_123_strides_0 = const()[name = tensor<string, []>("obj_123_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_123_pad_0 = const()[name = tensor<string, []>("obj_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_123_dilations_0 = const()[name = tensor<string, []>("obj_123_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_123_groups_0 = const()[name = tensor<string, []>("obj_123_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_8_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239844672)))]; |
| tensor<fp16, [768]> layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241024384)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_123_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("obj_123_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor<string, []>("inputs_53_cast_fp16")]; |
| tensor<int32, [1]> out_53_axes_0 = const()[name = tensor<string, []>("out_53_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_1994_to_fp16 = const()[name = tensor<string, []>("op_1994_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1994_to_fp16, x = inputs_53_cast_fp16)[name = tensor<string, []>("out_53_cast_fp16")]; |
| tensor<fp16, [768]> input_85_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_85_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241025984)))]; |
| tensor<fp16, [768]> input_85_beta_0_to_fp16 = const()[name = tensor<string, []>("input_85_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241027584)))]; |
| tensor<fp16, []> input_85_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_85_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_85_cast_fp16 = batch_norm(beta = input_85_beta_0_to_fp16, epsilon = input_85_epsilon_0_to_fp16, gamma = input_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")]; |
| tensor<string, []> input_87_pad_type_0 = const()[name = tensor<string, []>("input_87_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_87_strides_0 = const()[name = tensor<string, []>("input_87_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_87_pad_0 = const()[name = tensor<string, []>("input_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_87_dilations_0 = const()[name = tensor<string, []>("input_87_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_87_groups_0 = const()[name = tensor<string, []>("input_87_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_8_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241029184)))]; |
| tensor<fp16, [3072]> layers_8_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245747840)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")]; |
| tensor<string, []> input_89_mode_0 = const()[name = tensor<string, []>("input_89_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")]; |
| tensor<string, []> hidden_states_19_pad_type_0 = const()[name = tensor<string, []>("hidden_states_19_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_19_strides_0 = const()[name = tensor<string, []>("hidden_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_19_pad_0 = const()[name = tensor<string, []>("hidden_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_19_dilations_0 = const()[name = tensor<string, []>("hidden_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_19_groups_0 = const()[name = tensor<string, []>("hidden_states_19_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_8_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245754048)))]; |
| tensor<fp16, [768]> layers_8_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250472704)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("hidden_states_19_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor<string, []>("inputs_55_cast_fp16")]; |
| tensor<int32, []> var_2030 = const()[name = tensor<string, []>("op_2030"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_55_axes_0 = const()[name = tensor<string, []>("out_55_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2056_to_fp16 = const()[name = tensor<string, []>("op_2056_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_2056_to_fp16, x = inputs_55_cast_fp16)[name = tensor<string, []>("out_55_cast_fp16")]; |
| tensor<fp16, [768]> obj_127_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_127_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250474304)))]; |
| tensor<fp16, [768]> obj_127_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_127_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250475904)))]; |
| tensor<fp16, []> obj_127_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_127_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor<string, []>("obj_127_cast_fp16")]; |
| tensor<string, []> query_37_pad_type_0 = const()[name = tensor<string, []>("query_37_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_37_strides_0 = const()[name = tensor<string, []>("query_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_37_pad_0 = const()[name = tensor<string, []>("query_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_37_dilations_0 = const()[name = tensor<string, []>("query_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_37_groups_0 = const()[name = tensor<string, []>("query_37_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250477504)))]; |
| tensor<fp16, [768]> layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251657216)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")]; |
| tensor<string, []> current_key_19_pad_type_0 = const()[name = tensor<string, []>("current_key_19_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_19_strides_0 = const()[name = tensor<string, []>("current_key_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_19_pad_0 = const()[name = tensor<string, []>("current_key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_19_dilations_0 = const()[name = tensor<string, []>("current_key_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_19_groups_0 = const()[name = tensor<string, []>("current_key_19_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251658816)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_19_cast_fp16 = conv(dilations = current_key_19_dilations_0, groups = current_key_19_groups_0, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = current_key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor<string, []>("current_key_19_cast_fp16")]; |
| tensor<string, []> current_value_19_pad_type_0 = const()[name = tensor<string, []>("current_value_19_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_19_strides_0 = const()[name = tensor<string, []>("current_value_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_19_pad_0 = const()[name = tensor<string, []>("current_value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_19_dilations_0 = const()[name = tensor<string, []>("current_value_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_19_groups_0 = const()[name = tensor<string, []>("current_value_19_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252838528)))]; |
| tensor<fp16, [768]> layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254018240)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = current_value_19_dilations_0, groups = current_value_19_groups_0, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = current_value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor<string, []>("current_value_19_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2094_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_2094_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2096_cast_fp16 = mul(x = var_63_cast_fp16_9, y = var_161_cast_fp16)[name = tensor<string, []>("op_2096_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_37_cast_fp16 = add(x = var_2094_cast_fp16, y = var_2096_cast_fp16)[name = tensor<string, []>("key_37_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2098_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_2098_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2100_cast_fp16 = mul(x = var_78_cast_fp16_9, y = var_161_cast_fp16)[name = tensor<string, []>("op_2100_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_37_cast_fp16 = add(x = var_2098_cast_fp16, y = var_2100_cast_fp16)[name = tensor<string, []>("value_37_cast_fp16")]; |
| tensor<int32, [4]> var_2103 = const()[name = tensor<string, []>("op_2103"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_37_cast_fp16 = reshape(shape = var_2103, x = query_37_cast_fp16)[name = tensor<string, []>("mh_q_37_cast_fp16")]; |
| tensor<fp16, []> var_2105_to_fp16 = const()[name = tensor<string, []>("op_2105_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_2106_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2105_to_fp16)[name = tensor<string, []>("op_2106_cast_fp16")]; |
| tensor<int32, [4]> var_2107 = const()[name = tensor<string, []>("op_2107"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_2108_cast_fp16 = reshape(shape = var_2107, x = key_37_cast_fp16)[name = tensor<string, []>("op_2108_cast_fp16")]; |
| tensor<bool, []> mh_w_55_transpose_x_0 = const()[name = tensor<string, []>("mh_w_55_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_55_transpose_y_0 = const()[name = tensor<string, []>("mh_w_55_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_2106_cast_fp16, y = var_2108_cast_fp16)[name = tensor<string, []>("mh_w_55_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_57_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_2116_cast_fp16 = softmax(axis = var_2030, x = mh_w_57_cast_fp16)[name = tensor<string, []>("op_2116_cast_fp16")]; |
| tensor<int32, [4]> var_2117 = const()[name = tensor<string, []>("op_2117"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_2118_cast_fp16 = reshape(shape = var_2117, x = value_37_cast_fp16)[name = tensor<string, []>("op_2118_cast_fp16")]; |
| tensor<bool, []> attn_37_transpose_x_0 = const()[name = tensor<string, []>("attn_37_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_37_transpose_y_0 = const()[name = tensor<string, []>("attn_37_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2118_cast_fp16, y = var_2116_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")]; |
| tensor<int32, [4]> var_2121 = const()[name = tensor<string, []>("op_2121"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_91_cast_fp16 = reshape(shape = var_2121, x = attn_37_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")]; |
| tensor<string, []> obj_133_pad_type_0 = const()[name = tensor<string, []>("obj_133_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_133_strides_0 = const()[name = tensor<string, []>("obj_133_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_133_pad_0 = const()[name = tensor<string, []>("obj_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_133_dilations_0 = const()[name = tensor<string, []>("obj_133_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_133_groups_0 = const()[name = tensor<string, []>("obj_133_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254019840)))]; |
| tensor<fp16, [768]> layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255199552)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_133_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_133_dilations_0, groups = obj_133_groups_0, pad = obj_133_pad_0, pad_type = obj_133_pad_type_0, strides = obj_133_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("obj_133_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor<string, []>("inputs_57_cast_fp16")]; |
| tensor<int32, [1]> out_57_axes_0 = const()[name = tensor<string, []>("out_57_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2143_to_fp16 = const()[name = tensor<string, []>("op_2143_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_2143_to_fp16, x = inputs_57_cast_fp16)[name = tensor<string, []>("out_57_cast_fp16")]; |
| tensor<fp16, [768]> obj_135_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_135_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255201152)))]; |
| tensor<fp16, [768]> obj_135_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_135_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255202752)))]; |
| tensor<fp16, []> obj_135_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_135_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor<string, []>("obj_135_cast_fp16")]; |
| tensor<string, []> query_39_pad_type_0 = const()[name = tensor<string, []>("query_39_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_39_strides_0 = const()[name = tensor<string, []>("query_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_39_pad_0 = const()[name = tensor<string, []>("query_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_39_dilations_0 = const()[name = tensor<string, []>("query_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_39_groups_0 = const()[name = tensor<string, []>("query_39_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_9_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255204352)))]; |
| tensor<fp16, [768]> layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256384064)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor<string, []>("query_39_cast_fp16")]; |
| tensor<string, []> key_39_pad_type_0 = const()[name = tensor<string, []>("key_39_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_39_strides_0 = const()[name = tensor<string, []>("key_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_39_pad_0 = const()[name = tensor<string, []>("key_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_39_dilations_0 = const()[name = tensor<string, []>("key_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_39_groups_0 = const()[name = tensor<string, []>("key_39_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_9_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256385664)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_39_cast_fp16")]; |
| tensor<string, []> value_39_pad_type_0 = const()[name = tensor<string, []>("value_39_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_39_strides_0 = const()[name = tensor<string, []>("value_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_39_pad_0 = const()[name = tensor<string, []>("value_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_39_dilations_0 = const()[name = tensor<string, []>("value_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_39_groups_0 = const()[name = tensor<string, []>("value_39_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_9_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257565376)))]; |
| tensor<fp16, [768]> layers_9_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258745088)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_39_cast_fp16")]; |
| tensor<int32, [4]> var_2178 = const()[name = tensor<string, []>("op_2178"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_39_cast_fp16 = reshape(shape = var_2178, x = query_39_cast_fp16)[name = tensor<string, []>("mh_q_39_cast_fp16")]; |
| tensor<fp16, []> var_2180_to_fp16 = const()[name = tensor<string, []>("op_2180_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_2181_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2180_to_fp16)[name = tensor<string, []>("op_2181_cast_fp16")]; |
| tensor<int32, [4]> var_2182 = const()[name = tensor<string, []>("op_2182"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_2183_cast_fp16 = reshape(shape = var_2182, x = key_39_cast_fp16)[name = tensor<string, []>("op_2183_cast_fp16")]; |
| tensor<bool, []> mh_w_59_transpose_x_0 = const()[name = tensor<string, []>("mh_w_59_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_59_transpose_y_0 = const()[name = tensor<string, []>("mh_w_59_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_2181_cast_fp16, y = var_2183_cast_fp16)[name = tensor<string, []>("mh_w_59_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_139_cast_fp16 = softmax(axis = var_2030, x = mh_w_59_cast_fp16)[name = tensor<string, []>("obj_139_cast_fp16")]; |
| tensor<int32, [4]> var_2187 = const()[name = tensor<string, []>("op_2187"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_2188_cast_fp16 = reshape(shape = var_2187, x = value_39_cast_fp16)[name = tensor<string, []>("op_2188_cast_fp16")]; |
| tensor<bool, []> attn_39_transpose_x_0 = const()[name = tensor<string, []>("attn_39_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_39_transpose_y_0 = const()[name = tensor<string, []>("attn_39_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2188_cast_fp16, y = obj_139_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")]; |
| tensor<int32, [4]> var_2191 = const()[name = tensor<string, []>("op_2191"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_93_cast_fp16 = reshape(shape = var_2191, x = attn_39_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")]; |
| tensor<string, []> obj_137_pad_type_0 = const()[name = tensor<string, []>("obj_137_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_137_strides_0 = const()[name = tensor<string, []>("obj_137_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_137_pad_0 = const()[name = tensor<string, []>("obj_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_137_dilations_0 = const()[name = tensor<string, []>("obj_137_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_137_groups_0 = const()[name = tensor<string, []>("obj_137_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_9_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258746688)))]; |
| tensor<fp16, [768]> layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259926400)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_137_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = obj_137_dilations_0, groups = obj_137_groups_0, pad = obj_137_pad_0, pad_type = obj_137_pad_type_0, strides = obj_137_strides_0, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("obj_137_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor<string, []>("inputs_59_cast_fp16")]; |
| tensor<int32, [1]> out_59_axes_0 = const()[name = tensor<string, []>("out_59_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2212_to_fp16 = const()[name = tensor<string, []>("op_2212_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_2212_to_fp16, x = inputs_59_cast_fp16)[name = tensor<string, []>("out_59_cast_fp16")]; |
| tensor<fp16, [768]> input_95_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_95_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259928000)))]; |
| tensor<fp16, [768]> input_95_beta_0_to_fp16 = const()[name = tensor<string, []>("input_95_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259929600)))]; |
| tensor<fp16, []> input_95_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_95_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")]; |
| tensor<string, []> input_97_pad_type_0 = const()[name = tensor<string, []>("input_97_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_97_strides_0 = const()[name = tensor<string, []>("input_97_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_97_pad_0 = const()[name = tensor<string, []>("input_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_97_dilations_0 = const()[name = tensor<string, []>("input_97_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_97_groups_0 = const()[name = tensor<string, []>("input_97_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_9_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259931200)))]; |
| tensor<fp16, [3072]> layers_9_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264649856)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")]; |
| tensor<string, []> input_99_mode_0 = const()[name = tensor<string, []>("input_99_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")]; |
| tensor<string, []> hidden_states_21_pad_type_0 = const()[name = tensor<string, []>("hidden_states_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_21_strides_0 = const()[name = tensor<string, []>("hidden_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_21_pad_0 = const()[name = tensor<string, []>("hidden_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_21_dilations_0 = const()[name = tensor<string, []>("hidden_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_21_groups_0 = const()[name = tensor<string, []>("hidden_states_21_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_9_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264656064)))]; |
| tensor<fp16, [768]> layers_9_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269374720)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("hidden_states_21_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor<string, []>("inputs_61_cast_fp16")]; |
| tensor<int32, []> var_2248 = const()[name = tensor<string, []>("op_2248"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_61_axes_0 = const()[name = tensor<string, []>("out_61_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2274_to_fp16 = const()[name = tensor<string, []>("op_2274_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_2274_to_fp16, x = inputs_61_cast_fp16)[name = tensor<string, []>("out_61_cast_fp16")]; |
| tensor<fp16, [768]> obj_141_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_141_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269376320)))]; |
| tensor<fp16, [768]> obj_141_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_141_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269377920)))]; |
| tensor<fp16, []> obj_141_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_141_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor<string, []>("obj_141_cast_fp16")]; |
| tensor<string, []> query_41_pad_type_0 = const()[name = tensor<string, []>("query_41_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_41_strides_0 = const()[name = tensor<string, []>("query_41_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_41_pad_0 = const()[name = tensor<string, []>("query_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_41_dilations_0 = const()[name = tensor<string, []>("query_41_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_41_groups_0 = const()[name = tensor<string, []>("query_41_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269379520)))]; |
| tensor<fp16, [768]> layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270559232)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")]; |
| tensor<string, []> current_key_21_pad_type_0 = const()[name = tensor<string, []>("current_key_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_21_strides_0 = const()[name = tensor<string, []>("current_key_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_21_pad_0 = const()[name = tensor<string, []>("current_key_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_21_dilations_0 = const()[name = tensor<string, []>("current_key_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_21_groups_0 = const()[name = tensor<string, []>("current_key_21_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270560832)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_21_cast_fp16 = conv(dilations = current_key_21_dilations_0, groups = current_key_21_groups_0, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = current_key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor<string, []>("current_key_21_cast_fp16")]; |
| tensor<string, []> current_value_21_pad_type_0 = const()[name = tensor<string, []>("current_value_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_21_strides_0 = const()[name = tensor<string, []>("current_value_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_21_pad_0 = const()[name = tensor<string, []>("current_value_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_21_dilations_0 = const()[name = tensor<string, []>("current_value_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_21_groups_0 = const()[name = tensor<string, []>("current_value_21_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271740544)))]; |
| tensor<fp16, [768]> layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272920256)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = current_value_21_dilations_0, groups = current_value_21_groups_0, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = current_value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor<string, []>("current_value_21_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2312_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_2312_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2314_cast_fp16 = mul(x = var_63_cast_fp16_10, y = var_161_cast_fp16)[name = tensor<string, []>("op_2314_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_41_cast_fp16 = add(x = var_2312_cast_fp16, y = var_2314_cast_fp16)[name = tensor<string, []>("key_41_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2316_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_2316_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2318_cast_fp16 = mul(x = var_78_cast_fp16_10, y = var_161_cast_fp16)[name = tensor<string, []>("op_2318_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_41_cast_fp16 = add(x = var_2316_cast_fp16, y = var_2318_cast_fp16)[name = tensor<string, []>("value_41_cast_fp16")]; |
| tensor<int32, [4]> var_2321 = const()[name = tensor<string, []>("op_2321"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_41_cast_fp16 = reshape(shape = var_2321, x = query_41_cast_fp16)[name = tensor<string, []>("mh_q_41_cast_fp16")]; |
| tensor<fp16, []> var_2323_to_fp16 = const()[name = tensor<string, []>("op_2323_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_2324_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2323_to_fp16)[name = tensor<string, []>("op_2324_cast_fp16")]; |
| tensor<int32, [4]> var_2325 = const()[name = tensor<string, []>("op_2325"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_2326_cast_fp16 = reshape(shape = var_2325, x = key_41_cast_fp16)[name = tensor<string, []>("op_2326_cast_fp16")]; |
| tensor<bool, []> mh_w_61_transpose_x_0 = const()[name = tensor<string, []>("mh_w_61_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_61_transpose_y_0 = const()[name = tensor<string, []>("mh_w_61_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_2324_cast_fp16, y = var_2326_cast_fp16)[name = tensor<string, []>("mh_w_61_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_63_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_2334_cast_fp16 = softmax(axis = var_2248, x = mh_w_63_cast_fp16)[name = tensor<string, []>("op_2334_cast_fp16")]; |
| tensor<int32, [4]> var_2335 = const()[name = tensor<string, []>("op_2335"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_2336_cast_fp16 = reshape(shape = var_2335, x = value_41_cast_fp16)[name = tensor<string, []>("op_2336_cast_fp16")]; |
| tensor<bool, []> attn_41_transpose_x_0 = const()[name = tensor<string, []>("attn_41_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_41_transpose_y_0 = const()[name = tensor<string, []>("attn_41_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2336_cast_fp16, y = var_2334_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")]; |
| tensor<int32, [4]> var_2339 = const()[name = tensor<string, []>("op_2339"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_101_cast_fp16 = reshape(shape = var_2339, x = attn_41_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")]; |
| tensor<string, []> obj_147_pad_type_0 = const()[name = tensor<string, []>("obj_147_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_147_strides_0 = const()[name = tensor<string, []>("obj_147_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_147_pad_0 = const()[name = tensor<string, []>("obj_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_147_dilations_0 = const()[name = tensor<string, []>("obj_147_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_147_groups_0 = const()[name = tensor<string, []>("obj_147_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272921856)))]; |
| tensor<fp16, [768]> layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274101568)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_147_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_147_dilations_0, groups = obj_147_groups_0, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = obj_147_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("obj_147_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor<string, []>("inputs_63_cast_fp16")]; |
| tensor<int32, [1]> out_63_axes_0 = const()[name = tensor<string, []>("out_63_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2361_to_fp16 = const()[name = tensor<string, []>("op_2361_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2361_to_fp16, x = inputs_63_cast_fp16)[name = tensor<string, []>("out_63_cast_fp16")]; |
| tensor<fp16, [768]> obj_149_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_149_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274103168)))]; |
| tensor<fp16, [768]> obj_149_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_149_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274104768)))]; |
| tensor<fp16, []> obj_149_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_149_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor<string, []>("obj_149_cast_fp16")]; |
| tensor<string, []> query_43_pad_type_0 = const()[name = tensor<string, []>("query_43_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_43_strides_0 = const()[name = tensor<string, []>("query_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_43_pad_0 = const()[name = tensor<string, []>("query_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_43_dilations_0 = const()[name = tensor<string, []>("query_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_43_groups_0 = const()[name = tensor<string, []>("query_43_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_10_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274106368)))]; |
| tensor<fp16, [768]> layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275286080)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor<string, []>("query_43_cast_fp16")]; |
| tensor<string, []> key_43_pad_type_0 = const()[name = tensor<string, []>("key_43_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_43_strides_0 = const()[name = tensor<string, []>("key_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_43_pad_0 = const()[name = tensor<string, []>("key_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_43_dilations_0 = const()[name = tensor<string, []>("key_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_43_groups_0 = const()[name = tensor<string, []>("key_43_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_10_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275287680)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_43_cast_fp16")]; |
| tensor<string, []> value_43_pad_type_0 = const()[name = tensor<string, []>("value_43_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_43_strides_0 = const()[name = tensor<string, []>("value_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_43_pad_0 = const()[name = tensor<string, []>("value_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_43_dilations_0 = const()[name = tensor<string, []>("value_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_43_groups_0 = const()[name = tensor<string, []>("value_43_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_10_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276467392)))]; |
| tensor<fp16, [768]> layers_10_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277647104)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_43_cast_fp16")]; |
| tensor<int32, [4]> var_2396 = const()[name = tensor<string, []>("op_2396"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_43_cast_fp16 = reshape(shape = var_2396, x = query_43_cast_fp16)[name = tensor<string, []>("mh_q_43_cast_fp16")]; |
| tensor<fp16, []> var_2398_to_fp16 = const()[name = tensor<string, []>("op_2398_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_2399_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2398_to_fp16)[name = tensor<string, []>("op_2399_cast_fp16")]; |
| tensor<int32, [4]> var_2400 = const()[name = tensor<string, []>("op_2400"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_2401_cast_fp16 = reshape(shape = var_2400, x = key_43_cast_fp16)[name = tensor<string, []>("op_2401_cast_fp16")]; |
| tensor<bool, []> mh_w_65_transpose_x_0 = const()[name = tensor<string, []>("mh_w_65_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_65_transpose_y_0 = const()[name = tensor<string, []>("mh_w_65_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_2399_cast_fp16, y = var_2401_cast_fp16)[name = tensor<string, []>("mh_w_65_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_153_cast_fp16 = softmax(axis = var_2248, x = mh_w_65_cast_fp16)[name = tensor<string, []>("obj_153_cast_fp16")]; |
| tensor<int32, [4]> var_2405 = const()[name = tensor<string, []>("op_2405"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_2406_cast_fp16 = reshape(shape = var_2405, x = value_43_cast_fp16)[name = tensor<string, []>("op_2406_cast_fp16")]; |
| tensor<bool, []> attn_43_transpose_x_0 = const()[name = tensor<string, []>("attn_43_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_43_transpose_y_0 = const()[name = tensor<string, []>("attn_43_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2406_cast_fp16, y = obj_153_cast_fp16)[name = tensor<string, []>("attn_43_cast_fp16")]; |
| tensor<int32, [4]> var_2409 = const()[name = tensor<string, []>("op_2409"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_103_cast_fp16 = reshape(shape = var_2409, x = attn_43_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")]; |
| tensor<string, []> obj_151_pad_type_0 = const()[name = tensor<string, []>("obj_151_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_151_strides_0 = const()[name = tensor<string, []>("obj_151_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_151_pad_0 = const()[name = tensor<string, []>("obj_151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_151_dilations_0 = const()[name = tensor<string, []>("obj_151_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_151_groups_0 = const()[name = tensor<string, []>("obj_151_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_10_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277648704)))]; |
| tensor<fp16, [768]> layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278828416)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_151_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = obj_151_dilations_0, groups = obj_151_groups_0, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = obj_151_strides_0, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("obj_151_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor<string, []>("inputs_65_cast_fp16")]; |
| tensor<int32, [1]> out_65_axes_0 = const()[name = tensor<string, []>("out_65_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2430_to_fp16 = const()[name = tensor<string, []>("op_2430_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2430_to_fp16, x = inputs_65_cast_fp16)[name = tensor<string, []>("out_65_cast_fp16")]; |
| tensor<fp16, [768]> input_105_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_105_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278830016)))]; |
| tensor<fp16, [768]> input_105_beta_0_to_fp16 = const()[name = tensor<string, []>("input_105_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278831616)))]; |
| tensor<fp16, []> input_105_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_105_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_105_cast_fp16 = batch_norm(beta = input_105_beta_0_to_fp16, epsilon = input_105_epsilon_0_to_fp16, gamma = input_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")]; |
| tensor<string, []> input_107_pad_type_0 = const()[name = tensor<string, []>("input_107_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_107_strides_0 = const()[name = tensor<string, []>("input_107_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_107_pad_0 = const()[name = tensor<string, []>("input_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_107_dilations_0 = const()[name = tensor<string, []>("input_107_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_107_groups_0 = const()[name = tensor<string, []>("input_107_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_10_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278833216)))]; |
| tensor<fp16, [3072]> layers_10_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283551872)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")]; |
| tensor<string, []> input_109_mode_0 = const()[name = tensor<string, []>("input_109_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")]; |
| tensor<string, []> hidden_states_23_pad_type_0 = const()[name = tensor<string, []>("hidden_states_23_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = tensor<string, []>("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = tensor<string, []>("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = tensor<string, []>("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_23_groups_0 = const()[name = tensor<string, []>("hidden_states_23_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_10_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283558080)))]; |
| tensor<fp16, [768]> layers_10_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288276736)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("hidden_states_23_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor<string, []>("inputs_67_cast_fp16")]; |
| tensor<int32, []> var_2466 = const()[name = tensor<string, []>("op_2466"), val = tensor<int32, []>(3)]; |
| tensor<int32, [1]> out_67_axes_0 = const()[name = tensor<string, []>("out_67_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2492_to_fp16 = const()[name = tensor<string, []>("op_2492_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2492_to_fp16, x = inputs_67_cast_fp16)[name = tensor<string, []>("out_67_cast_fp16")]; |
| tensor<fp16, [768]> obj_155_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_155_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288278336)))]; |
| tensor<fp16, [768]> obj_155_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_155_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288279936)))]; |
| tensor<fp16, []> obj_155_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_155_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor<string, []>("obj_155_cast_fp16")]; |
| tensor<string, []> query_45_pad_type_0 = const()[name = tensor<string, []>("query_45_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_45_strides_0 = const()[name = tensor<string, []>("query_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_45_pad_0 = const()[name = tensor<string, []>("query_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_45_dilations_0 = const()[name = tensor<string, []>("query_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_45_groups_0 = const()[name = tensor<string, []>("query_45_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288281536)))]; |
| tensor<fp16, [768]> layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289461248)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")]; |
| tensor<string, []> current_key_pad_type_0 = const()[name = tensor<string, []>("current_key_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_key_strides_0 = const()[name = tensor<string, []>("current_key_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_key_pad_0 = const()[name = tensor<string, []>("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_key_dilations_0 = const()[name = tensor<string, []>("current_key_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_key_groups_0 = const()[name = tensor<string, []>("current_key_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289462848)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")]; |
| tensor<string, []> current_value_pad_type_0 = const()[name = tensor<string, []>("current_value_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> current_value_strides_0 = const()[name = tensor<string, []>("current_value_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> current_value_pad_0 = const()[name = tensor<string, []>("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> current_value_dilations_0 = const()[name = tensor<string, []>("current_value_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> current_value_groups_0 = const()[name = tensor<string, []>("current_value_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290642560)))]; |
| tensor<fp16, [768]> layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291822272)))]; |
| tensor<fp16, [1, 768, 1, 1]> current_value_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2530_cast_fp16 = mul(x = current_key_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_2530_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2532_cast_fp16 = mul(x = var_63_cast_fp16_11, y = var_161_cast_fp16)[name = tensor<string, []>("op_2532_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> key_45_cast_fp16 = add(x = var_2530_cast_fp16, y = var_2532_cast_fp16)[name = tensor<string, []>("key_45_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2534_cast_fp16 = mul(x = current_value_cast_fp16, y = var_158_cast_fp16)[name = tensor<string, []>("op_2534_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> var_2536_cast_fp16 = mul(x = var_78_cast_fp16_11, y = var_161_cast_fp16)[name = tensor<string, []>("op_2536_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 448]> value_45_cast_fp16 = add(x = var_2534_cast_fp16, y = var_2536_cast_fp16)[name = tensor<string, []>("value_45_cast_fp16")]; |
| tensor<int32, [4]> var_2539 = const()[name = tensor<string, []>("op_2539"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_45_cast_fp16 = reshape(shape = var_2539, x = query_45_cast_fp16)[name = tensor<string, []>("mh_q_45_cast_fp16")]; |
| tensor<fp16, []> var_2541_to_fp16 = const()[name = tensor<string, []>("op_2541_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_2542_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2541_to_fp16)[name = tensor<string, []>("op_2542_cast_fp16")]; |
| tensor<int32, [4]> var_2543 = const()[name = tensor<string, []>("op_2543"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_2544_cast_fp16 = reshape(shape = var_2543, x = key_45_cast_fp16)[name = tensor<string, []>("op_2544_cast_fp16")]; |
| tensor<bool, []> mh_w_67_transpose_x_0 = const()[name = tensor<string, []>("mh_w_67_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_67_transpose_y_0 = const()[name = tensor<string, []>("mh_w_67_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_2542_cast_fp16, y = var_2544_cast_fp16)[name = tensor<string, []>("mh_w_67_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_179_cast_fp16)[name = tensor<string, []>("mh_w_69_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 448]> var_2552_cast_fp16 = softmax(axis = var_2466, x = mh_w_69_cast_fp16)[name = tensor<string, []>("op_2552_cast_fp16")]; |
| tensor<int32, [4]> var_2553 = const()[name = tensor<string, []>("op_2553"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 448]> var_2554_cast_fp16 = reshape(shape = var_2553, x = value_45_cast_fp16)[name = tensor<string, []>("op_2554_cast_fp16")]; |
| tensor<bool, []> attn_45_transpose_x_0 = const()[name = tensor<string, []>("attn_45_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_45_transpose_y_0 = const()[name = tensor<string, []>("attn_45_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2554_cast_fp16, y = var_2552_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")]; |
| tensor<int32, [4]> var_2557 = const()[name = tensor<string, []>("op_2557"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_111_cast_fp16 = reshape(shape = var_2557, x = attn_45_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")]; |
| tensor<string, []> obj_161_pad_type_0 = const()[name = tensor<string, []>("obj_161_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_161_strides_0 = const()[name = tensor<string, []>("obj_161_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_161_pad_0 = const()[name = tensor<string, []>("obj_161_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_161_dilations_0 = const()[name = tensor<string, []>("obj_161_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_161_groups_0 = const()[name = tensor<string, []>("obj_161_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291823872)))]; |
| tensor<fp16, [768]> layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293003584)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_161_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_161_dilations_0, groups = obj_161_groups_0, pad = obj_161_pad_0, pad_type = obj_161_pad_type_0, strides = obj_161_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("obj_161_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor<string, []>("inputs_69_cast_fp16")]; |
| tensor<int32, [1]> out_69_axes_0 = const()[name = tensor<string, []>("out_69_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2579_to_fp16 = const()[name = tensor<string, []>("op_2579_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2579_to_fp16, x = inputs_69_cast_fp16)[name = tensor<string, []>("out_69_cast_fp16")]; |
| tensor<fp16, [768]> obj_163_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_163_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293005184)))]; |
| tensor<fp16, [768]> obj_163_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_163_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293006784)))]; |
| tensor<fp16, []> obj_163_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_163_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor<string, []>("obj_163_cast_fp16")]; |
| tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> query_strides_0 = const()[name = tensor<string, []>("query_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_dilations_0 = const()[name = tensor<string, []>("query_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> query_groups_0 = const()[name = tensor<string, []>("query_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_11_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293008384)))]; |
| tensor<fp16, [768]> layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294188096)))]; |
| tensor<fp16, [1, 768, 1, 1]> query_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor<string, []>("query_cast_fp16")]; |
| tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> key_strides_0 = const()[name = tensor<string, []>("key_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_dilations_0 = const()[name = tensor<string, []>("key_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> key_groups_0 = const()[name = tensor<string, []>("key_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_11_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294189696)))]; |
| tensor<fp16, [1, 768, 1, 1500]> key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")]; |
| tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> value_strides_0 = const()[name = tensor<string, []>("value_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_dilations_0 = const()[name = tensor<string, []>("value_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> value_groups_0 = const()[name = tensor<string, []>("value_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_11_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295369408)))]; |
| tensor<fp16, [768]> layers_11_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296549120)))]; |
| tensor<fp16, [1, 768, 1, 1500]> value_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")]; |
| tensor<int32, [4]> var_2614 = const()[name = tensor<string, []>("op_2614"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_2614, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")]; |
| tensor<fp16, []> var_2616_to_fp16 = const()[name = tensor<string, []>("op_2616_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 12, 64, 1]> var_2617_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_2616_to_fp16)[name = tensor<string, []>("op_2617_cast_fp16")]; |
| tensor<int32, [4]> var_2618 = const()[name = tensor<string, []>("op_2618"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_2619_cast_fp16 = reshape(shape = var_2618, x = key_cast_fp16)[name = tensor<string, []>("op_2619_cast_fp16")]; |
| tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 12, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_2617_cast_fp16, y = var_2619_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")]; |
| tensor<fp16, [1, 12, 1, 1500]> obj_167_cast_fp16 = softmax(axis = var_2466, x = mh_w_cast_fp16)[name = tensor<string, []>("obj_167_cast_fp16")]; |
| tensor<int32, [4]> var_2623 = const()[name = tensor<string, []>("op_2623"), val = tensor<int32, [4]>([1, 12, 64, -1])]; |
| tensor<fp16, [1, 12, 64, 1500]> var_2624_cast_fp16 = reshape(shape = var_2623, x = value_cast_fp16)[name = tensor<string, []>("op_2624_cast_fp16")]; |
| tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 12, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_2624_cast_fp16, y = obj_167_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")]; |
| tensor<int32, [4]> var_2627 = const()[name = tensor<string, []>("op_2627"), val = tensor<int32, [4]>([1, 768, 1, -1])]; |
| tensor<fp16, [1, 768, 1, 1]> input_113_cast_fp16 = reshape(shape = var_2627, x = attn_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")]; |
| tensor<string, []> obj_165_pad_type_0 = const()[name = tensor<string, []>("obj_165_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> obj_165_strides_0 = const()[name = tensor<string, []>("obj_165_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> obj_165_pad_0 = const()[name = tensor<string, []>("obj_165_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> obj_165_dilations_0 = const()[name = tensor<string, []>("obj_165_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> obj_165_groups_0 = const()[name = tensor<string, []>("obj_165_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 768, 1, 1]> layers_11_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296550720)))]; |
| tensor<fp16, [768]> layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297730432)))]; |
| tensor<fp16, [1, 768, 1, 1]> obj_165_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = obj_165_dilations_0, groups = obj_165_groups_0, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = obj_165_strides_0, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("obj_165_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor<string, []>("inputs_71_cast_fp16")]; |
| tensor<int32, [1]> out_71_axes_0 = const()[name = tensor<string, []>("out_71_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2645_to_fp16 = const()[name = tensor<string, []>("op_2645_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2645_to_fp16, x = inputs_71_cast_fp16)[name = tensor<string, []>("out_71_cast_fp16")]; |
| tensor<fp16, [768]> input_115_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_115_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297732032)))]; |
| tensor<fp16, [768]> input_115_beta_0_to_fp16 = const()[name = tensor<string, []>("input_115_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297733632)))]; |
| tensor<fp16, []> input_115_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_115_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")]; |
| tensor<string, []> input_117_pad_type_0 = const()[name = tensor<string, []>("input_117_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> input_117_strides_0 = const()[name = tensor<string, []>("input_117_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_117_pad_0 = const()[name = tensor<string, []>("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_117_dilations_0 = const()[name = tensor<string, []>("input_117_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> input_117_groups_0 = const()[name = tensor<string, []>("input_117_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [3072, 768, 1, 1]> layers_11_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297735232)))]; |
| tensor<fp16, [3072]> layers_11_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302453888)))]; |
| tensor<fp16, [1, 3072, 1, 1]> input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")]; |
| tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 3072, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_117_cast_fp16)[name = tensor<string, []>("input_cast_fp16")]; |
| tensor<string, []> hidden_states_25_pad_type_0 = const()[name = tensor<string, []>("hidden_states_25_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [2]> hidden_states_25_strides_0 = const()[name = tensor<string, []>("hidden_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_25_pad_0 = const()[name = tensor<string, []>("hidden_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_25_dilations_0 = const()[name = tensor<string, []>("hidden_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, []> hidden_states_25_groups_0 = const()[name = tensor<string, []>("hidden_states_25_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [768, 3072, 1, 1]> layers_11_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302460096)))]; |
| tensor<fp16, [768]> layers_11_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307178752)))]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_25_cast_fp16")]; |
| tensor<fp16, [1, 768, 1, 1]> inputs_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")]; |
| tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, []> var_2687_to_fp16 = const()[name = tensor<string, []>("op_2687_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_2687_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")]; |
| tensor<fp16, [768]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307180352)))]; |
| tensor<fp16, [768]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307181952)))]; |
| tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 768, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")]; |
| tensor<int32, [1]> var_2698_axes_0 = const()[name = tensor<string, []>("op_2698_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 768, 1]> var_2698_cast_fp16 = squeeze(axes = var_2698_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_2698_cast_fp16")]; |
| tensor<int32, [3]> var_2701_perm_0 = const()[name = tensor<string, []>("op_2701_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [51865]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51865]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307183552)))]; |
| tensor<fp16, [1, 1, 768]> var_2701_cast_fp16 = transpose(perm = var_2701_perm_0, x = var_2698_cast_fp16)[name = tensor<string, []>("transpose_0")]; |
| tensor<fp16, [1, 1, 51865]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_2701_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")]; |
| tensor<int32, []> var_2705 = const()[name = tensor<string, []>("op_2705"), val = tensor<int32, []>(1)]; |
| tensor<bool, []> obj_171_interleave_0 = const()[name = tensor<string, []>("obj_171_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 9216, 1, 1]> key_cache_updates = concat(axis = var_2705, interleave = obj_171_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_11_cast_fp16, current_key_13_cast_fp16, current_key_15_cast_fp16, current_key_17_cast_fp16, current_key_19_cast_fp16, current_key_21_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_171_cast_fp16")]; |
| tensor<int32, []> var_2708 = const()[name = tensor<string, []>("op_2708"), val = tensor<int32, []>(1)]; |
| tensor<bool, []> obj_173_interleave_0 = const()[name = tensor<string, []>("obj_173_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 9216, 1, 1]> value_cache_updates = concat(axis = var_2708, interleave = obj_173_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_11_cast_fp16, current_value_13_cast_fp16, current_value_15_cast_fp16, current_value_17_cast_fp16, current_value_19_cast_fp16, current_value_21_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_173_cast_fp16")]; |
| tensor<int32, [4]> var_2719_begin_0 = const()[name = tensor<string, []>("op_2719_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])]; |
| tensor<int32, [4]> var_2719_end_0 = const()[name = tensor<string, []>("op_2719_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1500])]; |
| tensor<bool, [4]> var_2719_end_mask_0 = const()[name = tensor<string, []>("op_2719_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2719_cast_fp16 = slice_by_index(begin = var_2719_begin_0, end = var_2719_end_0, end_mask = var_2719_end_mask_0, x = obj_83_cast_fp16)[name = tensor<string, []>("op_2719_cast_fp16")]; |
| tensor<int32, [4]> var_2722_begin_0 = const()[name = tensor<string, []>("op_2722_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2722_end_0 = const()[name = tensor<string, []>("op_2722_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2722_end_mask_0 = const()[name = tensor<string, []>("op_2722_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2722_squeeze_mask_0 = const()[name = tensor<string, []>("op_2722_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2722_cast_fp16 = slice_by_index(begin = var_2722_begin_0, end = var_2722_end_0, end_mask = var_2722_end_mask_0, squeeze_mask = var_2722_squeeze_mask_0, x = var_2719_cast_fp16)[name = tensor<string, []>("op_2722_cast_fp16")]; |
| tensor<int32, [4]> var_2737_begin_0 = const()[name = tensor<string, []>("op_2737_begin_0"), val = tensor<int32, [4]>([0, 9, 0, 0])]; |
| tensor<int32, [4]> var_2737_end_0 = const()[name = tensor<string, []>("op_2737_end_0"), val = tensor<int32, [4]>([1, 10, 1, 1500])]; |
| tensor<bool, [4]> var_2737_end_mask_0 = const()[name = tensor<string, []>("op_2737_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2737_cast_fp16 = slice_by_index(begin = var_2737_begin_0, end = var_2737_end_0, end_mask = var_2737_end_mask_0, x = obj_83_cast_fp16)[name = tensor<string, []>("op_2737_cast_fp16")]; |
| tensor<int32, [4]> var_2740_begin_0 = const()[name = tensor<string, []>("op_2740_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2740_end_0 = const()[name = tensor<string, []>("op_2740_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2740_end_mask_0 = const()[name = tensor<string, []>("op_2740_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2740_squeeze_mask_0 = const()[name = tensor<string, []>("op_2740_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2740_cast_fp16 = slice_by_index(begin = var_2740_begin_0, end = var_2740_end_0, end_mask = var_2740_end_mask_0, squeeze_mask = var_2740_squeeze_mask_0, x = var_2737_cast_fp16)[name = tensor<string, []>("op_2740_cast_fp16")]; |
| tensor<int32, [4]> var_2755_begin_0 = const()[name = tensor<string, []>("op_2755_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2755_end_0 = const()[name = tensor<string, []>("op_2755_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2755_end_mask_0 = const()[name = tensor<string, []>("op_2755_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2755_cast_fp16 = slice_by_index(begin = var_2755_begin_0, end = var_2755_end_0, end_mask = var_2755_end_mask_0, x = obj_125_cast_fp16)[name = tensor<string, []>("op_2755_cast_fp16")]; |
| tensor<int32, [4]> var_2758_begin_0 = const()[name = tensor<string, []>("op_2758_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2758_end_0 = const()[name = tensor<string, []>("op_2758_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2758_end_mask_0 = const()[name = tensor<string, []>("op_2758_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2758_squeeze_mask_0 = const()[name = tensor<string, []>("op_2758_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2758_cast_fp16 = slice_by_index(begin = var_2758_begin_0, end = var_2758_end_0, end_mask = var_2758_end_mask_0, squeeze_mask = var_2758_squeeze_mask_0, x = var_2755_cast_fp16)[name = tensor<string, []>("op_2758_cast_fp16")]; |
| tensor<int32, [4]> var_2773_begin_0 = const()[name = tensor<string, []>("op_2773_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])]; |
| tensor<int32, [4]> var_2773_end_0 = const()[name = tensor<string, []>("op_2773_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])]; |
| tensor<bool, [4]> var_2773_end_mask_0 = const()[name = tensor<string, []>("op_2773_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2773_cast_fp16 = slice_by_index(begin = var_2773_begin_0, end = var_2773_end_0, end_mask = var_2773_end_mask_0, x = obj_125_cast_fp16)[name = tensor<string, []>("op_2773_cast_fp16")]; |
| tensor<int32, [4]> var_2776_begin_0 = const()[name = tensor<string, []>("op_2776_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2776_end_0 = const()[name = tensor<string, []>("op_2776_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2776_end_mask_0 = const()[name = tensor<string, []>("op_2776_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2776_squeeze_mask_0 = const()[name = tensor<string, []>("op_2776_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2776_cast_fp16 = slice_by_index(begin = var_2776_begin_0, end = var_2776_end_0, end_mask = var_2776_end_mask_0, squeeze_mask = var_2776_squeeze_mask_0, x = var_2773_cast_fp16)[name = tensor<string, []>("op_2776_cast_fp16")]; |
| tensor<int32, [4]> var_2791_begin_0 = const()[name = tensor<string, []>("op_2791_begin_0"), val = tensor<int32, [4]>([0, 7, 0, 0])]; |
| tensor<int32, [4]> var_2791_end_0 = const()[name = tensor<string, []>("op_2791_end_0"), val = tensor<int32, [4]>([1, 8, 1, 1500])]; |
| tensor<bool, [4]> var_2791_end_mask_0 = const()[name = tensor<string, []>("op_2791_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2791_cast_fp16 = slice_by_index(begin = var_2791_begin_0, end = var_2791_end_0, end_mask = var_2791_end_mask_0, x = obj_125_cast_fp16)[name = tensor<string, []>("op_2791_cast_fp16")]; |
| tensor<int32, [4]> var_2794_begin_0 = const()[name = tensor<string, []>("op_2794_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2794_end_0 = const()[name = tensor<string, []>("op_2794_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2794_end_mask_0 = const()[name = tensor<string, []>("op_2794_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2794_squeeze_mask_0 = const()[name = tensor<string, []>("op_2794_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, squeeze_mask = var_2794_squeeze_mask_0, x = var_2791_cast_fp16)[name = tensor<string, []>("op_2794_cast_fp16")]; |
| tensor<int32, [4]> var_2809_begin_0 = const()[name = tensor<string, []>("op_2809_begin_0"), val = tensor<int32, [4]>([0, 8, 0, 0])]; |
| tensor<int32, [4]> var_2809_end_0 = const()[name = tensor<string, []>("op_2809_end_0"), val = tensor<int32, [4]>([1, 9, 1, 1500])]; |
| tensor<bool, [4]> var_2809_end_mask_0 = const()[name = tensor<string, []>("op_2809_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2809_cast_fp16 = slice_by_index(begin = var_2809_begin_0, end = var_2809_end_0, end_mask = var_2809_end_mask_0, x = obj_125_cast_fp16)[name = tensor<string, []>("op_2809_cast_fp16")]; |
| tensor<int32, [4]> var_2812_begin_0 = const()[name = tensor<string, []>("op_2812_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2812_end_0 = const()[name = tensor<string, []>("op_2812_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2812_end_mask_0 = const()[name = tensor<string, []>("op_2812_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2812_squeeze_mask_0 = const()[name = tensor<string, []>("op_2812_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, squeeze_mask = var_2812_squeeze_mask_0, x = var_2809_cast_fp16)[name = tensor<string, []>("op_2812_cast_fp16")]; |
| tensor<int32, [4]> var_2827_begin_0 = const()[name = tensor<string, []>("op_2827_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2827_end_0 = const()[name = tensor<string, []>("op_2827_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2827_end_mask_0 = const()[name = tensor<string, []>("op_2827_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2827_cast_fp16 = slice_by_index(begin = var_2827_begin_0, end = var_2827_end_0, end_mask = var_2827_end_mask_0, x = obj_139_cast_fp16)[name = tensor<string, []>("op_2827_cast_fp16")]; |
| tensor<int32, [4]> var_2830_begin_0 = const()[name = tensor<string, []>("op_2830_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2830_end_0 = const()[name = tensor<string, []>("op_2830_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2830_end_mask_0 = const()[name = tensor<string, []>("op_2830_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2830_squeeze_mask_0 = const()[name = tensor<string, []>("op_2830_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2830_cast_fp16 = slice_by_index(begin = var_2830_begin_0, end = var_2830_end_0, end_mask = var_2830_end_mask_0, squeeze_mask = var_2830_squeeze_mask_0, x = var_2827_cast_fp16)[name = tensor<string, []>("op_2830_cast_fp16")]; |
| tensor<int32, [4]> var_2845_begin_0 = const()[name = tensor<string, []>("op_2845_begin_0"), val = tensor<int32, [4]>([0, 7, 0, 0])]; |
| tensor<int32, [4]> var_2845_end_0 = const()[name = tensor<string, []>("op_2845_end_0"), val = tensor<int32, [4]>([1, 8, 1, 1500])]; |
| tensor<bool, [4]> var_2845_end_mask_0 = const()[name = tensor<string, []>("op_2845_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2845_cast_fp16 = slice_by_index(begin = var_2845_begin_0, end = var_2845_end_0, end_mask = var_2845_end_mask_0, x = obj_139_cast_fp16)[name = tensor<string, []>("op_2845_cast_fp16")]; |
| tensor<int32, [4]> var_2848_begin_0 = const()[name = tensor<string, []>("op_2848_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2848_end_0 = const()[name = tensor<string, []>("op_2848_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2848_end_mask_0 = const()[name = tensor<string, []>("op_2848_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2848_squeeze_mask_0 = const()[name = tensor<string, []>("op_2848_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2848_cast_fp16 = slice_by_index(begin = var_2848_begin_0, end = var_2848_end_0, end_mask = var_2848_end_mask_0, squeeze_mask = var_2848_squeeze_mask_0, x = var_2845_cast_fp16)[name = tensor<string, []>("op_2848_cast_fp16")]; |
| tensor<int32, [4]> var_2863_begin_0 = const()[name = tensor<string, []>("op_2863_begin_0"), val = tensor<int32, [4]>([0, 9, 0, 0])]; |
| tensor<int32, [4]> var_2863_end_0 = const()[name = tensor<string, []>("op_2863_end_0"), val = tensor<int32, [4]>([1, 10, 1, 1500])]; |
| tensor<bool, [4]> var_2863_end_mask_0 = const()[name = tensor<string, []>("op_2863_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2863_cast_fp16 = slice_by_index(begin = var_2863_begin_0, end = var_2863_end_0, end_mask = var_2863_end_mask_0, x = obj_139_cast_fp16)[name = tensor<string, []>("op_2863_cast_fp16")]; |
| tensor<int32, [4]> var_2866_begin_0 = const()[name = tensor<string, []>("op_2866_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2866_end_0 = const()[name = tensor<string, []>("op_2866_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2866_end_mask_0 = const()[name = tensor<string, []>("op_2866_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2866_squeeze_mask_0 = const()[name = tensor<string, []>("op_2866_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2866_cast_fp16 = slice_by_index(begin = var_2866_begin_0, end = var_2866_end_0, end_mask = var_2866_end_mask_0, squeeze_mask = var_2866_squeeze_mask_0, x = var_2863_cast_fp16)[name = tensor<string, []>("op_2866_cast_fp16")]; |
| tensor<int32, [4]> var_2881_begin_0 = const()[name = tensor<string, []>("op_2881_begin_0"), val = tensor<int32, [4]>([0, 5, 0, 0])]; |
| tensor<int32, [4]> var_2881_end_0 = const()[name = tensor<string, []>("op_2881_end_0"), val = tensor<int32, [4]>([1, 6, 1, 1500])]; |
| tensor<bool, [4]> var_2881_end_mask_0 = const()[name = tensor<string, []>("op_2881_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| tensor<fp16, [1, 1, 1, 1500]> var_2881_cast_fp16 = slice_by_index(begin = var_2881_begin_0, end = var_2881_end_0, end_mask = var_2881_end_mask_0, x = obj_153_cast_fp16)[name = tensor<string, []>("op_2881_cast_fp16")]; |
| tensor<int32, [4]> var_2884_begin_0 = const()[name = tensor<string, []>("op_2884_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_2884_end_0 = const()[name = tensor<string, []>("op_2884_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| tensor<bool, [4]> var_2884_end_mask_0 = const()[name = tensor<string, []>("op_2884_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| tensor<bool, [4]> var_2884_squeeze_mask_0 = const()[name = tensor<string, []>("op_2884_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| tensor<fp16, [1, 1, 1500]> var_2884_cast_fp16 = slice_by_index(begin = var_2884_begin_0, end = var_2884_end_0, end_mask = var_2884_end_mask_0, squeeze_mask = var_2884_squeeze_mask_0, x = var_2881_cast_fp16)[name = tensor<string, []>("op_2884_cast_fp16")]; |
| tensor<int32, []> var_2891 = const()[name = tensor<string, []>("op_2891"), val = tensor<int32, []>(1)]; |
| tensor<bool, []> var_2892_interleave_0 = const()[name = tensor<string, []>("op_2892_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 10, 1500]> var_2892_cast_fp16 = concat(axis = var_2891, interleave = var_2892_interleave_0, values = (var_2722_cast_fp16, var_2740_cast_fp16, var_2758_cast_fp16, var_2776_cast_fp16, var_2794_cast_fp16, var_2812_cast_fp16, var_2830_cast_fp16, var_2848_cast_fp16, var_2866_cast_fp16, var_2884_cast_fp16))[name = tensor<string, []>("op_2892_cast_fp16")]; |
| tensor<bool, []> var_2895 = const()[name = tensor<string, []>("op_2895"), val = tensor<bool, []>(false)]; |
| tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1, 1500]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_2895, x = var_2892_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")]; |
| } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); |
| } |
