program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.1.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})] { func main(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { tensor var_40_axis_0 = const()[name = tensor("op_40_axis_0"), val = tensor(0)]; tensor var_40_batch_dims_0 = const()[name = tensor("op_40_batch_dims_0"), val = tensor(0)]; tensor var_40_validate_indices_0 = const()[name = tensor("op_40_validate_indices_0"), val = tensor(false)]; tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; tensor var_40_cast_fp16 = gather(axis = var_40_axis_0, batch_dims = var_40_batch_dims_0, indices = input_ids, validate_indices = var_40_validate_indices_0, x = embed_tokens_weight_to_fp16)[name = tensor("op_40_cast_fp16")]; tensor var_44_axis_0 = const()[name = tensor("op_44_axis_0"), val = tensor(0)]; tensor var_44_batch_dims_0 = const()[name = tensor("op_44_batch_dims_0"), val = tensor(0)]; tensor var_44_validate_indices_0 = const()[name = tensor("op_44_validate_indices_0"), val = tensor(false)]; tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79664768)))]; tensor cache_length_to_int16_dtype_0 = const()[name = tensor("cache_length_to_int16_dtype_0"), val = tensor("int16")]; tensor cast_180 = cast(dtype = cache_length_to_int16_dtype_0, x = cache_length)[name = tensor("cast_180")]; tensor var_44_cast_fp16_cast_int16 = gather(axis = var_44_axis_0, batch_dims = var_44_batch_dims_0, indices = cast_180, validate_indices = var_44_validate_indices_0, x = embed_positions_weight_to_fp16)[name = tensor("op_44_cast_fp16_cast_int16")]; tensor hidden_states_1_cast_fp16 = add(x = var_40_cast_fp16, y = var_44_cast_fp16_cast_int16)[name = tensor("hidden_states_1_cast_fp16")]; tensor var_58_axes_0 = const()[name = tensor("op_58_axes_0"), val = tensor([2])]; tensor var_58_cast_fp16 = expand_dims(axes = var_58_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_58_cast_fp16")]; tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_58_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768])]; tensor var_63_axis_0 = const()[name = tensor("op_63_axis_0"), val = tensor(1)]; tensor var_63_cast_fp16_0, tensor var_63_cast_fp16_1, tensor var_63_cast_fp16_2, tensor var_63_cast_fp16_3, tensor var_63_cast_fp16_4, tensor var_63_cast_fp16_5, tensor var_63_cast_fp16_6, tensor var_63_cast_fp16_7, tensor var_63_cast_fp16_8, tensor var_63_cast_fp16_9, tensor var_63_cast_fp16_10, tensor var_63_cast_fp16_11 = split(axis = var_63_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_63_cast_fp16")]; tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768])]; tensor var_78_axis_0 = const()[name = tensor("op_78_axis_0"), val = tensor(1)]; tensor var_78_cast_fp16_0, tensor var_78_cast_fp16_1, tensor var_78_cast_fp16_2, tensor var_78_cast_fp16_3, tensor var_78_cast_fp16_4, tensor var_78_cast_fp16_5, tensor var_78_cast_fp16_6, tensor var_78_cast_fp16_7, tensor var_78_cast_fp16_8, tensor var_78_cast_fp16_9, tensor var_78_cast_fp16_10, tensor var_78_cast_fp16_11 = split(axis = var_78_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_78_cast_fp16")]; tensor var_96 = const()[name = tensor("op_96"), val = tensor(3)]; tensor var_103 = const()[name = tensor("op_103"), val = tensor(1)]; tensor var_104 = const()[name = tensor("op_104"), val = tensor(true)]; tensor var_116 = const()[name = tensor("op_116"), val = tensor([1])]; tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_116, keep_dims = var_104, x = inputs_1_cast_fp16)[name = tensor("channels_mean_1_cast_fp16")]; tensor zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor("zero_mean_1_cast_fp16")]; tensor zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor("zero_mean_sq_1_cast_fp16")]; tensor var_120 = const()[name = tensor("op_120"), val = tensor([1])]; tensor var_121_cast_fp16 = reduce_mean(axes = var_120, keep_dims = var_104, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_121_cast_fp16")]; tensor var_122_to_fp16 = const()[name = tensor("op_122_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_123_cast_fp16 = add(x = var_121_cast_fp16, y = var_122_to_fp16)[name = tensor("op_123_cast_fp16")]; tensor denom_1_epsilon_0 = const()[name = tensor("denom_1_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0, x = var_123_cast_fp16)[name = tensor("denom_1_cast_fp16")]; tensor out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80352960)))]; tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80354560)))]; tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80356160)))]; tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80357760)))]; tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("obj_1_cast_fp16")]; tensor var_138 = const()[name = tensor("op_138"), val = tensor([1, 1])]; tensor var_140 = const()[name = tensor("op_140"), val = tensor([1, 1])]; tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("custom")]; tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80359360)))]; tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81539072)))]; tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_140, groups = var_103, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_138, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; tensor var_144 = const()[name = tensor("op_144"), val = tensor([1, 1])]; tensor var_146 = const()[name = tensor("op_146"), val = tensor([1, 1])]; tensor current_key_1_pad_type_0 = const()[name = tensor("current_key_1_pad_type_0"), val = tensor("custom")]; tensor current_key_1_pad_0 = const()[name = tensor("current_key_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81540672)))]; tensor current_key_1_cast_fp16 = conv(dilations = var_146, groups = var_103, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_144, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; tensor var_151 = const()[name = tensor("op_151"), val = tensor([1, 1])]; tensor var_153 = const()[name = tensor("op_153"), val = tensor([1, 1])]; tensor current_value_1_pad_type_0 = const()[name = tensor("current_value_1_pad_type_0"), val = tensor("custom")]; tensor current_value_1_pad_0 = const()[name = tensor("current_value_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82720384)))]; tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83900096)))]; tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_153, groups = var_103, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_151, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; tensor var_157_axes_0 = const()[name = tensor("op_157_axes_0"), val = tensor([1])]; tensor var_157_cast_fp16 = expand_dims(axes = var_157_axes_0, x = kv_cache_update_mask)[name = tensor("op_157_cast_fp16")]; tensor var_158_axes_0 = const()[name = tensor("op_158_axes_0"), val = tensor([2])]; tensor var_158_cast_fp16 = expand_dims(axes = var_158_axes_0, x = var_157_cast_fp16)[name = tensor("op_158_cast_fp16")]; tensor var_160_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_160_cast_fp16")]; tensor var_97_to_fp16 = const()[name = tensor("op_97_to_fp16"), val = tensor(0x1p+0)]; tensor var_161_cast_fp16 = sub(x = var_97_to_fp16, y = var_158_cast_fp16)[name = tensor("op_161_cast_fp16")]; tensor var_162_cast_fp16 = mul(x = var_63_cast_fp16_0, y = var_161_cast_fp16)[name = tensor("op_162_cast_fp16")]; tensor key_1_cast_fp16 = add(x = var_160_cast_fp16, y = var_162_cast_fp16)[name = tensor("key_1_cast_fp16")]; tensor var_164_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_164_cast_fp16")]; tensor var_166_cast_fp16 = mul(x = var_78_cast_fp16_0, y = var_161_cast_fp16)[name = tensor("op_166_cast_fp16")]; tensor value_1_cast_fp16 = add(x = var_164_cast_fp16, y = var_166_cast_fp16)[name = tensor("value_1_cast_fp16")]; tensor var_169 = const()[name = tensor("op_169"), val = tensor([1, 12, 64, -1])]; tensor var_170_cast_fp16 = reshape(shape = var_169, x = query_1_cast_fp16)[name = tensor("op_170_cast_fp16")]; tensor var_171_to_fp16 = const()[name = tensor("op_171_to_fp16"), val = tensor(0x1p-3)]; tensor var_172_cast_fp16 = mul(x = var_170_cast_fp16, y = var_171_to_fp16)[name = tensor("op_172_cast_fp16")]; tensor var_173 = const()[name = tensor("op_173"), val = tensor([1, 12, 64, -1])]; tensor var_174_cast_fp16 = reshape(shape = var_173, x = key_1_cast_fp16)[name = tensor("op_174_cast_fp16")]; tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_1_cast_fp16")]; tensor var_178_axes_0 = const()[name = tensor("op_178_axes_0"), val = tensor([1])]; tensor var_178_cast_fp16 = expand_dims(axes = var_178_axes_0, x = decoder_key_padding_mask)[name = tensor("op_178_cast_fp16")]; tensor var_179_axes_0 = const()[name = tensor("op_179_axes_0"), val = tensor([2])]; tensor var_179_cast_fp16 = expand_dims(axes = var_179_axes_0, x = var_178_cast_fp16)[name = tensor("op_179_cast_fp16")]; tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; tensor var_182_cast_fp16 = softmax(axis = var_96, x = mh_w_3_cast_fp16)[name = tensor("op_182_cast_fp16")]; tensor var_183 = const()[name = tensor("op_183"), val = tensor([1, 12, 64, -1])]; tensor var_184_cast_fp16 = reshape(shape = var_183, x = value_1_cast_fp16)[name = tensor("op_184_cast_fp16")]; tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; tensor 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("attn_1_cast_fp16")]; tensor var_187 = const()[name = tensor("op_187"), val = tensor([1, 768, 1, -1])]; tensor input_1_cast_fp16 = reshape(shape = var_187, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; tensor var_191 = const()[name = tensor("op_191"), val = tensor([1, 1])]; tensor var_193 = const()[name = tensor("op_193"), val = tensor([1, 1])]; tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("custom")]; tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83901696)))]; tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85081408)))]; tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_193, groups = var_103, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_191, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("obj_7_cast_fp16")]; tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; tensor var_203 = const()[name = tensor("op_203"), val = tensor([1])]; tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_203, keep_dims = var_104, x = inputs_3_cast_fp16)[name = tensor("channels_mean_3_cast_fp16")]; tensor zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor("zero_mean_3_cast_fp16")]; tensor zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor("zero_mean_sq_3_cast_fp16")]; tensor var_207 = const()[name = tensor("op_207"), val = tensor([1])]; tensor var_208_cast_fp16 = reduce_mean(axes = var_207, keep_dims = var_104, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_208_cast_fp16")]; tensor var_209_to_fp16 = const()[name = tensor("op_209_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_210_cast_fp16 = add(x = var_208_cast_fp16, y = var_209_to_fp16)[name = tensor("op_210_cast_fp16")]; tensor denom_3_epsilon_0 = const()[name = tensor("denom_3_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0, x = var_210_cast_fp16)[name = tensor("denom_3_cast_fp16")]; tensor out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85083008)))]; tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85084608)))]; tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("obj_9_cast_fp16")]; tensor var_225 = const()[name = tensor("op_225"), val = tensor([1, 1])]; tensor var_227 = const()[name = tensor("op_227"), val = tensor([1, 1])]; tensor query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("custom")]; tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85086208)))]; tensor layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86265920)))]; tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_227, groups = var_103, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_225, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1])]; tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 1])]; tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("custom")]; tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86267520)))]; tensor key_3_cast_fp16 = conv(dilations = var_233, groups = var_103, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_231, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 1])]; tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 1])]; tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("custom")]; tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87447232)))]; tensor layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88626944)))]; tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_240, groups = var_103, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_238, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 12, 64, -1])]; tensor var_245_cast_fp16 = reshape(shape = var_244, x = query_3_cast_fp16)[name = tensor("op_245_cast_fp16")]; tensor var_246_to_fp16 = const()[name = tensor("op_246_to_fp16"), val = tensor(0x1p-3)]; tensor var_247_cast_fp16 = mul(x = var_245_cast_fp16, y = var_246_to_fp16)[name = tensor("op_247_cast_fp16")]; tensor var_248 = const()[name = tensor("op_248"), val = tensor([1, 12, 64, -1])]; tensor var_249_cast_fp16 = reshape(shape = var_248, x = key_3_cast_fp16)[name = tensor("op_249_cast_fp16")]; tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_5_cast_fp16")]; tensor var_252_cast_fp16 = softmax(axis = var_96, x = mh_w_5_cast_fp16)[name = tensor("op_252_cast_fp16")]; tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 12, 64, -1])]; tensor var_254_cast_fp16 = reshape(shape = var_253, x = value_3_cast_fp16)[name = tensor("op_254_cast_fp16")]; tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; tensor 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 = var_252_cast_fp16)[name = tensor("attn_3_cast_fp16")]; tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 768, 1, -1])]; tensor input_3_cast_fp16 = reshape(shape = var_257, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; tensor var_261 = const()[name = tensor("op_261"), val = tensor([1, 1])]; tensor var_263 = const()[name = tensor("op_263"), val = tensor([1, 1])]; tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("custom")]; tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88628544)))]; tensor layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89808256)))]; tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_263, groups = var_103, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_261, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("obj_11_cast_fp16")]; tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; tensor var_269 = const()[name = tensor("op_269"), val = tensor([1])]; tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_269, keep_dims = var_104, x = inputs_5_cast_fp16)[name = tensor("channels_mean_5_cast_fp16")]; tensor zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor("zero_mean_5_cast_fp16")]; tensor zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor("zero_mean_sq_5_cast_fp16")]; tensor var_273 = const()[name = tensor("op_273"), val = tensor([1])]; tensor var_274_cast_fp16 = reduce_mean(axes = var_273, keep_dims = var_104, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_274_cast_fp16")]; tensor var_275_to_fp16 = const()[name = tensor("op_275_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_276_cast_fp16 = add(x = var_274_cast_fp16, y = var_275_to_fp16)[name = tensor("op_276_cast_fp16")]; tensor denom_5_epsilon_0 = const()[name = tensor("denom_5_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0, x = var_276_cast_fp16)[name = tensor("denom_5_cast_fp16")]; tensor out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; tensor input_5_gamma_0_to_fp16 = const()[name = tensor("input_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89809856)))]; tensor input_5_beta_0_to_fp16 = const()[name = tensor("input_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89811456)))]; tensor input_5_epsilon_0_to_fp16 = const()[name = tensor("input_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_5_cast_fp16")]; tensor var_287 = const()[name = tensor("op_287"), val = tensor([1, 1])]; tensor var_289 = const()[name = tensor("op_289"), val = tensor([1, 1])]; tensor input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("custom")]; tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_fc1_weight_to_fp16 = const()[name = tensor("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89813056)))]; tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94531712)))]; tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_289, groups = var_103, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_287, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("EXACT")]; tensor input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; tensor var_295 = const()[name = tensor("op_295"), val = tensor([1, 1])]; tensor var_297 = const()[name = tensor("op_297"), val = tensor([1, 1])]; tensor hidden_states_3_pad_type_0 = const()[name = tensor("hidden_states_3_pad_type_0"), val = tensor("custom")]; tensor hidden_states_3_pad_0 = const()[name = tensor("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_fc2_weight_to_fp16 = const()[name = tensor("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94537920)))]; tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99256576)))]; tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_297, groups = var_103, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_295, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; tensor var_310 = const()[name = tensor("op_310"), val = tensor(3)]; tensor var_317 = const()[name = tensor("op_317"), val = tensor(1)]; tensor var_318 = const()[name = tensor("op_318"), val = tensor(true)]; tensor var_330 = const()[name = tensor("op_330"), val = tensor([1])]; tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_330, keep_dims = var_318, x = inputs_7_cast_fp16)[name = tensor("channels_mean_7_cast_fp16")]; tensor zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor("zero_mean_7_cast_fp16")]; tensor zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor("zero_mean_sq_7_cast_fp16")]; tensor var_334 = const()[name = tensor("op_334"), val = tensor([1])]; tensor var_335_cast_fp16 = reduce_mean(axes = var_334, keep_dims = var_318, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_335_cast_fp16")]; tensor var_336_to_fp16 = const()[name = tensor("op_336_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_337_cast_fp16 = add(x = var_335_cast_fp16, y = var_336_to_fp16)[name = tensor("op_337_cast_fp16")]; tensor denom_7_epsilon_0 = const()[name = tensor("denom_7_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0, x = var_337_cast_fp16)[name = tensor("denom_7_cast_fp16")]; tensor out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; tensor obj_13_gamma_0_to_fp16 = const()[name = tensor("obj_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99258176)))]; tensor obj_13_beta_0_to_fp16 = const()[name = tensor("obj_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99259776)))]; tensor obj_13_epsilon_0_to_fp16 = const()[name = tensor("obj_13_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_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("obj_13_cast_fp16")]; tensor var_352 = const()[name = tensor("op_352"), val = tensor([1, 1])]; tensor var_354 = const()[name = tensor("op_354"), val = tensor([1, 1])]; tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("custom")]; tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99261376)))]; tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100441088)))]; tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_354, groups = var_317, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_352, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("query_5_cast_fp16")]; tensor var_358 = const()[name = tensor("op_358"), val = tensor([1, 1])]; tensor var_360 = const()[name = tensor("op_360"), val = tensor([1, 1])]; tensor current_key_3_pad_type_0 = const()[name = tensor("current_key_3_pad_type_0"), val = tensor("custom")]; tensor current_key_3_pad_0 = const()[name = tensor("current_key_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100442688)))]; tensor current_key_3_cast_fp16 = conv(dilations = var_360, groups = var_317, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = var_358, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1])]; tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1])]; tensor current_value_3_pad_type_0 = const()[name = tensor("current_value_3_pad_type_0"), val = tensor("custom")]; tensor current_value_3_pad_0 = const()[name = tensor("current_value_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101622400)))]; tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102802112)))]; tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_367, groups = var_317, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = var_365, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; tensor var_374_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_374_cast_fp16")]; tensor var_376_cast_fp16 = mul(x = var_63_cast_fp16_1, y = var_161_cast_fp16)[name = tensor("op_376_cast_fp16")]; tensor key_5_cast_fp16 = add(x = var_374_cast_fp16, y = var_376_cast_fp16)[name = tensor("key_5_cast_fp16")]; tensor var_378_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_378_cast_fp16")]; tensor var_380_cast_fp16 = mul(x = var_78_cast_fp16_1, y = var_161_cast_fp16)[name = tensor("op_380_cast_fp16")]; tensor value_5_cast_fp16 = add(x = var_378_cast_fp16, y = var_380_cast_fp16)[name = tensor("value_5_cast_fp16")]; tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 12, 64, -1])]; tensor var_384_cast_fp16 = reshape(shape = var_383, x = query_5_cast_fp16)[name = tensor("op_384_cast_fp16")]; tensor var_385_to_fp16 = const()[name = tensor("op_385_to_fp16"), val = tensor(0x1p-3)]; tensor var_386_cast_fp16 = mul(x = var_384_cast_fp16, y = var_385_to_fp16)[name = tensor("op_386_cast_fp16")]; tensor var_387 = const()[name = tensor("op_387"), val = tensor([1, 12, 64, -1])]; tensor var_388_cast_fp16 = reshape(shape = var_387, x = key_5_cast_fp16)[name = tensor("op_388_cast_fp16")]; tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_7_cast_fp16")]; tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; tensor var_396_cast_fp16 = softmax(axis = var_310, x = mh_w_9_cast_fp16)[name = tensor("op_396_cast_fp16")]; tensor var_397 = const()[name = tensor("op_397"), val = tensor([1, 12, 64, -1])]; tensor var_398_cast_fp16 = reshape(shape = var_397, x = value_5_cast_fp16)[name = tensor("op_398_cast_fp16")]; tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; tensor 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("attn_5_cast_fp16")]; tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 768, 1, -1])]; tensor input_11_cast_fp16 = reshape(shape = var_401, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; tensor var_405 = const()[name = tensor("op_405"), val = tensor([1, 1])]; tensor var_407 = const()[name = tensor("op_407"), val = tensor([1, 1])]; tensor obj_19_pad_type_0 = const()[name = tensor("obj_19_pad_type_0"), val = tensor("custom")]; tensor obj_19_pad_0 = const()[name = tensor("obj_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102803712)))]; tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103983424)))]; tensor obj_19_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_407, groups = var_317, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = var_405, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("obj_19_cast_fp16")]; tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_19_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; tensor var_417 = const()[name = tensor("op_417"), val = tensor([1])]; tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_417, keep_dims = var_318, x = inputs_9_cast_fp16)[name = tensor("channels_mean_9_cast_fp16")]; tensor zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor("zero_mean_9_cast_fp16")]; tensor zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor("zero_mean_sq_9_cast_fp16")]; tensor var_421 = const()[name = tensor("op_421"), val = tensor([1])]; tensor var_422_cast_fp16 = reduce_mean(axes = var_421, keep_dims = var_318, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_422_cast_fp16")]; tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_424_cast_fp16 = add(x = var_422_cast_fp16, y = var_423_to_fp16)[name = tensor("op_424_cast_fp16")]; tensor denom_9_epsilon_0 = const()[name = tensor("denom_9_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0, x = var_424_cast_fp16)[name = tensor("denom_9_cast_fp16")]; tensor out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; tensor obj_21_gamma_0_to_fp16 = const()[name = tensor("obj_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103985024)))]; tensor obj_21_beta_0_to_fp16 = const()[name = tensor("obj_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103986624)))]; tensor obj_21_epsilon_0_to_fp16 = const()[name = tensor("obj_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_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("obj_21_cast_fp16")]; tensor var_439 = const()[name = tensor("op_439"), val = tensor([1, 1])]; tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 1])]; tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("custom")]; tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103988224)))]; tensor layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105167936)))]; tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_441, groups = var_317, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_439, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("query_7_cast_fp16")]; tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 1])]; tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1])]; tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("custom")]; tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105169536)))]; tensor key_7_cast_fp16 = conv(dilations = var_447, groups = var_317, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = var_445, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_7_cast_fp16")]; tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 1])]; tensor var_454 = const()[name = tensor("op_454"), val = tensor([1, 1])]; tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("custom")]; tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106349248)))]; tensor layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107528960)))]; tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_454, groups = var_317, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = var_452, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_7_cast_fp16")]; tensor var_458 = const()[name = tensor("op_458"), val = tensor([1, 12, 64, -1])]; tensor var_459_cast_fp16 = reshape(shape = var_458, x = query_7_cast_fp16)[name = tensor("op_459_cast_fp16")]; tensor var_460_to_fp16 = const()[name = tensor("op_460_to_fp16"), val = tensor(0x1p-3)]; tensor var_461_cast_fp16 = mul(x = var_459_cast_fp16, y = var_460_to_fp16)[name = tensor("op_461_cast_fp16")]; tensor var_462 = const()[name = tensor("op_462"), val = tensor([1, 12, 64, -1])]; tensor var_463_cast_fp16 = reshape(shape = var_462, x = key_7_cast_fp16)[name = tensor("op_463_cast_fp16")]; tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_11_cast_fp16")]; tensor var_466_cast_fp16 = softmax(axis = var_310, x = mh_w_11_cast_fp16)[name = tensor("op_466_cast_fp16")]; tensor var_467 = const()[name = tensor("op_467"), val = tensor([1, 12, 64, -1])]; tensor var_468_cast_fp16 = reshape(shape = var_467, x = value_7_cast_fp16)[name = tensor("op_468_cast_fp16")]; tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; tensor 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 = var_466_cast_fp16)[name = tensor("attn_7_cast_fp16")]; tensor var_471 = const()[name = tensor("op_471"), val = tensor([1, 768, 1, -1])]; tensor input_13_cast_fp16 = reshape(shape = var_471, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; tensor var_475 = const()[name = tensor("op_475"), val = tensor([1, 1])]; tensor var_477 = const()[name = tensor("op_477"), val = tensor([1, 1])]; tensor obj_23_pad_type_0 = const()[name = tensor("obj_23_pad_type_0"), val = tensor("custom")]; tensor obj_23_pad_0 = const()[name = tensor("obj_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107530560)))]; tensor layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108710272)))]; tensor obj_23_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_477, groups = var_317, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = var_475, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("obj_23_cast_fp16")]; tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_23_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; tensor var_483 = const()[name = tensor("op_483"), val = tensor([1])]; tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_483, keep_dims = var_318, x = inputs_11_cast_fp16)[name = tensor("channels_mean_11_cast_fp16")]; tensor zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor("zero_mean_11_cast_fp16")]; tensor zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor("zero_mean_sq_11_cast_fp16")]; tensor var_487 = const()[name = tensor("op_487"), val = tensor([1])]; tensor var_488_cast_fp16 = reduce_mean(axes = var_487, keep_dims = var_318, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_488_cast_fp16")]; tensor var_489_to_fp16 = const()[name = tensor("op_489_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_490_cast_fp16 = add(x = var_488_cast_fp16, y = var_489_to_fp16)[name = tensor("op_490_cast_fp16")]; tensor denom_11_epsilon_0 = const()[name = tensor("denom_11_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0, x = var_490_cast_fp16)[name = tensor("denom_11_cast_fp16")]; tensor out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; tensor input_15_gamma_0_to_fp16 = const()[name = tensor("input_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108711872)))]; tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108713472)))]; tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_15_cast_fp16")]; tensor var_501 = const()[name = tensor("op_501"), val = tensor([1, 1])]; tensor var_503 = const()[name = tensor("op_503"), val = tensor([1, 1])]; tensor input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("custom")]; tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108715072)))]; tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113433728)))]; tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_503, groups = var_317, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_501, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; tensor var_509 = const()[name = tensor("op_509"), val = tensor([1, 1])]; tensor var_511 = const()[name = tensor("op_511"), val = tensor([1, 1])]; tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("custom")]; tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_fc2_weight_to_fp16 = const()[name = tensor("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113439936)))]; tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118158592)))]; tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_511, groups = var_317, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_509, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; tensor var_524 = const()[name = tensor("op_524"), val = tensor(3)]; tensor var_531 = const()[name = tensor("op_531"), val = tensor(1)]; tensor var_532 = const()[name = tensor("op_532"), val = tensor(true)]; tensor var_544 = const()[name = tensor("op_544"), val = tensor([1])]; tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_544, keep_dims = var_532, x = inputs_13_cast_fp16)[name = tensor("channels_mean_13_cast_fp16")]; tensor zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor("zero_mean_13_cast_fp16")]; tensor zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor("zero_mean_sq_13_cast_fp16")]; tensor var_548 = const()[name = tensor("op_548"), val = tensor([1])]; tensor var_549_cast_fp16 = reduce_mean(axes = var_548, keep_dims = var_532, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_549_cast_fp16")]; tensor var_550_to_fp16 = const()[name = tensor("op_550_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_551_cast_fp16 = add(x = var_549_cast_fp16, y = var_550_to_fp16)[name = tensor("op_551_cast_fp16")]; tensor denom_13_epsilon_0 = const()[name = tensor("denom_13_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0, x = var_551_cast_fp16)[name = tensor("denom_13_cast_fp16")]; tensor out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; tensor obj_25_gamma_0_to_fp16 = const()[name = tensor("obj_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118160192)))]; tensor obj_25_beta_0_to_fp16 = const()[name = tensor("obj_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118161792)))]; tensor obj_25_epsilon_0_to_fp16 = const()[name = tensor("obj_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_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("obj_25_cast_fp16")]; tensor var_566 = const()[name = tensor("op_566"), val = tensor([1, 1])]; tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("custom")]; tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118163392)))]; tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119343104)))]; tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_568, groups = var_531, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_566, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("query_9_cast_fp16")]; tensor var_572 = const()[name = tensor("op_572"), val = tensor([1, 1])]; tensor var_574 = const()[name = tensor("op_574"), val = tensor([1, 1])]; tensor current_key_5_pad_type_0 = const()[name = tensor("current_key_5_pad_type_0"), val = tensor("custom")]; tensor current_key_5_pad_0 = const()[name = tensor("current_key_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119344704)))]; tensor current_key_5_cast_fp16 = conv(dilations = var_574, groups = var_531, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_572, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; tensor var_579 = const()[name = tensor("op_579"), val = tensor([1, 1])]; tensor var_581 = const()[name = tensor("op_581"), val = tensor([1, 1])]; tensor current_value_5_pad_type_0 = const()[name = tensor("current_value_5_pad_type_0"), val = tensor("custom")]; tensor current_value_5_pad_0 = const()[name = tensor("current_value_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120524416)))]; tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121704128)))]; tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_581, groups = var_531, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_579, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; tensor var_588_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_588_cast_fp16")]; tensor var_590_cast_fp16 = mul(x = var_63_cast_fp16_2, y = var_161_cast_fp16)[name = tensor("op_590_cast_fp16")]; tensor key_9_cast_fp16 = add(x = var_588_cast_fp16, y = var_590_cast_fp16)[name = tensor("key_9_cast_fp16")]; tensor var_592_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_592_cast_fp16")]; tensor var_594_cast_fp16 = mul(x = var_78_cast_fp16_2, y = var_161_cast_fp16)[name = tensor("op_594_cast_fp16")]; tensor value_9_cast_fp16 = add(x = var_592_cast_fp16, y = var_594_cast_fp16)[name = tensor("value_9_cast_fp16")]; tensor var_597 = const()[name = tensor("op_597"), val = tensor([1, 12, 64, -1])]; tensor var_598_cast_fp16 = reshape(shape = var_597, x = query_9_cast_fp16)[name = tensor("op_598_cast_fp16")]; tensor var_599_to_fp16 = const()[name = tensor("op_599_to_fp16"), val = tensor(0x1p-3)]; tensor var_600_cast_fp16 = mul(x = var_598_cast_fp16, y = var_599_to_fp16)[name = tensor("op_600_cast_fp16")]; tensor var_601 = const()[name = tensor("op_601"), val = tensor([1, 12, 64, -1])]; tensor var_602_cast_fp16 = reshape(shape = var_601, x = key_9_cast_fp16)[name = tensor("op_602_cast_fp16")]; tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_13_cast_fp16")]; tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; tensor var_610_cast_fp16 = softmax(axis = var_524, x = mh_w_15_cast_fp16)[name = tensor("op_610_cast_fp16")]; tensor var_611 = const()[name = tensor("op_611"), val = tensor([1, 12, 64, -1])]; tensor var_612_cast_fp16 = reshape(shape = var_611, x = value_9_cast_fp16)[name = tensor("op_612_cast_fp16")]; tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; tensor 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("attn_9_cast_fp16")]; tensor var_615 = const()[name = tensor("op_615"), val = tensor([1, 768, 1, -1])]; tensor input_21_cast_fp16 = reshape(shape = var_615, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; tensor var_619 = const()[name = tensor("op_619"), val = tensor([1, 1])]; tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 1])]; tensor obj_31_pad_type_0 = const()[name = tensor("obj_31_pad_type_0"), val = tensor("custom")]; tensor obj_31_pad_0 = const()[name = tensor("obj_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121705728)))]; tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122885440)))]; tensor obj_31_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_621, groups = var_531, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = var_619, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("obj_31_cast_fp16")]; tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_31_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; tensor var_631 = const()[name = tensor("op_631"), val = tensor([1])]; tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_631, keep_dims = var_532, x = inputs_15_cast_fp16)[name = tensor("channels_mean_15_cast_fp16")]; tensor zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor("zero_mean_15_cast_fp16")]; tensor zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor("zero_mean_sq_15_cast_fp16")]; tensor var_635 = const()[name = tensor("op_635"), val = tensor([1])]; tensor var_636_cast_fp16 = reduce_mean(axes = var_635, keep_dims = var_532, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_636_cast_fp16")]; tensor var_637_to_fp16 = const()[name = tensor("op_637_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_638_cast_fp16 = add(x = var_636_cast_fp16, y = var_637_to_fp16)[name = tensor("op_638_cast_fp16")]; tensor denom_15_epsilon_0 = const()[name = tensor("denom_15_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0, x = var_638_cast_fp16)[name = tensor("denom_15_cast_fp16")]; tensor out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; tensor obj_33_gamma_0_to_fp16 = const()[name = tensor("obj_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122887040)))]; tensor obj_33_beta_0_to_fp16 = const()[name = tensor("obj_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122888640)))]; tensor obj_33_epsilon_0_to_fp16 = const()[name = tensor("obj_33_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_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("obj_33_cast_fp16")]; tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1])]; tensor var_655 = const()[name = tensor("op_655"), val = tensor([1, 1])]; tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("custom")]; tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122890240)))]; tensor layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124069952)))]; tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_655, groups = var_531, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_653, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor("query_11_cast_fp16")]; tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 1])]; tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 1])]; tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("custom")]; tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124071552)))]; tensor key_11_cast_fp16 = conv(dilations = var_661, groups = var_531, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_659, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; tensor var_666 = const()[name = tensor("op_666"), val = tensor([1, 1])]; tensor var_668 = const()[name = tensor("op_668"), val = tensor([1, 1])]; tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("custom")]; tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125251264)))]; tensor layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126430976)))]; tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_668, groups = var_531, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_666, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; tensor var_672 = const()[name = tensor("op_672"), val = tensor([1, 12, 64, -1])]; tensor var_673_cast_fp16 = reshape(shape = var_672, x = query_11_cast_fp16)[name = tensor("op_673_cast_fp16")]; tensor var_674_to_fp16 = const()[name = tensor("op_674_to_fp16"), val = tensor(0x1p-3)]; tensor var_675_cast_fp16 = mul(x = var_673_cast_fp16, y = var_674_to_fp16)[name = tensor("op_675_cast_fp16")]; tensor var_676 = const()[name = tensor("op_676"), val = tensor([1, 12, 64, -1])]; tensor var_677_cast_fp16 = reshape(shape = var_676, x = key_11_cast_fp16)[name = tensor("op_677_cast_fp16")]; tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_17_cast_fp16")]; tensor var_680_cast_fp16 = softmax(axis = var_524, x = mh_w_17_cast_fp16)[name = tensor("op_680_cast_fp16")]; tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 12, 64, -1])]; tensor var_682_cast_fp16 = reshape(shape = var_681, x = value_11_cast_fp16)[name = tensor("op_682_cast_fp16")]; tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; tensor 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 = var_680_cast_fp16)[name = tensor("attn_11_cast_fp16")]; tensor var_685 = const()[name = tensor("op_685"), val = tensor([1, 768, 1, -1])]; tensor input_23_cast_fp16 = reshape(shape = var_685, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; tensor var_689 = const()[name = tensor("op_689"), val = tensor([1, 1])]; tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 1])]; tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("custom")]; tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126432576)))]; tensor layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127612288)))]; tensor obj_35_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_691, groups = var_531, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_689, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("obj_35_cast_fp16")]; tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; tensor var_697 = const()[name = tensor("op_697"), val = tensor([1])]; tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_697, keep_dims = var_532, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; tensor var_701 = const()[name = tensor("op_701"), val = tensor([1])]; tensor var_702_cast_fp16 = reduce_mean(axes = var_701, keep_dims = var_532, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_702_cast_fp16")]; tensor var_703_to_fp16 = const()[name = tensor("op_703_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_704_cast_fp16 = add(x = var_702_cast_fp16, y = var_703_to_fp16)[name = tensor("op_704_cast_fp16")]; tensor denom_17_epsilon_0 = const()[name = tensor("denom_17_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0, x = var_704_cast_fp16)[name = tensor("denom_17_cast_fp16")]; tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; tensor input_25_gamma_0_to_fp16 = const()[name = tensor("input_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127613888)))]; tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127615488)))]; tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_25_cast_fp16")]; tensor var_715 = const()[name = tensor("op_715"), val = tensor([1, 1])]; tensor var_717 = const()[name = tensor("op_717"), val = tensor([1, 1])]; tensor input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("custom")]; tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127617088)))]; tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132335744)))]; tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_717, groups = var_531, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_715, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("EXACT")]; tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; tensor var_723 = const()[name = tensor("op_723"), val = tensor([1, 1])]; tensor var_725 = const()[name = tensor("op_725"), val = tensor([1, 1])]; tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("custom")]; tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_fc2_weight_to_fp16 = const()[name = tensor("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132341952)))]; tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137060608)))]; tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_725, groups = var_531, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_723, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; tensor var_738 = const()[name = tensor("op_738"), val = tensor(3)]; tensor var_745 = const()[name = tensor("op_745"), val = tensor(1)]; tensor var_746 = const()[name = tensor("op_746"), val = tensor(true)]; tensor var_758 = const()[name = tensor("op_758"), val = tensor([1])]; tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_758, keep_dims = var_746, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; tensor var_762 = const()[name = tensor("op_762"), val = tensor([1])]; tensor var_763_cast_fp16 = reduce_mean(axes = var_762, keep_dims = var_746, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_763_cast_fp16")]; tensor var_764_to_fp16 = const()[name = tensor("op_764_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_765_cast_fp16 = add(x = var_763_cast_fp16, y = var_764_to_fp16)[name = tensor("op_765_cast_fp16")]; tensor denom_19_epsilon_0 = const()[name = tensor("denom_19_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0, x = var_765_cast_fp16)[name = tensor("denom_19_cast_fp16")]; tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137062208)))]; tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137063808)))]; tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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_19_cast_fp16)[name = tensor("obj_37_cast_fp16")]; tensor var_780 = const()[name = tensor("op_780"), val = tensor([1, 1])]; tensor var_782 = const()[name = tensor("op_782"), val = tensor([1, 1])]; tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("custom")]; tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137065408)))]; tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138245120)))]; tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_782, groups = var_745, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_780, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_13_cast_fp16")]; tensor var_786 = const()[name = tensor("op_786"), val = tensor([1, 1])]; tensor var_788 = const()[name = tensor("op_788"), val = tensor([1, 1])]; tensor current_key_7_pad_type_0 = const()[name = tensor("current_key_7_pad_type_0"), val = tensor("custom")]; tensor current_key_7_pad_0 = const()[name = tensor("current_key_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138246720)))]; tensor current_key_7_cast_fp16 = conv(dilations = var_788, groups = var_745, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = var_786, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; tensor var_793 = const()[name = tensor("op_793"), val = tensor([1, 1])]; tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 1])]; tensor current_value_7_pad_type_0 = const()[name = tensor("current_value_7_pad_type_0"), val = tensor("custom")]; tensor current_value_7_pad_0 = const()[name = tensor("current_value_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139426432)))]; tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140606144)))]; tensor current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_795, groups = var_745, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = var_793, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; tensor var_802_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_802_cast_fp16")]; tensor var_804_cast_fp16 = mul(x = var_63_cast_fp16_3, y = var_161_cast_fp16)[name = tensor("op_804_cast_fp16")]; tensor key_13_cast_fp16 = add(x = var_802_cast_fp16, y = var_804_cast_fp16)[name = tensor("key_13_cast_fp16")]; tensor var_806_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_806_cast_fp16")]; tensor var_808_cast_fp16 = mul(x = var_78_cast_fp16_3, y = var_161_cast_fp16)[name = tensor("op_808_cast_fp16")]; tensor value_13_cast_fp16 = add(x = var_806_cast_fp16, y = var_808_cast_fp16)[name = tensor("value_13_cast_fp16")]; tensor var_811 = const()[name = tensor("op_811"), val = tensor([1, 12, 64, -1])]; tensor var_812_cast_fp16 = reshape(shape = var_811, x = query_13_cast_fp16)[name = tensor("op_812_cast_fp16")]; tensor var_813_to_fp16 = const()[name = tensor("op_813_to_fp16"), val = tensor(0x1p-3)]; tensor var_814_cast_fp16 = mul(x = var_812_cast_fp16, y = var_813_to_fp16)[name = tensor("op_814_cast_fp16")]; tensor var_815 = const()[name = tensor("op_815"), val = tensor([1, 12, 64, -1])]; tensor var_816_cast_fp16 = reshape(shape = var_815, x = key_13_cast_fp16)[name = tensor("op_816_cast_fp16")]; tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_19_cast_fp16")]; tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; tensor var_824_cast_fp16 = softmax(axis = var_738, x = mh_w_21_cast_fp16)[name = tensor("op_824_cast_fp16")]; tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 12, 64, -1])]; tensor var_826_cast_fp16 = reshape(shape = var_825, x = value_13_cast_fp16)[name = tensor("op_826_cast_fp16")]; tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; tensor 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("attn_13_cast_fp16")]; tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 768, 1, -1])]; tensor input_31_cast_fp16 = reshape(shape = var_829, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; tensor var_833 = const()[name = tensor("op_833"), val = tensor([1, 1])]; tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 1])]; tensor obj_43_pad_type_0 = const()[name = tensor("obj_43_pad_type_0"), val = tensor("custom")]; tensor obj_43_pad_0 = const()[name = tensor("obj_43_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140607744)))]; tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141787456)))]; tensor obj_43_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_835, groups = var_745, pad = obj_43_pad_0, pad_type = obj_43_pad_type_0, strides = var_833, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("obj_43_cast_fp16")]; tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_43_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; tensor var_845 = const()[name = tensor("op_845"), val = tensor([1])]; tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_845, keep_dims = var_746, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; tensor var_849 = const()[name = tensor("op_849"), val = tensor([1])]; tensor var_850_cast_fp16 = reduce_mean(axes = var_849, keep_dims = var_746, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_850_cast_fp16")]; tensor var_851_to_fp16 = const()[name = tensor("op_851_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_852_cast_fp16 = add(x = var_850_cast_fp16, y = var_851_to_fp16)[name = tensor("op_852_cast_fp16")]; tensor denom_21_epsilon_0 = const()[name = tensor("denom_21_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0, x = var_852_cast_fp16)[name = tensor("denom_21_cast_fp16")]; tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; tensor obj_45_gamma_0_to_fp16 = const()[name = tensor("obj_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141789056)))]; tensor obj_45_beta_0_to_fp16 = const()[name = tensor("obj_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141790656)))]; tensor obj_45_epsilon_0_to_fp16 = const()[name = tensor("obj_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_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("obj_45_cast_fp16")]; tensor var_867 = const()[name = tensor("op_867"), val = tensor([1, 1])]; tensor var_869 = const()[name = tensor("op_869"), val = tensor([1, 1])]; tensor query_15_pad_type_0 = const()[name = tensor("query_15_pad_type_0"), val = tensor("custom")]; tensor query_15_pad_0 = const()[name = tensor("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141792256)))]; tensor layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142971968)))]; tensor query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_869, groups = var_745, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = var_867, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor("query_15_cast_fp16")]; tensor var_873 = const()[name = tensor("op_873"), val = tensor([1, 1])]; tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 1])]; tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("custom")]; tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142973568)))]; tensor key_15_cast_fp16 = conv(dilations = var_875, groups = var_745, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = var_873, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_15_cast_fp16")]; tensor var_880 = const()[name = tensor("op_880"), val = tensor([1, 1])]; tensor var_882 = const()[name = tensor("op_882"), val = tensor([1, 1])]; tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("custom")]; tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144153280)))]; tensor layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145332992)))]; tensor value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_882, groups = var_745, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = var_880, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_15_cast_fp16")]; tensor var_886 = const()[name = tensor("op_886"), val = tensor([1, 12, 64, -1])]; tensor var_887_cast_fp16 = reshape(shape = var_886, x = query_15_cast_fp16)[name = tensor("op_887_cast_fp16")]; tensor var_888_to_fp16 = const()[name = tensor("op_888_to_fp16"), val = tensor(0x1p-3)]; tensor var_889_cast_fp16 = mul(x = var_887_cast_fp16, y = var_888_to_fp16)[name = tensor("op_889_cast_fp16")]; tensor var_890 = const()[name = tensor("op_890"), val = tensor([1, 12, 64, -1])]; tensor var_891_cast_fp16 = reshape(shape = var_890, x = key_15_cast_fp16)[name = tensor("op_891_cast_fp16")]; tensor mh_w_23_transpose_x_0 = const()[name = tensor("mh_w_23_transpose_x_0"), val = tensor(true)]; tensor mh_w_23_transpose_y_0 = const()[name = tensor("mh_w_23_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_23_cast_fp16")]; tensor var_894_cast_fp16 = softmax(axis = var_738, x = mh_w_23_cast_fp16)[name = tensor("op_894_cast_fp16")]; tensor var_895 = const()[name = tensor("op_895"), val = tensor([1, 12, 64, -1])]; tensor var_896_cast_fp16 = reshape(shape = var_895, x = value_15_cast_fp16)[name = tensor("op_896_cast_fp16")]; tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; tensor 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 = var_894_cast_fp16)[name = tensor("attn_15_cast_fp16")]; tensor var_899 = const()[name = tensor("op_899"), val = tensor([1, 768, 1, -1])]; tensor input_33_cast_fp16 = reshape(shape = var_899, x = attn_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; tensor var_903 = const()[name = tensor("op_903"), val = tensor([1, 1])]; tensor var_905 = const()[name = tensor("op_905"), val = tensor([1, 1])]; tensor obj_47_pad_type_0 = const()[name = tensor("obj_47_pad_type_0"), val = tensor("custom")]; tensor obj_47_pad_0 = const()[name = tensor("obj_47_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145334592)))]; tensor layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146514304)))]; tensor obj_47_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_905, groups = var_745, pad = obj_47_pad_0, pad_type = obj_47_pad_type_0, strides = var_903, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_47_cast_fp16")]; tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_47_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; tensor var_911 = const()[name = tensor("op_911"), val = tensor([1])]; tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_911, keep_dims = var_746, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; tensor var_915 = const()[name = tensor("op_915"), val = tensor([1])]; tensor var_916_cast_fp16 = reduce_mean(axes = var_915, keep_dims = var_746, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_916_cast_fp16")]; tensor var_917_to_fp16 = const()[name = tensor("op_917_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_918_cast_fp16 = add(x = var_916_cast_fp16, y = var_917_to_fp16)[name = tensor("op_918_cast_fp16")]; tensor denom_23_epsilon_0 = const()[name = tensor("denom_23_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0, x = var_918_cast_fp16)[name = tensor("denom_23_cast_fp16")]; tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146515904)))]; tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146517504)))]; tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_35_cast_fp16")]; tensor var_929 = const()[name = tensor("op_929"), val = tensor([1, 1])]; tensor var_931 = const()[name = tensor("op_931"), val = tensor([1, 1])]; tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("custom")]; tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_fc1_weight_to_fp16 = const()[name = tensor("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146519104)))]; tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151237760)))]; tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_931, groups = var_745, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_929, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; tensor input_39_mode_0 = const()[name = tensor("input_39_mode_0"), val = tensor("EXACT")]; tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; tensor var_937 = const()[name = tensor("op_937"), val = tensor([1, 1])]; tensor var_939 = const()[name = tensor("op_939"), val = tensor([1, 1])]; tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("custom")]; tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_fc2_weight_to_fp16 = const()[name = tensor("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151243968)))]; tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155962624)))]; tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_939, groups = var_745, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_937, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; tensor var_952 = const()[name = tensor("op_952"), val = tensor(3)]; tensor var_959 = const()[name = tensor("op_959"), val = tensor(1)]; tensor var_960 = const()[name = tensor("op_960"), val = tensor(true)]; tensor var_972 = const()[name = tensor("op_972"), val = tensor([1])]; tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_972, keep_dims = var_960, x = inputs_25_cast_fp16)[name = tensor("channels_mean_25_cast_fp16")]; tensor zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor("zero_mean_25_cast_fp16")]; tensor zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor("zero_mean_sq_25_cast_fp16")]; tensor var_976 = const()[name = tensor("op_976"), val = tensor([1])]; tensor var_977_cast_fp16 = reduce_mean(axes = var_976, keep_dims = var_960, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_977_cast_fp16")]; tensor var_978_to_fp16 = const()[name = tensor("op_978_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_979_cast_fp16 = add(x = var_977_cast_fp16, y = var_978_to_fp16)[name = tensor("op_979_cast_fp16")]; tensor denom_25_epsilon_0 = const()[name = tensor("denom_25_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0, x = var_979_cast_fp16)[name = tensor("denom_25_cast_fp16")]; tensor out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; tensor obj_49_gamma_0_to_fp16 = const()[name = tensor("obj_49_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155964224)))]; tensor obj_49_beta_0_to_fp16 = const()[name = tensor("obj_49_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155965824)))]; tensor obj_49_epsilon_0_to_fp16 = const()[name = tensor("obj_49_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_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("obj_49_cast_fp16")]; tensor var_994 = const()[name = tensor("op_994"), val = tensor([1, 1])]; tensor var_996 = const()[name = tensor("op_996"), val = tensor([1, 1])]; tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("custom")]; tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155967424)))]; tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157147136)))]; tensor query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = var_996, groups = var_959, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = var_994, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("query_17_cast_fp16")]; tensor var_1000 = const()[name = tensor("op_1000"), val = tensor([1, 1])]; tensor var_1002 = const()[name = tensor("op_1002"), val = tensor([1, 1])]; tensor current_key_9_pad_type_0 = const()[name = tensor("current_key_9_pad_type_0"), val = tensor("custom")]; tensor current_key_9_pad_0 = const()[name = tensor("current_key_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157148736)))]; tensor current_key_9_cast_fp16 = conv(dilations = var_1002, groups = var_959, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = var_1000, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; tensor var_1007 = const()[name = tensor("op_1007"), val = tensor([1, 1])]; tensor var_1009 = const()[name = tensor("op_1009"), val = tensor([1, 1])]; tensor current_value_9_pad_type_0 = const()[name = tensor("current_value_9_pad_type_0"), val = tensor("custom")]; tensor current_value_9_pad_0 = const()[name = tensor("current_value_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158328448)))]; tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159508160)))]; tensor current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = var_1009, groups = var_959, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = var_1007, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; tensor var_1016_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1016_cast_fp16")]; tensor var_1018_cast_fp16 = mul(x = var_63_cast_fp16_4, y = var_161_cast_fp16)[name = tensor("op_1018_cast_fp16")]; tensor key_17_cast_fp16 = add(x = var_1016_cast_fp16, y = var_1018_cast_fp16)[name = tensor("key_17_cast_fp16")]; tensor var_1020_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1020_cast_fp16")]; tensor var_1022_cast_fp16 = mul(x = var_78_cast_fp16_4, y = var_161_cast_fp16)[name = tensor("op_1022_cast_fp16")]; tensor value_17_cast_fp16 = add(x = var_1020_cast_fp16, y = var_1022_cast_fp16)[name = tensor("value_17_cast_fp16")]; tensor var_1025 = const()[name = tensor("op_1025"), val = tensor([1, 12, 64, -1])]; tensor var_1026_cast_fp16 = reshape(shape = var_1025, x = query_17_cast_fp16)[name = tensor("op_1026_cast_fp16")]; tensor var_1027_to_fp16 = const()[name = tensor("op_1027_to_fp16"), val = tensor(0x1p-3)]; tensor var_1028_cast_fp16 = mul(x = var_1026_cast_fp16, y = var_1027_to_fp16)[name = tensor("op_1028_cast_fp16")]; tensor var_1029 = const()[name = tensor("op_1029"), val = tensor([1, 12, 64, -1])]; tensor var_1030_cast_fp16 = reshape(shape = var_1029, x = key_17_cast_fp16)[name = tensor("op_1030_cast_fp16")]; tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_25_cast_fp16")]; tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; tensor var_1038_cast_fp16 = softmax(axis = var_952, x = mh_w_27_cast_fp16)[name = tensor("op_1038_cast_fp16")]; tensor var_1039 = const()[name = tensor("op_1039"), val = tensor([1, 12, 64, -1])]; tensor var_1040_cast_fp16 = reshape(shape = var_1039, x = value_17_cast_fp16)[name = tensor("op_1040_cast_fp16")]; tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; tensor 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("attn_17_cast_fp16")]; tensor var_1043 = const()[name = tensor("op_1043"), val = tensor([1, 768, 1, -1])]; tensor input_41_cast_fp16 = reshape(shape = var_1043, x = attn_17_cast_fp16)[name = tensor("input_41_cast_fp16")]; tensor var_1047 = const()[name = tensor("op_1047"), val = tensor([1, 1])]; tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; tensor obj_55_pad_type_0 = const()[name = tensor("obj_55_pad_type_0"), val = tensor("custom")]; tensor obj_55_pad_0 = const()[name = tensor("obj_55_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159509760)))]; tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160689472)))]; tensor obj_55_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = var_1049, groups = var_959, pad = obj_55_pad_0, pad_type = obj_55_pad_type_0, strides = var_1047, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("obj_55_cast_fp16")]; tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_55_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([1])]; tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1059, keep_dims = var_960, x = inputs_27_cast_fp16)[name = tensor("channels_mean_27_cast_fp16")]; tensor zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor("zero_mean_27_cast_fp16")]; tensor zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor("zero_mean_sq_27_cast_fp16")]; tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1])]; tensor var_1064_cast_fp16 = reduce_mean(axes = var_1063, keep_dims = var_960, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1064_cast_fp16")]; tensor var_1065_to_fp16 = const()[name = tensor("op_1065_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1066_cast_fp16 = add(x = var_1064_cast_fp16, y = var_1065_to_fp16)[name = tensor("op_1066_cast_fp16")]; tensor denom_27_epsilon_0 = const()[name = tensor("denom_27_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0, x = var_1066_cast_fp16)[name = tensor("denom_27_cast_fp16")]; tensor out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160691072)))]; tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160692672)))]; tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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_27_cast_fp16)[name = tensor("obj_57_cast_fp16")]; tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([1, 1])]; tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([1, 1])]; tensor query_19_pad_type_0 = const()[name = tensor("query_19_pad_type_0"), val = tensor("custom")]; tensor query_19_pad_0 = const()[name = tensor("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160694272)))]; tensor layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161873984)))]; tensor query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = var_1083, groups = var_959, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = var_1081, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("query_19_cast_fp16")]; tensor var_1087 = const()[name = tensor("op_1087"), val = tensor([1, 1])]; tensor var_1089 = const()[name = tensor("op_1089"), val = tensor([1, 1])]; tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("custom")]; tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161875584)))]; tensor key_19_cast_fp16 = conv(dilations = var_1089, groups = var_959, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = var_1087, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_19_cast_fp16")]; tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([1, 1])]; tensor var_1096 = const()[name = tensor("op_1096"), val = tensor([1, 1])]; tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("custom")]; tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163055296)))]; tensor layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164235008)))]; tensor value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = var_1096, groups = var_959, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = var_1094, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_19_cast_fp16")]; tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1, 12, 64, -1])]; tensor var_1101_cast_fp16 = reshape(shape = var_1100, x = query_19_cast_fp16)[name = tensor("op_1101_cast_fp16")]; tensor var_1102_to_fp16 = const()[name = tensor("op_1102_to_fp16"), val = tensor(0x1p-3)]; tensor var_1103_cast_fp16 = mul(x = var_1101_cast_fp16, y = var_1102_to_fp16)[name = tensor("op_1103_cast_fp16")]; tensor var_1104 = const()[name = tensor("op_1104"), val = tensor([1, 12, 64, -1])]; tensor var_1105_cast_fp16 = reshape(shape = var_1104, x = key_19_cast_fp16)[name = tensor("op_1105_cast_fp16")]; tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_29_cast_fp16")]; tensor var_1108_cast_fp16 = softmax(axis = var_952, x = mh_w_29_cast_fp16)[name = tensor("op_1108_cast_fp16")]; tensor var_1109 = const()[name = tensor("op_1109"), val = tensor([1, 12, 64, -1])]; tensor var_1110_cast_fp16 = reshape(shape = var_1109, x = value_19_cast_fp16)[name = tensor("op_1110_cast_fp16")]; tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; tensor 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 = var_1108_cast_fp16)[name = tensor("attn_19_cast_fp16")]; tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1, 768, 1, -1])]; tensor input_43_cast_fp16 = reshape(shape = var_1113, x = attn_19_cast_fp16)[name = tensor("input_43_cast_fp16")]; tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([1, 1])]; tensor var_1119 = const()[name = tensor("op_1119"), val = tensor([1, 1])]; tensor obj_59_pad_type_0 = const()[name = tensor("obj_59_pad_type_0"), val = tensor("custom")]; tensor obj_59_pad_0 = const()[name = tensor("obj_59_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164236608)))]; tensor layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165416320)))]; tensor obj_59_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = var_1119, groups = var_959, pad = obj_59_pad_0, pad_type = obj_59_pad_type_0, strides = var_1117, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("obj_59_cast_fp16")]; tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_59_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([1])]; tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_1125, keep_dims = var_960, x = inputs_29_cast_fp16)[name = tensor("channels_mean_29_cast_fp16")]; tensor zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor("zero_mean_29_cast_fp16")]; tensor zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor("zero_mean_sq_29_cast_fp16")]; tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([1])]; tensor var_1130_cast_fp16 = reduce_mean(axes = var_1129, keep_dims = var_960, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_1130_cast_fp16")]; tensor var_1131_to_fp16 = const()[name = tensor("op_1131_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1132_cast_fp16 = add(x = var_1130_cast_fp16, y = var_1131_to_fp16)[name = tensor("op_1132_cast_fp16")]; tensor denom_29_epsilon_0 = const()[name = tensor("denom_29_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0, x = var_1132_cast_fp16)[name = tensor("denom_29_cast_fp16")]; tensor out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; tensor input_45_gamma_0_to_fp16 = const()[name = tensor("input_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165417920)))]; tensor input_45_beta_0_to_fp16 = const()[name = tensor("input_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165419520)))]; tensor input_45_epsilon_0_to_fp16 = const()[name = tensor("input_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_45_cast_fp16")]; tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1, 1])]; tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([1, 1])]; tensor input_47_pad_type_0 = const()[name = tensor("input_47_pad_type_0"), val = tensor("custom")]; tensor input_47_pad_0 = const()[name = tensor("input_47_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_fc1_weight_to_fp16 = const()[name = tensor("layers_4_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165421120)))]; tensor layers_4_fc1_bias_to_fp16 = const()[name = tensor("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170139776)))]; tensor input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = var_1145, groups = var_959, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = var_1143, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 1])]; tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 1])]; tensor hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("custom")]; tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_fc2_weight_to_fp16 = const()[name = tensor("layers_4_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170145984)))]; tensor layers_4_fc2_bias_to_fp16 = const()[name = tensor("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174864640)))]; tensor hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = var_1153, groups = var_959, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = var_1151, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; tensor var_1166 = const()[name = tensor("op_1166"), val = tensor(3)]; tensor var_1173 = const()[name = tensor("op_1173"), val = tensor(1)]; tensor var_1174 = const()[name = tensor("op_1174"), val = tensor(true)]; tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1])]; tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_1186, keep_dims = var_1174, x = inputs_31_cast_fp16)[name = tensor("channels_mean_31_cast_fp16")]; tensor zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor("zero_mean_31_cast_fp16")]; tensor zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor("zero_mean_sq_31_cast_fp16")]; tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1])]; tensor var_1191_cast_fp16 = reduce_mean(axes = var_1190, keep_dims = var_1174, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_1191_cast_fp16")]; tensor var_1192_to_fp16 = const()[name = tensor("op_1192_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1193_cast_fp16 = add(x = var_1191_cast_fp16, y = var_1192_to_fp16)[name = tensor("op_1193_cast_fp16")]; tensor denom_31_epsilon_0 = const()[name = tensor("denom_31_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0, x = var_1193_cast_fp16)[name = tensor("denom_31_cast_fp16")]; tensor out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; tensor obj_61_gamma_0_to_fp16 = const()[name = tensor("obj_61_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174866240)))]; tensor obj_61_beta_0_to_fp16 = const()[name = tensor("obj_61_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174867840)))]; tensor obj_61_epsilon_0_to_fp16 = const()[name = tensor("obj_61_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_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("obj_61_cast_fp16")]; tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([1, 1])]; tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([1, 1])]; tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("custom")]; tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174869440)))]; tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176049152)))]; tensor query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = var_1210, groups = var_1173, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = var_1208, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("query_21_cast_fp16")]; tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([1, 1])]; tensor var_1216 = const()[name = tensor("op_1216"), val = tensor([1, 1])]; tensor current_key_11_pad_type_0 = const()[name = tensor("current_key_11_pad_type_0"), val = tensor("custom")]; tensor current_key_11_pad_0 = const()[name = tensor("current_key_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176050752)))]; tensor current_key_11_cast_fp16 = conv(dilations = var_1216, groups = var_1173, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = var_1214, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([1, 1])]; tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 1])]; tensor current_value_11_pad_type_0 = const()[name = tensor("current_value_11_pad_type_0"), val = tensor("custom")]; tensor current_value_11_pad_0 = const()[name = tensor("current_value_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177230464)))]; tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178410176)))]; tensor current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = var_1223, groups = var_1173, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = var_1221, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; tensor var_1230_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1230_cast_fp16")]; tensor var_1232_cast_fp16 = mul(x = var_63_cast_fp16_5, y = var_161_cast_fp16)[name = tensor("op_1232_cast_fp16")]; tensor key_21_cast_fp16 = add(x = var_1230_cast_fp16, y = var_1232_cast_fp16)[name = tensor("key_21_cast_fp16")]; tensor var_1234_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1234_cast_fp16")]; tensor var_1236_cast_fp16 = mul(x = var_78_cast_fp16_5, y = var_161_cast_fp16)[name = tensor("op_1236_cast_fp16")]; tensor value_21_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1236_cast_fp16)[name = tensor("value_21_cast_fp16")]; tensor var_1239 = const()[name = tensor("op_1239"), val = tensor([1, 12, 64, -1])]; tensor var_1240_cast_fp16 = reshape(shape = var_1239, x = query_21_cast_fp16)[name = tensor("op_1240_cast_fp16")]; tensor var_1241_to_fp16 = const()[name = tensor("op_1241_to_fp16"), val = tensor(0x1p-3)]; tensor var_1242_cast_fp16 = mul(x = var_1240_cast_fp16, y = var_1241_to_fp16)[name = tensor("op_1242_cast_fp16")]; tensor var_1243 = const()[name = tensor("op_1243"), val = tensor([1, 12, 64, -1])]; tensor var_1244_cast_fp16 = reshape(shape = var_1243, x = key_21_cast_fp16)[name = tensor("op_1244_cast_fp16")]; tensor mh_w_31_transpose_x_0 = const()[name = tensor("mh_w_31_transpose_x_0"), val = tensor(true)]; tensor mh_w_31_transpose_y_0 = const()[name = tensor("mh_w_31_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_31_cast_fp16")]; tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; tensor var_1252_cast_fp16 = softmax(axis = var_1166, x = mh_w_33_cast_fp16)[name = tensor("op_1252_cast_fp16")]; tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([1, 12, 64, -1])]; tensor var_1254_cast_fp16 = reshape(shape = var_1253, x = value_21_cast_fp16)[name = tensor("op_1254_cast_fp16")]; tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; tensor 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("attn_21_cast_fp16")]; tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([1, 768, 1, -1])]; tensor input_51_cast_fp16 = reshape(shape = var_1257, x = attn_21_cast_fp16)[name = tensor("input_51_cast_fp16")]; tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([1, 1])]; tensor var_1263 = const()[name = tensor("op_1263"), val = tensor([1, 1])]; tensor obj_67_pad_type_0 = const()[name = tensor("obj_67_pad_type_0"), val = tensor("custom")]; tensor obj_67_pad_0 = const()[name = tensor("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178411776)))]; tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179591488)))]; tensor obj_67_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = var_1263, groups = var_1173, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = var_1261, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("obj_67_cast_fp16")]; tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; tensor var_1273 = const()[name = tensor("op_1273"), val = tensor([1])]; tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_1273, keep_dims = var_1174, x = inputs_33_cast_fp16)[name = tensor("channels_mean_33_cast_fp16")]; tensor zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor("zero_mean_33_cast_fp16")]; tensor zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor("zero_mean_sq_33_cast_fp16")]; tensor var_1277 = const()[name = tensor("op_1277"), val = tensor([1])]; tensor var_1278_cast_fp16 = reduce_mean(axes = var_1277, keep_dims = var_1174, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_1278_cast_fp16")]; tensor var_1279_to_fp16 = const()[name = tensor("op_1279_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1280_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1279_to_fp16)[name = tensor("op_1280_cast_fp16")]; tensor denom_33_epsilon_0 = const()[name = tensor("denom_33_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0, x = var_1280_cast_fp16)[name = tensor("denom_33_cast_fp16")]; tensor out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; tensor obj_69_gamma_0_to_fp16 = const()[name = tensor("obj_69_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179593088)))]; tensor obj_69_beta_0_to_fp16 = const()[name = tensor("obj_69_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179594688)))]; tensor obj_69_epsilon_0_to_fp16 = const()[name = tensor("obj_69_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_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("obj_69_cast_fp16")]; tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([1, 1])]; tensor var_1297 = const()[name = tensor("op_1297"), val = tensor([1, 1])]; tensor query_23_pad_type_0 = const()[name = tensor("query_23_pad_type_0"), val = tensor("custom")]; tensor query_23_pad_0 = const()[name = tensor("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179596288)))]; tensor layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180776000)))]; tensor query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = var_1297, groups = var_1173, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = var_1295, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor("query_23_cast_fp16")]; tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1])]; tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 1])]; tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("custom")]; tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180777600)))]; tensor key_23_cast_fp16 = conv(dilations = var_1303, groups = var_1173, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = var_1301, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_23_cast_fp16")]; tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([1, 1])]; tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([1, 1])]; tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("custom")]; tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181957312)))]; tensor layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183137024)))]; tensor value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = var_1310, groups = var_1173, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = var_1308, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_23_cast_fp16")]; tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 12, 64, -1])]; tensor var_1315_cast_fp16 = reshape(shape = var_1314, x = query_23_cast_fp16)[name = tensor("op_1315_cast_fp16")]; tensor var_1316_to_fp16 = const()[name = tensor("op_1316_to_fp16"), val = tensor(0x1p-3)]; tensor var_1317_cast_fp16 = mul(x = var_1315_cast_fp16, y = var_1316_to_fp16)[name = tensor("op_1317_cast_fp16")]; tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([1, 12, 64, -1])]; tensor var_1319_cast_fp16 = reshape(shape = var_1318, x = key_23_cast_fp16)[name = tensor("op_1319_cast_fp16")]; tensor mh_w_35_transpose_x_0 = const()[name = tensor("mh_w_35_transpose_x_0"), val = tensor(true)]; tensor mh_w_35_transpose_y_0 = const()[name = tensor("mh_w_35_transpose_y_0"), val = tensor(false)]; tensor 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("mh_w_35_cast_fp16")]; tensor var_1322_cast_fp16 = softmax(axis = var_1166, x = mh_w_35_cast_fp16)[name = tensor("op_1322_cast_fp16")]; tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([1, 12, 64, -1])]; tensor var_1324_cast_fp16 = reshape(shape = var_1323, x = value_23_cast_fp16)[name = tensor("op_1324_cast_fp16")]; tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; tensor 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 = var_1322_cast_fp16)[name = tensor("attn_23_cast_fp16")]; tensor var_1327 = const()[name = tensor("op_1327"), val = tensor([1, 768, 1, -1])]; tensor input_53_cast_fp16 = reshape(shape = var_1327, x = attn_23_cast_fp16)[name = tensor("input_53_cast_fp16")]; tensor var_1331 = const()[name = tensor("op_1331"), val = tensor([1, 1])]; tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1])]; tensor obj_71_pad_type_0 = const()[name = tensor("obj_71_pad_type_0"), val = tensor("custom")]; tensor obj_71_pad_0 = const()[name = tensor("obj_71_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183138624)))]; tensor layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184318336)))]; tensor obj_71_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = var_1333, groups = var_1173, pad = obj_71_pad_0, pad_type = obj_71_pad_type_0, strides = var_1331, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("obj_71_cast_fp16")]; tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_71_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; tensor var_1339 = const()[name = tensor("op_1339"), val = tensor([1])]; tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_1339, keep_dims = var_1174, x = inputs_35_cast_fp16)[name = tensor("channels_mean_35_cast_fp16")]; tensor zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor("zero_mean_35_cast_fp16")]; tensor zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor("zero_mean_sq_35_cast_fp16")]; tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([1])]; tensor var_1344_cast_fp16 = reduce_mean(axes = var_1343, keep_dims = var_1174, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_1344_cast_fp16")]; tensor var_1345_to_fp16 = const()[name = tensor("op_1345_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1346_cast_fp16 = add(x = var_1344_cast_fp16, y = var_1345_to_fp16)[name = tensor("op_1346_cast_fp16")]; tensor denom_35_epsilon_0 = const()[name = tensor("denom_35_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0, x = var_1346_cast_fp16)[name = tensor("denom_35_cast_fp16")]; tensor out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; tensor input_55_gamma_0_to_fp16 = const()[name = tensor("input_55_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184319936)))]; tensor input_55_beta_0_to_fp16 = const()[name = tensor("input_55_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184321536)))]; tensor input_55_epsilon_0_to_fp16 = const()[name = tensor("input_55_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_55_cast_fp16")]; tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([1, 1])]; tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([1, 1])]; tensor input_57_pad_type_0 = const()[name = tensor("input_57_pad_type_0"), val = tensor("custom")]; tensor input_57_pad_0 = const()[name = tensor("input_57_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_fc1_weight_to_fp16 = const()[name = tensor("layers_5_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184323136)))]; tensor layers_5_fc1_bias_to_fp16 = const()[name = tensor("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189041792)))]; tensor input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = var_1359, groups = var_1173, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = var_1357, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; tensor var_1365 = const()[name = tensor("op_1365"), val = tensor([1, 1])]; tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 1])]; tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("custom")]; tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_fc2_weight_to_fp16 = const()[name = tensor("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189048000)))]; tensor layers_5_fc2_bias_to_fp16 = const()[name = tensor("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193766656)))]; tensor hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = var_1367, groups = var_1173, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = var_1365, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; tensor var_1380 = const()[name = tensor("op_1380"), val = tensor(3)]; tensor var_1387 = const()[name = tensor("op_1387"), val = tensor(1)]; tensor var_1388 = const()[name = tensor("op_1388"), val = tensor(true)]; tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1])]; tensor channels_mean_37_cast_fp16 = reduce_mean(axes = var_1400, keep_dims = var_1388, x = inputs_37_cast_fp16)[name = tensor("channels_mean_37_cast_fp16")]; tensor zero_mean_37_cast_fp16 = sub(x = inputs_37_cast_fp16, y = channels_mean_37_cast_fp16)[name = tensor("zero_mean_37_cast_fp16")]; tensor zero_mean_sq_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = zero_mean_37_cast_fp16)[name = tensor("zero_mean_sq_37_cast_fp16")]; tensor var_1404 = const()[name = tensor("op_1404"), val = tensor([1])]; tensor var_1405_cast_fp16 = reduce_mean(axes = var_1404, keep_dims = var_1388, x = zero_mean_sq_37_cast_fp16)[name = tensor("op_1405_cast_fp16")]; tensor var_1406_to_fp16 = const()[name = tensor("op_1406_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1407_cast_fp16 = add(x = var_1405_cast_fp16, y = var_1406_to_fp16)[name = tensor("op_1407_cast_fp16")]; tensor denom_37_epsilon_0 = const()[name = tensor("denom_37_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0, x = var_1407_cast_fp16)[name = tensor("denom_37_cast_fp16")]; tensor out_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = denom_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; tensor obj_73_gamma_0_to_fp16 = const()[name = tensor("obj_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193768256)))]; tensor obj_73_beta_0_to_fp16 = const()[name = tensor("obj_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193769856)))]; tensor obj_73_epsilon_0_to_fp16 = const()[name = tensor("obj_73_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_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("obj_73_cast_fp16")]; tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([1, 1])]; tensor var_1424 = const()[name = tensor("op_1424"), val = tensor([1, 1])]; tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("custom")]; tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193771456)))]; tensor layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194951168)))]; tensor query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = var_1424, groups = var_1387, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = var_1422, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("query_25_cast_fp16")]; tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([1, 1])]; tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 1])]; tensor current_key_13_pad_type_0 = const()[name = tensor("current_key_13_pad_type_0"), val = tensor("custom")]; tensor current_key_13_pad_0 = const()[name = tensor("current_key_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194952768)))]; tensor current_key_13_cast_fp16 = conv(dilations = var_1430, groups = var_1387, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = var_1428, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; tensor var_1435 = const()[name = tensor("op_1435"), val = tensor([1, 1])]; tensor var_1437 = const()[name = tensor("op_1437"), val = tensor([1, 1])]; tensor current_value_13_pad_type_0 = const()[name = tensor("current_value_13_pad_type_0"), val = tensor("custom")]; tensor current_value_13_pad_0 = const()[name = tensor("current_value_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196132480)))]; tensor layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197312192)))]; tensor current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = var_1437, groups = var_1387, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = var_1435, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; tensor var_1444_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1444_cast_fp16")]; tensor var_1446_cast_fp16 = mul(x = var_63_cast_fp16_6, y = var_161_cast_fp16)[name = tensor("op_1446_cast_fp16")]; tensor key_25_cast_fp16 = add(x = var_1444_cast_fp16, y = var_1446_cast_fp16)[name = tensor("key_25_cast_fp16")]; tensor var_1448_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1448_cast_fp16")]; tensor var_1450_cast_fp16 = mul(x = var_78_cast_fp16_6, y = var_161_cast_fp16)[name = tensor("op_1450_cast_fp16")]; tensor value_25_cast_fp16 = add(x = var_1448_cast_fp16, y = var_1450_cast_fp16)[name = tensor("value_25_cast_fp16")]; tensor var_1453 = const()[name = tensor("op_1453"), val = tensor([1, 12, 64, -1])]; tensor var_1454_cast_fp16 = reshape(shape = var_1453, x = query_25_cast_fp16)[name = tensor("op_1454_cast_fp16")]; tensor var_1455_to_fp16 = const()[name = tensor("op_1455_to_fp16"), val = tensor(0x1p-3)]; tensor var_1456_cast_fp16 = mul(x = var_1454_cast_fp16, y = var_1455_to_fp16)[name = tensor("op_1456_cast_fp16")]; tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 12, 64, -1])]; tensor var_1458_cast_fp16 = reshape(shape = var_1457, x = key_25_cast_fp16)[name = tensor("op_1458_cast_fp16")]; tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1456_cast_fp16, y = var_1458_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; tensor var_1466_cast_fp16 = softmax(axis = var_1380, x = mh_w_39_cast_fp16)[name = tensor("op_1466_cast_fp16")]; tensor var_1467 = const()[name = tensor("op_1467"), val = tensor([1, 12, 64, -1])]; tensor var_1468_cast_fp16 = reshape(shape = var_1467, x = value_25_cast_fp16)[name = tensor("op_1468_cast_fp16")]; tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1468_cast_fp16, y = var_1466_cast_fp16)[name = tensor("attn_25_cast_fp16")]; tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([1, 768, 1, -1])]; tensor input_61_cast_fp16 = reshape(shape = var_1471, x = attn_25_cast_fp16)[name = tensor("input_61_cast_fp16")]; tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([1, 1])]; tensor var_1477 = const()[name = tensor("op_1477"), val = tensor([1, 1])]; tensor obj_79_pad_type_0 = const()[name = tensor("obj_79_pad_type_0"), val = tensor("custom")]; tensor obj_79_pad_0 = const()[name = tensor("obj_79_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197313792)))]; tensor layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198493504)))]; tensor obj_79_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = var_1477, groups = var_1387, pad = obj_79_pad_0, pad_type = obj_79_pad_type_0, strides = var_1475, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("obj_79_cast_fp16")]; tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_79_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; tensor var_1487 = const()[name = tensor("op_1487"), val = tensor([1])]; tensor channels_mean_39_cast_fp16 = reduce_mean(axes = var_1487, keep_dims = var_1388, x = inputs_39_cast_fp16)[name = tensor("channels_mean_39_cast_fp16")]; tensor zero_mean_39_cast_fp16 = sub(x = inputs_39_cast_fp16, y = channels_mean_39_cast_fp16)[name = tensor("zero_mean_39_cast_fp16")]; tensor zero_mean_sq_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = zero_mean_39_cast_fp16)[name = tensor("zero_mean_sq_39_cast_fp16")]; tensor var_1491 = const()[name = tensor("op_1491"), val = tensor([1])]; tensor var_1492_cast_fp16 = reduce_mean(axes = var_1491, keep_dims = var_1388, x = zero_mean_sq_39_cast_fp16)[name = tensor("op_1492_cast_fp16")]; tensor var_1493_to_fp16 = const()[name = tensor("op_1493_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1494_cast_fp16 = add(x = var_1492_cast_fp16, y = var_1493_to_fp16)[name = tensor("op_1494_cast_fp16")]; tensor denom_39_epsilon_0 = const()[name = tensor("denom_39_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0, x = var_1494_cast_fp16)[name = tensor("denom_39_cast_fp16")]; tensor out_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = denom_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; tensor obj_81_gamma_0_to_fp16 = const()[name = tensor("obj_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198495104)))]; tensor obj_81_beta_0_to_fp16 = const()[name = tensor("obj_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198496704)))]; tensor obj_81_epsilon_0_to_fp16 = const()[name = tensor("obj_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_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("obj_81_cast_fp16")]; tensor var_1509 = const()[name = tensor("op_1509"), val = tensor([1, 1])]; tensor var_1511 = const()[name = tensor("op_1511"), val = tensor([1, 1])]; tensor query_27_pad_type_0 = const()[name = tensor("query_27_pad_type_0"), val = tensor("custom")]; tensor query_27_pad_0 = const()[name = tensor("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198498304)))]; tensor layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199678016)))]; tensor query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = var_1511, groups = var_1387, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = var_1509, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor("query_27_cast_fp16")]; tensor var_1515 = const()[name = tensor("op_1515"), val = tensor([1, 1])]; tensor var_1517 = const()[name = tensor("op_1517"), val = tensor([1, 1])]; tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("custom")]; tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199679616)))]; tensor key_27_cast_fp16 = conv(dilations = var_1517, groups = var_1387, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = var_1515, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_27_cast_fp16")]; tensor var_1522 = const()[name = tensor("op_1522"), val = tensor([1, 1])]; tensor var_1524 = const()[name = tensor("op_1524"), val = tensor([1, 1])]; tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("custom")]; tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200859328)))]; tensor layers_6_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202039040)))]; tensor value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = var_1524, groups = var_1387, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = var_1522, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_27_cast_fp16")]; tensor var_1528 = const()[name = tensor("op_1528"), val = tensor([1, 12, 64, -1])]; tensor var_1529_cast_fp16 = reshape(shape = var_1528, x = query_27_cast_fp16)[name = tensor("op_1529_cast_fp16")]; tensor var_1530_to_fp16 = const()[name = tensor("op_1530_to_fp16"), val = tensor(0x1p-3)]; tensor var_1531_cast_fp16 = mul(x = var_1529_cast_fp16, y = var_1530_to_fp16)[name = tensor("op_1531_cast_fp16")]; tensor var_1532 = const()[name = tensor("op_1532"), val = tensor([1, 12, 64, -1])]; tensor var_1533_cast_fp16 = reshape(shape = var_1532, x = key_27_cast_fp16)[name = tensor("op_1533_cast_fp16")]; tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1531_cast_fp16, y = var_1533_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; tensor var_1536_cast_fp16 = softmax(axis = var_1380, x = mh_w_41_cast_fp16)[name = tensor("op_1536_cast_fp16")]; tensor var_1537 = const()[name = tensor("op_1537"), val = tensor([1, 12, 64, -1])]; tensor var_1538_cast_fp16 = reshape(shape = var_1537, x = value_27_cast_fp16)[name = tensor("op_1538_cast_fp16")]; tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1538_cast_fp16, y = var_1536_cast_fp16)[name = tensor("attn_27_cast_fp16")]; tensor var_1541 = const()[name = tensor("op_1541"), val = tensor([1, 768, 1, -1])]; tensor input_63_cast_fp16 = reshape(shape = var_1541, x = attn_27_cast_fp16)[name = tensor("input_63_cast_fp16")]; tensor var_1545 = const()[name = tensor("op_1545"), val = tensor([1, 1])]; tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 1])]; tensor obj_83_pad_type_0 = const()[name = tensor("obj_83_pad_type_0"), val = tensor("custom")]; tensor obj_83_pad_0 = const()[name = tensor("obj_83_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202040640)))]; tensor layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203220352)))]; tensor obj_83_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = var_1547, groups = var_1387, pad = obj_83_pad_0, pad_type = obj_83_pad_type_0, strides = var_1545, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("obj_83_cast_fp16")]; tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_83_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; tensor var_1553 = const()[name = tensor("op_1553"), val = tensor([1])]; tensor channels_mean_41_cast_fp16 = reduce_mean(axes = var_1553, keep_dims = var_1388, x = inputs_41_cast_fp16)[name = tensor("channels_mean_41_cast_fp16")]; tensor zero_mean_41_cast_fp16 = sub(x = inputs_41_cast_fp16, y = channels_mean_41_cast_fp16)[name = tensor("zero_mean_41_cast_fp16")]; tensor zero_mean_sq_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = zero_mean_41_cast_fp16)[name = tensor("zero_mean_sq_41_cast_fp16")]; tensor var_1557 = const()[name = tensor("op_1557"), val = tensor([1])]; tensor var_1558_cast_fp16 = reduce_mean(axes = var_1557, keep_dims = var_1388, x = zero_mean_sq_41_cast_fp16)[name = tensor("op_1558_cast_fp16")]; tensor var_1559_to_fp16 = const()[name = tensor("op_1559_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1560_cast_fp16 = add(x = var_1558_cast_fp16, y = var_1559_to_fp16)[name = tensor("op_1560_cast_fp16")]; tensor denom_41_epsilon_0 = const()[name = tensor("denom_41_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0, x = var_1560_cast_fp16)[name = tensor("denom_41_cast_fp16")]; tensor out_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = denom_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; tensor input_65_gamma_0_to_fp16 = const()[name = tensor("input_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203221952)))]; tensor input_65_beta_0_to_fp16 = const()[name = tensor("input_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203223552)))]; tensor input_65_epsilon_0_to_fp16 = const()[name = tensor("input_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_65_cast_fp16")]; tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([1, 1])]; tensor var_1573 = const()[name = tensor("op_1573"), val = tensor([1, 1])]; tensor input_67_pad_type_0 = const()[name = tensor("input_67_pad_type_0"), val = tensor("custom")]; tensor input_67_pad_0 = const()[name = tensor("input_67_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_fc1_weight_to_fp16 = const()[name = tensor("layers_6_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203225152)))]; tensor layers_6_fc1_bias_to_fp16 = const()[name = tensor("layers_6_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207943808)))]; tensor input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = var_1573, groups = var_1387, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = var_1571, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("input_67_cast_fp16")]; tensor input_69_mode_0 = const()[name = tensor("input_69_mode_0"), val = tensor("EXACT")]; tensor input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; tensor var_1579 = const()[name = tensor("op_1579"), val = tensor([1, 1])]; tensor var_1581 = const()[name = tensor("op_1581"), val = tensor([1, 1])]; tensor hidden_states_15_pad_type_0 = const()[name = tensor("hidden_states_15_pad_type_0"), val = tensor("custom")]; tensor hidden_states_15_pad_0 = const()[name = tensor("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_fc2_weight_to_fp16 = const()[name = tensor("layers_6_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207950016)))]; tensor layers_6_fc2_bias_to_fp16 = const()[name = tensor("layers_6_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212668672)))]; tensor hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = var_1581, groups = var_1387, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = var_1579, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("hidden_states_15_cast_fp16")]; tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; tensor var_1594 = const()[name = tensor("op_1594"), val = tensor(3)]; tensor var_1601 = const()[name = tensor("op_1601"), val = tensor(1)]; tensor var_1602 = const()[name = tensor("op_1602"), val = tensor(true)]; tensor var_1614 = const()[name = tensor("op_1614"), val = tensor([1])]; tensor channels_mean_43_cast_fp16 = reduce_mean(axes = var_1614, keep_dims = var_1602, x = inputs_43_cast_fp16)[name = tensor("channels_mean_43_cast_fp16")]; tensor zero_mean_43_cast_fp16 = sub(x = inputs_43_cast_fp16, y = channels_mean_43_cast_fp16)[name = tensor("zero_mean_43_cast_fp16")]; tensor zero_mean_sq_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = zero_mean_43_cast_fp16)[name = tensor("zero_mean_sq_43_cast_fp16")]; tensor var_1618 = const()[name = tensor("op_1618"), val = tensor([1])]; tensor var_1619_cast_fp16 = reduce_mean(axes = var_1618, keep_dims = var_1602, x = zero_mean_sq_43_cast_fp16)[name = tensor("op_1619_cast_fp16")]; tensor var_1620_to_fp16 = const()[name = tensor("op_1620_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1621_cast_fp16 = add(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = tensor("op_1621_cast_fp16")]; tensor denom_43_epsilon_0 = const()[name = tensor("denom_43_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0, x = var_1621_cast_fp16)[name = tensor("denom_43_cast_fp16")]; tensor out_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = denom_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212670272)))]; tensor obj_85_beta_0_to_fp16 = const()[name = tensor("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212671872)))]; tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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_43_cast_fp16)[name = tensor("obj_85_cast_fp16")]; tensor var_1636 = const()[name = tensor("op_1636"), val = tensor([1, 1])]; tensor var_1638 = const()[name = tensor("op_1638"), val = tensor([1, 1])]; tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("custom")]; tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212673472)))]; tensor layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213853184)))]; tensor query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = var_1638, groups = var_1601, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = var_1636, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("query_29_cast_fp16")]; tensor var_1642 = const()[name = tensor("op_1642"), val = tensor([1, 1])]; tensor var_1644 = const()[name = tensor("op_1644"), val = tensor([1, 1])]; tensor current_key_15_pad_type_0 = const()[name = tensor("current_key_15_pad_type_0"), val = tensor("custom")]; tensor current_key_15_pad_0 = const()[name = tensor("current_key_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213854784)))]; tensor current_key_15_cast_fp16 = conv(dilations = var_1644, groups = var_1601, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = var_1642, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; tensor var_1649 = const()[name = tensor("op_1649"), val = tensor([1, 1])]; tensor var_1651 = const()[name = tensor("op_1651"), val = tensor([1, 1])]; tensor current_value_15_pad_type_0 = const()[name = tensor("current_value_15_pad_type_0"), val = tensor("custom")]; tensor current_value_15_pad_0 = const()[name = tensor("current_value_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215034496)))]; tensor layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216214208)))]; tensor current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = var_1651, groups = var_1601, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = var_1649, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; tensor var_1658_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1658_cast_fp16")]; tensor var_1660_cast_fp16 = mul(x = var_63_cast_fp16_7, y = var_161_cast_fp16)[name = tensor("op_1660_cast_fp16")]; tensor key_29_cast_fp16 = add(x = var_1658_cast_fp16, y = var_1660_cast_fp16)[name = tensor("key_29_cast_fp16")]; tensor var_1662_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1662_cast_fp16")]; tensor var_1664_cast_fp16 = mul(x = var_78_cast_fp16_7, y = var_161_cast_fp16)[name = tensor("op_1664_cast_fp16")]; tensor value_29_cast_fp16 = add(x = var_1662_cast_fp16, y = var_1664_cast_fp16)[name = tensor("value_29_cast_fp16")]; tensor var_1667 = const()[name = tensor("op_1667"), val = tensor([1, 12, 64, -1])]; tensor var_1668_cast_fp16 = reshape(shape = var_1667, x = query_29_cast_fp16)[name = tensor("op_1668_cast_fp16")]; tensor var_1669_to_fp16 = const()[name = tensor("op_1669_to_fp16"), val = tensor(0x1p-3)]; tensor var_1670_cast_fp16 = mul(x = var_1668_cast_fp16, y = var_1669_to_fp16)[name = tensor("op_1670_cast_fp16")]; tensor var_1671 = const()[name = tensor("op_1671"), val = tensor([1, 12, 64, -1])]; tensor var_1672_cast_fp16 = reshape(shape = var_1671, x = key_29_cast_fp16)[name = tensor("op_1672_cast_fp16")]; tensor mh_w_43_transpose_x_0 = const()[name = tensor("mh_w_43_transpose_x_0"), val = tensor(true)]; tensor mh_w_43_transpose_y_0 = const()[name = tensor("mh_w_43_transpose_y_0"), val = tensor(false)]; tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_1670_cast_fp16, y = var_1672_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; tensor var_1680_cast_fp16 = softmax(axis = var_1594, x = mh_w_45_cast_fp16)[name = tensor("op_1680_cast_fp16")]; tensor var_1681 = const()[name = tensor("op_1681"), val = tensor([1, 12, 64, -1])]; tensor var_1682_cast_fp16 = reshape(shape = var_1681, x = value_29_cast_fp16)[name = tensor("op_1682_cast_fp16")]; tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1682_cast_fp16, y = var_1680_cast_fp16)[name = tensor("attn_29_cast_fp16")]; tensor var_1685 = const()[name = tensor("op_1685"), val = tensor([1, 768, 1, -1])]; tensor input_71_cast_fp16 = reshape(shape = var_1685, x = attn_29_cast_fp16)[name = tensor("input_71_cast_fp16")]; tensor var_1689 = const()[name = tensor("op_1689"), val = tensor([1, 1])]; tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1])]; tensor obj_91_pad_type_0 = const()[name = tensor("obj_91_pad_type_0"), val = tensor("custom")]; tensor obj_91_pad_0 = const()[name = tensor("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216215808)))]; tensor layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217395520)))]; tensor obj_91_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = var_1691, groups = var_1601, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = var_1689, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("obj_91_cast_fp16")]; tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; tensor var_1701 = const()[name = tensor("op_1701"), val = tensor([1])]; tensor channels_mean_45_cast_fp16 = reduce_mean(axes = var_1701, keep_dims = var_1602, x = inputs_45_cast_fp16)[name = tensor("channels_mean_45_cast_fp16")]; tensor zero_mean_45_cast_fp16 = sub(x = inputs_45_cast_fp16, y = channels_mean_45_cast_fp16)[name = tensor("zero_mean_45_cast_fp16")]; tensor zero_mean_sq_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = zero_mean_45_cast_fp16)[name = tensor("zero_mean_sq_45_cast_fp16")]; tensor var_1705 = const()[name = tensor("op_1705"), val = tensor([1])]; tensor var_1706_cast_fp16 = reduce_mean(axes = var_1705, keep_dims = var_1602, x = zero_mean_sq_45_cast_fp16)[name = tensor("op_1706_cast_fp16")]; tensor var_1707_to_fp16 = const()[name = tensor("op_1707_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1708_cast_fp16 = add(x = var_1706_cast_fp16, y = var_1707_to_fp16)[name = tensor("op_1708_cast_fp16")]; tensor denom_45_epsilon_0 = const()[name = tensor("denom_45_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0, x = var_1708_cast_fp16)[name = tensor("denom_45_cast_fp16")]; tensor out_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = denom_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217397120)))]; tensor obj_93_beta_0_to_fp16 = const()[name = tensor("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217398720)))]; tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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_45_cast_fp16)[name = tensor("obj_93_cast_fp16")]; tensor var_1723 = const()[name = tensor("op_1723"), val = tensor([1, 1])]; tensor var_1725 = const()[name = tensor("op_1725"), val = tensor([1, 1])]; tensor query_31_pad_type_0 = const()[name = tensor("query_31_pad_type_0"), val = tensor("custom")]; tensor query_31_pad_0 = const()[name = tensor("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217400320)))]; tensor layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218580032)))]; tensor query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = var_1725, groups = var_1601, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = var_1723, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("query_31_cast_fp16")]; tensor var_1729 = const()[name = tensor("op_1729"), val = tensor([1, 1])]; tensor var_1731 = const()[name = tensor("op_1731"), val = tensor([1, 1])]; tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("custom")]; tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218581632)))]; tensor key_31_cast_fp16 = conv(dilations = var_1731, groups = var_1601, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = var_1729, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_31_cast_fp16")]; tensor var_1736 = const()[name = tensor("op_1736"), val = tensor([1, 1])]; tensor var_1738 = const()[name = tensor("op_1738"), val = tensor([1, 1])]; tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("custom")]; tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219761344)))]; tensor layers_7_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220941056)))]; tensor value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = var_1738, groups = var_1601, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = var_1736, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_31_cast_fp16")]; tensor var_1742 = const()[name = tensor("op_1742"), val = tensor([1, 12, 64, -1])]; tensor var_1743_cast_fp16 = reshape(shape = var_1742, x = query_31_cast_fp16)[name = tensor("op_1743_cast_fp16")]; tensor var_1744_to_fp16 = const()[name = tensor("op_1744_to_fp16"), val = tensor(0x1p-3)]; tensor var_1745_cast_fp16 = mul(x = var_1743_cast_fp16, y = var_1744_to_fp16)[name = tensor("op_1745_cast_fp16")]; tensor var_1746 = const()[name = tensor("op_1746"), val = tensor([1, 12, 64, -1])]; tensor var_1747_cast_fp16 = reshape(shape = var_1746, x = key_31_cast_fp16)[name = tensor("op_1747_cast_fp16")]; tensor mh_w_47_transpose_x_0 = const()[name = tensor("mh_w_47_transpose_x_0"), val = tensor(true)]; tensor mh_w_47_transpose_y_0 = const()[name = tensor("mh_w_47_transpose_y_0"), val = tensor(false)]; tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_1745_cast_fp16, y = var_1747_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; tensor var_1750_cast_fp16 = softmax(axis = var_1594, x = mh_w_47_cast_fp16)[name = tensor("op_1750_cast_fp16")]; tensor var_1751 = const()[name = tensor("op_1751"), val = tensor([1, 12, 64, -1])]; tensor var_1752_cast_fp16 = reshape(shape = var_1751, x = value_31_cast_fp16)[name = tensor("op_1752_cast_fp16")]; tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_1752_cast_fp16, y = var_1750_cast_fp16)[name = tensor("attn_31_cast_fp16")]; tensor var_1755 = const()[name = tensor("op_1755"), val = tensor([1, 768, 1, -1])]; tensor input_73_cast_fp16 = reshape(shape = var_1755, x = attn_31_cast_fp16)[name = tensor("input_73_cast_fp16")]; tensor var_1759 = const()[name = tensor("op_1759"), val = tensor([1, 1])]; tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([1, 1])]; tensor obj_95_pad_type_0 = const()[name = tensor("obj_95_pad_type_0"), val = tensor("custom")]; tensor obj_95_pad_0 = const()[name = tensor("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220942656)))]; tensor layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222122368)))]; tensor obj_95_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = var_1761, groups = var_1601, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = var_1759, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("obj_95_cast_fp16")]; tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; tensor var_1767 = const()[name = tensor("op_1767"), val = tensor([1])]; tensor channels_mean_47_cast_fp16 = reduce_mean(axes = var_1767, keep_dims = var_1602, x = inputs_47_cast_fp16)[name = tensor("channels_mean_47_cast_fp16")]; tensor zero_mean_47_cast_fp16 = sub(x = inputs_47_cast_fp16, y = channels_mean_47_cast_fp16)[name = tensor("zero_mean_47_cast_fp16")]; tensor zero_mean_sq_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = zero_mean_47_cast_fp16)[name = tensor("zero_mean_sq_47_cast_fp16")]; tensor var_1771 = const()[name = tensor("op_1771"), val = tensor([1])]; tensor var_1772_cast_fp16 = reduce_mean(axes = var_1771, keep_dims = var_1602, x = zero_mean_sq_47_cast_fp16)[name = tensor("op_1772_cast_fp16")]; tensor var_1773_to_fp16 = const()[name = tensor("op_1773_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1774_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1773_to_fp16)[name = tensor("op_1774_cast_fp16")]; tensor denom_47_epsilon_0 = const()[name = tensor("denom_47_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0, x = var_1774_cast_fp16)[name = tensor("denom_47_cast_fp16")]; tensor out_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = denom_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; tensor input_75_gamma_0_to_fp16 = const()[name = tensor("input_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222123968)))]; tensor input_75_beta_0_to_fp16 = const()[name = tensor("input_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222125568)))]; tensor input_75_epsilon_0_to_fp16 = const()[name = tensor("input_75_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_75_cast_fp16")]; tensor var_1785 = const()[name = tensor("op_1785"), val = tensor([1, 1])]; tensor var_1787 = const()[name = tensor("op_1787"), val = tensor([1, 1])]; tensor input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("custom")]; tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_fc1_weight_to_fp16 = const()[name = tensor("layers_7_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222127168)))]; tensor layers_7_fc1_bias_to_fp16 = const()[name = tensor("layers_7_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226845824)))]; tensor input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = var_1787, groups = var_1601, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = var_1785, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; tensor input_79_mode_0 = const()[name = tensor("input_79_mode_0"), val = tensor("EXACT")]; tensor input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; tensor var_1793 = const()[name = tensor("op_1793"), val = tensor([1, 1])]; tensor var_1795 = const()[name = tensor("op_1795"), val = tensor([1, 1])]; tensor hidden_states_17_pad_type_0 = const()[name = tensor("hidden_states_17_pad_type_0"), val = tensor("custom")]; tensor hidden_states_17_pad_0 = const()[name = tensor("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_fc2_weight_to_fp16 = const()[name = tensor("layers_7_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226852032)))]; tensor layers_7_fc2_bias_to_fp16 = const()[name = tensor("layers_7_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231570688)))]; tensor hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = var_1795, groups = var_1601, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = var_1793, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("hidden_states_17_cast_fp16")]; tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; tensor var_1808 = const()[name = tensor("op_1808"), val = tensor(3)]; tensor var_1815 = const()[name = tensor("op_1815"), val = tensor(1)]; tensor var_1816 = const()[name = tensor("op_1816"), val = tensor(true)]; tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([1])]; tensor channels_mean_49_cast_fp16 = reduce_mean(axes = var_1828, keep_dims = var_1816, x = inputs_49_cast_fp16)[name = tensor("channels_mean_49_cast_fp16")]; tensor zero_mean_49_cast_fp16 = sub(x = inputs_49_cast_fp16, y = channels_mean_49_cast_fp16)[name = tensor("zero_mean_49_cast_fp16")]; tensor zero_mean_sq_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = zero_mean_49_cast_fp16)[name = tensor("zero_mean_sq_49_cast_fp16")]; tensor var_1832 = const()[name = tensor("op_1832"), val = tensor([1])]; tensor var_1833_cast_fp16 = reduce_mean(axes = var_1832, keep_dims = var_1816, x = zero_mean_sq_49_cast_fp16)[name = tensor("op_1833_cast_fp16")]; tensor var_1834_to_fp16 = const()[name = tensor("op_1834_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1835_cast_fp16 = add(x = var_1833_cast_fp16, y = var_1834_to_fp16)[name = tensor("op_1835_cast_fp16")]; tensor denom_49_epsilon_0 = const()[name = tensor("denom_49_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0, x = var_1835_cast_fp16)[name = tensor("denom_49_cast_fp16")]; tensor out_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = denom_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; tensor obj_97_gamma_0_to_fp16 = const()[name = tensor("obj_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231572288)))]; tensor obj_97_beta_0_to_fp16 = const()[name = tensor("obj_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231573888)))]; tensor obj_97_epsilon_0_to_fp16 = const()[name = tensor("obj_97_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_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("obj_97_cast_fp16")]; tensor var_1850 = const()[name = tensor("op_1850"), val = tensor([1, 1])]; tensor var_1852 = const()[name = tensor("op_1852"), val = tensor([1, 1])]; tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("custom")]; tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231575488)))]; tensor layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232755200)))]; tensor query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = var_1852, groups = var_1815, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = var_1850, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("query_33_cast_fp16")]; tensor var_1856 = const()[name = tensor("op_1856"), val = tensor([1, 1])]; tensor var_1858 = const()[name = tensor("op_1858"), val = tensor([1, 1])]; tensor current_key_17_pad_type_0 = const()[name = tensor("current_key_17_pad_type_0"), val = tensor("custom")]; tensor current_key_17_pad_0 = const()[name = tensor("current_key_17_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232756800)))]; tensor current_key_17_cast_fp16 = conv(dilations = var_1858, groups = var_1815, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = var_1856, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; tensor var_1863 = const()[name = tensor("op_1863"), val = tensor([1, 1])]; tensor var_1865 = const()[name = tensor("op_1865"), val = tensor([1, 1])]; tensor current_value_17_pad_type_0 = const()[name = tensor("current_value_17_pad_type_0"), val = tensor("custom")]; tensor current_value_17_pad_0 = const()[name = tensor("current_value_17_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233936512)))]; tensor layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235116224)))]; tensor current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = var_1865, groups = var_1815, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = var_1863, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; tensor var_1872_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1872_cast_fp16")]; tensor var_1874_cast_fp16 = mul(x = var_63_cast_fp16_8, y = var_161_cast_fp16)[name = tensor("op_1874_cast_fp16")]; tensor key_33_cast_fp16 = add(x = var_1872_cast_fp16, y = var_1874_cast_fp16)[name = tensor("key_33_cast_fp16")]; tensor var_1876_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1876_cast_fp16")]; tensor var_1878_cast_fp16 = mul(x = var_78_cast_fp16_8, y = var_161_cast_fp16)[name = tensor("op_1878_cast_fp16")]; tensor value_33_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1878_cast_fp16)[name = tensor("value_33_cast_fp16")]; tensor var_1881 = const()[name = tensor("op_1881"), val = tensor([1, 12, 64, -1])]; tensor var_1882_cast_fp16 = reshape(shape = var_1881, x = query_33_cast_fp16)[name = tensor("op_1882_cast_fp16")]; tensor var_1883_to_fp16 = const()[name = tensor("op_1883_to_fp16"), val = tensor(0x1p-3)]; tensor var_1884_cast_fp16 = mul(x = var_1882_cast_fp16, y = var_1883_to_fp16)[name = tensor("op_1884_cast_fp16")]; tensor var_1885 = const()[name = tensor("op_1885"), val = tensor([1, 12, 64, -1])]; tensor var_1886_cast_fp16 = reshape(shape = var_1885, x = key_33_cast_fp16)[name = tensor("op_1886_cast_fp16")]; tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_1884_cast_fp16, y = var_1886_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; tensor var_1894_cast_fp16 = softmax(axis = var_1808, x = mh_w_51_cast_fp16)[name = tensor("op_1894_cast_fp16")]; tensor var_1895 = const()[name = tensor("op_1895"), val = tensor([1, 12, 64, -1])]; tensor var_1896_cast_fp16 = reshape(shape = var_1895, x = value_33_cast_fp16)[name = tensor("op_1896_cast_fp16")]; tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_1896_cast_fp16, y = var_1894_cast_fp16)[name = tensor("attn_33_cast_fp16")]; tensor var_1899 = const()[name = tensor("op_1899"), val = tensor([1, 768, 1, -1])]; tensor input_81_cast_fp16 = reshape(shape = var_1899, x = attn_33_cast_fp16)[name = tensor("input_81_cast_fp16")]; tensor var_1903 = const()[name = tensor("op_1903"), val = tensor([1, 1])]; tensor var_1905 = const()[name = tensor("op_1905"), val = tensor([1, 1])]; tensor obj_103_pad_type_0 = const()[name = tensor("obj_103_pad_type_0"), val = tensor("custom")]; tensor obj_103_pad_0 = const()[name = tensor("obj_103_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235117824)))]; tensor layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236297536)))]; tensor obj_103_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = var_1905, groups = var_1815, pad = obj_103_pad_0, pad_type = obj_103_pad_type_0, strides = var_1903, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("obj_103_cast_fp16")]; tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_103_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; tensor var_1915 = const()[name = tensor("op_1915"), val = tensor([1])]; tensor channels_mean_51_cast_fp16 = reduce_mean(axes = var_1915, keep_dims = var_1816, x = inputs_51_cast_fp16)[name = tensor("channels_mean_51_cast_fp16")]; tensor zero_mean_51_cast_fp16 = sub(x = inputs_51_cast_fp16, y = channels_mean_51_cast_fp16)[name = tensor("zero_mean_51_cast_fp16")]; tensor zero_mean_sq_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = zero_mean_51_cast_fp16)[name = tensor("zero_mean_sq_51_cast_fp16")]; tensor var_1919 = const()[name = tensor("op_1919"), val = tensor([1])]; tensor var_1920_cast_fp16 = reduce_mean(axes = var_1919, keep_dims = var_1816, x = zero_mean_sq_51_cast_fp16)[name = tensor("op_1920_cast_fp16")]; tensor var_1921_to_fp16 = const()[name = tensor("op_1921_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1922_cast_fp16 = add(x = var_1920_cast_fp16, y = var_1921_to_fp16)[name = tensor("op_1922_cast_fp16")]; tensor denom_51_epsilon_0 = const()[name = tensor("denom_51_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0, x = var_1922_cast_fp16)[name = tensor("denom_51_cast_fp16")]; tensor out_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = denom_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; tensor obj_105_gamma_0_to_fp16 = const()[name = tensor("obj_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236299136)))]; tensor obj_105_beta_0_to_fp16 = const()[name = tensor("obj_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236300736)))]; tensor obj_105_epsilon_0_to_fp16 = const()[name = tensor("obj_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_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("obj_105_cast_fp16")]; tensor var_1937 = const()[name = tensor("op_1937"), val = tensor([1, 1])]; tensor var_1939 = const()[name = tensor("op_1939"), val = tensor([1, 1])]; tensor query_35_pad_type_0 = const()[name = tensor("query_35_pad_type_0"), val = tensor("custom")]; tensor query_35_pad_0 = const()[name = tensor("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236302336)))]; tensor layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237482048)))]; tensor query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = var_1939, groups = var_1815, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = var_1937, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor("query_35_cast_fp16")]; tensor var_1943 = const()[name = tensor("op_1943"), val = tensor([1, 1])]; tensor var_1945 = const()[name = tensor("op_1945"), val = tensor([1, 1])]; tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("custom")]; tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237483648)))]; tensor key_35_cast_fp16 = conv(dilations = var_1945, groups = var_1815, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = var_1943, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_35_cast_fp16")]; tensor var_1950 = const()[name = tensor("op_1950"), val = tensor([1, 1])]; tensor var_1952 = const()[name = tensor("op_1952"), val = tensor([1, 1])]; tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("custom")]; tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238663360)))]; tensor layers_8_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239843072)))]; tensor value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = var_1952, groups = var_1815, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = var_1950, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_35_cast_fp16")]; tensor var_1956 = const()[name = tensor("op_1956"), val = tensor([1, 12, 64, -1])]; tensor var_1957_cast_fp16 = reshape(shape = var_1956, x = query_35_cast_fp16)[name = tensor("op_1957_cast_fp16")]; tensor var_1958_to_fp16 = const()[name = tensor("op_1958_to_fp16"), val = tensor(0x1p-3)]; tensor var_1959_cast_fp16 = mul(x = var_1957_cast_fp16, y = var_1958_to_fp16)[name = tensor("op_1959_cast_fp16")]; tensor var_1960 = const()[name = tensor("op_1960"), val = tensor([1, 12, 64, -1])]; tensor var_1961_cast_fp16 = reshape(shape = var_1960, x = key_35_cast_fp16)[name = tensor("op_1961_cast_fp16")]; tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_1959_cast_fp16, y = var_1961_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; tensor var_1964_cast_fp16 = softmax(axis = var_1808, x = mh_w_53_cast_fp16)[name = tensor("op_1964_cast_fp16")]; tensor var_1965 = const()[name = tensor("op_1965"), val = tensor([1, 12, 64, -1])]; tensor var_1966_cast_fp16 = reshape(shape = var_1965, x = value_35_cast_fp16)[name = tensor("op_1966_cast_fp16")]; tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_1966_cast_fp16, y = var_1964_cast_fp16)[name = tensor("attn_35_cast_fp16")]; tensor var_1969 = const()[name = tensor("op_1969"), val = tensor([1, 768, 1, -1])]; tensor input_83_cast_fp16 = reshape(shape = var_1969, x = attn_35_cast_fp16)[name = tensor("input_83_cast_fp16")]; tensor var_1973 = const()[name = tensor("op_1973"), val = tensor([1, 1])]; tensor var_1975 = const()[name = tensor("op_1975"), val = tensor([1, 1])]; tensor obj_107_pad_type_0 = const()[name = tensor("obj_107_pad_type_0"), val = tensor("custom")]; tensor obj_107_pad_0 = const()[name = tensor("obj_107_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239844672)))]; tensor layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241024384)))]; tensor obj_107_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = var_1975, groups = var_1815, pad = obj_107_pad_0, pad_type = obj_107_pad_type_0, strides = var_1973, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("obj_107_cast_fp16")]; tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_107_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; tensor var_1981 = const()[name = tensor("op_1981"), val = tensor([1])]; tensor channels_mean_53_cast_fp16 = reduce_mean(axes = var_1981, keep_dims = var_1816, x = inputs_53_cast_fp16)[name = tensor("channels_mean_53_cast_fp16")]; tensor zero_mean_53_cast_fp16 = sub(x = inputs_53_cast_fp16, y = channels_mean_53_cast_fp16)[name = tensor("zero_mean_53_cast_fp16")]; tensor zero_mean_sq_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = zero_mean_53_cast_fp16)[name = tensor("zero_mean_sq_53_cast_fp16")]; tensor var_1985 = const()[name = tensor("op_1985"), val = tensor([1])]; tensor var_1986_cast_fp16 = reduce_mean(axes = var_1985, keep_dims = var_1816, x = zero_mean_sq_53_cast_fp16)[name = tensor("op_1986_cast_fp16")]; tensor var_1987_to_fp16 = const()[name = tensor("op_1987_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1988_cast_fp16 = add(x = var_1986_cast_fp16, y = var_1987_to_fp16)[name = tensor("op_1988_cast_fp16")]; tensor denom_53_epsilon_0 = const()[name = tensor("denom_53_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0, x = var_1988_cast_fp16)[name = tensor("denom_53_cast_fp16")]; tensor out_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = denom_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; tensor input_85_gamma_0_to_fp16 = const()[name = tensor("input_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241025984)))]; tensor input_85_beta_0_to_fp16 = const()[name = tensor("input_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241027584)))]; tensor input_85_epsilon_0_to_fp16 = const()[name = tensor("input_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_85_cast_fp16")]; tensor var_1999 = const()[name = tensor("op_1999"), val = tensor([1, 1])]; tensor var_2001 = const()[name = tensor("op_2001"), val = tensor([1, 1])]; tensor input_87_pad_type_0 = const()[name = tensor("input_87_pad_type_0"), val = tensor("custom")]; tensor input_87_pad_0 = const()[name = tensor("input_87_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_fc1_weight_to_fp16 = const()[name = tensor("layers_8_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241029184)))]; tensor layers_8_fc1_bias_to_fp16 = const()[name = tensor("layers_8_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245747840)))]; tensor input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = var_2001, groups = var_1815, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = var_1999, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; tensor input_89_mode_0 = const()[name = tensor("input_89_mode_0"), val = tensor("EXACT")]; tensor input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; tensor var_2007 = const()[name = tensor("op_2007"), val = tensor([1, 1])]; tensor var_2009 = const()[name = tensor("op_2009"), val = tensor([1, 1])]; tensor hidden_states_19_pad_type_0 = const()[name = tensor("hidden_states_19_pad_type_0"), val = tensor("custom")]; tensor hidden_states_19_pad_0 = const()[name = tensor("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_fc2_weight_to_fp16 = const()[name = tensor("layers_8_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245754048)))]; tensor layers_8_fc2_bias_to_fp16 = const()[name = tensor("layers_8_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250472704)))]; tensor hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = var_2009, groups = var_1815, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = var_2007, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; tensor var_2022 = const()[name = tensor("op_2022"), val = tensor(3)]; tensor var_2029 = const()[name = tensor("op_2029"), val = tensor(1)]; tensor var_2030 = const()[name = tensor("op_2030"), val = tensor(true)]; tensor var_2042 = const()[name = tensor("op_2042"), val = tensor([1])]; tensor channels_mean_55_cast_fp16 = reduce_mean(axes = var_2042, keep_dims = var_2030, x = inputs_55_cast_fp16)[name = tensor("channels_mean_55_cast_fp16")]; tensor zero_mean_55_cast_fp16 = sub(x = inputs_55_cast_fp16, y = channels_mean_55_cast_fp16)[name = tensor("zero_mean_55_cast_fp16")]; tensor zero_mean_sq_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = zero_mean_55_cast_fp16)[name = tensor("zero_mean_sq_55_cast_fp16")]; tensor var_2046 = const()[name = tensor("op_2046"), val = tensor([1])]; tensor var_2047_cast_fp16 = reduce_mean(axes = var_2046, keep_dims = var_2030, x = zero_mean_sq_55_cast_fp16)[name = tensor("op_2047_cast_fp16")]; tensor var_2048_to_fp16 = const()[name = tensor("op_2048_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2049_cast_fp16 = add(x = var_2047_cast_fp16, y = var_2048_to_fp16)[name = tensor("op_2049_cast_fp16")]; tensor denom_55_epsilon_0 = const()[name = tensor("denom_55_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0, x = var_2049_cast_fp16)[name = tensor("denom_55_cast_fp16")]; tensor out_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = denom_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; tensor obj_109_gamma_0_to_fp16 = const()[name = tensor("obj_109_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250474304)))]; tensor obj_109_beta_0_to_fp16 = const()[name = tensor("obj_109_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250475904)))]; tensor obj_109_epsilon_0_to_fp16 = const()[name = tensor("obj_109_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_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("obj_109_cast_fp16")]; tensor var_2064 = const()[name = tensor("op_2064"), val = tensor([1, 1])]; tensor var_2066 = const()[name = tensor("op_2066"), val = tensor([1, 1])]; tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("custom")]; tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250477504)))]; tensor layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251657216)))]; tensor query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = var_2066, groups = var_2029, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = var_2064, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("query_37_cast_fp16")]; tensor var_2070 = const()[name = tensor("op_2070"), val = tensor([1, 1])]; tensor var_2072 = const()[name = tensor("op_2072"), val = tensor([1, 1])]; tensor current_key_19_pad_type_0 = const()[name = tensor("current_key_19_pad_type_0"), val = tensor("custom")]; tensor current_key_19_pad_0 = const()[name = tensor("current_key_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251658816)))]; tensor current_key_19_cast_fp16 = conv(dilations = var_2072, groups = var_2029, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = var_2070, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; tensor var_2077 = const()[name = tensor("op_2077"), val = tensor([1, 1])]; tensor var_2079 = const()[name = tensor("op_2079"), val = tensor([1, 1])]; tensor current_value_19_pad_type_0 = const()[name = tensor("current_value_19_pad_type_0"), val = tensor("custom")]; tensor current_value_19_pad_0 = const()[name = tensor("current_value_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252838528)))]; tensor layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254018240)))]; tensor current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = var_2079, groups = var_2029, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = var_2077, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; tensor var_2086_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_2086_cast_fp16")]; tensor var_2088_cast_fp16 = mul(x = var_63_cast_fp16_9, y = var_161_cast_fp16)[name = tensor("op_2088_cast_fp16")]; tensor key_37_cast_fp16 = add(x = var_2086_cast_fp16, y = var_2088_cast_fp16)[name = tensor("key_37_cast_fp16")]; tensor var_2090_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_2090_cast_fp16")]; tensor var_2092_cast_fp16 = mul(x = var_78_cast_fp16_9, y = var_161_cast_fp16)[name = tensor("op_2092_cast_fp16")]; tensor value_37_cast_fp16 = add(x = var_2090_cast_fp16, y = var_2092_cast_fp16)[name = tensor("value_37_cast_fp16")]; tensor var_2095 = const()[name = tensor("op_2095"), val = tensor([1, 12, 64, -1])]; tensor var_2096_cast_fp16 = reshape(shape = var_2095, x = query_37_cast_fp16)[name = tensor("op_2096_cast_fp16")]; tensor var_2097_to_fp16 = const()[name = tensor("op_2097_to_fp16"), val = tensor(0x1p-3)]; tensor var_2098_cast_fp16 = mul(x = var_2096_cast_fp16, y = var_2097_to_fp16)[name = tensor("op_2098_cast_fp16")]; tensor var_2099 = const()[name = tensor("op_2099"), val = tensor([1, 12, 64, -1])]; tensor var_2100_cast_fp16 = reshape(shape = var_2099, x = key_37_cast_fp16)[name = tensor("op_2100_cast_fp16")]; tensor mh_w_55_transpose_x_0 = const()[name = tensor("mh_w_55_transpose_x_0"), val = tensor(true)]; tensor mh_w_55_transpose_y_0 = const()[name = tensor("mh_w_55_transpose_y_0"), val = tensor(false)]; tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_2098_cast_fp16, y = var_2100_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; tensor var_2108_cast_fp16 = softmax(axis = var_2022, x = mh_w_57_cast_fp16)[name = tensor("op_2108_cast_fp16")]; tensor var_2109 = const()[name = tensor("op_2109"), val = tensor([1, 12, 64, -1])]; tensor var_2110_cast_fp16 = reshape(shape = var_2109, x = value_37_cast_fp16)[name = tensor("op_2110_cast_fp16")]; tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2110_cast_fp16, y = var_2108_cast_fp16)[name = tensor("attn_37_cast_fp16")]; tensor var_2113 = const()[name = tensor("op_2113"), val = tensor([1, 768, 1, -1])]; tensor input_91_cast_fp16 = reshape(shape = var_2113, x = attn_37_cast_fp16)[name = tensor("input_91_cast_fp16")]; tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 1])]; tensor var_2119 = const()[name = tensor("op_2119"), val = tensor([1, 1])]; tensor obj_115_pad_type_0 = const()[name = tensor("obj_115_pad_type_0"), val = tensor("custom")]; tensor obj_115_pad_0 = const()[name = tensor("obj_115_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254019840)))]; tensor layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255199552)))]; tensor obj_115_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = var_2119, groups = var_2029, pad = obj_115_pad_0, pad_type = obj_115_pad_type_0, strides = var_2117, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("obj_115_cast_fp16")]; tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_115_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; tensor var_2129 = const()[name = tensor("op_2129"), val = tensor([1])]; tensor channels_mean_57_cast_fp16 = reduce_mean(axes = var_2129, keep_dims = var_2030, x = inputs_57_cast_fp16)[name = tensor("channels_mean_57_cast_fp16")]; tensor zero_mean_57_cast_fp16 = sub(x = inputs_57_cast_fp16, y = channels_mean_57_cast_fp16)[name = tensor("zero_mean_57_cast_fp16")]; tensor zero_mean_sq_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = zero_mean_57_cast_fp16)[name = tensor("zero_mean_sq_57_cast_fp16")]; tensor var_2133 = const()[name = tensor("op_2133"), val = tensor([1])]; tensor var_2134_cast_fp16 = reduce_mean(axes = var_2133, keep_dims = var_2030, x = zero_mean_sq_57_cast_fp16)[name = tensor("op_2134_cast_fp16")]; tensor var_2135_to_fp16 = const()[name = tensor("op_2135_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2136_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2135_to_fp16)[name = tensor("op_2136_cast_fp16")]; tensor denom_57_epsilon_0 = const()[name = tensor("denom_57_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0, x = var_2136_cast_fp16)[name = tensor("denom_57_cast_fp16")]; tensor out_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = denom_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; tensor obj_117_gamma_0_to_fp16 = const()[name = tensor("obj_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255201152)))]; tensor obj_117_beta_0_to_fp16 = const()[name = tensor("obj_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255202752)))]; tensor obj_117_epsilon_0_to_fp16 = const()[name = tensor("obj_117_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_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("obj_117_cast_fp16")]; tensor var_2151 = const()[name = tensor("op_2151"), val = tensor([1, 1])]; tensor var_2153 = const()[name = tensor("op_2153"), val = tensor([1, 1])]; tensor query_39_pad_type_0 = const()[name = tensor("query_39_pad_type_0"), val = tensor("custom")]; tensor query_39_pad_0 = const()[name = tensor("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255204352)))]; tensor layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256384064)))]; tensor query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = var_2153, groups = var_2029, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = var_2151, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor("query_39_cast_fp16")]; tensor var_2157 = const()[name = tensor("op_2157"), val = tensor([1, 1])]; tensor var_2159 = const()[name = tensor("op_2159"), val = tensor([1, 1])]; tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("custom")]; tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256385664)))]; tensor key_39_cast_fp16 = conv(dilations = var_2159, groups = var_2029, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = var_2157, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_39_cast_fp16")]; tensor var_2164 = const()[name = tensor("op_2164"), val = tensor([1, 1])]; tensor var_2166 = const()[name = tensor("op_2166"), val = tensor([1, 1])]; tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("custom")]; tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257565376)))]; tensor layers_9_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258745088)))]; tensor value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = var_2166, groups = var_2029, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = var_2164, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_39_cast_fp16")]; tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1, 12, 64, -1])]; tensor var_2171_cast_fp16 = reshape(shape = var_2170, x = query_39_cast_fp16)[name = tensor("op_2171_cast_fp16")]; tensor var_2172_to_fp16 = const()[name = tensor("op_2172_to_fp16"), val = tensor(0x1p-3)]; tensor var_2173_cast_fp16 = mul(x = var_2171_cast_fp16, y = var_2172_to_fp16)[name = tensor("op_2173_cast_fp16")]; tensor var_2174 = const()[name = tensor("op_2174"), val = tensor([1, 12, 64, -1])]; tensor var_2175_cast_fp16 = reshape(shape = var_2174, x = key_39_cast_fp16)[name = tensor("op_2175_cast_fp16")]; tensor mh_w_59_transpose_x_0 = const()[name = tensor("mh_w_59_transpose_x_0"), val = tensor(true)]; tensor mh_w_59_transpose_y_0 = const()[name = tensor("mh_w_59_transpose_y_0"), val = tensor(false)]; tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_2173_cast_fp16, y = var_2175_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; tensor var_2178_cast_fp16 = softmax(axis = var_2022, x = mh_w_59_cast_fp16)[name = tensor("op_2178_cast_fp16")]; tensor var_2179 = const()[name = tensor("op_2179"), val = tensor([1, 12, 64, -1])]; tensor var_2180_cast_fp16 = reshape(shape = var_2179, x = value_39_cast_fp16)[name = tensor("op_2180_cast_fp16")]; tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2180_cast_fp16, y = var_2178_cast_fp16)[name = tensor("attn_39_cast_fp16")]; tensor var_2183 = const()[name = tensor("op_2183"), val = tensor([1, 768, 1, -1])]; tensor input_93_cast_fp16 = reshape(shape = var_2183, x = attn_39_cast_fp16)[name = tensor("input_93_cast_fp16")]; tensor var_2187 = const()[name = tensor("op_2187"), val = tensor([1, 1])]; tensor var_2189 = const()[name = tensor("op_2189"), val = tensor([1, 1])]; tensor obj_119_pad_type_0 = const()[name = tensor("obj_119_pad_type_0"), val = tensor("custom")]; tensor obj_119_pad_0 = const()[name = tensor("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258746688)))]; tensor layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259926400)))]; tensor obj_119_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = var_2189, groups = var_2029, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = var_2187, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("obj_119_cast_fp16")]; tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; tensor var_2195 = const()[name = tensor("op_2195"), val = tensor([1])]; tensor channels_mean_59_cast_fp16 = reduce_mean(axes = var_2195, keep_dims = var_2030, x = inputs_59_cast_fp16)[name = tensor("channels_mean_59_cast_fp16")]; tensor zero_mean_59_cast_fp16 = sub(x = inputs_59_cast_fp16, y = channels_mean_59_cast_fp16)[name = tensor("zero_mean_59_cast_fp16")]; tensor zero_mean_sq_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = zero_mean_59_cast_fp16)[name = tensor("zero_mean_sq_59_cast_fp16")]; tensor var_2199 = const()[name = tensor("op_2199"), val = tensor([1])]; tensor var_2200_cast_fp16 = reduce_mean(axes = var_2199, keep_dims = var_2030, x = zero_mean_sq_59_cast_fp16)[name = tensor("op_2200_cast_fp16")]; tensor var_2201_to_fp16 = const()[name = tensor("op_2201_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2202_cast_fp16 = add(x = var_2200_cast_fp16, y = var_2201_to_fp16)[name = tensor("op_2202_cast_fp16")]; tensor denom_59_epsilon_0 = const()[name = tensor("denom_59_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0, x = var_2202_cast_fp16)[name = tensor("denom_59_cast_fp16")]; tensor out_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = denom_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; tensor input_95_gamma_0_to_fp16 = const()[name = tensor("input_95_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259928000)))]; tensor input_95_beta_0_to_fp16 = const()[name = tensor("input_95_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259929600)))]; tensor input_95_epsilon_0_to_fp16 = const()[name = tensor("input_95_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_95_cast_fp16")]; tensor var_2213 = const()[name = tensor("op_2213"), val = tensor([1, 1])]; tensor var_2215 = const()[name = tensor("op_2215"), val = tensor([1, 1])]; tensor input_97_pad_type_0 = const()[name = tensor("input_97_pad_type_0"), val = tensor("custom")]; tensor input_97_pad_0 = const()[name = tensor("input_97_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_fc1_weight_to_fp16 = const()[name = tensor("layers_9_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259931200)))]; tensor layers_9_fc1_bias_to_fp16 = const()[name = tensor("layers_9_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264649856)))]; tensor input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = var_2215, groups = var_2029, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = var_2213, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("input_97_cast_fp16")]; tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("EXACT")]; tensor input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1])]; tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([1, 1])]; tensor hidden_states_21_pad_type_0 = const()[name = tensor("hidden_states_21_pad_type_0"), val = tensor("custom")]; tensor hidden_states_21_pad_0 = const()[name = tensor("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_fc2_weight_to_fp16 = const()[name = tensor("layers_9_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264656064)))]; tensor layers_9_fc2_bias_to_fp16 = const()[name = tensor("layers_9_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269374720)))]; tensor hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = var_2223, groups = var_2029, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = var_2221, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("hidden_states_21_cast_fp16")]; tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; tensor var_2236 = const()[name = tensor("op_2236"), val = tensor(3)]; tensor var_2243 = const()[name = tensor("op_2243"), val = tensor(1)]; tensor var_2244 = const()[name = tensor("op_2244"), val = tensor(true)]; tensor var_2256 = const()[name = tensor("op_2256"), val = tensor([1])]; tensor channels_mean_61_cast_fp16 = reduce_mean(axes = var_2256, keep_dims = var_2244, x = inputs_61_cast_fp16)[name = tensor("channels_mean_61_cast_fp16")]; tensor zero_mean_61_cast_fp16 = sub(x = inputs_61_cast_fp16, y = channels_mean_61_cast_fp16)[name = tensor("zero_mean_61_cast_fp16")]; tensor zero_mean_sq_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = zero_mean_61_cast_fp16)[name = tensor("zero_mean_sq_61_cast_fp16")]; tensor var_2260 = const()[name = tensor("op_2260"), val = tensor([1])]; tensor var_2261_cast_fp16 = reduce_mean(axes = var_2260, keep_dims = var_2244, x = zero_mean_sq_61_cast_fp16)[name = tensor("op_2261_cast_fp16")]; tensor var_2262_to_fp16 = const()[name = tensor("op_2262_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2263_cast_fp16 = add(x = var_2261_cast_fp16, y = var_2262_to_fp16)[name = tensor("op_2263_cast_fp16")]; tensor denom_61_epsilon_0 = const()[name = tensor("denom_61_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0, x = var_2263_cast_fp16)[name = tensor("denom_61_cast_fp16")]; tensor out_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = denom_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269376320)))]; tensor obj_121_beta_0_to_fp16 = const()[name = tensor("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269377920)))]; tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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_61_cast_fp16)[name = tensor("obj_121_cast_fp16")]; tensor var_2278 = const()[name = tensor("op_2278"), val = tensor([1, 1])]; tensor var_2280 = const()[name = tensor("op_2280"), val = tensor([1, 1])]; tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("custom")]; tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269379520)))]; tensor layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270559232)))]; tensor query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = var_2280, groups = var_2243, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = var_2278, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("query_41_cast_fp16")]; tensor var_2284 = const()[name = tensor("op_2284"), val = tensor([1, 1])]; tensor var_2286 = const()[name = tensor("op_2286"), val = tensor([1, 1])]; tensor current_key_21_pad_type_0 = const()[name = tensor("current_key_21_pad_type_0"), val = tensor("custom")]; tensor current_key_21_pad_0 = const()[name = tensor("current_key_21_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270560832)))]; tensor current_key_21_cast_fp16 = conv(dilations = var_2286, groups = var_2243, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = var_2284, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1])]; tensor var_2293 = const()[name = tensor("op_2293"), val = tensor([1, 1])]; tensor current_value_21_pad_type_0 = const()[name = tensor("current_value_21_pad_type_0"), val = tensor("custom")]; tensor current_value_21_pad_0 = const()[name = tensor("current_value_21_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271740544)))]; tensor layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272920256)))]; tensor current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = var_2293, groups = var_2243, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = var_2291, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; tensor var_2300_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_2300_cast_fp16")]; tensor var_2302_cast_fp16 = mul(x = var_63_cast_fp16_10, y = var_161_cast_fp16)[name = tensor("op_2302_cast_fp16")]; tensor key_41_cast_fp16 = add(x = var_2300_cast_fp16, y = var_2302_cast_fp16)[name = tensor("key_41_cast_fp16")]; tensor var_2304_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_2304_cast_fp16")]; tensor var_2306_cast_fp16 = mul(x = var_78_cast_fp16_10, y = var_161_cast_fp16)[name = tensor("op_2306_cast_fp16")]; tensor value_41_cast_fp16 = add(x = var_2304_cast_fp16, y = var_2306_cast_fp16)[name = tensor("value_41_cast_fp16")]; tensor var_2309 = const()[name = tensor("op_2309"), val = tensor([1, 12, 64, -1])]; tensor var_2310_cast_fp16 = reshape(shape = var_2309, x = query_41_cast_fp16)[name = tensor("op_2310_cast_fp16")]; tensor var_2311_to_fp16 = const()[name = tensor("op_2311_to_fp16"), val = tensor(0x1p-3)]; tensor var_2312_cast_fp16 = mul(x = var_2310_cast_fp16, y = var_2311_to_fp16)[name = tensor("op_2312_cast_fp16")]; tensor var_2313 = const()[name = tensor("op_2313"), val = tensor([1, 12, 64, -1])]; tensor var_2314_cast_fp16 = reshape(shape = var_2313, x = key_41_cast_fp16)[name = tensor("op_2314_cast_fp16")]; tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_2312_cast_fp16, y = var_2314_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; tensor var_2322_cast_fp16 = softmax(axis = var_2236, x = mh_w_63_cast_fp16)[name = tensor("op_2322_cast_fp16")]; tensor var_2323 = const()[name = tensor("op_2323"), val = tensor([1, 12, 64, -1])]; tensor var_2324_cast_fp16 = reshape(shape = var_2323, x = value_41_cast_fp16)[name = tensor("op_2324_cast_fp16")]; tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2324_cast_fp16, y = var_2322_cast_fp16)[name = tensor("attn_41_cast_fp16")]; tensor var_2327 = const()[name = tensor("op_2327"), val = tensor([1, 768, 1, -1])]; tensor input_101_cast_fp16 = reshape(shape = var_2327, x = attn_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; tensor var_2331 = const()[name = tensor("op_2331"), val = tensor([1, 1])]; tensor var_2333 = const()[name = tensor("op_2333"), val = tensor([1, 1])]; tensor obj_127_pad_type_0 = const()[name = tensor("obj_127_pad_type_0"), val = tensor("custom")]; tensor obj_127_pad_0 = const()[name = tensor("obj_127_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272921856)))]; tensor layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274101568)))]; tensor obj_127_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = var_2333, groups = var_2243, pad = obj_127_pad_0, pad_type = obj_127_pad_type_0, strides = var_2331, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("obj_127_cast_fp16")]; tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_127_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; tensor var_2343 = const()[name = tensor("op_2343"), val = tensor([1])]; tensor channels_mean_63_cast_fp16 = reduce_mean(axes = var_2343, keep_dims = var_2244, x = inputs_63_cast_fp16)[name = tensor("channels_mean_63_cast_fp16")]; tensor zero_mean_63_cast_fp16 = sub(x = inputs_63_cast_fp16, y = channels_mean_63_cast_fp16)[name = tensor("zero_mean_63_cast_fp16")]; tensor zero_mean_sq_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = zero_mean_63_cast_fp16)[name = tensor("zero_mean_sq_63_cast_fp16")]; tensor var_2347 = const()[name = tensor("op_2347"), val = tensor([1])]; tensor var_2348_cast_fp16 = reduce_mean(axes = var_2347, keep_dims = var_2244, x = zero_mean_sq_63_cast_fp16)[name = tensor("op_2348_cast_fp16")]; tensor var_2349_to_fp16 = const()[name = tensor("op_2349_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2350_cast_fp16 = add(x = var_2348_cast_fp16, y = var_2349_to_fp16)[name = tensor("op_2350_cast_fp16")]; tensor denom_63_epsilon_0 = const()[name = tensor("denom_63_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0, x = var_2350_cast_fp16)[name = tensor("denom_63_cast_fp16")]; tensor out_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = denom_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; tensor obj_129_gamma_0_to_fp16 = const()[name = tensor("obj_129_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274103168)))]; tensor obj_129_beta_0_to_fp16 = const()[name = tensor("obj_129_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274104768)))]; tensor obj_129_epsilon_0_to_fp16 = const()[name = tensor("obj_129_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_129_cast_fp16 = batch_norm(beta = obj_129_beta_0_to_fp16, epsilon = obj_129_epsilon_0_to_fp16, gamma = obj_129_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("obj_129_cast_fp16")]; tensor var_2365 = const()[name = tensor("op_2365"), val = tensor([1, 1])]; tensor var_2367 = const()[name = tensor("op_2367"), val = tensor([1, 1])]; tensor query_43_pad_type_0 = const()[name = tensor("query_43_pad_type_0"), val = tensor("custom")]; tensor query_43_pad_0 = const()[name = tensor("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274106368)))]; tensor layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275286080)))]; tensor query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = var_2367, groups = var_2243, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = var_2365, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_129_cast_fp16)[name = tensor("query_43_cast_fp16")]; tensor var_2371 = const()[name = tensor("op_2371"), val = tensor([1, 1])]; tensor var_2373 = const()[name = tensor("op_2373"), val = tensor([1, 1])]; tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("custom")]; tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275287680)))]; tensor key_43_cast_fp16 = conv(dilations = var_2373, groups = var_2243, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = var_2371, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_43_cast_fp16")]; tensor var_2378 = const()[name = tensor("op_2378"), val = tensor([1, 1])]; tensor var_2380 = const()[name = tensor("op_2380"), val = tensor([1, 1])]; tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("custom")]; tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276467392)))]; tensor layers_10_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277647104)))]; tensor value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = var_2380, groups = var_2243, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = var_2378, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_43_cast_fp16")]; tensor var_2384 = const()[name = tensor("op_2384"), val = tensor([1, 12, 64, -1])]; tensor var_2385_cast_fp16 = reshape(shape = var_2384, x = query_43_cast_fp16)[name = tensor("op_2385_cast_fp16")]; tensor var_2386_to_fp16 = const()[name = tensor("op_2386_to_fp16"), val = tensor(0x1p-3)]; tensor var_2387_cast_fp16 = mul(x = var_2385_cast_fp16, y = var_2386_to_fp16)[name = tensor("op_2387_cast_fp16")]; tensor var_2388 = const()[name = tensor("op_2388"), val = tensor([1, 12, 64, -1])]; tensor var_2389_cast_fp16 = reshape(shape = var_2388, x = key_43_cast_fp16)[name = tensor("op_2389_cast_fp16")]; tensor mh_w_65_transpose_x_0 = const()[name = tensor("mh_w_65_transpose_x_0"), val = tensor(true)]; tensor mh_w_65_transpose_y_0 = const()[name = tensor("mh_w_65_transpose_y_0"), val = tensor(false)]; tensor mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_2387_cast_fp16, y = var_2389_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; tensor var_2392_cast_fp16 = softmax(axis = var_2236, x = mh_w_65_cast_fp16)[name = tensor("op_2392_cast_fp16")]; tensor var_2393 = const()[name = tensor("op_2393"), val = tensor([1, 12, 64, -1])]; tensor var_2394_cast_fp16 = reshape(shape = var_2393, x = value_43_cast_fp16)[name = tensor("op_2394_cast_fp16")]; tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2394_cast_fp16, y = var_2392_cast_fp16)[name = tensor("attn_43_cast_fp16")]; tensor var_2397 = const()[name = tensor("op_2397"), val = tensor([1, 768, 1, -1])]; tensor input_103_cast_fp16 = reshape(shape = var_2397, x = attn_43_cast_fp16)[name = tensor("input_103_cast_fp16")]; tensor var_2401 = const()[name = tensor("op_2401"), val = tensor([1, 1])]; tensor var_2403 = const()[name = tensor("op_2403"), val = tensor([1, 1])]; tensor obj_131_pad_type_0 = const()[name = tensor("obj_131_pad_type_0"), val = tensor("custom")]; tensor obj_131_pad_0 = const()[name = tensor("obj_131_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277648704)))]; tensor layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278828416)))]; tensor obj_131_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = var_2403, groups = var_2243, pad = obj_131_pad_0, pad_type = obj_131_pad_type_0, strides = var_2401, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("obj_131_cast_fp16")]; tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_131_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; tensor var_2409 = const()[name = tensor("op_2409"), val = tensor([1])]; tensor channels_mean_65_cast_fp16 = reduce_mean(axes = var_2409, keep_dims = var_2244, x = inputs_65_cast_fp16)[name = tensor("channels_mean_65_cast_fp16")]; tensor zero_mean_65_cast_fp16 = sub(x = inputs_65_cast_fp16, y = channels_mean_65_cast_fp16)[name = tensor("zero_mean_65_cast_fp16")]; tensor zero_mean_sq_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = zero_mean_65_cast_fp16)[name = tensor("zero_mean_sq_65_cast_fp16")]; tensor var_2413 = const()[name = tensor("op_2413"), val = tensor([1])]; tensor var_2414_cast_fp16 = reduce_mean(axes = var_2413, keep_dims = var_2244, x = zero_mean_sq_65_cast_fp16)[name = tensor("op_2414_cast_fp16")]; tensor var_2415_to_fp16 = const()[name = tensor("op_2415_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2416_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2415_to_fp16)[name = tensor("op_2416_cast_fp16")]; tensor denom_65_epsilon_0 = const()[name = tensor("denom_65_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0, x = var_2416_cast_fp16)[name = tensor("denom_65_cast_fp16")]; tensor out_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = denom_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; tensor input_105_gamma_0_to_fp16 = const()[name = tensor("input_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278830016)))]; tensor input_105_beta_0_to_fp16 = const()[name = tensor("input_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278831616)))]; tensor input_105_epsilon_0_to_fp16 = const()[name = tensor("input_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_105_cast_fp16")]; tensor var_2427 = const()[name = tensor("op_2427"), val = tensor([1, 1])]; tensor var_2429 = const()[name = tensor("op_2429"), val = tensor([1, 1])]; tensor input_107_pad_type_0 = const()[name = tensor("input_107_pad_type_0"), val = tensor("custom")]; tensor input_107_pad_0 = const()[name = tensor("input_107_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_fc1_weight_to_fp16 = const()[name = tensor("layers_10_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278833216)))]; tensor layers_10_fc1_bias_to_fp16 = const()[name = tensor("layers_10_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283551872)))]; tensor input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = var_2429, groups = var_2243, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = var_2427, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; tensor var_2435 = const()[name = tensor("op_2435"), val = tensor([1, 1])]; tensor var_2437 = const()[name = tensor("op_2437"), val = tensor([1, 1])]; tensor hidden_states_23_pad_type_0 = const()[name = tensor("hidden_states_23_pad_type_0"), val = tensor("custom")]; tensor hidden_states_23_pad_0 = const()[name = tensor("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_fc2_weight_to_fp16 = const()[name = tensor("layers_10_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283558080)))]; tensor layers_10_fc2_bias_to_fp16 = const()[name = tensor("layers_10_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288276736)))]; tensor hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = var_2437, groups = var_2243, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = var_2435, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("hidden_states_23_cast_fp16")]; tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; tensor var_2450 = const()[name = tensor("op_2450"), val = tensor(3)]; tensor var_2457 = const()[name = tensor("op_2457"), val = tensor(1)]; tensor var_2458 = const()[name = tensor("op_2458"), val = tensor(true)]; tensor var_2470 = const()[name = tensor("op_2470"), val = tensor([1])]; tensor channels_mean_67_cast_fp16 = reduce_mean(axes = var_2470, keep_dims = var_2458, x = inputs_67_cast_fp16)[name = tensor("channels_mean_67_cast_fp16")]; tensor zero_mean_67_cast_fp16 = sub(x = inputs_67_cast_fp16, y = channels_mean_67_cast_fp16)[name = tensor("zero_mean_67_cast_fp16")]; tensor zero_mean_sq_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = zero_mean_67_cast_fp16)[name = tensor("zero_mean_sq_67_cast_fp16")]; tensor var_2474 = const()[name = tensor("op_2474"), val = tensor([1])]; tensor var_2475_cast_fp16 = reduce_mean(axes = var_2474, keep_dims = var_2458, x = zero_mean_sq_67_cast_fp16)[name = tensor("op_2475_cast_fp16")]; tensor var_2476_to_fp16 = const()[name = tensor("op_2476_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2477_cast_fp16 = add(x = var_2475_cast_fp16, y = var_2476_to_fp16)[name = tensor("op_2477_cast_fp16")]; tensor denom_67_epsilon_0 = const()[name = tensor("denom_67_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0, x = var_2477_cast_fp16)[name = tensor("denom_67_cast_fp16")]; tensor out_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = denom_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; tensor obj_133_gamma_0_to_fp16 = const()[name = tensor("obj_133_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288278336)))]; tensor obj_133_beta_0_to_fp16 = const()[name = tensor("obj_133_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288279936)))]; tensor obj_133_epsilon_0_to_fp16 = const()[name = tensor("obj_133_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_133_cast_fp16 = batch_norm(beta = obj_133_beta_0_to_fp16, epsilon = obj_133_epsilon_0_to_fp16, gamma = obj_133_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("obj_133_cast_fp16")]; tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 1])]; tensor var_2494 = const()[name = tensor("op_2494"), val = tensor([1, 1])]; tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("custom")]; tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288281536)))]; tensor layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289461248)))]; tensor query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = var_2494, groups = var_2457, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = var_2492, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_133_cast_fp16)[name = tensor("query_45_cast_fp16")]; tensor var_2498 = const()[name = tensor("op_2498"), val = tensor([1, 1])]; tensor var_2500 = const()[name = tensor("op_2500"), val = tensor([1, 1])]; tensor current_key_pad_type_0 = const()[name = tensor("current_key_pad_type_0"), val = tensor("custom")]; tensor current_key_pad_0 = const()[name = tensor("current_key_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289462848)))]; tensor current_key_cast_fp16 = conv(dilations = var_2500, groups = var_2457, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_2498, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_133_cast_fp16)[name = tensor("current_key_cast_fp16")]; tensor var_2505 = const()[name = tensor("op_2505"), val = tensor([1, 1])]; tensor var_2507 = const()[name = tensor("op_2507"), val = tensor([1, 1])]; tensor current_value_pad_type_0 = const()[name = tensor("current_value_pad_type_0"), val = tensor("custom")]; tensor current_value_pad_0 = const()[name = tensor("current_value_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290642560)))]; tensor layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291822272)))]; tensor current_value_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = var_2507, groups = var_2457, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_2505, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_133_cast_fp16)[name = tensor("current_value_cast_fp16")]; tensor var_2514_cast_fp16 = mul(x = current_key_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_2514_cast_fp16")]; tensor var_2516_cast_fp16 = mul(x = var_63_cast_fp16_11, y = var_161_cast_fp16)[name = tensor("op_2516_cast_fp16")]; tensor key_45_cast_fp16 = add(x = var_2514_cast_fp16, y = var_2516_cast_fp16)[name = tensor("key_45_cast_fp16")]; tensor var_2518_cast_fp16 = mul(x = current_value_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_2518_cast_fp16")]; tensor var_2520_cast_fp16 = mul(x = var_78_cast_fp16_11, y = var_161_cast_fp16)[name = tensor("op_2520_cast_fp16")]; tensor value_45_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2520_cast_fp16)[name = tensor("value_45_cast_fp16")]; tensor var_2523 = const()[name = tensor("op_2523"), val = tensor([1, 12, 64, -1])]; tensor var_2524_cast_fp16 = reshape(shape = var_2523, x = query_45_cast_fp16)[name = tensor("op_2524_cast_fp16")]; tensor var_2525_to_fp16 = const()[name = tensor("op_2525_to_fp16"), val = tensor(0x1p-3)]; tensor var_2526_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_2525_to_fp16)[name = tensor("op_2526_cast_fp16")]; tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1, 12, 64, -1])]; tensor var_2528_cast_fp16 = reshape(shape = var_2527, x = key_45_cast_fp16)[name = tensor("op_2528_cast_fp16")]; tensor mh_w_67_transpose_x_0 = const()[name = tensor("mh_w_67_transpose_x_0"), val = tensor(true)]; tensor mh_w_67_transpose_y_0 = const()[name = tensor("mh_w_67_transpose_y_0"), val = tensor(false)]; tensor mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_2526_cast_fp16, y = var_2528_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_179_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; tensor var_2536_cast_fp16 = softmax(axis = var_2450, x = mh_w_69_cast_fp16)[name = tensor("op_2536_cast_fp16")]; tensor var_2537 = const()[name = tensor("op_2537"), val = tensor([1, 12, 64, -1])]; tensor var_2538_cast_fp16 = reshape(shape = var_2537, x = value_45_cast_fp16)[name = tensor("op_2538_cast_fp16")]; tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2538_cast_fp16, y = var_2536_cast_fp16)[name = tensor("attn_45_cast_fp16")]; tensor var_2541 = const()[name = tensor("op_2541"), val = tensor([1, 768, 1, -1])]; tensor input_111_cast_fp16 = reshape(shape = var_2541, x = attn_45_cast_fp16)[name = tensor("input_111_cast_fp16")]; tensor var_2545 = const()[name = tensor("op_2545"), val = tensor([1, 1])]; tensor var_2547 = const()[name = tensor("op_2547"), val = tensor([1, 1])]; tensor obj_139_pad_type_0 = const()[name = tensor("obj_139_pad_type_0"), val = tensor("custom")]; tensor obj_139_pad_0 = const()[name = tensor("obj_139_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291823872)))]; tensor layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293003584)))]; tensor obj_139_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = var_2547, groups = var_2457, pad = obj_139_pad_0, pad_type = obj_139_pad_type_0, strides = var_2545, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("obj_139_cast_fp16")]; tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_139_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; tensor var_2557 = const()[name = tensor("op_2557"), val = tensor([1])]; tensor channels_mean_69_cast_fp16 = reduce_mean(axes = var_2557, keep_dims = var_2458, x = inputs_69_cast_fp16)[name = tensor("channels_mean_69_cast_fp16")]; tensor zero_mean_69_cast_fp16 = sub(x = inputs_69_cast_fp16, y = channels_mean_69_cast_fp16)[name = tensor("zero_mean_69_cast_fp16")]; tensor zero_mean_sq_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = zero_mean_69_cast_fp16)[name = tensor("zero_mean_sq_69_cast_fp16")]; tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([1])]; tensor var_2562_cast_fp16 = reduce_mean(axes = var_2561, keep_dims = var_2458, x = zero_mean_sq_69_cast_fp16)[name = tensor("op_2562_cast_fp16")]; tensor var_2563_to_fp16 = const()[name = tensor("op_2563_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2564_cast_fp16 = add(x = var_2562_cast_fp16, y = var_2563_to_fp16)[name = tensor("op_2564_cast_fp16")]; tensor denom_69_epsilon_0 = const()[name = tensor("denom_69_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0, x = var_2564_cast_fp16)[name = tensor("denom_69_cast_fp16")]; tensor out_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = denom_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; tensor obj_141_gamma_0_to_fp16 = const()[name = tensor("obj_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293005184)))]; tensor obj_141_beta_0_to_fp16 = const()[name = tensor("obj_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293006784)))]; tensor obj_141_epsilon_0_to_fp16 = const()[name = tensor("obj_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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_69_cast_fp16)[name = tensor("obj_141_cast_fp16")]; tensor var_2579 = const()[name = tensor("op_2579"), val = tensor([1, 1])]; tensor var_2581 = const()[name = tensor("op_2581"), val = tensor([1, 1])]; tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("custom")]; tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293008384)))]; tensor layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294188096)))]; tensor query_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = var_2581, groups = var_2457, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_2579, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("query_cast_fp16")]; tensor var_2585 = const()[name = tensor("op_2585"), val = tensor([1, 1])]; tensor var_2587 = const()[name = tensor("op_2587"), val = tensor([1, 1])]; tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("custom")]; tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294189696)))]; tensor key_cast_fp16 = conv(dilations = var_2587, groups = var_2457, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_2585, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; tensor var_2592 = const()[name = tensor("op_2592"), val = tensor([1, 1])]; tensor var_2594 = const()[name = tensor("op_2594"), val = tensor([1, 1])]; tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("custom")]; tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295369408)))]; tensor layers_11_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296549120)))]; tensor value_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = var_2594, groups = var_2457, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_2592, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; tensor var_2598 = const()[name = tensor("op_2598"), val = tensor([1, 12, 64, -1])]; tensor var_2599_cast_fp16 = reshape(shape = var_2598, x = query_cast_fp16)[name = tensor("op_2599_cast_fp16")]; tensor var_2600_to_fp16 = const()[name = tensor("op_2600_to_fp16"), val = tensor(0x1p-3)]; tensor var_2601_cast_fp16 = mul(x = var_2599_cast_fp16, y = var_2600_to_fp16)[name = tensor("op_2601_cast_fp16")]; tensor var_2602 = const()[name = tensor("op_2602"), val = tensor([1, 12, 64, -1])]; tensor var_2603_cast_fp16 = reshape(shape = var_2602, x = key_cast_fp16)[name = tensor("op_2603_cast_fp16")]; tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_2601_cast_fp16, y = var_2603_cast_fp16)[name = tensor("mh_w_cast_fp16")]; tensor var_2606_cast_fp16 = softmax(axis = var_2450, x = mh_w_cast_fp16)[name = tensor("op_2606_cast_fp16")]; tensor var_2607 = const()[name = tensor("op_2607"), val = tensor([1, 12, 64, -1])]; tensor var_2608_cast_fp16 = reshape(shape = var_2607, x = value_cast_fp16)[name = tensor("op_2608_cast_fp16")]; tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_2608_cast_fp16, y = var_2606_cast_fp16)[name = tensor("attn_cast_fp16")]; tensor var_2611 = const()[name = tensor("op_2611"), val = tensor([1, 768, 1, -1])]; tensor input_113_cast_fp16 = reshape(shape = var_2611, x = attn_cast_fp16)[name = tensor("input_113_cast_fp16")]; tensor var_2615 = const()[name = tensor("op_2615"), val = tensor([1, 1])]; tensor var_2617 = const()[name = tensor("op_2617"), val = tensor([1, 1])]; tensor obj_143_pad_type_0 = const()[name = tensor("obj_143_pad_type_0"), val = tensor("custom")]; tensor obj_143_pad_0 = const()[name = tensor("obj_143_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296550720)))]; tensor layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297730432)))]; tensor obj_143_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = var_2617, groups = var_2457, pad = obj_143_pad_0, pad_type = obj_143_pad_type_0, strides = var_2615, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("obj_143_cast_fp16")]; tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_143_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; tensor var_2623 = const()[name = tensor("op_2623"), val = tensor([1])]; tensor channels_mean_71_cast_fp16 = reduce_mean(axes = var_2623, keep_dims = var_2458, x = inputs_71_cast_fp16)[name = tensor("channels_mean_71_cast_fp16")]; tensor zero_mean_71_cast_fp16 = sub(x = inputs_71_cast_fp16, y = channels_mean_71_cast_fp16)[name = tensor("zero_mean_71_cast_fp16")]; tensor zero_mean_sq_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = zero_mean_71_cast_fp16)[name = tensor("zero_mean_sq_71_cast_fp16")]; tensor var_2627 = const()[name = tensor("op_2627"), val = tensor([1])]; tensor var_2628_cast_fp16 = reduce_mean(axes = var_2627, keep_dims = var_2458, x = zero_mean_sq_71_cast_fp16)[name = tensor("op_2628_cast_fp16")]; tensor var_2629_to_fp16 = const()[name = tensor("op_2629_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2630_cast_fp16 = add(x = var_2628_cast_fp16, y = var_2629_to_fp16)[name = tensor("op_2630_cast_fp16")]; tensor denom_71_epsilon_0 = const()[name = tensor("denom_71_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0, x = var_2630_cast_fp16)[name = tensor("denom_71_cast_fp16")]; tensor out_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = denom_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; tensor input_115_gamma_0_to_fp16 = const()[name = tensor("input_115_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297732032)))]; tensor input_115_beta_0_to_fp16 = const()[name = tensor("input_115_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297733632)))]; tensor input_115_epsilon_0_to_fp16 = const()[name = tensor("input_115_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("input_115_cast_fp16")]; tensor var_2641 = const()[name = tensor("op_2641"), val = tensor([1, 1])]; tensor var_2643 = const()[name = tensor("op_2643"), val = tensor([1, 1])]; tensor input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("custom")]; tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_fc1_weight_to_fp16 = const()[name = tensor("layers_11_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297735232)))]; tensor layers_11_fc1_bias_to_fp16 = const()[name = tensor("layers_11_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302453888)))]; tensor input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = var_2643, groups = var_2457, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = var_2641, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_117_cast_fp16)[name = tensor("input_cast_fp16")]; tensor var_2649 = const()[name = tensor("op_2649"), val = tensor([1, 1])]; tensor var_2651 = const()[name = tensor("op_2651"), val = tensor([1, 1])]; tensor hidden_states_25_pad_type_0 = const()[name = tensor("hidden_states_25_pad_type_0"), val = tensor("custom")]; tensor hidden_states_25_pad_0 = const()[name = tensor("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_fc2_weight_to_fp16 = const()[name = tensor("layers_11_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302460096)))]; tensor layers_11_fc2_bias_to_fp16 = const()[name = tensor("layers_11_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307178752)))]; tensor hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = var_2651, groups = var_2457, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = var_2649, weight = layers_11_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; tensor inputs_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_cast_fp16")]; tensor var_2661 = const()[name = tensor("op_2661"), val = tensor(true)]; tensor var_2665 = const()[name = tensor("op_2665"), val = tensor([1])]; tensor channels_mean_cast_fp16 = reduce_mean(axes = var_2665, keep_dims = var_2661, x = inputs_cast_fp16)[name = tensor("channels_mean_cast_fp16")]; tensor zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor("zero_mean_cast_fp16")]; tensor zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor("zero_mean_sq_cast_fp16")]; tensor var_2669 = const()[name = tensor("op_2669"), val = tensor([1])]; tensor var_2670_cast_fp16 = reduce_mean(axes = var_2669, keep_dims = var_2661, x = zero_mean_sq_cast_fp16)[name = tensor("op_2670_cast_fp16")]; tensor var_2671_to_fp16 = const()[name = tensor("op_2671_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2672_cast_fp16 = add(x = var_2670_cast_fp16, y = var_2671_to_fp16)[name = tensor("op_2672_cast_fp16")]; tensor denom_epsilon_0 = const()[name = tensor("denom_epsilon_0"), val = tensor(0x1.197998p-40)]; tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0, x = var_2672_cast_fp16)[name = tensor("denom_cast_fp16")]; tensor out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor("out_cast_fp16")]; tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307180352)))]; tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307181952)))]; tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor 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("hidden_states_cast_fp16")]; tensor var_2682_axes_0 = const()[name = tensor("op_2682_axes_0"), val = tensor([2])]; tensor var_2682_cast_fp16 = squeeze(axes = var_2682_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_2682_cast_fp16")]; tensor var_2685_perm_0 = const()[name = tensor("op_2685_perm_0"), val = tensor([0, 2, 1])]; tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307183552)))]; tensor transpose_0 = transpose(perm = var_2685_perm_0, x = var_2682_cast_fp16)[name = tensor("transpose_0")]; tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor("linear_0_cast_fp16")]; tensor var_2689 = const()[name = tensor("op_2689"), val = tensor(1)]; tensor obj_147_interleave_0 = const()[name = tensor("obj_147_interleave_0"), val = tensor(false)]; tensor key_cache_updates = concat(axis = var_2689, interleave = obj_147_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("obj_147_cast_fp16")]; tensor var_2692 = const()[name = tensor("op_2692"), val = tensor(1)]; tensor obj_interleave_0 = const()[name = tensor("obj_interleave_0"), val = tensor(false)]; tensor value_cache_updates = concat(axis = var_2692, interleave = obj_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("obj_cast_fp16")]; } -> (logits, key_cache_updates, value_cache_updates); }