diff --git "a/openai_whisper-medium.en/TextDecoder.mlmodelc/model.mil" "b/openai_whisper-medium.en/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-medium.en/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,3728 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.2"}})] +{ + 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_64_axis_0 = const()[name = tensor("op_64_axis_0"), val = tensor(0)]; + tensor var_64_batch_dims_0 = const()[name = tensor("op_64_batch_dims_0"), val = tensor(0)]; + 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_64_cast_fp16 = gather(axis = var_64_axis_0, batch_dims = var_64_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_64_cast_fp16")]; + tensor var_68_axis_0 = const()[name = tensor("op_68_axis_0"), val = tensor(0)]; + tensor var_68_batch_dims_0 = const()[name = tensor("op_68_batch_dims_0"), val = tensor(0)]; + 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(106217600)))]; + tensor var_68_cast_fp16 = gather(axis = var_68_axis_0, batch_dims = var_68_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_68_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_64_cast_fp16, y = var_68_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_82_axes_0 = const()[name = tensor("op_82_axes_0"), val = tensor([2])]; + tensor var_82_cast_fp16 = expand_dims(axes = var_82_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_82_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_82_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024])]; + tensor var_87_axis_0 = const()[name = tensor("op_87_axis_0"), val = tensor(1)]; + tensor var_87_cast_fp16_0, tensor var_87_cast_fp16_1, tensor var_87_cast_fp16_2, tensor var_87_cast_fp16_3, tensor var_87_cast_fp16_4, tensor var_87_cast_fp16_5, tensor var_87_cast_fp16_6, tensor var_87_cast_fp16_7, tensor var_87_cast_fp16_8, tensor var_87_cast_fp16_9, tensor var_87_cast_fp16_10, tensor var_87_cast_fp16_11, tensor var_87_cast_fp16_12, tensor var_87_cast_fp16_13, tensor var_87_cast_fp16_14, tensor var_87_cast_fp16_15, tensor var_87_cast_fp16_16, tensor var_87_cast_fp16_17, tensor var_87_cast_fp16_18, tensor var_87_cast_fp16_19, tensor var_87_cast_fp16_20, tensor var_87_cast_fp16_21, tensor var_87_cast_fp16_22, tensor var_87_cast_fp16_23 = split(axis = var_87_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_87_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024])]; + tensor var_114_axis_0 = const()[name = tensor("op_114_axis_0"), val = tensor(1)]; + tensor var_114_cast_fp16_0, tensor var_114_cast_fp16_1, tensor var_114_cast_fp16_2, tensor var_114_cast_fp16_3, tensor var_114_cast_fp16_4, tensor var_114_cast_fp16_5, tensor var_114_cast_fp16_6, tensor var_114_cast_fp16_7, tensor var_114_cast_fp16_8, tensor var_114_cast_fp16_9, tensor var_114_cast_fp16_10, tensor var_114_cast_fp16_11, tensor var_114_cast_fp16_12, tensor var_114_cast_fp16_13, tensor var_114_cast_fp16_14, tensor var_114_cast_fp16_15, tensor var_114_cast_fp16_16, tensor var_114_cast_fp16_17, tensor var_114_cast_fp16_18, tensor var_114_cast_fp16_19, tensor var_114_cast_fp16_20, tensor var_114_cast_fp16_21, tensor var_114_cast_fp16_22, tensor var_114_cast_fp16_23 = split(axis = var_114_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_114_cast_fp16")]; + tensor var_144 = const()[name = tensor("op_144"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_169_to_fp16 = const()[name = tensor("op_169_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_169_to_fp16, x = inputs_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(107135168)))]; + 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(107137280)))]; + 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(107139392)))]; + 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(107141504)))]; + 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 query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("valid")]; + tensor query_1_strides_0 = const()[name = tensor("query_1_strides_0"), val = tensor([1, 1])]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_1_dilations_0 = const()[name = tensor("query_1_dilations_0"), val = tensor([1, 1])]; + tensor query_1_groups_0 = const()[name = tensor("query_1_groups_0"), val = tensor(1)]; + 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(107143616)))]; + 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(109240832)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor current_key_1_pad_type_0 = const()[name = tensor("current_key_1_pad_type_0"), val = tensor("valid")]; + tensor current_key_1_strides_0 = const()[name = tensor("current_key_1_strides_0"), val = tensor([1, 1])]; + tensor current_key_1_pad_0 = const()[name = tensor("current_key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_1_dilations_0 = const()[name = tensor("current_key_1_dilations_0"), val = tensor([1, 1])]; + tensor current_key_1_groups_0 = const()[name = tensor("current_key_1_groups_0"), val = tensor(1)]; + 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(109242944)))]; + tensor current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor current_value_1_pad_type_0 = const()[name = tensor("current_value_1_pad_type_0"), val = tensor("valid")]; + tensor current_value_1_strides_0 = const()[name = tensor("current_value_1_strides_0"), val = tensor([1, 1])]; + tensor current_value_1_pad_0 = const()[name = tensor("current_value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_1_dilations_0 = const()[name = tensor("current_value_1_dilations_0"), val = tensor([1, 1])]; + tensor current_value_1_groups_0 = const()[name = tensor("current_value_1_groups_0"), val = tensor(1)]; + 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(111340160)))]; + 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(113437376)))]; + tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_204_axes_0 = const()[name = tensor("op_204_axes_0"), val = tensor([1])]; + tensor var_204_cast_fp16 = expand_dims(axes = var_204_axes_0, x = kv_cache_update_mask)[name = tensor("op_204_cast_fp16")]; + tensor var_205_axes_0 = const()[name = tensor("op_205_axes_0"), val = tensor([2])]; + tensor var_205_cast_fp16 = expand_dims(axes = var_205_axes_0, x = var_204_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor var_145_to_fp16 = const()[name = tensor("op_145_to_fp16"), val = tensor(0x1p+0)]; + tensor var_207_cast_fp16 = sub(x = var_145_to_fp16, y = var_205_cast_fp16)[name = tensor("op_207_cast_fp16")]; + tensor var_208_cast_fp16 = mul(x = var_87_cast_fp16_0, y = var_207_cast_fp16)[name = tensor("op_208_cast_fp16")]; + tensor var_209_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_209_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_208_cast_fp16, y = var_209_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_212_cast_fp16 = mul(x = var_114_cast_fp16_0, y = var_207_cast_fp16)[name = tensor("op_212_cast_fp16")]; + tensor var_213_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_213_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_212_cast_fp16, y = var_213_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_217, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_219_to_fp16 = const()[name = tensor("op_219_to_fp16"), val = tensor(0x1p-3)]; + tensor var_220_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_219_to_fp16)[name = tensor("op_220_cast_fp16")]; + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 16, 64, 448])]; + tensor var_224_cast_fp16 = reshape(shape = var_223, x = key_1_cast_fp16)[name = tensor("op_224_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_220_cast_fp16, y = var_224_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_228_axes_0 = const()[name = tensor("op_228_axes_0"), val = tensor([1])]; + tensor var_228_cast_fp16 = expand_dims(axes = var_228_axes_0, x = decoder_key_padding_mask)[name = tensor("op_228_cast_fp16")]; + tensor var_229_axes_0 = const()[name = tensor("op_229_axes_0"), val = tensor([2])]; + tensor var_229_cast_fp16 = expand_dims(axes = var_229_axes_0, x = var_228_cast_fp16)[name = tensor("op_229_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_232_cast_fp16 = softmax(axis = var_144, x = mh_w_3_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 16, 64, 448])]; + tensor var_234_cast_fp16 = reshape(shape = var_233, x = value_1_cast_fp16)[name = tensor("op_234_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_234_cast_fp16, y = var_232_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 1024, 1, 1])]; + tensor input_1_cast_fp16 = reshape(shape = var_237, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("valid")]; + tensor obj_7_strides_0 = const()[name = tensor("obj_7_strides_0"), val = tensor([1, 1])]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_7_dilations_0 = const()[name = tensor("obj_7_dilations_0"), val = tensor([1, 1])]; + tensor obj_7_groups_0 = const()[name = tensor("obj_7_groups_0"), val = tensor(1)]; + 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(113439488)))]; + 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(115536704)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("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 out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([1])]; + tensor var_259_to_fp16 = const()[name = tensor("op_259_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_259_to_fp16, x = inputs_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(115538816)))]; + 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(115540928)))]; + 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 query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("valid")]; + tensor query_3_strides_0 = const()[name = tensor("query_3_strides_0"), val = tensor([1, 1])]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_3_dilations_0 = const()[name = tensor("query_3_dilations_0"), val = tensor([1, 1])]; + tensor query_3_groups_0 = const()[name = tensor("query_3_groups_0"), val = tensor(1)]; + 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(115543040)))]; + 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(117640256)))]; + tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("valid")]; + tensor key_3_strides_0 = const()[name = tensor("key_3_strides_0"), val = tensor([1, 1])]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_3_dilations_0 = const()[name = tensor("key_3_dilations_0"), val = tensor([1, 1])]; + tensor key_3_groups_0 = const()[name = tensor("key_3_groups_0"), val = tensor(1)]; + 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(117642368)))]; + tensor key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("valid")]; + tensor value_3_strides_0 = const()[name = tensor("value_3_strides_0"), val = tensor([1, 1])]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_3_dilations_0 = const()[name = tensor("value_3_dilations_0"), val = tensor([1, 1])]; + tensor value_3_groups_0 = const()[name = tensor("value_3_groups_0"), val = tensor(1)]; + 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(119739584)))]; + 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(121836800)))]; + tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; + tensor var_295 = const()[name = tensor("op_295"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_295, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_297_to_fp16 = const()[name = tensor("op_297_to_fp16"), val = tensor(0x1p-3)]; + tensor var_298_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_297_to_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_301 = const()[name = tensor("op_301"), val = tensor([1, 16, 64, 1500])]; + tensor var_302_cast_fp16 = reshape(shape = var_301, x = key_3_cast_fp16)[name = tensor("op_302_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_298_cast_fp16, y = var_302_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_144, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_306 = const()[name = tensor("op_306"), val = tensor([1, 16, 64, 1500])]; + tensor var_307_cast_fp16 = reshape(shape = var_306, x = value_3_cast_fp16)[name = tensor("op_307_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_307_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_310 = const()[name = tensor("op_310"), val = tensor([1, 1024, 1, 1])]; + tensor input_3_cast_fp16 = reshape(shape = var_310, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("valid")]; + tensor obj_11_strides_0 = const()[name = tensor("obj_11_strides_0"), val = tensor([1, 1])]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_11_dilations_0 = const()[name = tensor("obj_11_dilations_0"), val = tensor([1, 1])]; + tensor obj_11_groups_0 = const()[name = tensor("obj_11_groups_0"), val = tensor(1)]; + 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(121838912)))]; + 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(123936128)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("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 out_5_axes_0 = const()[name = tensor("out_5_axes_0"), val = tensor([1])]; + tensor var_328_to_fp16 = const()[name = tensor("op_328_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_328_to_fp16, x = inputs_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(123938240)))]; + 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(123940352)))]; + 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 input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("valid")]; + tensor input_7_strides_0 = const()[name = tensor("input_7_strides_0"), val = tensor([1, 1])]; + tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_7_dilations_0 = const()[name = tensor("input_7_dilations_0"), val = tensor([1, 1])]; + tensor input_7_groups_0 = const()[name = tensor("input_7_groups_0"), val = tensor(1)]; + 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(123942464)))]; + 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(132331136)))]; + tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("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 hidden_states_3_pad_type_0 = const()[name = tensor("hidden_states_3_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_3_strides_0 = const()[name = tensor("hidden_states_3_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_3_pad_0 = const()[name = tensor("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_3_dilations_0 = const()[name = tensor("hidden_states_3_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_3_groups_0 = const()[name = tensor("hidden_states_3_groups_0"), val = tensor(1)]; + 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(132339392)))]; + 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(140728064)))]; + tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("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_363 = const()[name = tensor("op_363"), val = tensor(3)]; + tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; + tensor var_388_to_fp16 = const()[name = tensor("op_388_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_388_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140730176)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140732288)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("valid")]; + tensor query_5_strides_0 = const()[name = tensor("query_5_strides_0"), val = tensor([1, 1])]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_5_dilations_0 = const()[name = tensor("query_5_dilations_0"), val = tensor([1, 1])]; + tensor query_5_groups_0 = const()[name = tensor("query_5_groups_0"), val = tensor(1)]; + 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(140734400)))]; + 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(142831616)))]; + tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor current_key_3_pad_type_0 = const()[name = tensor("current_key_3_pad_type_0"), val = tensor("valid")]; + tensor current_key_3_strides_0 = const()[name = tensor("current_key_3_strides_0"), val = tensor([1, 1])]; + tensor current_key_3_pad_0 = const()[name = tensor("current_key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_3_dilations_0 = const()[name = tensor("current_key_3_dilations_0"), val = tensor([1, 1])]; + tensor current_key_3_groups_0 = const()[name = tensor("current_key_3_groups_0"), val = tensor(1)]; + 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(142833728)))]; + tensor current_key_3_cast_fp16 = conv(dilations = current_key_3_dilations_0, groups = current_key_3_groups_0, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = current_key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor current_value_3_pad_type_0 = const()[name = tensor("current_value_3_pad_type_0"), val = tensor("valid")]; + tensor current_value_3_strides_0 = const()[name = tensor("current_value_3_strides_0"), val = tensor([1, 1])]; + tensor current_value_3_pad_0 = const()[name = tensor("current_value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_3_dilations_0 = const()[name = tensor("current_value_3_dilations_0"), val = tensor([1, 1])]; + tensor current_value_3_groups_0 = const()[name = tensor("current_value_3_groups_0"), val = tensor(1)]; + 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(144930944)))]; + 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(147028160)))]; + tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_3_dilations_0, groups = current_value_3_groups_0, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = current_value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_427_cast_fp16 = mul(x = var_87_cast_fp16_1, y = var_207_cast_fp16)[name = tensor("op_427_cast_fp16")]; + tensor var_428_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_428_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_427_cast_fp16, y = var_428_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_431_cast_fp16 = mul(x = var_114_cast_fp16_1, y = var_207_cast_fp16)[name = tensor("op_431_cast_fp16")]; + tensor var_432_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_432_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_431_cast_fp16, y = var_432_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_436 = const()[name = tensor("op_436"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_436, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_438_to_fp16 = const()[name = tensor("op_438_to_fp16"), val = tensor(0x1p-3)]; + tensor var_439_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_438_to_fp16)[name = tensor("op_439_cast_fp16")]; + tensor var_442 = const()[name = tensor("op_442"), val = tensor([1, 16, 64, 448])]; + tensor var_443_cast_fp16 = reshape(shape = var_442, x = key_5_cast_fp16)[name = tensor("op_443_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_439_cast_fp16, y = var_443_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_451_cast_fp16 = softmax(axis = var_363, x = mh_w_9_cast_fp16)[name = tensor("op_451_cast_fp16")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 16, 64, 448])]; + tensor var_453_cast_fp16 = reshape(shape = var_452, x = value_5_cast_fp16)[name = tensor("op_453_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_453_cast_fp16, y = var_451_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_456 = const()[name = tensor("op_456"), val = tensor([1, 1024, 1, 1])]; + tensor input_11_cast_fp16 = reshape(shape = var_456, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor obj_21_pad_type_0 = const()[name = tensor("obj_21_pad_type_0"), val = tensor("valid")]; + tensor obj_21_strides_0 = const()[name = tensor("obj_21_strides_0"), val = tensor([1, 1])]; + tensor obj_21_pad_0 = const()[name = tensor("obj_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_21_dilations_0 = const()[name = tensor("obj_21_dilations_0"), val = tensor([1, 1])]; + tensor obj_21_groups_0 = const()[name = tensor("obj_21_groups_0"), val = tensor(1)]; + 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(147030272)))]; + 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(149127488)))]; + tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([1])]; + tensor var_478_to_fp16 = const()[name = tensor("op_478_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_478_to_fp16, x = inputs_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149129600)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149131712)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("valid")]; + tensor query_7_strides_0 = const()[name = tensor("query_7_strides_0"), val = tensor([1, 1])]; + tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_7_dilations_0 = const()[name = tensor("query_7_dilations_0"), val = tensor([1, 1])]; + tensor query_7_groups_0 = const()[name = tensor("query_7_groups_0"), val = tensor(1)]; + 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(149133824)))]; + 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(151231040)))]; + tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("valid")]; + tensor key_7_strides_0 = const()[name = tensor("key_7_strides_0"), val = tensor([1, 1])]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_7_dilations_0 = const()[name = tensor("key_7_dilations_0"), val = tensor([1, 1])]; + tensor key_7_groups_0 = const()[name = tensor("key_7_groups_0"), val = tensor(1)]; + 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(151233152)))]; + tensor key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_7_cast_fp16")]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("valid")]; + tensor value_7_strides_0 = const()[name = tensor("value_7_strides_0"), val = tensor([1, 1])]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_7_dilations_0 = const()[name = tensor("value_7_dilations_0"), val = tensor([1, 1])]; + tensor value_7_groups_0 = const()[name = tensor("value_7_groups_0"), val = tensor(1)]; + 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(153330368)))]; + 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(155427584)))]; + tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_7_cast_fp16")]; + tensor var_514 = const()[name = tensor("op_514"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_514, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_516_to_fp16 = const()[name = tensor("op_516_to_fp16"), val = tensor(0x1p-3)]; + tensor var_517_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_516_to_fp16)[name = tensor("op_517_cast_fp16")]; + tensor var_520 = const()[name = tensor("op_520"), val = tensor([1, 16, 64, 1500])]; + tensor var_521_cast_fp16 = reshape(shape = var_520, x = key_7_cast_fp16)[name = tensor("op_521_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_517_cast_fp16, y = var_521_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_363, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_525 = const()[name = tensor("op_525"), val = tensor([1, 16, 64, 1500])]; + tensor var_526_cast_fp16 = reshape(shape = var_525, x = value_7_cast_fp16)[name = tensor("op_526_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_526_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_529 = const()[name = tensor("op_529"), val = tensor([1, 1024, 1, 1])]; + tensor input_13_cast_fp16 = reshape(shape = var_529, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor obj_25_pad_type_0 = const()[name = tensor("obj_25_pad_type_0"), val = tensor("valid")]; + tensor obj_25_strides_0 = const()[name = tensor("obj_25_strides_0"), val = tensor([1, 1])]; + tensor obj_25_pad_0 = const()[name = tensor("obj_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_25_dilations_0 = const()[name = tensor("obj_25_dilations_0"), val = tensor([1, 1])]; + tensor obj_25_groups_0 = const()[name = tensor("obj_25_groups_0"), val = tensor(1)]; + 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(155429696)))]; + 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(157526912)))]; + tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_547_to_fp16, x = inputs_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(157529024)))]; + 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(157531136)))]; + 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 input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("valid")]; + tensor input_17_strides_0 = const()[name = tensor("input_17_strides_0"), val = tensor([1, 1])]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_17_dilations_0 = const()[name = tensor("input_17_dilations_0"), val = tensor([1, 1])]; + tensor input_17_groups_0 = const()[name = tensor("input_17_groups_0"), val = tensor(1)]; + 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(157533248)))]; + 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(165921920)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("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 hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_5_strides_0 = const()[name = tensor("hidden_states_5_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_5_dilations_0 = const()[name = tensor("hidden_states_5_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_5_groups_0 = const()[name = tensor("hidden_states_5_groups_0"), val = tensor(1)]; + 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(165930176)))]; + 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(174318848)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("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_582 = const()[name = tensor("op_582"), val = tensor(3)]; + tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; + tensor var_607_to_fp16 = const()[name = tensor("op_607_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_607_to_fp16, x = inputs_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174320960)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174323072)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("valid")]; + tensor query_9_strides_0 = const()[name = tensor("query_9_strides_0"), val = tensor([1, 1])]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_9_dilations_0 = const()[name = tensor("query_9_dilations_0"), val = tensor([1, 1])]; + tensor query_9_groups_0 = const()[name = tensor("query_9_groups_0"), val = tensor(1)]; + 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(174325184)))]; + 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(176422400)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor current_key_5_pad_type_0 = const()[name = tensor("current_key_5_pad_type_0"), val = tensor("valid")]; + tensor current_key_5_strides_0 = const()[name = tensor("current_key_5_strides_0"), val = tensor([1, 1])]; + tensor current_key_5_pad_0 = const()[name = tensor("current_key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_5_dilations_0 = const()[name = tensor("current_key_5_dilations_0"), val = tensor([1, 1])]; + tensor current_key_5_groups_0 = const()[name = tensor("current_key_5_groups_0"), val = tensor(1)]; + 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(176424512)))]; + tensor current_key_5_cast_fp16 = conv(dilations = current_key_5_dilations_0, groups = current_key_5_groups_0, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = current_key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor current_value_5_pad_type_0 = const()[name = tensor("current_value_5_pad_type_0"), val = tensor("valid")]; + tensor current_value_5_strides_0 = const()[name = tensor("current_value_5_strides_0"), val = tensor([1, 1])]; + tensor current_value_5_pad_0 = const()[name = tensor("current_value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_5_dilations_0 = const()[name = tensor("current_value_5_dilations_0"), val = tensor([1, 1])]; + tensor current_value_5_groups_0 = const()[name = tensor("current_value_5_groups_0"), val = tensor(1)]; + 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(178521728)))]; + 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(180618944)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = current_value_5_dilations_0, groups = current_value_5_groups_0, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = current_value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_646_cast_fp16 = mul(x = var_87_cast_fp16_2, y = var_207_cast_fp16)[name = tensor("op_646_cast_fp16")]; + tensor var_647_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_646_cast_fp16, y = var_647_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_650_cast_fp16 = mul(x = var_114_cast_fp16_2, y = var_207_cast_fp16)[name = tensor("op_650_cast_fp16")]; + tensor var_651_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_651_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_650_cast_fp16, y = var_651_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_655 = const()[name = tensor("op_655"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_655, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_657_to_fp16 = const()[name = tensor("op_657_to_fp16"), val = tensor(0x1p-3)]; + tensor var_658_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_657_to_fp16)[name = tensor("op_658_cast_fp16")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 16, 64, 448])]; + tensor var_662_cast_fp16 = reshape(shape = var_661, x = key_9_cast_fp16)[name = tensor("op_662_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_658_cast_fp16, y = var_662_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_670_cast_fp16 = softmax(axis = var_582, x = mh_w_15_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor var_671 = const()[name = tensor("op_671"), val = tensor([1, 16, 64, 448])]; + tensor var_672_cast_fp16 = reshape(shape = var_671, x = value_9_cast_fp16)[name = tensor("op_672_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_672_cast_fp16, y = var_670_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([1, 1024, 1, 1])]; + tensor input_21_cast_fp16 = reshape(shape = var_675, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("valid")]; + tensor obj_35_strides_0 = const()[name = tensor("obj_35_strides_0"), val = tensor([1, 1])]; + tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_35_dilations_0 = const()[name = tensor("obj_35_dilations_0"), val = tensor([1, 1])]; + tensor obj_35_groups_0 = const()[name = tensor("obj_35_groups_0"), val = tensor(1)]; + 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(180621056)))]; + 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(182718272)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([1])]; + tensor var_697_to_fp16 = const()[name = tensor("op_697_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_697_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_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(182720384)))]; + 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(182722496)))]; + 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_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("valid")]; + tensor query_11_strides_0 = const()[name = tensor("query_11_strides_0"), val = tensor([1, 1])]; + tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_11_dilations_0 = const()[name = tensor("query_11_dilations_0"), val = tensor([1, 1])]; + tensor query_11_groups_0 = const()[name = tensor("query_11_groups_0"), val = tensor(1)]; + 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(182724608)))]; + 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(184821824)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("valid")]; + tensor key_11_strides_0 = const()[name = tensor("key_11_strides_0"), val = tensor([1, 1])]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_11_dilations_0 = const()[name = tensor("key_11_dilations_0"), val = tensor([1, 1])]; + tensor key_11_groups_0 = const()[name = tensor("key_11_groups_0"), val = tensor(1)]; + 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(184823936)))]; + tensor key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("valid")]; + tensor value_11_strides_0 = const()[name = tensor("value_11_strides_0"), val = tensor([1, 1])]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_11_dilations_0 = const()[name = tensor("value_11_dilations_0"), val = tensor([1, 1])]; + tensor value_11_groups_0 = const()[name = tensor("value_11_groups_0"), val = tensor(1)]; + 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(186921152)))]; + 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(189018368)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; + tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_733, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_735_to_fp16 = const()[name = tensor("op_735_to_fp16"), val = tensor(0x1p-3)]; + tensor var_736_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_735_to_fp16)[name = tensor("op_736_cast_fp16")]; + tensor var_739 = const()[name = tensor("op_739"), val = tensor([1, 16, 64, 1500])]; + tensor var_740_cast_fp16 = reshape(shape = var_739, x = key_11_cast_fp16)[name = tensor("op_740_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_736_cast_fp16, y = var_740_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_582, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_744 = const()[name = tensor("op_744"), val = tensor([1, 16, 64, 1500])]; + tensor var_745_cast_fp16 = reshape(shape = var_744, x = value_11_cast_fp16)[name = tensor("op_745_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_745_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_748 = const()[name = tensor("op_748"), val = tensor([1, 1024, 1, 1])]; + tensor input_23_cast_fp16 = reshape(shape = var_748, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor obj_39_pad_type_0 = const()[name = tensor("obj_39_pad_type_0"), val = tensor("valid")]; + tensor obj_39_strides_0 = const()[name = tensor("obj_39_strides_0"), val = tensor([1, 1])]; + tensor obj_39_pad_0 = const()[name = tensor("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_39_dilations_0 = const()[name = tensor("obj_39_dilations_0"), val = tensor([1, 1])]; + tensor obj_39_groups_0 = const()[name = tensor("obj_39_groups_0"), val = tensor(1)]; + 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(189020480)))]; + 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(191117696)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor out_17_axes_0 = const()[name = tensor("out_17_axes_0"), val = tensor([1])]; + tensor var_766_to_fp16 = const()[name = tensor("op_766_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_766_to_fp16, x = inputs_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(191119808)))]; + 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(191121920)))]; + 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 input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("valid")]; + tensor input_27_strides_0 = const()[name = tensor("input_27_strides_0"), val = tensor([1, 1])]; + tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_27_dilations_0 = const()[name = tensor("input_27_dilations_0"), val = tensor([1, 1])]; + tensor input_27_groups_0 = const()[name = tensor("input_27_groups_0"), val = tensor(1)]; + 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(191124032)))]; + 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(199512704)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("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 hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_7_strides_0 = const()[name = tensor("hidden_states_7_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_7_dilations_0 = const()[name = tensor("hidden_states_7_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_7_groups_0 = const()[name = tensor("hidden_states_7_groups_0"), val = tensor(1)]; + 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(199520960)))]; + 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(207909632)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("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_801 = const()[name = tensor("op_801"), val = tensor(3)]; + tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; + tensor var_826_to_fp16 = const()[name = tensor("op_826_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_826_to_fp16, x = inputs_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207911744)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207913856)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("valid")]; + tensor query_13_strides_0 = const()[name = tensor("query_13_strides_0"), val = tensor([1, 1])]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_13_dilations_0 = const()[name = tensor("query_13_dilations_0"), val = tensor([1, 1])]; + tensor query_13_groups_0 = const()[name = tensor("query_13_groups_0"), val = tensor(1)]; + 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(207915968)))]; + 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(210013184)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor current_key_7_pad_type_0 = const()[name = tensor("current_key_7_pad_type_0"), val = tensor("valid")]; + tensor current_key_7_strides_0 = const()[name = tensor("current_key_7_strides_0"), val = tensor([1, 1])]; + tensor current_key_7_pad_0 = const()[name = tensor("current_key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_7_dilations_0 = const()[name = tensor("current_key_7_dilations_0"), val = tensor([1, 1])]; + tensor current_key_7_groups_0 = const()[name = tensor("current_key_7_groups_0"), val = tensor(1)]; + 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(210015296)))]; + tensor current_key_7_cast_fp16 = conv(dilations = current_key_7_dilations_0, groups = current_key_7_groups_0, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = current_key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; + tensor current_value_7_pad_type_0 = const()[name = tensor("current_value_7_pad_type_0"), val = tensor("valid")]; + tensor current_value_7_strides_0 = const()[name = tensor("current_value_7_strides_0"), val = tensor([1, 1])]; + tensor current_value_7_pad_0 = const()[name = tensor("current_value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_7_dilations_0 = const()[name = tensor("current_value_7_dilations_0"), val = tensor([1, 1])]; + tensor current_value_7_groups_0 = const()[name = tensor("current_value_7_groups_0"), val = tensor(1)]; + 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(212112512)))]; + 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(214209728)))]; + tensor current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = current_value_7_dilations_0, groups = current_value_7_groups_0, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = current_value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; + tensor var_865_cast_fp16 = mul(x = var_87_cast_fp16_3, y = var_207_cast_fp16)[name = tensor("op_865_cast_fp16")]; + tensor var_866_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_866_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_865_cast_fp16, y = var_866_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_869_cast_fp16 = mul(x = var_114_cast_fp16_3, y = var_207_cast_fp16)[name = tensor("op_869_cast_fp16")]; + tensor var_870_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_870_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_869_cast_fp16, y = var_870_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_874 = const()[name = tensor("op_874"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_874, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_876_to_fp16 = const()[name = tensor("op_876_to_fp16"), val = tensor(0x1p-3)]; + tensor var_877_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_876_to_fp16)[name = tensor("op_877_cast_fp16")]; + tensor var_880 = const()[name = tensor("op_880"), val = tensor([1, 16, 64, 448])]; + tensor var_881_cast_fp16 = reshape(shape = var_880, x = key_13_cast_fp16)[name = tensor("op_881_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_877_cast_fp16, y = var_881_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_889_cast_fp16 = softmax(axis = var_801, x = mh_w_21_cast_fp16)[name = tensor("op_889_cast_fp16")]; + tensor var_890 = const()[name = tensor("op_890"), val = tensor([1, 16, 64, 448])]; + tensor var_891_cast_fp16 = reshape(shape = var_890, x = value_13_cast_fp16)[name = tensor("op_891_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_891_cast_fp16, y = var_889_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_894 = const()[name = tensor("op_894"), val = tensor([1, 1024, 1, 1])]; + tensor input_31_cast_fp16 = reshape(shape = var_894, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor obj_49_pad_type_0 = const()[name = tensor("obj_49_pad_type_0"), val = tensor("valid")]; + tensor obj_49_strides_0 = const()[name = tensor("obj_49_strides_0"), val = tensor([1, 1])]; + tensor obj_49_pad_0 = const()[name = tensor("obj_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_49_dilations_0 = const()[name = tensor("obj_49_dilations_0"), val = tensor([1, 1])]; + tensor obj_49_groups_0 = const()[name = tensor("obj_49_groups_0"), val = tensor(1)]; + 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(214211840)))]; + 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(216309056)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([1])]; + tensor var_916_to_fp16 = const()[name = tensor("op_916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_916_to_fp16, x = inputs_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216311168)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216313280)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor query_15_pad_type_0 = const()[name = tensor("query_15_pad_type_0"), val = tensor("valid")]; + tensor query_15_strides_0 = const()[name = tensor("query_15_strides_0"), val = tensor([1, 1])]; + tensor query_15_pad_0 = const()[name = tensor("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_15_dilations_0 = const()[name = tensor("query_15_dilations_0"), val = tensor([1, 1])]; + tensor query_15_groups_0 = const()[name = tensor("query_15_groups_0"), val = tensor(1)]; + 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(216315392)))]; + 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(218412608)))]; + tensor query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("valid")]; + tensor key_15_strides_0 = const()[name = tensor("key_15_strides_0"), val = tensor([1, 1])]; + tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_15_dilations_0 = const()[name = tensor("key_15_dilations_0"), val = tensor([1, 1])]; + tensor key_15_groups_0 = const()[name = tensor("key_15_groups_0"), val = tensor(1)]; + 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(218414720)))]; + tensor key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_15_cast_fp16")]; + tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("valid")]; + tensor value_15_strides_0 = const()[name = tensor("value_15_strides_0"), val = tensor([1, 1])]; + tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_15_dilations_0 = const()[name = tensor("value_15_dilations_0"), val = tensor([1, 1])]; + tensor value_15_groups_0 = const()[name = tensor("value_15_groups_0"), val = tensor(1)]; + 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(220511936)))]; + 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(222609152)))]; + tensor value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_15_cast_fp16")]; + tensor var_952 = const()[name = tensor("op_952"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_952, x = query_15_cast_fp16)[name = tensor("mh_q_15_cast_fp16")]; + tensor var_954_to_fp16 = const()[name = tensor("op_954_to_fp16"), val = tensor(0x1p-3)]; + tensor var_955_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_954_to_fp16)[name = tensor("op_955_cast_fp16")]; + tensor var_958 = const()[name = tensor("op_958"), val = tensor([1, 16, 64, 1500])]; + tensor var_959_cast_fp16 = reshape(shape = var_958, x = key_15_cast_fp16)[name = tensor("op_959_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_955_cast_fp16, y = var_959_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_801, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_963 = const()[name = tensor("op_963"), val = tensor([1, 16, 64, 1500])]; + tensor var_964_cast_fp16 = reshape(shape = var_963, x = value_15_cast_fp16)[name = tensor("op_964_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_964_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_967 = const()[name = tensor("op_967"), val = tensor([1, 1024, 1, 1])]; + tensor input_33_cast_fp16 = reshape(shape = var_967, x = attn_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor obj_53_pad_type_0 = const()[name = tensor("obj_53_pad_type_0"), val = tensor("valid")]; + tensor obj_53_strides_0 = const()[name = tensor("obj_53_strides_0"), val = tensor([1, 1])]; + tensor obj_53_pad_0 = const()[name = tensor("obj_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_53_dilations_0 = const()[name = tensor("obj_53_dilations_0"), val = tensor([1, 1])]; + tensor obj_53_groups_0 = const()[name = tensor("obj_53_groups_0"), val = tensor(1)]; + 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(222611264)))]; + 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(224708480)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor out_23_axes_0 = const()[name = tensor("out_23_axes_0"), val = tensor([1])]; + tensor var_985_to_fp16 = const()[name = tensor("op_985_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_985_to_fp16, x = inputs_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(224710592)))]; + 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(224712704)))]; + 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 input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("valid")]; + tensor input_37_strides_0 = const()[name = tensor("input_37_strides_0"), val = tensor([1, 1])]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_37_dilations_0 = const()[name = tensor("input_37_dilations_0"), val = tensor([1, 1])]; + tensor input_37_groups_0 = const()[name = tensor("input_37_groups_0"), val = tensor(1)]; + 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(224714816)))]; + 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(233103488)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("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 hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_9_strides_0 = const()[name = tensor("hidden_states_9_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_9_dilations_0 = const()[name = tensor("hidden_states_9_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_9_groups_0 = const()[name = tensor("hidden_states_9_groups_0"), val = tensor(1)]; + 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(233111744)))]; + 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(241500416)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("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_1020 = const()[name = tensor("op_1020"), val = tensor(3)]; + tensor out_25_axes_0 = const()[name = tensor("out_25_axes_0"), val = tensor([1])]; + tensor var_1045_to_fp16 = const()[name = tensor("op_1045_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_1045_to_fp16, x = inputs_25_cast_fp16)[name = tensor("out_25_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(241502528)))]; + 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(241504640)))]; + 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_25_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("valid")]; + tensor query_17_strides_0 = const()[name = tensor("query_17_strides_0"), val = tensor([1, 1])]; + tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_17_dilations_0 = const()[name = tensor("query_17_dilations_0"), val = tensor([1, 1])]; + tensor query_17_groups_0 = const()[name = tensor("query_17_groups_0"), val = tensor(1)]; + 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(241506752)))]; + 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(243603968)))]; + tensor query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor current_key_9_pad_type_0 = const()[name = tensor("current_key_9_pad_type_0"), val = tensor("valid")]; + tensor current_key_9_strides_0 = const()[name = tensor("current_key_9_strides_0"), val = tensor([1, 1])]; + tensor current_key_9_pad_0 = const()[name = tensor("current_key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_9_dilations_0 = const()[name = tensor("current_key_9_dilations_0"), val = tensor([1, 1])]; + tensor current_key_9_groups_0 = const()[name = tensor("current_key_9_groups_0"), val = tensor(1)]; + 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(243606080)))]; + tensor current_key_9_cast_fp16 = conv(dilations = current_key_9_dilations_0, groups = current_key_9_groups_0, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = current_key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; + tensor current_value_9_pad_type_0 = const()[name = tensor("current_value_9_pad_type_0"), val = tensor("valid")]; + tensor current_value_9_strides_0 = const()[name = tensor("current_value_9_strides_0"), val = tensor([1, 1])]; + tensor current_value_9_pad_0 = const()[name = tensor("current_value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_9_dilations_0 = const()[name = tensor("current_value_9_dilations_0"), val = tensor([1, 1])]; + tensor current_value_9_groups_0 = const()[name = tensor("current_value_9_groups_0"), val = tensor(1)]; + 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(245703296)))]; + 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(247800512)))]; + tensor current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = current_value_9_dilations_0, groups = current_value_9_groups_0, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = current_value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; + tensor var_1084_cast_fp16 = mul(x = var_87_cast_fp16_4, y = var_207_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor var_1085_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_1084_cast_fp16, y = var_1085_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1088_cast_fp16 = mul(x = var_114_cast_fp16_4, y = var_207_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_1093, x = query_17_cast_fp16)[name = tensor("mh_q_17_cast_fp16")]; + tensor var_1095_to_fp16 = const()[name = tensor("op_1095_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1096_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1095_to_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1099 = const()[name = tensor("op_1099"), val = tensor([1, 16, 64, 448])]; + tensor var_1100_cast_fp16 = reshape(shape = var_1099, x = key_17_cast_fp16)[name = tensor("op_1100_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_1096_cast_fp16, y = var_1100_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1108_cast_fp16 = softmax(axis = var_1020, x = mh_w_27_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1109 = const()[name = tensor("op_1109"), val = tensor([1, 16, 64, 448])]; + tensor var_1110_cast_fp16 = reshape(shape = var_1109, x = value_17_cast_fp16)[name = tensor("op_1110_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_1110_cast_fp16, y = var_1108_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1, 1024, 1, 1])]; + tensor input_41_cast_fp16 = reshape(shape = var_1113, x = attn_17_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor obj_63_pad_type_0 = const()[name = tensor("obj_63_pad_type_0"), val = tensor("valid")]; + tensor obj_63_strides_0 = const()[name = tensor("obj_63_strides_0"), val = tensor([1, 1])]; + tensor obj_63_pad_0 = const()[name = tensor("obj_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_63_dilations_0 = const()[name = tensor("obj_63_dilations_0"), val = tensor([1, 1])]; + tensor obj_63_groups_0 = const()[name = tensor("obj_63_groups_0"), val = tensor(1)]; + 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(247802624)))]; + 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(249899840)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_63_dilations_0, groups = obj_63_groups_0, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = obj_63_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([1])]; + tensor var_1135_to_fp16 = const()[name = tensor("op_1135_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1135_to_fp16, x = inputs_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249901952)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249904064)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor query_19_pad_type_0 = const()[name = tensor("query_19_pad_type_0"), val = tensor("valid")]; + tensor query_19_strides_0 = const()[name = tensor("query_19_strides_0"), val = tensor([1, 1])]; + tensor query_19_pad_0 = const()[name = tensor("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_19_dilations_0 = const()[name = tensor("query_19_dilations_0"), val = tensor([1, 1])]; + tensor query_19_groups_0 = const()[name = tensor("query_19_groups_0"), val = tensor(1)]; + 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(249906176)))]; + 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(252003392)))]; + tensor query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("valid")]; + tensor key_19_strides_0 = const()[name = tensor("key_19_strides_0"), val = tensor([1, 1])]; + tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_19_dilations_0 = const()[name = tensor("key_19_dilations_0"), val = tensor([1, 1])]; + tensor key_19_groups_0 = const()[name = tensor("key_19_groups_0"), val = tensor(1)]; + 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(252005504)))]; + tensor key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_19_cast_fp16")]; + tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("valid")]; + tensor value_19_strides_0 = const()[name = tensor("value_19_strides_0"), val = tensor([1, 1])]; + tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_19_dilations_0 = const()[name = tensor("value_19_dilations_0"), val = tensor([1, 1])]; + tensor value_19_groups_0 = const()[name = tensor("value_19_groups_0"), val = tensor(1)]; + 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(254102720)))]; + 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(256199936)))]; + tensor value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_19_cast_fp16")]; + tensor var_1171 = const()[name = tensor("op_1171"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_1171, x = query_19_cast_fp16)[name = tensor("mh_q_19_cast_fp16")]; + tensor var_1173_to_fp16 = const()[name = tensor("op_1173_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1174_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1173_to_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([1, 16, 64, 1500])]; + tensor var_1178_cast_fp16 = reshape(shape = var_1177, x = key_19_cast_fp16)[name = tensor("op_1178_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_1174_cast_fp16, y = var_1178_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_1020, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 16, 64, 1500])]; + tensor var_1183_cast_fp16 = reshape(shape = var_1182, x = value_19_cast_fp16)[name = tensor("op_1183_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_1183_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 1024, 1, 1])]; + tensor input_43_cast_fp16 = reshape(shape = var_1186, x = attn_19_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor obj_67_pad_type_0 = const()[name = tensor("obj_67_pad_type_0"), val = tensor("valid")]; + tensor obj_67_strides_0 = const()[name = tensor("obj_67_strides_0"), val = tensor([1, 1])]; + tensor obj_67_pad_0 = const()[name = tensor("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_67_dilations_0 = const()[name = tensor("obj_67_dilations_0"), val = tensor([1, 1])]; + tensor obj_67_groups_0 = const()[name = tensor("obj_67_groups_0"), val = tensor(1)]; + 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(256202048)))]; + 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(258299264)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor out_29_axes_0 = const()[name = tensor("out_29_axes_0"), val = tensor([1])]; + tensor var_1204_to_fp16 = const()[name = tensor("op_1204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1204_to_fp16, x = inputs_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(258301376)))]; + 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(258303488)))]; + 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 input_47_pad_type_0 = const()[name = tensor("input_47_pad_type_0"), val = tensor("valid")]; + tensor input_47_strides_0 = const()[name = tensor("input_47_strides_0"), val = tensor([1, 1])]; + tensor input_47_pad_0 = const()[name = tensor("input_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_47_dilations_0 = const()[name = tensor("input_47_dilations_0"), val = tensor([1, 1])]; + tensor input_47_groups_0 = const()[name = tensor("input_47_groups_0"), val = tensor(1)]; + 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(258305600)))]; + 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(266694272)))]; + tensor input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("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 hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_11_strides_0 = const()[name = tensor("hidden_states_11_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_11_dilations_0 = const()[name = tensor("hidden_states_11_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_11_groups_0 = const()[name = tensor("hidden_states_11_groups_0"), val = tensor(1)]; + 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(266702528)))]; + 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(275091200)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("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_1239 = const()[name = tensor("op_1239"), val = tensor(3)]; + tensor out_31_axes_0 = const()[name = tensor("out_31_axes_0"), val = tensor([1])]; + tensor var_1264_to_fp16 = const()[name = tensor("op_1264_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1264_to_fp16, x = inputs_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor obj_71_gamma_0_to_fp16 = const()[name = tensor("obj_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275093312)))]; + tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275095424)))]; + tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("valid")]; + tensor query_21_strides_0 = const()[name = tensor("query_21_strides_0"), val = tensor([1, 1])]; + tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_21_dilations_0 = const()[name = tensor("query_21_dilations_0"), val = tensor([1, 1])]; + tensor query_21_groups_0 = const()[name = tensor("query_21_groups_0"), val = tensor(1)]; + 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(275097536)))]; + 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(277194752)))]; + tensor query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor current_key_11_pad_type_0 = const()[name = tensor("current_key_11_pad_type_0"), val = tensor("valid")]; + tensor current_key_11_strides_0 = const()[name = tensor("current_key_11_strides_0"), val = tensor([1, 1])]; + tensor current_key_11_pad_0 = const()[name = tensor("current_key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_11_dilations_0 = const()[name = tensor("current_key_11_dilations_0"), val = tensor([1, 1])]; + tensor current_key_11_groups_0 = const()[name = tensor("current_key_11_groups_0"), val = tensor(1)]; + 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(277196864)))]; + tensor current_key_11_cast_fp16 = conv(dilations = current_key_11_dilations_0, groups = current_key_11_groups_0, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = current_key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; + tensor current_value_11_pad_type_0 = const()[name = tensor("current_value_11_pad_type_0"), val = tensor("valid")]; + tensor current_value_11_strides_0 = const()[name = tensor("current_value_11_strides_0"), val = tensor([1, 1])]; + tensor current_value_11_pad_0 = const()[name = tensor("current_value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_11_dilations_0 = const()[name = tensor("current_value_11_dilations_0"), val = tensor([1, 1])]; + tensor current_value_11_groups_0 = const()[name = tensor("current_value_11_groups_0"), val = tensor(1)]; + 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(279294080)))]; + 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(281391296)))]; + tensor current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = current_value_11_dilations_0, groups = current_value_11_groups_0, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = current_value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; + tensor var_1303_cast_fp16 = mul(x = var_87_cast_fp16_5, y = var_207_cast_fp16)[name = tensor("op_1303_cast_fp16")]; + tensor var_1304_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1304_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_1303_cast_fp16, y = var_1304_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1307_cast_fp16 = mul(x = var_114_cast_fp16_5, y = var_207_cast_fp16)[name = tensor("op_1307_cast_fp16")]; + tensor var_1308_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_1307_cast_fp16, y = var_1308_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_1312, x = query_21_cast_fp16)[name = tensor("mh_q_21_cast_fp16")]; + tensor var_1314_to_fp16 = const()[name = tensor("op_1314_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1315_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1314_to_fp16)[name = tensor("op_1315_cast_fp16")]; + tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([1, 16, 64, 448])]; + tensor var_1319_cast_fp16 = reshape(shape = var_1318, x = key_21_cast_fp16)[name = tensor("op_1319_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_1315_cast_fp16, y = var_1319_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_1327_cast_fp16 = softmax(axis = var_1239, x = mh_w_33_cast_fp16)[name = tensor("op_1327_cast_fp16")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([1, 16, 64, 448])]; + tensor var_1329_cast_fp16 = reshape(shape = var_1328, x = value_21_cast_fp16)[name = tensor("op_1329_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_1329_cast_fp16, y = var_1327_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 1024, 1, 1])]; + tensor input_51_cast_fp16 = reshape(shape = var_1332, x = attn_21_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor obj_77_pad_type_0 = const()[name = tensor("obj_77_pad_type_0"), val = tensor("valid")]; + tensor obj_77_strides_0 = const()[name = tensor("obj_77_strides_0"), val = tensor([1, 1])]; + tensor obj_77_pad_0 = const()[name = tensor("obj_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_77_dilations_0 = const()[name = tensor("obj_77_dilations_0"), val = tensor([1, 1])]; + tensor obj_77_groups_0 = const()[name = tensor("obj_77_groups_0"), val = tensor(1)]; + 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(281393408)))]; + 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(283490624)))]; + tensor obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_77_dilations_0, groups = obj_77_groups_0, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = obj_77_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([1])]; + tensor var_1354_to_fp16 = const()[name = tensor("op_1354_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1354_to_fp16, x = inputs_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_79_gamma_0_to_fp16 = const()[name = tensor("obj_79_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283492736)))]; + tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283494848)))]; + tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor query_23_pad_type_0 = const()[name = tensor("query_23_pad_type_0"), val = tensor("valid")]; + tensor query_23_strides_0 = const()[name = tensor("query_23_strides_0"), val = tensor([1, 1])]; + tensor query_23_pad_0 = const()[name = tensor("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_23_dilations_0 = const()[name = tensor("query_23_dilations_0"), val = tensor([1, 1])]; + tensor query_23_groups_0 = const()[name = tensor("query_23_groups_0"), val = tensor(1)]; + 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(283496960)))]; + 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(285594176)))]; + tensor query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("valid")]; + tensor key_23_strides_0 = const()[name = tensor("key_23_strides_0"), val = tensor([1, 1])]; + tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_23_dilations_0 = const()[name = tensor("key_23_dilations_0"), val = tensor([1, 1])]; + tensor key_23_groups_0 = const()[name = tensor("key_23_groups_0"), val = tensor(1)]; + 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(285596288)))]; + tensor key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_23_cast_fp16")]; + tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("valid")]; + tensor value_23_strides_0 = const()[name = tensor("value_23_strides_0"), val = tensor([1, 1])]; + tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_23_dilations_0 = const()[name = tensor("value_23_dilations_0"), val = tensor([1, 1])]; + tensor value_23_groups_0 = const()[name = tensor("value_23_groups_0"), val = tensor(1)]; + 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(287693504)))]; + 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(289790720)))]; + tensor value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_23_cast_fp16")]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_23_cast_fp16 = reshape(shape = var_1390, x = query_23_cast_fp16)[name = tensor("mh_q_23_cast_fp16")]; + tensor var_1392_to_fp16 = const()[name = tensor("op_1392_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1393_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1392_to_fp16)[name = tensor("op_1393_cast_fp16")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([1, 16, 64, 1500])]; + tensor var_1397_cast_fp16 = reshape(shape = var_1396, x = key_23_cast_fp16)[name = tensor("op_1397_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_1393_cast_fp16, y = var_1397_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_1239, x = mh_w_35_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([1, 16, 64, 1500])]; + tensor var_1402_cast_fp16 = reshape(shape = var_1401, x = value_23_cast_fp16)[name = tensor("op_1402_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_1402_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([1, 1024, 1, 1])]; + tensor input_53_cast_fp16 = reshape(shape = var_1405, x = attn_23_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor obj_81_pad_type_0 = const()[name = tensor("obj_81_pad_type_0"), val = tensor("valid")]; + tensor obj_81_strides_0 = const()[name = tensor("obj_81_strides_0"), val = tensor([1, 1])]; + tensor obj_81_pad_0 = const()[name = tensor("obj_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_81_dilations_0 = const()[name = tensor("obj_81_dilations_0"), val = tensor([1, 1])]; + tensor obj_81_groups_0 = const()[name = tensor("obj_81_groups_0"), val = tensor(1)]; + 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(289792832)))]; + 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(291890048)))]; + tensor obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = obj_81_dilations_0, groups = obj_81_groups_0, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = obj_81_strides_0, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor out_35_axes_0 = const()[name = tensor("out_35_axes_0"), val = tensor([1])]; + tensor var_1423_to_fp16 = const()[name = tensor("op_1423_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1423_to_fp16, x = inputs_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(291892160)))]; + 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(291894272)))]; + 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 input_57_pad_type_0 = const()[name = tensor("input_57_pad_type_0"), val = tensor("valid")]; + tensor input_57_strides_0 = const()[name = tensor("input_57_strides_0"), val = tensor([1, 1])]; + tensor input_57_pad_0 = const()[name = tensor("input_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_57_dilations_0 = const()[name = tensor("input_57_dilations_0"), val = tensor([1, 1])]; + tensor input_57_groups_0 = const()[name = tensor("input_57_groups_0"), val = tensor(1)]; + 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(291896384)))]; + 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(300285056)))]; + tensor input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("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 hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_13_strides_0 = const()[name = tensor("hidden_states_13_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_13_dilations_0 = const()[name = tensor("hidden_states_13_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_13_groups_0 = const()[name = tensor("hidden_states_13_groups_0"), val = tensor(1)]; + 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(300293312)))]; + 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(308681984)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("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_1458 = const()[name = tensor("op_1458"), val = tensor(3)]; + tensor out_37_axes_0 = const()[name = tensor("out_37_axes_0"), val = tensor([1])]; + tensor var_1483_to_fp16 = const()[name = tensor("op_1483_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1483_to_fp16, x = inputs_37_cast_fp16)[name = tensor("out_37_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(308684096)))]; + 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(308686208)))]; + 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_37_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("valid")]; + tensor query_25_strides_0 = const()[name = tensor("query_25_strides_0"), val = tensor([1, 1])]; + tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_25_dilations_0 = const()[name = tensor("query_25_dilations_0"), val = tensor([1, 1])]; + tensor query_25_groups_0 = const()[name = tensor("query_25_groups_0"), val = tensor(1)]; + 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(308688320)))]; + 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(310785536)))]; + tensor query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor current_key_13_pad_type_0 = const()[name = tensor("current_key_13_pad_type_0"), val = tensor("valid")]; + tensor current_key_13_strides_0 = const()[name = tensor("current_key_13_strides_0"), val = tensor([1, 1])]; + tensor current_key_13_pad_0 = const()[name = tensor("current_key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_13_dilations_0 = const()[name = tensor("current_key_13_dilations_0"), val = tensor([1, 1])]; + tensor current_key_13_groups_0 = const()[name = tensor("current_key_13_groups_0"), val = tensor(1)]; + 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(310787648)))]; + tensor current_key_13_cast_fp16 = conv(dilations = current_key_13_dilations_0, groups = current_key_13_groups_0, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = current_key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; + tensor current_value_13_pad_type_0 = const()[name = tensor("current_value_13_pad_type_0"), val = tensor("valid")]; + tensor current_value_13_strides_0 = const()[name = tensor("current_value_13_strides_0"), val = tensor([1, 1])]; + tensor current_value_13_pad_0 = const()[name = tensor("current_value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_13_dilations_0 = const()[name = tensor("current_value_13_dilations_0"), val = tensor([1, 1])]; + tensor current_value_13_groups_0 = const()[name = tensor("current_value_13_groups_0"), val = tensor(1)]; + 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(312884864)))]; + 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(314982080)))]; + tensor current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = current_value_13_dilations_0, groups = current_value_13_groups_0, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = current_value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; + tensor var_1522_cast_fp16 = mul(x = var_87_cast_fp16_6, y = var_207_cast_fp16)[name = tensor("op_1522_cast_fp16")]; + tensor var_1523_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1523_cast_fp16")]; + tensor key_25_cast_fp16 = add(x = var_1522_cast_fp16, y = var_1523_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_1526_cast_fp16 = mul(x = var_114_cast_fp16_6, y = var_207_cast_fp16)[name = tensor("op_1526_cast_fp16")]; + tensor var_1527_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor value_25_cast_fp16 = add(x = var_1526_cast_fp16, y = var_1527_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_1531 = const()[name = tensor("op_1531"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_25_cast_fp16 = reshape(shape = var_1531, x = query_25_cast_fp16)[name = tensor("mh_q_25_cast_fp16")]; + tensor var_1533_to_fp16 = const()[name = tensor("op_1533_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1534_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1533_to_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor var_1537 = const()[name = tensor("op_1537"), val = tensor([1, 16, 64, 448])]; + tensor var_1538_cast_fp16 = reshape(shape = var_1537, x = key_25_cast_fp16)[name = tensor("op_1538_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_1534_cast_fp16, y = var_1538_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_1546_cast_fp16 = softmax(axis = var_1458, x = mh_w_39_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 16, 64, 448])]; + tensor var_1548_cast_fp16 = reshape(shape = var_1547, x = value_25_cast_fp16)[name = tensor("op_1548_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_1548_cast_fp16, y = var_1546_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_1551 = const()[name = tensor("op_1551"), val = tensor([1, 1024, 1, 1])]; + tensor input_61_cast_fp16 = reshape(shape = var_1551, x = attn_25_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor obj_91_pad_type_0 = const()[name = tensor("obj_91_pad_type_0"), val = tensor("valid")]; + tensor obj_91_strides_0 = const()[name = tensor("obj_91_strides_0"), val = tensor([1, 1])]; + tensor obj_91_pad_0 = const()[name = tensor("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_91_dilations_0 = const()[name = tensor("obj_91_dilations_0"), val = tensor([1, 1])]; + tensor obj_91_groups_0 = const()[name = tensor("obj_91_groups_0"), val = tensor(1)]; + 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(314984192)))]; + 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(317081408)))]; + tensor obj_91_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor out_39_axes_0 = const()[name = tensor("out_39_axes_0"), val = tensor([1])]; + tensor var_1573_to_fp16 = const()[name = tensor("op_1573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1573_to_fp16, x = inputs_39_cast_fp16)[name = tensor("out_39_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(317083520)))]; + 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(317085632)))]; + 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_39_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor query_27_pad_type_0 = const()[name = tensor("query_27_pad_type_0"), val = tensor("valid")]; + tensor query_27_strides_0 = const()[name = tensor("query_27_strides_0"), val = tensor([1, 1])]; + tensor query_27_pad_0 = const()[name = tensor("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_27_dilations_0 = const()[name = tensor("query_27_dilations_0"), val = tensor([1, 1])]; + tensor query_27_groups_0 = const()[name = tensor("query_27_groups_0"), val = tensor(1)]; + 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(317087744)))]; + 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(319184960)))]; + tensor query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("valid")]; + tensor key_27_strides_0 = const()[name = tensor("key_27_strides_0"), val = tensor([1, 1])]; + tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_27_dilations_0 = const()[name = tensor("key_27_dilations_0"), val = tensor([1, 1])]; + tensor key_27_groups_0 = const()[name = tensor("key_27_groups_0"), val = tensor(1)]; + 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(319187072)))]; + tensor key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_27_cast_fp16")]; + tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("valid")]; + tensor value_27_strides_0 = const()[name = tensor("value_27_strides_0"), val = tensor([1, 1])]; + tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_27_dilations_0 = const()[name = tensor("value_27_dilations_0"), val = tensor([1, 1])]; + tensor value_27_groups_0 = const()[name = tensor("value_27_groups_0"), val = tensor(1)]; + 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(321284288)))]; + 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(323381504)))]; + tensor value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_27_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_27_cast_fp16 = reshape(shape = var_1609, x = query_27_cast_fp16)[name = tensor("mh_q_27_cast_fp16")]; + tensor var_1611_to_fp16 = const()[name = tensor("op_1611_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1612_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1611_to_fp16)[name = tensor("op_1612_cast_fp16")]; + tensor var_1615 = const()[name = tensor("op_1615"), val = tensor([1, 16, 64, 1500])]; + tensor var_1616_cast_fp16 = reshape(shape = var_1615, x = key_27_cast_fp16)[name = tensor("op_1616_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_1612_cast_fp16, y = var_1616_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor obj_97_cast_fp16 = softmax(axis = var_1458, x = mh_w_41_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor var_1620 = const()[name = tensor("op_1620"), val = tensor([1, 16, 64, 1500])]; + tensor var_1621_cast_fp16 = reshape(shape = var_1620, x = value_27_cast_fp16)[name = tensor("op_1621_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_1621_cast_fp16, y = obj_97_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_1624 = const()[name = tensor("op_1624"), val = tensor([1, 1024, 1, 1])]; + tensor input_63_cast_fp16 = reshape(shape = var_1624, x = attn_27_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor obj_95_pad_type_0 = const()[name = tensor("obj_95_pad_type_0"), val = tensor("valid")]; + tensor obj_95_strides_0 = const()[name = tensor("obj_95_strides_0"), val = tensor([1, 1])]; + tensor obj_95_pad_0 = const()[name = tensor("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_95_dilations_0 = const()[name = tensor("obj_95_dilations_0"), val = tensor([1, 1])]; + tensor obj_95_groups_0 = const()[name = tensor("obj_95_groups_0"), val = tensor(1)]; + 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(323383616)))]; + 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(325480832)))]; + tensor obj_95_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor out_41_axes_0 = const()[name = tensor("out_41_axes_0"), val = tensor([1])]; + tensor var_1642_to_fp16 = const()[name = tensor("op_1642_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1642_to_fp16, x = inputs_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(325482944)))]; + 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(325485056)))]; + 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 input_67_pad_type_0 = const()[name = tensor("input_67_pad_type_0"), val = tensor("valid")]; + tensor input_67_strides_0 = const()[name = tensor("input_67_strides_0"), val = tensor([1, 1])]; + tensor input_67_pad_0 = const()[name = tensor("input_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_67_dilations_0 = const()[name = tensor("input_67_dilations_0"), val = tensor([1, 1])]; + tensor input_67_groups_0 = const()[name = tensor("input_67_groups_0"), val = tensor(1)]; + 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(325487168)))]; + 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(333875840)))]; + tensor input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("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 hidden_states_15_pad_type_0 = const()[name = tensor("hidden_states_15_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_15_strides_0 = const()[name = tensor("hidden_states_15_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_0 = const()[name = tensor("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_15_dilations_0 = const()[name = tensor("hidden_states_15_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_15_groups_0 = const()[name = tensor("hidden_states_15_groups_0"), val = tensor(1)]; + 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(333884096)))]; + 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(342272768)))]; + tensor hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("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_1677 = const()[name = tensor("op_1677"), val = tensor(3)]; + tensor out_43_axes_0 = const()[name = tensor("out_43_axes_0"), val = tensor([1])]; + tensor var_1702_to_fp16 = const()[name = tensor("op_1702_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1702_to_fp16, x = inputs_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor obj_99_gamma_0_to_fp16 = const()[name = tensor("obj_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342274880)))]; + tensor obj_99_beta_0_to_fp16 = const()[name = tensor("obj_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342276992)))]; + tensor obj_99_epsilon_0_to_fp16 = const()[name = tensor("obj_99_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("valid")]; + tensor query_29_strides_0 = const()[name = tensor("query_29_strides_0"), val = tensor([1, 1])]; + tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_29_dilations_0 = const()[name = tensor("query_29_dilations_0"), val = tensor([1, 1])]; + tensor query_29_groups_0 = const()[name = tensor("query_29_groups_0"), val = tensor(1)]; + 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(342279104)))]; + 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(344376320)))]; + tensor query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor current_key_15_pad_type_0 = const()[name = tensor("current_key_15_pad_type_0"), val = tensor("valid")]; + tensor current_key_15_strides_0 = const()[name = tensor("current_key_15_strides_0"), val = tensor([1, 1])]; + tensor current_key_15_pad_0 = const()[name = tensor("current_key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_15_dilations_0 = const()[name = tensor("current_key_15_dilations_0"), val = tensor([1, 1])]; + tensor current_key_15_groups_0 = const()[name = tensor("current_key_15_groups_0"), val = tensor(1)]; + 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(344378432)))]; + tensor current_key_15_cast_fp16 = conv(dilations = current_key_15_dilations_0, groups = current_key_15_groups_0, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = current_key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; + tensor current_value_15_pad_type_0 = const()[name = tensor("current_value_15_pad_type_0"), val = tensor("valid")]; + tensor current_value_15_strides_0 = const()[name = tensor("current_value_15_strides_0"), val = tensor([1, 1])]; + tensor current_value_15_pad_0 = const()[name = tensor("current_value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_15_dilations_0 = const()[name = tensor("current_value_15_dilations_0"), val = tensor([1, 1])]; + tensor current_value_15_groups_0 = const()[name = tensor("current_value_15_groups_0"), val = tensor(1)]; + 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(346475648)))]; + 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(348572864)))]; + tensor current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = current_value_15_dilations_0, groups = current_value_15_groups_0, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = current_value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; + tensor var_1741_cast_fp16 = mul(x = var_87_cast_fp16_7, y = var_207_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor var_1742_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1742_cast_fp16")]; + tensor key_29_cast_fp16 = add(x = var_1741_cast_fp16, y = var_1742_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_1745_cast_fp16 = mul(x = var_114_cast_fp16_7, y = var_207_cast_fp16)[name = tensor("op_1745_cast_fp16")]; + tensor var_1746_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1746_cast_fp16")]; + tensor value_29_cast_fp16 = add(x = var_1745_cast_fp16, y = var_1746_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_1750 = const()[name = tensor("op_1750"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_29_cast_fp16 = reshape(shape = var_1750, x = query_29_cast_fp16)[name = tensor("mh_q_29_cast_fp16")]; + tensor var_1752_to_fp16 = const()[name = tensor("op_1752_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1753_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1752_to_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor var_1756 = const()[name = tensor("op_1756"), val = tensor([1, 16, 64, 448])]; + tensor var_1757_cast_fp16 = reshape(shape = var_1756, x = key_29_cast_fp16)[name = tensor("op_1757_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_1753_cast_fp16, y = var_1757_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_1765_cast_fp16 = softmax(axis = var_1677, x = mh_w_45_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor var_1766 = const()[name = tensor("op_1766"), val = tensor([1, 16, 64, 448])]; + tensor var_1767_cast_fp16 = reshape(shape = var_1766, x = value_29_cast_fp16)[name = tensor("op_1767_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_1767_cast_fp16, y = var_1765_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_1770 = const()[name = tensor("op_1770"), val = tensor([1, 1024, 1, 1])]; + tensor input_71_cast_fp16 = reshape(shape = var_1770, x = attn_29_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor obj_105_pad_type_0 = const()[name = tensor("obj_105_pad_type_0"), val = tensor("valid")]; + tensor obj_105_strides_0 = const()[name = tensor("obj_105_strides_0"), val = tensor([1, 1])]; + tensor obj_105_pad_0 = const()[name = tensor("obj_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_105_dilations_0 = const()[name = tensor("obj_105_dilations_0"), val = tensor([1, 1])]; + tensor obj_105_groups_0 = const()[name = tensor("obj_105_groups_0"), val = tensor(1)]; + 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(348574976)))]; + 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(350672192)))]; + tensor obj_105_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_105_dilations_0, groups = obj_105_groups_0, pad = obj_105_pad_0, pad_type = obj_105_pad_type_0, strides = obj_105_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("obj_105_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor out_45_axes_0 = const()[name = tensor("out_45_axes_0"), val = tensor([1])]; + tensor var_1792_to_fp16 = const()[name = tensor("op_1792_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1792_to_fp16, x = inputs_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_107_gamma_0_to_fp16 = const()[name = tensor("obj_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350674304)))]; + tensor obj_107_beta_0_to_fp16 = const()[name = tensor("obj_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350676416)))]; + tensor obj_107_epsilon_0_to_fp16 = const()[name = tensor("obj_107_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor query_31_pad_type_0 = const()[name = tensor("query_31_pad_type_0"), val = tensor("valid")]; + tensor query_31_strides_0 = const()[name = tensor("query_31_strides_0"), val = tensor([1, 1])]; + tensor query_31_pad_0 = const()[name = tensor("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_31_dilations_0 = const()[name = tensor("query_31_dilations_0"), val = tensor([1, 1])]; + tensor query_31_groups_0 = const()[name = tensor("query_31_groups_0"), val = tensor(1)]; + 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(350678528)))]; + 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(352775744)))]; + tensor query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("valid")]; + tensor key_31_strides_0 = const()[name = tensor("key_31_strides_0"), val = tensor([1, 1])]; + tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_31_dilations_0 = const()[name = tensor("key_31_dilations_0"), val = tensor([1, 1])]; + tensor key_31_groups_0 = const()[name = tensor("key_31_groups_0"), val = tensor(1)]; + 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(352777856)))]; + tensor key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_31_cast_fp16")]; + tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("valid")]; + tensor value_31_strides_0 = const()[name = tensor("value_31_strides_0"), val = tensor([1, 1])]; + tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_31_dilations_0 = const()[name = tensor("value_31_dilations_0"), val = tensor([1, 1])]; + tensor value_31_groups_0 = const()[name = tensor("value_31_groups_0"), val = tensor(1)]; + 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(354875072)))]; + 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(356972288)))]; + tensor value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_31_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_31_cast_fp16 = reshape(shape = var_1828, x = query_31_cast_fp16)[name = tensor("mh_q_31_cast_fp16")]; + tensor var_1830_to_fp16 = const()[name = tensor("op_1830_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1831_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_1830_to_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor var_1834 = const()[name = tensor("op_1834"), val = tensor([1, 16, 64, 1500])]; + tensor var_1835_cast_fp16 = reshape(shape = var_1834, x = key_31_cast_fp16)[name = tensor("op_1835_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_1831_cast_fp16, y = var_1835_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor obj_111_cast_fp16 = softmax(axis = var_1677, x = mh_w_47_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor var_1839 = const()[name = tensor("op_1839"), val = tensor([1, 16, 64, 1500])]; + tensor var_1840_cast_fp16 = reshape(shape = var_1839, x = value_31_cast_fp16)[name = tensor("op_1840_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_1840_cast_fp16, y = obj_111_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_1843 = const()[name = tensor("op_1843"), val = tensor([1, 1024, 1, 1])]; + tensor input_73_cast_fp16 = reshape(shape = var_1843, x = attn_31_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor obj_109_pad_type_0 = const()[name = tensor("obj_109_pad_type_0"), val = tensor("valid")]; + tensor obj_109_strides_0 = const()[name = tensor("obj_109_strides_0"), val = tensor([1, 1])]; + tensor obj_109_pad_0 = const()[name = tensor("obj_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_109_dilations_0 = const()[name = tensor("obj_109_dilations_0"), val = tensor([1, 1])]; + tensor obj_109_groups_0 = const()[name = tensor("obj_109_groups_0"), val = tensor(1)]; + 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(356974400)))]; + 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(359071616)))]; + tensor obj_109_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = obj_109_dilations_0, groups = obj_109_groups_0, pad = obj_109_pad_0, pad_type = obj_109_pad_type_0, strides = obj_109_strides_0, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("obj_109_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor out_47_axes_0 = const()[name = tensor("out_47_axes_0"), val = tensor([1])]; + tensor var_1861_to_fp16 = const()[name = tensor("op_1861_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1861_to_fp16, x = inputs_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(359073728)))]; + 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(359075840)))]; + 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 input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("valid")]; + tensor input_77_strides_0 = const()[name = tensor("input_77_strides_0"), val = tensor([1, 1])]; + tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_77_dilations_0 = const()[name = tensor("input_77_dilations_0"), val = tensor([1, 1])]; + tensor input_77_groups_0 = const()[name = tensor("input_77_groups_0"), val = tensor(1)]; + 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(359077952)))]; + 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(367466624)))]; + tensor input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("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 hidden_states_17_pad_type_0 = const()[name = tensor("hidden_states_17_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_17_strides_0 = const()[name = tensor("hidden_states_17_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_17_pad_0 = const()[name = tensor("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_17_dilations_0 = const()[name = tensor("hidden_states_17_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_17_groups_0 = const()[name = tensor("hidden_states_17_groups_0"), val = tensor(1)]; + 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(367474880)))]; + 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(375863552)))]; + tensor hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("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_1896 = const()[name = tensor("op_1896"), val = tensor(3)]; + tensor out_49_axes_0 = const()[name = tensor("out_49_axes_0"), val = tensor([1])]; + tensor var_1921_to_fp16 = const()[name = tensor("op_1921_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1921_to_fp16, x = inputs_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375865664)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375867776)))]; + tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("valid")]; + tensor query_33_strides_0 = const()[name = tensor("query_33_strides_0"), val = tensor([1, 1])]; + tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_33_dilations_0 = const()[name = tensor("query_33_dilations_0"), val = tensor([1, 1])]; + tensor query_33_groups_0 = const()[name = tensor("query_33_groups_0"), val = tensor(1)]; + 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(375869888)))]; + 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(377967104)))]; + tensor query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor current_key_17_pad_type_0 = const()[name = tensor("current_key_17_pad_type_0"), val = tensor("valid")]; + tensor current_key_17_strides_0 = const()[name = tensor("current_key_17_strides_0"), val = tensor([1, 1])]; + tensor current_key_17_pad_0 = const()[name = tensor("current_key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_17_dilations_0 = const()[name = tensor("current_key_17_dilations_0"), val = tensor([1, 1])]; + tensor current_key_17_groups_0 = const()[name = tensor("current_key_17_groups_0"), val = tensor(1)]; + 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(377969216)))]; + tensor current_key_17_cast_fp16 = conv(dilations = current_key_17_dilations_0, groups = current_key_17_groups_0, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = current_key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; + tensor current_value_17_pad_type_0 = const()[name = tensor("current_value_17_pad_type_0"), val = tensor("valid")]; + tensor current_value_17_strides_0 = const()[name = tensor("current_value_17_strides_0"), val = tensor([1, 1])]; + tensor current_value_17_pad_0 = const()[name = tensor("current_value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_17_dilations_0 = const()[name = tensor("current_value_17_dilations_0"), val = tensor([1, 1])]; + tensor current_value_17_groups_0 = const()[name = tensor("current_value_17_groups_0"), val = tensor(1)]; + 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(380066432)))]; + 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(382163648)))]; + tensor current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = current_value_17_dilations_0, groups = current_value_17_groups_0, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = current_value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = var_87_cast_fp16_8, y = var_207_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor var_1961_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1961_cast_fp16")]; + tensor key_33_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1961_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_1964_cast_fp16 = mul(x = var_114_cast_fp16_8, y = var_207_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor var_1965_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1965_cast_fp16")]; + tensor value_33_cast_fp16 = add(x = var_1964_cast_fp16, y = var_1965_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_1969 = const()[name = tensor("op_1969"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_33_cast_fp16 = reshape(shape = var_1969, x = query_33_cast_fp16)[name = tensor("mh_q_33_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor var_1975 = const()[name = tensor("op_1975"), val = tensor([1, 16, 64, 448])]; + tensor var_1976_cast_fp16 = reshape(shape = var_1975, x = key_33_cast_fp16)[name = tensor("op_1976_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_1972_cast_fp16, y = var_1976_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_1984_cast_fp16 = softmax(axis = var_1896, x = mh_w_51_cast_fp16)[name = tensor("op_1984_cast_fp16")]; + tensor var_1985 = const()[name = tensor("op_1985"), val = tensor([1, 16, 64, 448])]; + tensor var_1986_cast_fp16 = reshape(shape = var_1985, x = value_33_cast_fp16)[name = tensor("op_1986_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_1986_cast_fp16, y = var_1984_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_1989 = const()[name = tensor("op_1989"), val = tensor([1, 1024, 1, 1])]; + tensor input_81_cast_fp16 = reshape(shape = var_1989, x = attn_33_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor obj_119_pad_type_0 = const()[name = tensor("obj_119_pad_type_0"), val = tensor("valid")]; + tensor obj_119_strides_0 = const()[name = tensor("obj_119_strides_0"), val = tensor([1, 1])]; + tensor obj_119_pad_0 = const()[name = tensor("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_119_dilations_0 = const()[name = tensor("obj_119_dilations_0"), val = tensor([1, 1])]; + tensor obj_119_groups_0 = const()[name = tensor("obj_119_groups_0"), val = tensor(1)]; + 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(382165760)))]; + 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(384262976)))]; + tensor obj_119_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor out_51_axes_0 = const()[name = tensor("out_51_axes_0"), val = tensor([1])]; + tensor var_2011_to_fp16 = const()[name = tensor("op_2011_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_2011_to_fp16, x = inputs_51_cast_fp16)[name = tensor("out_51_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(384265088)))]; + 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(384267200)))]; + 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_51_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor query_35_pad_type_0 = const()[name = tensor("query_35_pad_type_0"), val = tensor("valid")]; + tensor query_35_strides_0 = const()[name = tensor("query_35_strides_0"), val = tensor([1, 1])]; + tensor query_35_pad_0 = const()[name = tensor("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_35_dilations_0 = const()[name = tensor("query_35_dilations_0"), val = tensor([1, 1])]; + tensor query_35_groups_0 = const()[name = tensor("query_35_groups_0"), val = tensor(1)]; + 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(384269312)))]; + 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(386366528)))]; + tensor query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("valid")]; + tensor key_35_strides_0 = const()[name = tensor("key_35_strides_0"), val = tensor([1, 1])]; + tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_35_dilations_0 = const()[name = tensor("key_35_dilations_0"), val = tensor([1, 1])]; + tensor key_35_groups_0 = const()[name = tensor("key_35_groups_0"), val = tensor(1)]; + 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(386368640)))]; + tensor key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_35_cast_fp16")]; + tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("valid")]; + tensor value_35_strides_0 = const()[name = tensor("value_35_strides_0"), val = tensor([1, 1])]; + tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_35_dilations_0 = const()[name = tensor("value_35_dilations_0"), val = tensor([1, 1])]; + tensor value_35_groups_0 = const()[name = tensor("value_35_groups_0"), val = tensor(1)]; + 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(388465856)))]; + 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(390563072)))]; + tensor value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_35_cast_fp16")]; + tensor var_2047 = const()[name = tensor("op_2047"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_35_cast_fp16 = reshape(shape = var_2047, x = query_35_cast_fp16)[name = tensor("mh_q_35_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2050_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_2049_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor var_2053 = const()[name = tensor("op_2053"), val = tensor([1, 16, 64, 1500])]; + tensor var_2054_cast_fp16 = reshape(shape = var_2053, x = key_35_cast_fp16)[name = tensor("op_2054_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_2050_cast_fp16, y = var_2054_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor obj_125_cast_fp16 = softmax(axis = var_1896, x = mh_w_53_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor var_2058 = const()[name = tensor("op_2058"), val = tensor([1, 16, 64, 1500])]; + tensor var_2059_cast_fp16 = reshape(shape = var_2058, x = value_35_cast_fp16)[name = tensor("op_2059_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_2059_cast_fp16, y = obj_125_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_2062 = const()[name = tensor("op_2062"), val = tensor([1, 1024, 1, 1])]; + tensor input_83_cast_fp16 = reshape(shape = var_2062, x = attn_35_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor obj_123_pad_type_0 = const()[name = tensor("obj_123_pad_type_0"), val = tensor("valid")]; + tensor obj_123_strides_0 = const()[name = tensor("obj_123_strides_0"), val = tensor([1, 1])]; + tensor obj_123_pad_0 = const()[name = tensor("obj_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_123_dilations_0 = const()[name = tensor("obj_123_dilations_0"), val = tensor([1, 1])]; + tensor obj_123_groups_0 = const()[name = tensor("obj_123_groups_0"), val = tensor(1)]; + 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(390565184)))]; + 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(392662400)))]; + tensor obj_123_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor out_53_axes_0 = const()[name = tensor("out_53_axes_0"), val = tensor([1])]; + tensor var_2080_to_fp16 = const()[name = tensor("op_2080_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_2080_to_fp16, x = inputs_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(392664512)))]; + 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(392666624)))]; + 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 input_87_pad_type_0 = const()[name = tensor("input_87_pad_type_0"), val = tensor("valid")]; + tensor input_87_strides_0 = const()[name = tensor("input_87_strides_0"), val = tensor([1, 1])]; + tensor input_87_pad_0 = const()[name = tensor("input_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_87_dilations_0 = const()[name = tensor("input_87_dilations_0"), val = tensor([1, 1])]; + tensor input_87_groups_0 = const()[name = tensor("input_87_groups_0"), val = tensor(1)]; + 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(392668736)))]; + 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(401057408)))]; + tensor input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("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 hidden_states_19_pad_type_0 = const()[name = tensor("hidden_states_19_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_19_strides_0 = const()[name = tensor("hidden_states_19_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_19_pad_0 = const()[name = tensor("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_19_dilations_0 = const()[name = tensor("hidden_states_19_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_19_groups_0 = const()[name = tensor("hidden_states_19_groups_0"), val = tensor(1)]; + 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(401065664)))]; + 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(409454336)))]; + tensor hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("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_2115 = const()[name = tensor("op_2115"), val = tensor(3)]; + tensor out_55_axes_0 = const()[name = tensor("out_55_axes_0"), val = tensor([1])]; + tensor var_2140_to_fp16 = const()[name = tensor("op_2140_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_2140_to_fp16, x = inputs_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor obj_127_gamma_0_to_fp16 = const()[name = tensor("obj_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409456448)))]; + tensor obj_127_beta_0_to_fp16 = const()[name = tensor("obj_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409458560)))]; + tensor obj_127_epsilon_0_to_fp16 = const()[name = tensor("obj_127_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("obj_127_cast_fp16")]; + tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("valid")]; + tensor query_37_strides_0 = const()[name = tensor("query_37_strides_0"), val = tensor([1, 1])]; + tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_37_dilations_0 = const()[name = tensor("query_37_dilations_0"), val = tensor([1, 1])]; + tensor query_37_groups_0 = const()[name = tensor("query_37_groups_0"), val = tensor(1)]; + 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(409460672)))]; + 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(411557888)))]; + tensor query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor current_key_19_pad_type_0 = const()[name = tensor("current_key_19_pad_type_0"), val = tensor("valid")]; + tensor current_key_19_strides_0 = const()[name = tensor("current_key_19_strides_0"), val = tensor([1, 1])]; + tensor current_key_19_pad_0 = const()[name = tensor("current_key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_19_dilations_0 = const()[name = tensor("current_key_19_dilations_0"), val = tensor([1, 1])]; + tensor current_key_19_groups_0 = const()[name = tensor("current_key_19_groups_0"), val = tensor(1)]; + 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(411560000)))]; + tensor current_key_19_cast_fp16 = conv(dilations = current_key_19_dilations_0, groups = current_key_19_groups_0, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = current_key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; + tensor current_value_19_pad_type_0 = const()[name = tensor("current_value_19_pad_type_0"), val = tensor("valid")]; + tensor current_value_19_strides_0 = const()[name = tensor("current_value_19_strides_0"), val = tensor([1, 1])]; + tensor current_value_19_pad_0 = const()[name = tensor("current_value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_19_dilations_0 = const()[name = tensor("current_value_19_dilations_0"), val = tensor([1, 1])]; + tensor current_value_19_groups_0 = const()[name = tensor("current_value_19_groups_0"), val = tensor(1)]; + 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(413657216)))]; + 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(415754432)))]; + tensor current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = current_value_19_dilations_0, groups = current_value_19_groups_0, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = current_value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; + tensor var_2179_cast_fp16 = mul(x = var_87_cast_fp16_9, y = var_207_cast_fp16)[name = tensor("op_2179_cast_fp16")]; + tensor var_2180_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2180_cast_fp16")]; + tensor key_37_cast_fp16 = add(x = var_2179_cast_fp16, y = var_2180_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_2183_cast_fp16 = mul(x = var_114_cast_fp16_9, y = var_207_cast_fp16)[name = tensor("op_2183_cast_fp16")]; + tensor var_2184_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2184_cast_fp16")]; + tensor value_37_cast_fp16 = add(x = var_2183_cast_fp16, y = var_2184_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_2188 = const()[name = tensor("op_2188"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_37_cast_fp16 = reshape(shape = var_2188, x = query_37_cast_fp16)[name = tensor("mh_q_37_cast_fp16")]; + tensor var_2190_to_fp16 = const()[name = tensor("op_2190_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2191_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2190_to_fp16)[name = tensor("op_2191_cast_fp16")]; + tensor var_2194 = const()[name = tensor("op_2194"), val = tensor([1, 16, 64, 448])]; + tensor var_2195_cast_fp16 = reshape(shape = var_2194, x = key_37_cast_fp16)[name = tensor("op_2195_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_2191_cast_fp16, y = var_2195_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_2203_cast_fp16 = softmax(axis = var_2115, x = mh_w_57_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor var_2204 = const()[name = tensor("op_2204"), val = tensor([1, 16, 64, 448])]; + tensor var_2205_cast_fp16 = reshape(shape = var_2204, x = value_37_cast_fp16)[name = tensor("op_2205_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_2205_cast_fp16, y = var_2203_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_2208 = const()[name = tensor("op_2208"), val = tensor([1, 1024, 1, 1])]; + tensor input_91_cast_fp16 = reshape(shape = var_2208, x = attn_37_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor obj_133_pad_type_0 = const()[name = tensor("obj_133_pad_type_0"), val = tensor("valid")]; + tensor obj_133_strides_0 = const()[name = tensor("obj_133_strides_0"), val = tensor([1, 1])]; + tensor obj_133_pad_0 = const()[name = tensor("obj_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_133_dilations_0 = const()[name = tensor("obj_133_dilations_0"), val = tensor([1, 1])]; + tensor obj_133_groups_0 = const()[name = tensor("obj_133_groups_0"), val = tensor(1)]; + 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(415756544)))]; + 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(417853760)))]; + tensor obj_133_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_133_dilations_0, groups = obj_133_groups_0, pad = obj_133_pad_0, pad_type = obj_133_pad_type_0, strides = obj_133_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("obj_133_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor out_57_axes_0 = const()[name = tensor("out_57_axes_0"), val = tensor([1])]; + tensor var_2230_to_fp16 = const()[name = tensor("op_2230_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_2230_to_fp16, x = inputs_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_135_gamma_0_to_fp16 = const()[name = tensor("obj_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417855872)))]; + tensor obj_135_beta_0_to_fp16 = const()[name = tensor("obj_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417857984)))]; + tensor obj_135_epsilon_0_to_fp16 = const()[name = tensor("obj_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_135_cast_fp16")]; + tensor query_39_pad_type_0 = const()[name = tensor("query_39_pad_type_0"), val = tensor("valid")]; + tensor query_39_strides_0 = const()[name = tensor("query_39_strides_0"), val = tensor([1, 1])]; + tensor query_39_pad_0 = const()[name = tensor("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_39_dilations_0 = const()[name = tensor("query_39_dilations_0"), val = tensor([1, 1])]; + tensor query_39_groups_0 = const()[name = tensor("query_39_groups_0"), val = tensor(1)]; + 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(417860096)))]; + 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(419957312)))]; + tensor query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("valid")]; + tensor key_39_strides_0 = const()[name = tensor("key_39_strides_0"), val = tensor([1, 1])]; + tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_39_dilations_0 = const()[name = tensor("key_39_dilations_0"), val = tensor([1, 1])]; + tensor key_39_groups_0 = const()[name = tensor("key_39_groups_0"), val = tensor(1)]; + 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(419959424)))]; + tensor key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_39_cast_fp16")]; + tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("valid")]; + tensor value_39_strides_0 = const()[name = tensor("value_39_strides_0"), val = tensor([1, 1])]; + tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_39_dilations_0 = const()[name = tensor("value_39_dilations_0"), val = tensor([1, 1])]; + tensor value_39_groups_0 = const()[name = tensor("value_39_groups_0"), val = tensor(1)]; + 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(422056640)))]; + 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(424153856)))]; + tensor value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_39_cast_fp16")]; + tensor var_2266 = const()[name = tensor("op_2266"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_39_cast_fp16 = reshape(shape = var_2266, x = query_39_cast_fp16)[name = tensor("mh_q_39_cast_fp16")]; + tensor var_2268_to_fp16 = const()[name = tensor("op_2268_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2269_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2268_to_fp16)[name = tensor("op_2269_cast_fp16")]; + tensor var_2272 = const()[name = tensor("op_2272"), val = tensor([1, 16, 64, 1500])]; + tensor var_2273_cast_fp16 = reshape(shape = var_2272, x = key_39_cast_fp16)[name = tensor("op_2273_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_2269_cast_fp16, y = var_2273_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor obj_139_cast_fp16 = softmax(axis = var_2115, x = mh_w_59_cast_fp16)[name = tensor("obj_139_cast_fp16")]; + tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1, 16, 64, 1500])]; + tensor var_2278_cast_fp16 = reshape(shape = var_2277, x = value_39_cast_fp16)[name = tensor("op_2278_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_2278_cast_fp16, y = obj_139_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_2281 = const()[name = tensor("op_2281"), val = tensor([1, 1024, 1, 1])]; + tensor input_93_cast_fp16 = reshape(shape = var_2281, x = attn_39_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor obj_137_pad_type_0 = const()[name = tensor("obj_137_pad_type_0"), val = tensor("valid")]; + tensor obj_137_strides_0 = const()[name = tensor("obj_137_strides_0"), val = tensor([1, 1])]; + tensor obj_137_pad_0 = const()[name = tensor("obj_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_137_dilations_0 = const()[name = tensor("obj_137_dilations_0"), val = tensor([1, 1])]; + tensor obj_137_groups_0 = const()[name = tensor("obj_137_groups_0"), val = tensor(1)]; + 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(424155968)))]; + 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(426253184)))]; + tensor obj_137_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = obj_137_dilations_0, groups = obj_137_groups_0, pad = obj_137_pad_0, pad_type = obj_137_pad_type_0, strides = obj_137_strides_0, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("obj_137_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor out_59_axes_0 = const()[name = tensor("out_59_axes_0"), val = tensor([1])]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_2299_to_fp16, x = inputs_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(426255296)))]; + 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(426257408)))]; + 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 input_97_pad_type_0 = const()[name = tensor("input_97_pad_type_0"), val = tensor("valid")]; + tensor input_97_strides_0 = const()[name = tensor("input_97_strides_0"), val = tensor([1, 1])]; + tensor input_97_pad_0 = const()[name = tensor("input_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_97_dilations_0 = const()[name = tensor("input_97_dilations_0"), val = tensor([1, 1])]; + tensor input_97_groups_0 = const()[name = tensor("input_97_groups_0"), val = tensor(1)]; + 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(426259520)))]; + 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(434648192)))]; + tensor input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("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 hidden_states_21_pad_type_0 = const()[name = tensor("hidden_states_21_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_21_strides_0 = const()[name = tensor("hidden_states_21_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_21_pad_0 = const()[name = tensor("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_21_dilations_0 = const()[name = tensor("hidden_states_21_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_21_groups_0 = const()[name = tensor("hidden_states_21_groups_0"), val = tensor(1)]; + 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(434656448)))]; + 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(443045120)))]; + tensor hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("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_2334 = const()[name = tensor("op_2334"), val = tensor(3)]; + tensor out_61_axes_0 = const()[name = tensor("out_61_axes_0"), val = tensor([1])]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_2359_to_fp16, x = inputs_61_cast_fp16)[name = tensor("out_61_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(443047232)))]; + 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(443049344)))]; + 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_61_cast_fp16)[name = tensor("obj_141_cast_fp16")]; + tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("valid")]; + tensor query_41_strides_0 = const()[name = tensor("query_41_strides_0"), val = tensor([1, 1])]; + tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_41_dilations_0 = const()[name = tensor("query_41_dilations_0"), val = tensor([1, 1])]; + tensor query_41_groups_0 = const()[name = tensor("query_41_groups_0"), val = tensor(1)]; + 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(443051456)))]; + 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(445148672)))]; + tensor query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor current_key_21_pad_type_0 = const()[name = tensor("current_key_21_pad_type_0"), val = tensor("valid")]; + tensor current_key_21_strides_0 = const()[name = tensor("current_key_21_strides_0"), val = tensor([1, 1])]; + tensor current_key_21_pad_0 = const()[name = tensor("current_key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_21_dilations_0 = const()[name = tensor("current_key_21_dilations_0"), val = tensor([1, 1])]; + tensor current_key_21_groups_0 = const()[name = tensor("current_key_21_groups_0"), val = tensor(1)]; + 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(445150784)))]; + tensor current_key_21_cast_fp16 = conv(dilations = current_key_21_dilations_0, groups = current_key_21_groups_0, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = current_key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; + tensor current_value_21_pad_type_0 = const()[name = tensor("current_value_21_pad_type_0"), val = tensor("valid")]; + tensor current_value_21_strides_0 = const()[name = tensor("current_value_21_strides_0"), val = tensor([1, 1])]; + tensor current_value_21_pad_0 = const()[name = tensor("current_value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_21_dilations_0 = const()[name = tensor("current_value_21_dilations_0"), val = tensor([1, 1])]; + tensor current_value_21_groups_0 = const()[name = tensor("current_value_21_groups_0"), val = tensor(1)]; + 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(447248000)))]; + 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(449345216)))]; + tensor current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = current_value_21_dilations_0, groups = current_value_21_groups_0, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = current_value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; + tensor var_2398_cast_fp16 = mul(x = var_87_cast_fp16_10, y = var_207_cast_fp16)[name = tensor("op_2398_cast_fp16")]; + tensor var_2399_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2399_cast_fp16")]; + tensor key_41_cast_fp16 = add(x = var_2398_cast_fp16, y = var_2399_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_2402_cast_fp16 = mul(x = var_114_cast_fp16_10, y = var_207_cast_fp16)[name = tensor("op_2402_cast_fp16")]; + tensor var_2403_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2403_cast_fp16")]; + tensor value_41_cast_fp16 = add(x = var_2402_cast_fp16, y = var_2403_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_2407 = const()[name = tensor("op_2407"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_41_cast_fp16 = reshape(shape = var_2407, x = query_41_cast_fp16)[name = tensor("mh_q_41_cast_fp16")]; + tensor var_2409_to_fp16 = const()[name = tensor("op_2409_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2410_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2409_to_fp16)[name = tensor("op_2410_cast_fp16")]; + tensor var_2413 = const()[name = tensor("op_2413"), val = tensor([1, 16, 64, 448])]; + tensor var_2414_cast_fp16 = reshape(shape = var_2413, x = key_41_cast_fp16)[name = tensor("op_2414_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_2410_cast_fp16, y = var_2414_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; + tensor var_2422_cast_fp16 = softmax(axis = var_2334, x = mh_w_63_cast_fp16)[name = tensor("op_2422_cast_fp16")]; + tensor var_2423 = const()[name = tensor("op_2423"), val = tensor([1, 16, 64, 448])]; + tensor var_2424_cast_fp16 = reshape(shape = var_2423, x = value_41_cast_fp16)[name = tensor("op_2424_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_2424_cast_fp16, y = var_2422_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_2427 = const()[name = tensor("op_2427"), val = tensor([1, 1024, 1, 1])]; + tensor input_101_cast_fp16 = reshape(shape = var_2427, x = attn_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor obj_147_pad_type_0 = const()[name = tensor("obj_147_pad_type_0"), val = tensor("valid")]; + tensor obj_147_strides_0 = const()[name = tensor("obj_147_strides_0"), val = tensor([1, 1])]; + tensor obj_147_pad_0 = const()[name = tensor("obj_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_147_dilations_0 = const()[name = tensor("obj_147_dilations_0"), val = tensor([1, 1])]; + tensor obj_147_groups_0 = const()[name = tensor("obj_147_groups_0"), val = tensor(1)]; + 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(449347328)))]; + 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(451444544)))]; + tensor obj_147_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_147_dilations_0, groups = obj_147_groups_0, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = obj_147_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("obj_147_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor out_63_axes_0 = const()[name = tensor("out_63_axes_0"), val = tensor([1])]; + tensor var_2449_to_fp16 = const()[name = tensor("op_2449_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2449_to_fp16, x = inputs_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor obj_149_gamma_0_to_fp16 = const()[name = tensor("obj_149_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451446656)))]; + tensor obj_149_beta_0_to_fp16 = const()[name = tensor("obj_149_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451448768)))]; + tensor obj_149_epsilon_0_to_fp16 = const()[name = tensor("obj_149_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("obj_149_cast_fp16")]; + tensor query_43_pad_type_0 = const()[name = tensor("query_43_pad_type_0"), val = tensor("valid")]; + tensor query_43_strides_0 = const()[name = tensor("query_43_strides_0"), val = tensor([1, 1])]; + tensor query_43_pad_0 = const()[name = tensor("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_43_dilations_0 = const()[name = tensor("query_43_dilations_0"), val = tensor([1, 1])]; + tensor query_43_groups_0 = const()[name = tensor("query_43_groups_0"), val = tensor(1)]; + 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(451450880)))]; + 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(453548096)))]; + tensor query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("valid")]; + tensor key_43_strides_0 = const()[name = tensor("key_43_strides_0"), val = tensor([1, 1])]; + tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_43_dilations_0 = const()[name = tensor("key_43_dilations_0"), val = tensor([1, 1])]; + tensor key_43_groups_0 = const()[name = tensor("key_43_groups_0"), val = tensor(1)]; + 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(453550208)))]; + tensor key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_43_cast_fp16")]; + tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("valid")]; + tensor value_43_strides_0 = const()[name = tensor("value_43_strides_0"), val = tensor([1, 1])]; + tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_43_dilations_0 = const()[name = tensor("value_43_dilations_0"), val = tensor([1, 1])]; + tensor value_43_groups_0 = const()[name = tensor("value_43_groups_0"), val = tensor(1)]; + 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(455647424)))]; + 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(457744640)))]; + tensor value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_43_cast_fp16")]; + tensor var_2485 = const()[name = tensor("op_2485"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_43_cast_fp16 = reshape(shape = var_2485, x = query_43_cast_fp16)[name = tensor("mh_q_43_cast_fp16")]; + tensor var_2487_to_fp16 = const()[name = tensor("op_2487_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2488_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2487_to_fp16)[name = tensor("op_2488_cast_fp16")]; + tensor var_2491 = const()[name = tensor("op_2491"), val = tensor([1, 16, 64, 1500])]; + tensor var_2492_cast_fp16 = reshape(shape = var_2491, x = key_43_cast_fp16)[name = tensor("op_2492_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_2488_cast_fp16, y = var_2492_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; + tensor obj_153_cast_fp16 = softmax(axis = var_2334, x = mh_w_65_cast_fp16)[name = tensor("obj_153_cast_fp16")]; + tensor var_2496 = const()[name = tensor("op_2496"), val = tensor([1, 16, 64, 1500])]; + tensor var_2497_cast_fp16 = reshape(shape = var_2496, x = value_43_cast_fp16)[name = tensor("op_2497_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_2497_cast_fp16, y = obj_153_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_2500 = const()[name = tensor("op_2500"), val = tensor([1, 1024, 1, 1])]; + tensor input_103_cast_fp16 = reshape(shape = var_2500, x = attn_43_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor obj_151_pad_type_0 = const()[name = tensor("obj_151_pad_type_0"), val = tensor("valid")]; + tensor obj_151_strides_0 = const()[name = tensor("obj_151_strides_0"), val = tensor([1, 1])]; + tensor obj_151_pad_0 = const()[name = tensor("obj_151_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_151_dilations_0 = const()[name = tensor("obj_151_dilations_0"), val = tensor([1, 1])]; + tensor obj_151_groups_0 = const()[name = tensor("obj_151_groups_0"), val = tensor(1)]; + 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(457746752)))]; + 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(459843968)))]; + tensor obj_151_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = obj_151_dilations_0, groups = obj_151_groups_0, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = obj_151_strides_0, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("obj_151_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor out_65_axes_0 = const()[name = tensor("out_65_axes_0"), val = tensor([1])]; + tensor var_2518_to_fp16 = const()[name = tensor("op_2518_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2518_to_fp16, x = inputs_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(459846080)))]; + 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(459848192)))]; + 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 input_107_pad_type_0 = const()[name = tensor("input_107_pad_type_0"), val = tensor("valid")]; + tensor input_107_strides_0 = const()[name = tensor("input_107_strides_0"), val = tensor([1, 1])]; + tensor input_107_pad_0 = const()[name = tensor("input_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_107_dilations_0 = const()[name = tensor("input_107_dilations_0"), val = tensor([1, 1])]; + tensor input_107_groups_0 = const()[name = tensor("input_107_groups_0"), val = tensor(1)]; + 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(459850304)))]; + 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(468238976)))]; + tensor input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("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 hidden_states_23_pad_type_0 = const()[name = tensor("hidden_states_23_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_23_strides_0 = const()[name = tensor("hidden_states_23_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_0 = const()[name = tensor("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_23_dilations_0 = const()[name = tensor("hidden_states_23_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_23_groups_0 = const()[name = tensor("hidden_states_23_groups_0"), val = tensor(1)]; + 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(468247232)))]; + 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(476635904)))]; + tensor hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("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_2553 = const()[name = tensor("op_2553"), val = tensor(3)]; + tensor out_67_axes_0 = const()[name = tensor("out_67_axes_0"), val = tensor([1])]; + tensor var_2578_to_fp16 = const()[name = tensor("op_2578_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2578_to_fp16, x = inputs_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor obj_155_gamma_0_to_fp16 = const()[name = tensor("obj_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476638016)))]; + tensor obj_155_beta_0_to_fp16 = const()[name = tensor("obj_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476640128)))]; + tensor obj_155_epsilon_0_to_fp16 = const()[name = tensor("obj_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("obj_155_cast_fp16")]; + tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("valid")]; + tensor query_45_strides_0 = const()[name = tensor("query_45_strides_0"), val = tensor([1, 1])]; + tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_45_dilations_0 = const()[name = tensor("query_45_dilations_0"), val = tensor([1, 1])]; + tensor query_45_groups_0 = const()[name = tensor("query_45_groups_0"), val = tensor(1)]; + 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(476642240)))]; + 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(478739456)))]; + tensor query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor current_key_23_pad_type_0 = const()[name = tensor("current_key_23_pad_type_0"), val = tensor("valid")]; + tensor current_key_23_strides_0 = const()[name = tensor("current_key_23_strides_0"), val = tensor([1, 1])]; + tensor current_key_23_pad_0 = const()[name = tensor("current_key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_23_dilations_0 = const()[name = tensor("current_key_23_dilations_0"), val = tensor([1, 1])]; + tensor current_key_23_groups_0 = const()[name = tensor("current_key_23_groups_0"), val = tensor(1)]; + 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(478741568)))]; + tensor current_key_23_cast_fp16 = conv(dilations = current_key_23_dilations_0, groups = current_key_23_groups_0, pad = current_key_23_pad_0, pad_type = current_key_23_pad_type_0, strides = current_key_23_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_key_23_cast_fp16")]; + tensor current_value_23_pad_type_0 = const()[name = tensor("current_value_23_pad_type_0"), val = tensor("valid")]; + tensor current_value_23_strides_0 = const()[name = tensor("current_value_23_strides_0"), val = tensor([1, 1])]; + tensor current_value_23_pad_0 = const()[name = tensor("current_value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_23_dilations_0 = const()[name = tensor("current_value_23_dilations_0"), val = tensor([1, 1])]; + tensor current_value_23_groups_0 = const()[name = tensor("current_value_23_groups_0"), val = tensor(1)]; + 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(480838784)))]; + 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(482936000)))]; + tensor current_value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = current_value_23_dilations_0, groups = current_value_23_groups_0, pad = current_value_23_pad_0, pad_type = current_value_23_pad_type_0, strides = current_value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_value_23_cast_fp16")]; + tensor var_2617_cast_fp16 = mul(x = var_87_cast_fp16_11, y = var_207_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor var_2618_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2618_cast_fp16")]; + tensor key_45_cast_fp16 = add(x = var_2617_cast_fp16, y = var_2618_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_2621_cast_fp16 = mul(x = var_114_cast_fp16_11, y = var_207_cast_fp16)[name = tensor("op_2621_cast_fp16")]; + tensor var_2622_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2622_cast_fp16")]; + tensor value_45_cast_fp16 = add(x = var_2621_cast_fp16, y = var_2622_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_2626 = const()[name = tensor("op_2626"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_45_cast_fp16 = reshape(shape = var_2626, x = query_45_cast_fp16)[name = tensor("mh_q_45_cast_fp16")]; + tensor var_2628_to_fp16 = const()[name = tensor("op_2628_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2629_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2628_to_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor var_2632 = const()[name = tensor("op_2632"), val = tensor([1, 16, 64, 448])]; + tensor var_2633_cast_fp16 = reshape(shape = var_2632, x = key_45_cast_fp16)[name = tensor("op_2633_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_2629_cast_fp16, y = var_2633_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; + tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; + tensor var_2641_cast_fp16 = softmax(axis = var_2553, x = mh_w_69_cast_fp16)[name = tensor("op_2641_cast_fp16")]; + tensor var_2642 = const()[name = tensor("op_2642"), val = tensor([1, 16, 64, 448])]; + tensor var_2643_cast_fp16 = reshape(shape = var_2642, x = value_45_cast_fp16)[name = tensor("op_2643_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_2643_cast_fp16, y = var_2641_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_2646 = const()[name = tensor("op_2646"), val = tensor([1, 1024, 1, 1])]; + tensor input_111_cast_fp16 = reshape(shape = var_2646, x = attn_45_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor obj_161_pad_type_0 = const()[name = tensor("obj_161_pad_type_0"), val = tensor("valid")]; + tensor obj_161_strides_0 = const()[name = tensor("obj_161_strides_0"), val = tensor([1, 1])]; + tensor obj_161_pad_0 = const()[name = tensor("obj_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_161_dilations_0 = const()[name = tensor("obj_161_dilations_0"), val = tensor([1, 1])]; + tensor obj_161_groups_0 = const()[name = tensor("obj_161_groups_0"), val = tensor(1)]; + 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(482938112)))]; + 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(485035328)))]; + tensor obj_161_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_161_dilations_0, groups = obj_161_groups_0, pad = obj_161_pad_0, pad_type = obj_161_pad_type_0, strides = obj_161_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("obj_161_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor out_69_axes_0 = const()[name = tensor("out_69_axes_0"), val = tensor([1])]; + tensor var_2668_to_fp16 = const()[name = tensor("op_2668_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2668_to_fp16, x = inputs_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_163_gamma_0_to_fp16 = const()[name = tensor("obj_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485037440)))]; + tensor obj_163_beta_0_to_fp16 = const()[name = tensor("obj_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485039552)))]; + tensor obj_163_epsilon_0_to_fp16 = const()[name = tensor("obj_163_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_163_cast_fp16")]; + tensor query_47_pad_type_0 = const()[name = tensor("query_47_pad_type_0"), val = tensor("valid")]; + tensor query_47_strides_0 = const()[name = tensor("query_47_strides_0"), val = tensor([1, 1])]; + tensor query_47_pad_0 = const()[name = tensor("query_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_47_dilations_0 = const()[name = tensor("query_47_dilations_0"), val = tensor([1, 1])]; + tensor query_47_groups_0 = const()[name = tensor("query_47_groups_0"), val = tensor(1)]; + 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(485041664)))]; + 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(487138880)))]; + tensor query_47_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = query_47_dilations_0, groups = query_47_groups_0, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = query_47_strides_0, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor("query_47_cast_fp16")]; + tensor key_47_pad_type_0 = const()[name = tensor("key_47_pad_type_0"), val = tensor("valid")]; + tensor key_47_strides_0 = const()[name = tensor("key_47_strides_0"), val = tensor([1, 1])]; + tensor key_47_pad_0 = const()[name = tensor("key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_47_dilations_0 = const()[name = tensor("key_47_dilations_0"), val = tensor([1, 1])]; + tensor key_47_groups_0 = const()[name = tensor("key_47_groups_0"), val = tensor(1)]; + 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(487140992)))]; + tensor key_47_cast_fp16 = conv(dilations = key_47_dilations_0, groups = key_47_groups_0, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = key_47_strides_0, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_47_cast_fp16")]; + tensor value_47_pad_type_0 = const()[name = tensor("value_47_pad_type_0"), val = tensor("valid")]; + tensor value_47_strides_0 = const()[name = tensor("value_47_strides_0"), val = tensor([1, 1])]; + tensor value_47_pad_0 = const()[name = tensor("value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_47_dilations_0 = const()[name = tensor("value_47_dilations_0"), val = tensor([1, 1])]; + tensor value_47_groups_0 = const()[name = tensor("value_47_groups_0"), val = tensor(1)]; + 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(489238208)))]; + 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(491335424)))]; + tensor value_47_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = value_47_dilations_0, groups = value_47_groups_0, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = value_47_strides_0, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_47_cast_fp16")]; + tensor var_2704 = const()[name = tensor("op_2704"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_47_cast_fp16 = reshape(shape = var_2704, x = query_47_cast_fp16)[name = tensor("mh_q_47_cast_fp16")]; + tensor var_2706_to_fp16 = const()[name = tensor("op_2706_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2707_cast_fp16 = mul(x = mh_q_47_cast_fp16, y = var_2706_to_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor var_2710 = const()[name = tensor("op_2710"), val = tensor([1, 16, 64, 1500])]; + tensor var_2711_cast_fp16 = reshape(shape = var_2710, x = key_47_cast_fp16)[name = tensor("op_2711_cast_fp16")]; + tensor mh_w_71_transpose_x_0 = const()[name = tensor("mh_w_71_transpose_x_0"), val = tensor(true)]; + tensor mh_w_71_transpose_y_0 = const()[name = tensor("mh_w_71_transpose_y_0"), val = tensor(false)]; + tensor mh_w_71_cast_fp16 = matmul(transpose_x = mh_w_71_transpose_x_0, transpose_y = mh_w_71_transpose_y_0, x = var_2707_cast_fp16, y = var_2711_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; + tensor obj_167_cast_fp16 = softmax(axis = var_2553, x = mh_w_71_cast_fp16)[name = tensor("obj_167_cast_fp16")]; + tensor var_2715 = const()[name = tensor("op_2715"), val = tensor([1, 16, 64, 1500])]; + tensor var_2716_cast_fp16 = reshape(shape = var_2715, x = value_47_cast_fp16)[name = tensor("op_2716_cast_fp16")]; + tensor attn_47_transpose_x_0 = const()[name = tensor("attn_47_transpose_x_0"), val = tensor(false)]; + tensor attn_47_transpose_y_0 = const()[name = tensor("attn_47_transpose_y_0"), val = tensor(true)]; + tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2716_cast_fp16, y = obj_167_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_2719 = const()[name = tensor("op_2719"), val = tensor([1, 1024, 1, 1])]; + tensor input_113_cast_fp16 = reshape(shape = var_2719, x = attn_47_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor obj_165_pad_type_0 = const()[name = tensor("obj_165_pad_type_0"), val = tensor("valid")]; + tensor obj_165_strides_0 = const()[name = tensor("obj_165_strides_0"), val = tensor([1, 1])]; + tensor obj_165_pad_0 = const()[name = tensor("obj_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_165_dilations_0 = const()[name = tensor("obj_165_dilations_0"), val = tensor([1, 1])]; + tensor obj_165_groups_0 = const()[name = tensor("obj_165_groups_0"), val = tensor(1)]; + 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(491337536)))]; + 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(493434752)))]; + tensor obj_165_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = obj_165_dilations_0, groups = obj_165_groups_0, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = obj_165_strides_0, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("obj_165_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor out_71_axes_0 = const()[name = tensor("out_71_axes_0"), val = tensor([1])]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2740_to_fp16, x = inputs_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(493436864)))]; + 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(493438976)))]; + 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 input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("valid")]; + tensor input_117_strides_0 = const()[name = tensor("input_117_strides_0"), val = tensor([1, 1])]; + tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_117_dilations_0 = const()[name = tensor("input_117_dilations_0"), val = tensor([1, 1])]; + tensor input_117_groups_0 = const()[name = tensor("input_117_groups_0"), val = tensor(1)]; + 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(493441088)))]; + 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(501829760)))]; + tensor input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor hidden_states_25_pad_type_0 = const()[name = tensor("hidden_states_25_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_25_strides_0 = const()[name = tensor("hidden_states_25_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_25_pad_0 = const()[name = tensor("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_25_dilations_0 = const()[name = tensor("hidden_states_25_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_25_groups_0 = const()[name = tensor("hidden_states_25_groups_0"), val = tensor(1)]; + 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(501838016)))]; + 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(510226688)))]; + tensor hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_2776 = const()[name = tensor("op_2776"), val = tensor(3)]; + tensor out_73_axes_0 = const()[name = tensor("out_73_axes_0"), val = tensor([1])]; + tensor var_2801_to_fp16 = const()[name = tensor("op_2801_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2801_to_fp16, x = inputs_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_169_gamma_0_to_fp16 = const()[name = tensor("obj_169_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510228800)))]; + tensor obj_169_beta_0_to_fp16 = const()[name = tensor("obj_169_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510230912)))]; + tensor obj_169_epsilon_0_to_fp16 = const()[name = tensor("obj_169_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_169_cast_fp16 = batch_norm(beta = obj_169_beta_0_to_fp16, epsilon = obj_169_epsilon_0_to_fp16, gamma = obj_169_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_169_cast_fp16")]; + tensor query_49_pad_type_0 = const()[name = tensor("query_49_pad_type_0"), val = tensor("valid")]; + tensor query_49_strides_0 = const()[name = tensor("query_49_strides_0"), val = tensor([1, 1])]; + tensor query_49_pad_0 = const()[name = tensor("query_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_49_dilations_0 = const()[name = tensor("query_49_dilations_0"), val = tensor([1, 1])]; + tensor query_49_groups_0 = const()[name = tensor("query_49_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510233024)))]; + tensor layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512330240)))]; + tensor query_49_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor current_key_25_pad_type_0 = const()[name = tensor("current_key_25_pad_type_0"), val = tensor("valid")]; + tensor current_key_25_strides_0 = const()[name = tensor("current_key_25_strides_0"), val = tensor([1, 1])]; + tensor current_key_25_pad_0 = const()[name = tensor("current_key_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_25_dilations_0 = const()[name = tensor("current_key_25_dilations_0"), val = tensor([1, 1])]; + tensor current_key_25_groups_0 = const()[name = tensor("current_key_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512332352)))]; + tensor current_key_25_cast_fp16 = conv(dilations = current_key_25_dilations_0, groups = current_key_25_groups_0, pad = current_key_25_pad_0, pad_type = current_key_25_pad_type_0, strides = current_key_25_strides_0, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_key_25_cast_fp16")]; + tensor current_value_25_pad_type_0 = const()[name = tensor("current_value_25_pad_type_0"), val = tensor("valid")]; + tensor current_value_25_strides_0 = const()[name = tensor("current_value_25_strides_0"), val = tensor([1, 1])]; + tensor current_value_25_pad_0 = const()[name = tensor("current_value_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_25_dilations_0 = const()[name = tensor("current_value_25_dilations_0"), val = tensor([1, 1])]; + tensor current_value_25_groups_0 = const()[name = tensor("current_value_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514429568)))]; + tensor layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(516526784)))]; + tensor current_value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = current_value_25_dilations_0, groups = current_value_25_groups_0, pad = current_value_25_pad_0, pad_type = current_value_25_pad_type_0, strides = current_value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_value_25_cast_fp16")]; + tensor var_2840_cast_fp16 = mul(x = var_87_cast_fp16_12, y = var_207_cast_fp16)[name = tensor("op_2840_cast_fp16")]; + tensor var_2841_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2841_cast_fp16")]; + tensor key_49_cast_fp16 = add(x = var_2840_cast_fp16, y = var_2841_cast_fp16)[name = tensor("key_49_cast_fp16")]; + tensor var_2844_cast_fp16 = mul(x = var_114_cast_fp16_12, y = var_207_cast_fp16)[name = tensor("op_2844_cast_fp16")]; + tensor var_2845_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2845_cast_fp16")]; + tensor value_49_cast_fp16 = add(x = var_2844_cast_fp16, y = var_2845_cast_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_49_cast_fp16 = reshape(shape = var_2849, x = query_49_cast_fp16)[name = tensor("mh_q_49_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2852_cast_fp16 = mul(x = mh_q_49_cast_fp16, y = var_2851_to_fp16)[name = tensor("op_2852_cast_fp16")]; + tensor var_2855 = const()[name = tensor("op_2855"), val = tensor([1, 16, 64, 448])]; + tensor var_2856_cast_fp16 = reshape(shape = var_2855, x = key_49_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor mh_w_73_transpose_x_0 = const()[name = tensor("mh_w_73_transpose_x_0"), val = tensor(true)]; + tensor mh_w_73_transpose_y_0 = const()[name = tensor("mh_w_73_transpose_y_0"), val = tensor(false)]; + tensor mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_2852_cast_fp16, y = var_2856_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; + tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; + tensor var_2864_cast_fp16 = softmax(axis = var_2776, x = mh_w_75_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor([1, 16, 64, 448])]; + tensor var_2866_cast_fp16 = reshape(shape = var_2865, x = value_49_cast_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor attn_49_transpose_x_0 = const()[name = tensor("attn_49_transpose_x_0"), val = tensor(false)]; + tensor attn_49_transpose_y_0 = const()[name = tensor("attn_49_transpose_y_0"), val = tensor(true)]; + tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_2866_cast_fp16, y = var_2864_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_2869 = const()[name = tensor("op_2869"), val = tensor([1, 1024, 1, 1])]; + tensor input_121_cast_fp16 = reshape(shape = var_2869, x = attn_49_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor obj_175_pad_type_0 = const()[name = tensor("obj_175_pad_type_0"), val = tensor("valid")]; + tensor obj_175_strides_0 = const()[name = tensor("obj_175_strides_0"), val = tensor([1, 1])]; + tensor obj_175_pad_0 = const()[name = tensor("obj_175_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_175_dilations_0 = const()[name = tensor("obj_175_dilations_0"), val = tensor([1, 1])]; + tensor obj_175_groups_0 = const()[name = tensor("obj_175_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(516528896)))]; + tensor layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518626112)))]; + tensor obj_175_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = obj_175_dilations_0, groups = obj_175_groups_0, pad = obj_175_pad_0, pad_type = obj_175_pad_type_0, strides = obj_175_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("obj_175_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_175_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor out_75_axes_0 = const()[name = tensor("out_75_axes_0"), val = tensor([1])]; + tensor var_2891_to_fp16 = const()[name = tensor("op_2891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2891_to_fp16, x = inputs_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor obj_177_gamma_0_to_fp16 = const()[name = tensor("obj_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518628224)))]; + tensor obj_177_beta_0_to_fp16 = const()[name = tensor("obj_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518630336)))]; + tensor obj_177_epsilon_0_to_fp16 = const()[name = tensor("obj_177_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_177_cast_fp16 = batch_norm(beta = obj_177_beta_0_to_fp16, epsilon = obj_177_epsilon_0_to_fp16, gamma = obj_177_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("obj_177_cast_fp16")]; + tensor query_51_pad_type_0 = const()[name = tensor("query_51_pad_type_0"), val = tensor("valid")]; + tensor query_51_strides_0 = const()[name = tensor("query_51_strides_0"), val = tensor([1, 1])]; + tensor query_51_pad_0 = const()[name = tensor("query_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_51_dilations_0 = const()[name = tensor("query_51_dilations_0"), val = tensor([1, 1])]; + tensor query_51_groups_0 = const()[name = tensor("query_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518632448)))]; + tensor layers_12_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520729664)))]; + tensor query_51_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_bias_to_fp16, dilations = query_51_dilations_0, groups = query_51_groups_0, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = query_51_strides_0, weight = layers_12_encoder_attn_q_proj_weight_to_fp16, x = obj_177_cast_fp16)[name = tensor("query_51_cast_fp16")]; + tensor key_51_pad_type_0 = const()[name = tensor("key_51_pad_type_0"), val = tensor("valid")]; + tensor key_51_strides_0 = const()[name = tensor("key_51_strides_0"), val = tensor([1, 1])]; + tensor key_51_pad_0 = const()[name = tensor("key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_51_dilations_0 = const()[name = tensor("key_51_dilations_0"), val = tensor([1, 1])]; + tensor key_51_groups_0 = const()[name = tensor("key_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520731776)))]; + tensor key_51_cast_fp16 = conv(dilations = key_51_dilations_0, groups = key_51_groups_0, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = key_51_strides_0, weight = layers_12_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_51_cast_fp16")]; + tensor value_51_pad_type_0 = const()[name = tensor("value_51_pad_type_0"), val = tensor("valid")]; + tensor value_51_strides_0 = const()[name = tensor("value_51_strides_0"), val = tensor([1, 1])]; + tensor value_51_pad_0 = const()[name = tensor("value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_51_dilations_0 = const()[name = tensor("value_51_dilations_0"), val = tensor([1, 1])]; + tensor value_51_groups_0 = const()[name = tensor("value_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522828992)))]; + tensor layers_12_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524926208)))]; + tensor value_51_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_bias_to_fp16, dilations = value_51_dilations_0, groups = value_51_groups_0, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = value_51_strides_0, weight = layers_12_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_51_cast_fp16")]; + tensor var_2927 = const()[name = tensor("op_2927"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_51_cast_fp16 = reshape(shape = var_2927, x = query_51_cast_fp16)[name = tensor("mh_q_51_cast_fp16")]; + tensor var_2929_to_fp16 = const()[name = tensor("op_2929_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2930_cast_fp16 = mul(x = mh_q_51_cast_fp16, y = var_2929_to_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor var_2933 = const()[name = tensor("op_2933"), val = tensor([1, 16, 64, 1500])]; + tensor var_2934_cast_fp16 = reshape(shape = var_2933, x = key_51_cast_fp16)[name = tensor("op_2934_cast_fp16")]; + tensor mh_w_77_transpose_x_0 = const()[name = tensor("mh_w_77_transpose_x_0"), val = tensor(true)]; + tensor mh_w_77_transpose_y_0 = const()[name = tensor("mh_w_77_transpose_y_0"), val = tensor(false)]; + tensor mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_2930_cast_fp16, y = var_2934_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; + tensor obj_181_cast_fp16 = softmax(axis = var_2776, x = mh_w_77_cast_fp16)[name = tensor("obj_181_cast_fp16")]; + tensor var_2938 = const()[name = tensor("op_2938"), val = tensor([1, 16, 64, 1500])]; + tensor var_2939_cast_fp16 = reshape(shape = var_2938, x = value_51_cast_fp16)[name = tensor("op_2939_cast_fp16")]; + tensor attn_51_transpose_x_0 = const()[name = tensor("attn_51_transpose_x_0"), val = tensor(false)]; + tensor attn_51_transpose_y_0 = const()[name = tensor("attn_51_transpose_y_0"), val = tensor(true)]; + tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_2939_cast_fp16, y = obj_181_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_2942 = const()[name = tensor("op_2942"), val = tensor([1, 1024, 1, 1])]; + tensor input_123_cast_fp16 = reshape(shape = var_2942, x = attn_51_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor obj_179_pad_type_0 = const()[name = tensor("obj_179_pad_type_0"), val = tensor("valid")]; + tensor obj_179_strides_0 = const()[name = tensor("obj_179_strides_0"), val = tensor([1, 1])]; + tensor obj_179_pad_0 = const()[name = tensor("obj_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_179_dilations_0 = const()[name = tensor("obj_179_dilations_0"), val = tensor([1, 1])]; + tensor obj_179_groups_0 = const()[name = tensor("obj_179_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524928320)))]; + tensor layers_12_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527025536)))]; + tensor obj_179_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_bias_to_fp16, dilations = obj_179_dilations_0, groups = obj_179_groups_0, pad = obj_179_pad_0, pad_type = obj_179_pad_type_0, strides = obj_179_strides_0, weight = layers_12_encoder_attn_o_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("obj_179_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_179_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor out_77_axes_0 = const()[name = tensor("out_77_axes_0"), val = tensor([1])]; + tensor var_2960_to_fp16 = const()[name = tensor("op_2960_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2960_to_fp16, x = inputs_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor input_125_gamma_0_to_fp16 = const()[name = tensor("input_125_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527027648)))]; + tensor input_125_beta_0_to_fp16 = const()[name = tensor("input_125_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527029760)))]; + tensor input_125_epsilon_0_to_fp16 = const()[name = tensor("input_125_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = batch_norm(beta = input_125_beta_0_to_fp16, epsilon = input_125_epsilon_0_to_fp16, gamma = input_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor input_127_pad_type_0 = const()[name = tensor("input_127_pad_type_0"), val = tensor("valid")]; + tensor input_127_strides_0 = const()[name = tensor("input_127_strides_0"), val = tensor([1, 1])]; + tensor input_127_pad_0 = const()[name = tensor("input_127_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_127_dilations_0 = const()[name = tensor("input_127_dilations_0"), val = tensor([1, 1])]; + tensor input_127_groups_0 = const()[name = tensor("input_127_groups_0"), val = tensor(1)]; + tensor layers_12_fc1_weight_to_fp16 = const()[name = tensor("layers_12_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527031872)))]; + tensor layers_12_fc1_bias_to_fp16 = const()[name = tensor("layers_12_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535420544)))]; + tensor input_127_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = input_127_dilations_0, groups = input_127_groups_0, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = input_127_strides_0, weight = layers_12_fc1_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_mode_0 = const()[name = tensor("input_129_mode_0"), val = tensor("EXACT")]; + tensor input_129_cast_fp16 = gelu(mode = input_129_mode_0, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor hidden_states_27_pad_type_0 = const()[name = tensor("hidden_states_27_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_27_strides_0 = const()[name = tensor("hidden_states_27_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_27_pad_0 = const()[name = tensor("hidden_states_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_27_dilations_0 = const()[name = tensor("hidden_states_27_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_27_groups_0 = const()[name = tensor("hidden_states_27_groups_0"), val = tensor(1)]; + tensor layers_12_fc2_weight_to_fp16 = const()[name = tensor("layers_12_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535428800)))]; + tensor layers_12_fc2_bias_to_fp16 = const()[name = tensor("layers_12_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543817472)))]; + tensor hidden_states_27_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_12_fc2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("hidden_states_27_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor var_2995 = const()[name = tensor("op_2995"), val = tensor(3)]; + tensor out_79_axes_0 = const()[name = tensor("out_79_axes_0"), val = tensor([1])]; + tensor var_3020_to_fp16 = const()[name = tensor("op_3020_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_3020_to_fp16, x = inputs_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor obj_183_gamma_0_to_fp16 = const()[name = tensor("obj_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543819584)))]; + tensor obj_183_beta_0_to_fp16 = const()[name = tensor("obj_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543821696)))]; + tensor obj_183_epsilon_0_to_fp16 = const()[name = tensor("obj_183_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_183_cast_fp16 = batch_norm(beta = obj_183_beta_0_to_fp16, epsilon = obj_183_epsilon_0_to_fp16, gamma = obj_183_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("obj_183_cast_fp16")]; + tensor query_53_pad_type_0 = const()[name = tensor("query_53_pad_type_0"), val = tensor("valid")]; + tensor query_53_strides_0 = const()[name = tensor("query_53_strides_0"), val = tensor([1, 1])]; + tensor query_53_pad_0 = const()[name = tensor("query_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_53_dilations_0 = const()[name = tensor("query_53_dilations_0"), val = tensor([1, 1])]; + tensor query_53_groups_0 = const()[name = tensor("query_53_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543823808)))]; + tensor layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545921024)))]; + tensor query_53_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor current_key_27_pad_type_0 = const()[name = tensor("current_key_27_pad_type_0"), val = tensor("valid")]; + tensor current_key_27_strides_0 = const()[name = tensor("current_key_27_strides_0"), val = tensor([1, 1])]; + tensor current_key_27_pad_0 = const()[name = tensor("current_key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_27_dilations_0 = const()[name = tensor("current_key_27_dilations_0"), val = tensor([1, 1])]; + tensor current_key_27_groups_0 = const()[name = tensor("current_key_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545923136)))]; + tensor current_key_27_cast_fp16 = conv(dilations = current_key_27_dilations_0, groups = current_key_27_groups_0, pad = current_key_27_pad_0, pad_type = current_key_27_pad_type_0, strides = current_key_27_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_key_27_cast_fp16")]; + tensor current_value_27_pad_type_0 = const()[name = tensor("current_value_27_pad_type_0"), val = tensor("valid")]; + tensor current_value_27_strides_0 = const()[name = tensor("current_value_27_strides_0"), val = tensor([1, 1])]; + tensor current_value_27_pad_0 = const()[name = tensor("current_value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_27_dilations_0 = const()[name = tensor("current_value_27_dilations_0"), val = tensor([1, 1])]; + tensor current_value_27_groups_0 = const()[name = tensor("current_value_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(548020352)))]; + tensor layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550117568)))]; + tensor current_value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = current_value_27_dilations_0, groups = current_value_27_groups_0, pad = current_value_27_pad_0, pad_type = current_value_27_pad_type_0, strides = current_value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_value_27_cast_fp16")]; + tensor var_3059_cast_fp16 = mul(x = var_87_cast_fp16_13, y = var_207_cast_fp16)[name = tensor("op_3059_cast_fp16")]; + tensor var_3060_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3060_cast_fp16")]; + tensor key_53_cast_fp16 = add(x = var_3059_cast_fp16, y = var_3060_cast_fp16)[name = tensor("key_53_cast_fp16")]; + tensor var_3063_cast_fp16 = mul(x = var_114_cast_fp16_13, y = var_207_cast_fp16)[name = tensor("op_3063_cast_fp16")]; + tensor var_3064_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3064_cast_fp16")]; + tensor value_53_cast_fp16 = add(x = var_3063_cast_fp16, y = var_3064_cast_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_3068 = const()[name = tensor("op_3068"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_53_cast_fp16 = reshape(shape = var_3068, x = query_53_cast_fp16)[name = tensor("mh_q_53_cast_fp16")]; + tensor var_3070_to_fp16 = const()[name = tensor("op_3070_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3071_cast_fp16 = mul(x = mh_q_53_cast_fp16, y = var_3070_to_fp16)[name = tensor("op_3071_cast_fp16")]; + tensor var_3074 = const()[name = tensor("op_3074"), val = tensor([1, 16, 64, 448])]; + tensor var_3075_cast_fp16 = reshape(shape = var_3074, x = key_53_cast_fp16)[name = tensor("op_3075_cast_fp16")]; + tensor mh_w_79_transpose_x_0 = const()[name = tensor("mh_w_79_transpose_x_0"), val = tensor(true)]; + tensor mh_w_79_transpose_y_0 = const()[name = tensor("mh_w_79_transpose_y_0"), val = tensor(false)]; + tensor mh_w_79_cast_fp16 = matmul(transpose_x = mh_w_79_transpose_x_0, transpose_y = mh_w_79_transpose_y_0, x = var_3071_cast_fp16, y = var_3075_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; + tensor mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; + tensor var_3083_cast_fp16 = softmax(axis = var_2995, x = mh_w_81_cast_fp16)[name = tensor("op_3083_cast_fp16")]; + tensor var_3084 = const()[name = tensor("op_3084"), val = tensor([1, 16, 64, 448])]; + tensor var_3085_cast_fp16 = reshape(shape = var_3084, x = value_53_cast_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor attn_53_transpose_x_0 = const()[name = tensor("attn_53_transpose_x_0"), val = tensor(false)]; + tensor attn_53_transpose_y_0 = const()[name = tensor("attn_53_transpose_y_0"), val = tensor(true)]; + tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3085_cast_fp16, y = var_3083_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_3088 = const()[name = tensor("op_3088"), val = tensor([1, 1024, 1, 1])]; + tensor input_131_cast_fp16 = reshape(shape = var_3088, x = attn_53_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor obj_189_pad_type_0 = const()[name = tensor("obj_189_pad_type_0"), val = tensor("valid")]; + tensor obj_189_strides_0 = const()[name = tensor("obj_189_strides_0"), val = tensor([1, 1])]; + tensor obj_189_pad_0 = const()[name = tensor("obj_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_189_dilations_0 = const()[name = tensor("obj_189_dilations_0"), val = tensor([1, 1])]; + tensor obj_189_groups_0 = const()[name = tensor("obj_189_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550119680)))]; + tensor layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552216896)))]; + tensor obj_189_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = obj_189_dilations_0, groups = obj_189_groups_0, pad = obj_189_pad_0, pad_type = obj_189_pad_type_0, strides = obj_189_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("obj_189_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_189_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor out_81_axes_0 = const()[name = tensor("out_81_axes_0"), val = tensor([1])]; + tensor var_3110_to_fp16 = const()[name = tensor("op_3110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_3110_to_fp16, x = inputs_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor obj_191_gamma_0_to_fp16 = const()[name = tensor("obj_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552219008)))]; + tensor obj_191_beta_0_to_fp16 = const()[name = tensor("obj_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552221120)))]; + tensor obj_191_epsilon_0_to_fp16 = const()[name = tensor("obj_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_191_cast_fp16 = batch_norm(beta = obj_191_beta_0_to_fp16, epsilon = obj_191_epsilon_0_to_fp16, gamma = obj_191_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("obj_191_cast_fp16")]; + tensor query_55_pad_type_0 = const()[name = tensor("query_55_pad_type_0"), val = tensor("valid")]; + tensor query_55_strides_0 = const()[name = tensor("query_55_strides_0"), val = tensor([1, 1])]; + tensor query_55_pad_0 = const()[name = tensor("query_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_55_dilations_0 = const()[name = tensor("query_55_dilations_0"), val = tensor([1, 1])]; + tensor query_55_groups_0 = const()[name = tensor("query_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552223232)))]; + tensor layers_13_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554320448)))]; + tensor query_55_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_bias_to_fp16, dilations = query_55_dilations_0, groups = query_55_groups_0, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = query_55_strides_0, weight = layers_13_encoder_attn_q_proj_weight_to_fp16, x = obj_191_cast_fp16)[name = tensor("query_55_cast_fp16")]; + tensor key_55_pad_type_0 = const()[name = tensor("key_55_pad_type_0"), val = tensor("valid")]; + tensor key_55_strides_0 = const()[name = tensor("key_55_strides_0"), val = tensor([1, 1])]; + tensor key_55_pad_0 = const()[name = tensor("key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_55_dilations_0 = const()[name = tensor("key_55_dilations_0"), val = tensor([1, 1])]; + tensor key_55_groups_0 = const()[name = tensor("key_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554322560)))]; + tensor key_55_cast_fp16 = conv(dilations = key_55_dilations_0, groups = key_55_groups_0, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = key_55_strides_0, weight = layers_13_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_55_cast_fp16")]; + tensor value_55_pad_type_0 = const()[name = tensor("value_55_pad_type_0"), val = tensor("valid")]; + tensor value_55_strides_0 = const()[name = tensor("value_55_strides_0"), val = tensor([1, 1])]; + tensor value_55_pad_0 = const()[name = tensor("value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_55_dilations_0 = const()[name = tensor("value_55_dilations_0"), val = tensor([1, 1])]; + tensor value_55_groups_0 = const()[name = tensor("value_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556419776)))]; + tensor layers_13_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(558516992)))]; + tensor value_55_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_bias_to_fp16, dilations = value_55_dilations_0, groups = value_55_groups_0, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = value_55_strides_0, weight = layers_13_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_55_cast_fp16")]; + tensor var_3146 = const()[name = tensor("op_3146"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_55_cast_fp16 = reshape(shape = var_3146, x = query_55_cast_fp16)[name = tensor("mh_q_55_cast_fp16")]; + tensor var_3148_to_fp16 = const()[name = tensor("op_3148_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3149_cast_fp16 = mul(x = mh_q_55_cast_fp16, y = var_3148_to_fp16)[name = tensor("op_3149_cast_fp16")]; + tensor var_3152 = const()[name = tensor("op_3152"), val = tensor([1, 16, 64, 1500])]; + tensor var_3153_cast_fp16 = reshape(shape = var_3152, x = key_55_cast_fp16)[name = tensor("op_3153_cast_fp16")]; + tensor mh_w_83_transpose_x_0 = const()[name = tensor("mh_w_83_transpose_x_0"), val = tensor(true)]; + tensor mh_w_83_transpose_y_0 = const()[name = tensor("mh_w_83_transpose_y_0"), val = tensor(false)]; + tensor mh_w_83_cast_fp16 = matmul(transpose_x = mh_w_83_transpose_x_0, transpose_y = mh_w_83_transpose_y_0, x = var_3149_cast_fp16, y = var_3153_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; + tensor obj_195_cast_fp16 = softmax(axis = var_2995, x = mh_w_83_cast_fp16)[name = tensor("obj_195_cast_fp16")]; + tensor var_3157 = const()[name = tensor("op_3157"), val = tensor([1, 16, 64, 1500])]; + tensor var_3158_cast_fp16 = reshape(shape = var_3157, x = value_55_cast_fp16)[name = tensor("op_3158_cast_fp16")]; + tensor attn_55_transpose_x_0 = const()[name = tensor("attn_55_transpose_x_0"), val = tensor(false)]; + tensor attn_55_transpose_y_0 = const()[name = tensor("attn_55_transpose_y_0"), val = tensor(true)]; + tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3158_cast_fp16, y = obj_195_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_3161 = const()[name = tensor("op_3161"), val = tensor([1, 1024, 1, 1])]; + tensor input_133_cast_fp16 = reshape(shape = var_3161, x = attn_55_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor obj_193_pad_type_0 = const()[name = tensor("obj_193_pad_type_0"), val = tensor("valid")]; + tensor obj_193_strides_0 = const()[name = tensor("obj_193_strides_0"), val = tensor([1, 1])]; + tensor obj_193_pad_0 = const()[name = tensor("obj_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_193_dilations_0 = const()[name = tensor("obj_193_dilations_0"), val = tensor([1, 1])]; + tensor obj_193_groups_0 = const()[name = tensor("obj_193_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(558519104)))]; + tensor layers_13_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560616320)))]; + tensor obj_193_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_bias_to_fp16, dilations = obj_193_dilations_0, groups = obj_193_groups_0, pad = obj_193_pad_0, pad_type = obj_193_pad_type_0, strides = obj_193_strides_0, weight = layers_13_encoder_attn_o_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("obj_193_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_193_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor out_83_axes_0 = const()[name = tensor("out_83_axes_0"), val = tensor([1])]; + tensor var_3179_to_fp16 = const()[name = tensor("op_3179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_3179_to_fp16, x = inputs_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor input_135_gamma_0_to_fp16 = const()[name = tensor("input_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560618432)))]; + tensor input_135_beta_0_to_fp16 = const()[name = tensor("input_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560620544)))]; + tensor input_135_epsilon_0_to_fp16 = const()[name = tensor("input_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = batch_norm(beta = input_135_beta_0_to_fp16, epsilon = input_135_epsilon_0_to_fp16, gamma = input_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_pad_type_0 = const()[name = tensor("input_137_pad_type_0"), val = tensor("valid")]; + tensor input_137_strides_0 = const()[name = tensor("input_137_strides_0"), val = tensor([1, 1])]; + tensor input_137_pad_0 = const()[name = tensor("input_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_137_dilations_0 = const()[name = tensor("input_137_dilations_0"), val = tensor([1, 1])]; + tensor input_137_groups_0 = const()[name = tensor("input_137_groups_0"), val = tensor(1)]; + tensor layers_13_fc1_weight_to_fp16 = const()[name = tensor("layers_13_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560622656)))]; + tensor layers_13_fc1_bias_to_fp16 = const()[name = tensor("layers_13_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569011328)))]; + tensor input_137_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = input_137_dilations_0, groups = input_137_groups_0, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = input_137_strides_0, weight = layers_13_fc1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_mode_0 = const()[name = tensor("input_139_mode_0"), val = tensor("EXACT")]; + tensor input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor hidden_states_29_pad_type_0 = const()[name = tensor("hidden_states_29_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_29_strides_0 = const()[name = tensor("hidden_states_29_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_29_pad_0 = const()[name = tensor("hidden_states_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_29_dilations_0 = const()[name = tensor("hidden_states_29_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_29_groups_0 = const()[name = tensor("hidden_states_29_groups_0"), val = tensor(1)]; + tensor layers_13_fc2_weight_to_fp16 = const()[name = tensor("layers_13_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569019584)))]; + tensor layers_13_fc2_bias_to_fp16 = const()[name = tensor("layers_13_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577408256)))]; + tensor hidden_states_29_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = layers_13_fc2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("hidden_states_29_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_3214 = const()[name = tensor("op_3214"), val = tensor(3)]; + tensor out_85_axes_0 = const()[name = tensor("out_85_axes_0"), val = tensor([1])]; + tensor var_3239_to_fp16 = const()[name = tensor("op_3239_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_3239_to_fp16, x = inputs_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor obj_197_gamma_0_to_fp16 = const()[name = tensor("obj_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577410368)))]; + tensor obj_197_beta_0_to_fp16 = const()[name = tensor("obj_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577412480)))]; + tensor obj_197_epsilon_0_to_fp16 = const()[name = tensor("obj_197_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_197_cast_fp16 = batch_norm(beta = obj_197_beta_0_to_fp16, epsilon = obj_197_epsilon_0_to_fp16, gamma = obj_197_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("obj_197_cast_fp16")]; + tensor query_57_pad_type_0 = const()[name = tensor("query_57_pad_type_0"), val = tensor("valid")]; + tensor query_57_strides_0 = const()[name = tensor("query_57_strides_0"), val = tensor([1, 1])]; + tensor query_57_pad_0 = const()[name = tensor("query_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_57_dilations_0 = const()[name = tensor("query_57_dilations_0"), val = tensor([1, 1])]; + tensor query_57_groups_0 = const()[name = tensor("query_57_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577414592)))]; + tensor layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579511808)))]; + tensor query_57_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor current_key_29_pad_type_0 = const()[name = tensor("current_key_29_pad_type_0"), val = tensor("valid")]; + tensor current_key_29_strides_0 = const()[name = tensor("current_key_29_strides_0"), val = tensor([1, 1])]; + tensor current_key_29_pad_0 = const()[name = tensor("current_key_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_29_dilations_0 = const()[name = tensor("current_key_29_dilations_0"), val = tensor([1, 1])]; + tensor current_key_29_groups_0 = const()[name = tensor("current_key_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579513920)))]; + tensor current_key_29_cast_fp16 = conv(dilations = current_key_29_dilations_0, groups = current_key_29_groups_0, pad = current_key_29_pad_0, pad_type = current_key_29_pad_type_0, strides = current_key_29_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_key_29_cast_fp16")]; + tensor current_value_29_pad_type_0 = const()[name = tensor("current_value_29_pad_type_0"), val = tensor("valid")]; + tensor current_value_29_strides_0 = const()[name = tensor("current_value_29_strides_0"), val = tensor([1, 1])]; + tensor current_value_29_pad_0 = const()[name = tensor("current_value_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_29_dilations_0 = const()[name = tensor("current_value_29_dilations_0"), val = tensor([1, 1])]; + tensor current_value_29_groups_0 = const()[name = tensor("current_value_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581611136)))]; + tensor layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583708352)))]; + tensor current_value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = current_value_29_dilations_0, groups = current_value_29_groups_0, pad = current_value_29_pad_0, pad_type = current_value_29_pad_type_0, strides = current_value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_value_29_cast_fp16")]; + tensor var_3278_cast_fp16 = mul(x = var_87_cast_fp16_14, y = var_207_cast_fp16)[name = tensor("op_3278_cast_fp16")]; + tensor var_3279_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3279_cast_fp16")]; + tensor key_57_cast_fp16 = add(x = var_3278_cast_fp16, y = var_3279_cast_fp16)[name = tensor("key_57_cast_fp16")]; + tensor var_3282_cast_fp16 = mul(x = var_114_cast_fp16_14, y = var_207_cast_fp16)[name = tensor("op_3282_cast_fp16")]; + tensor var_3283_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3283_cast_fp16")]; + tensor value_57_cast_fp16 = add(x = var_3282_cast_fp16, y = var_3283_cast_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_3287 = const()[name = tensor("op_3287"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_57_cast_fp16 = reshape(shape = var_3287, x = query_57_cast_fp16)[name = tensor("mh_q_57_cast_fp16")]; + tensor var_3289_to_fp16 = const()[name = tensor("op_3289_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3290_cast_fp16 = mul(x = mh_q_57_cast_fp16, y = var_3289_to_fp16)[name = tensor("op_3290_cast_fp16")]; + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([1, 16, 64, 448])]; + tensor var_3294_cast_fp16 = reshape(shape = var_3293, x = key_57_cast_fp16)[name = tensor("op_3294_cast_fp16")]; + tensor mh_w_85_transpose_x_0 = const()[name = tensor("mh_w_85_transpose_x_0"), val = tensor(true)]; + tensor mh_w_85_transpose_y_0 = const()[name = tensor("mh_w_85_transpose_y_0"), val = tensor(false)]; + tensor mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_3290_cast_fp16, y = var_3294_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; + tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; + tensor var_3302_cast_fp16 = softmax(axis = var_3214, x = mh_w_87_cast_fp16)[name = tensor("op_3302_cast_fp16")]; + tensor var_3303 = const()[name = tensor("op_3303"), val = tensor([1, 16, 64, 448])]; + tensor var_3304_cast_fp16 = reshape(shape = var_3303, x = value_57_cast_fp16)[name = tensor("op_3304_cast_fp16")]; + tensor attn_57_transpose_x_0 = const()[name = tensor("attn_57_transpose_x_0"), val = tensor(false)]; + tensor attn_57_transpose_y_0 = const()[name = tensor("attn_57_transpose_y_0"), val = tensor(true)]; + tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3304_cast_fp16, y = var_3302_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_3307 = const()[name = tensor("op_3307"), val = tensor([1, 1024, 1, 1])]; + tensor input_141_cast_fp16 = reshape(shape = var_3307, x = attn_57_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor obj_203_pad_type_0 = const()[name = tensor("obj_203_pad_type_0"), val = tensor("valid")]; + tensor obj_203_strides_0 = const()[name = tensor("obj_203_strides_0"), val = tensor([1, 1])]; + tensor obj_203_pad_0 = const()[name = tensor("obj_203_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_203_dilations_0 = const()[name = tensor("obj_203_dilations_0"), val = tensor([1, 1])]; + tensor obj_203_groups_0 = const()[name = tensor("obj_203_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583710464)))]; + tensor layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585807680)))]; + tensor obj_203_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = obj_203_dilations_0, groups = obj_203_groups_0, pad = obj_203_pad_0, pad_type = obj_203_pad_type_0, strides = obj_203_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("obj_203_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_203_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor out_87_axes_0 = const()[name = tensor("out_87_axes_0"), val = tensor([1])]; + tensor var_3329_to_fp16 = const()[name = tensor("op_3329_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_3329_to_fp16, x = inputs_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor obj_205_gamma_0_to_fp16 = const()[name = tensor("obj_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585809792)))]; + tensor obj_205_beta_0_to_fp16 = const()[name = tensor("obj_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585811904)))]; + tensor obj_205_epsilon_0_to_fp16 = const()[name = tensor("obj_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_205_cast_fp16 = batch_norm(beta = obj_205_beta_0_to_fp16, epsilon = obj_205_epsilon_0_to_fp16, gamma = obj_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("obj_205_cast_fp16")]; + tensor query_59_pad_type_0 = const()[name = tensor("query_59_pad_type_0"), val = tensor("valid")]; + tensor query_59_strides_0 = const()[name = tensor("query_59_strides_0"), val = tensor([1, 1])]; + tensor query_59_pad_0 = const()[name = tensor("query_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_59_dilations_0 = const()[name = tensor("query_59_dilations_0"), val = tensor([1, 1])]; + tensor query_59_groups_0 = const()[name = tensor("query_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585814016)))]; + tensor layers_14_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587911232)))]; + tensor query_59_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_bias_to_fp16, dilations = query_59_dilations_0, groups = query_59_groups_0, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = query_59_strides_0, weight = layers_14_encoder_attn_q_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("query_59_cast_fp16")]; + tensor key_59_pad_type_0 = const()[name = tensor("key_59_pad_type_0"), val = tensor("valid")]; + tensor key_59_strides_0 = const()[name = tensor("key_59_strides_0"), val = tensor([1, 1])]; + tensor key_59_pad_0 = const()[name = tensor("key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_59_dilations_0 = const()[name = tensor("key_59_dilations_0"), val = tensor([1, 1])]; + tensor key_59_groups_0 = const()[name = tensor("key_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587913344)))]; + tensor key_59_cast_fp16 = conv(dilations = key_59_dilations_0, groups = key_59_groups_0, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = key_59_strides_0, weight = layers_14_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_59_cast_fp16")]; + tensor value_59_pad_type_0 = const()[name = tensor("value_59_pad_type_0"), val = tensor("valid")]; + tensor value_59_strides_0 = const()[name = tensor("value_59_strides_0"), val = tensor([1, 1])]; + tensor value_59_pad_0 = const()[name = tensor("value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_59_dilations_0 = const()[name = tensor("value_59_dilations_0"), val = tensor([1, 1])]; + tensor value_59_groups_0 = const()[name = tensor("value_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(590010560)))]; + tensor layers_14_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592107776)))]; + tensor value_59_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_bias_to_fp16, dilations = value_59_dilations_0, groups = value_59_groups_0, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = value_59_strides_0, weight = layers_14_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_59_cast_fp16")]; + tensor var_3365 = const()[name = tensor("op_3365"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_59_cast_fp16 = reshape(shape = var_3365, x = query_59_cast_fp16)[name = tensor("mh_q_59_cast_fp16")]; + tensor var_3367_to_fp16 = const()[name = tensor("op_3367_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3368_cast_fp16 = mul(x = mh_q_59_cast_fp16, y = var_3367_to_fp16)[name = tensor("op_3368_cast_fp16")]; + tensor var_3371 = const()[name = tensor("op_3371"), val = tensor([1, 16, 64, 1500])]; + tensor var_3372_cast_fp16 = reshape(shape = var_3371, x = key_59_cast_fp16)[name = tensor("op_3372_cast_fp16")]; + tensor mh_w_89_transpose_x_0 = const()[name = tensor("mh_w_89_transpose_x_0"), val = tensor(true)]; + tensor mh_w_89_transpose_y_0 = const()[name = tensor("mh_w_89_transpose_y_0"), val = tensor(false)]; + tensor mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_3368_cast_fp16, y = var_3372_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; + tensor obj_209_cast_fp16 = softmax(axis = var_3214, x = mh_w_89_cast_fp16)[name = tensor("obj_209_cast_fp16")]; + tensor var_3376 = const()[name = tensor("op_3376"), val = tensor([1, 16, 64, 1500])]; + tensor var_3377_cast_fp16 = reshape(shape = var_3376, x = value_59_cast_fp16)[name = tensor("op_3377_cast_fp16")]; + tensor attn_59_transpose_x_0 = const()[name = tensor("attn_59_transpose_x_0"), val = tensor(false)]; + tensor attn_59_transpose_y_0 = const()[name = tensor("attn_59_transpose_y_0"), val = tensor(true)]; + tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3377_cast_fp16, y = obj_209_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_3380 = const()[name = tensor("op_3380"), val = tensor([1, 1024, 1, 1])]; + tensor input_143_cast_fp16 = reshape(shape = var_3380, x = attn_59_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor obj_207_pad_type_0 = const()[name = tensor("obj_207_pad_type_0"), val = tensor("valid")]; + tensor obj_207_strides_0 = const()[name = tensor("obj_207_strides_0"), val = tensor([1, 1])]; + tensor obj_207_pad_0 = const()[name = tensor("obj_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_207_dilations_0 = const()[name = tensor("obj_207_dilations_0"), val = tensor([1, 1])]; + tensor obj_207_groups_0 = const()[name = tensor("obj_207_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592109888)))]; + tensor layers_14_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594207104)))]; + tensor obj_207_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_bias_to_fp16, dilations = obj_207_dilations_0, groups = obj_207_groups_0, pad = obj_207_pad_0, pad_type = obj_207_pad_type_0, strides = obj_207_strides_0, weight = layers_14_encoder_attn_o_proj_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("obj_207_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_207_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor out_89_axes_0 = const()[name = tensor("out_89_axes_0"), val = tensor([1])]; + tensor var_3401_to_fp16 = const()[name = tensor("op_3401_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_3401_to_fp16, x = inputs_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor input_145_gamma_0_to_fp16 = const()[name = tensor("input_145_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594209216)))]; + tensor input_145_beta_0_to_fp16 = const()[name = tensor("input_145_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594211328)))]; + tensor input_145_epsilon_0_to_fp16 = const()[name = tensor("input_145_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_145_cast_fp16 = batch_norm(beta = input_145_beta_0_to_fp16, epsilon = input_145_epsilon_0_to_fp16, gamma = input_145_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_pad_type_0 = const()[name = tensor("input_147_pad_type_0"), val = tensor("valid")]; + tensor input_147_strides_0 = const()[name = tensor("input_147_strides_0"), val = tensor([1, 1])]; + tensor input_147_pad_0 = const()[name = tensor("input_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_147_dilations_0 = const()[name = tensor("input_147_dilations_0"), val = tensor([1, 1])]; + tensor input_147_groups_0 = const()[name = tensor("input_147_groups_0"), val = tensor(1)]; + tensor layers_14_fc1_weight_to_fp16 = const()[name = tensor("layers_14_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594213440)))]; + tensor layers_14_fc1_bias_to_fp16 = const()[name = tensor("layers_14_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(602602112)))]; + tensor input_147_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = layers_14_fc1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_mode_0 = const()[name = tensor("input_149_mode_0"), val = tensor("EXACT")]; + tensor input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor hidden_states_31_pad_type_0 = const()[name = tensor("hidden_states_31_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_31_strides_0 = const()[name = tensor("hidden_states_31_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_31_pad_0 = const()[name = tensor("hidden_states_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_31_dilations_0 = const()[name = tensor("hidden_states_31_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_31_groups_0 = const()[name = tensor("hidden_states_31_groups_0"), val = tensor(1)]; + tensor layers_14_fc2_weight_to_fp16 = const()[name = tensor("layers_14_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(602610368)))]; + tensor layers_14_fc2_bias_to_fp16 = const()[name = tensor("layers_14_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(610999040)))]; + tensor hidden_states_31_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = hidden_states_31_dilations_0, groups = hidden_states_31_groups_0, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = hidden_states_31_strides_0, weight = layers_14_fc2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor(3)]; + tensor out_91_axes_0 = const()[name = tensor("out_91_axes_0"), val = tensor([1])]; + tensor var_3462_to_fp16 = const()[name = tensor("op_3462_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_3462_to_fp16, x = inputs_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor obj_211_gamma_0_to_fp16 = const()[name = tensor("obj_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611001152)))]; + tensor obj_211_beta_0_to_fp16 = const()[name = tensor("obj_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611003264)))]; + tensor obj_211_epsilon_0_to_fp16 = const()[name = tensor("obj_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_211_cast_fp16 = batch_norm(beta = obj_211_beta_0_to_fp16, epsilon = obj_211_epsilon_0_to_fp16, gamma = obj_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("obj_211_cast_fp16")]; + tensor query_61_pad_type_0 = const()[name = tensor("query_61_pad_type_0"), val = tensor("valid")]; + tensor query_61_strides_0 = const()[name = tensor("query_61_strides_0"), val = tensor([1, 1])]; + tensor query_61_pad_0 = const()[name = tensor("query_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_61_dilations_0 = const()[name = tensor("query_61_dilations_0"), val = tensor([1, 1])]; + tensor query_61_groups_0 = const()[name = tensor("query_61_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611005376)))]; + tensor layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(613102592)))]; + tensor query_61_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor current_key_31_pad_type_0 = const()[name = tensor("current_key_31_pad_type_0"), val = tensor("valid")]; + tensor current_key_31_strides_0 = const()[name = tensor("current_key_31_strides_0"), val = tensor([1, 1])]; + tensor current_key_31_pad_0 = const()[name = tensor("current_key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_31_dilations_0 = const()[name = tensor("current_key_31_dilations_0"), val = tensor([1, 1])]; + tensor current_key_31_groups_0 = const()[name = tensor("current_key_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(613104704)))]; + tensor current_key_31_cast_fp16 = conv(dilations = current_key_31_dilations_0, groups = current_key_31_groups_0, pad = current_key_31_pad_0, pad_type = current_key_31_pad_type_0, strides = current_key_31_strides_0, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_key_31_cast_fp16")]; + tensor current_value_31_pad_type_0 = const()[name = tensor("current_value_31_pad_type_0"), val = tensor("valid")]; + tensor current_value_31_strides_0 = const()[name = tensor("current_value_31_strides_0"), val = tensor([1, 1])]; + tensor current_value_31_pad_0 = const()[name = tensor("current_value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_31_dilations_0 = const()[name = tensor("current_value_31_dilations_0"), val = tensor([1, 1])]; + tensor current_value_31_groups_0 = const()[name = tensor("current_value_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(615201920)))]; + tensor layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617299136)))]; + tensor current_value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = current_value_31_dilations_0, groups = current_value_31_groups_0, pad = current_value_31_pad_0, pad_type = current_value_31_pad_type_0, strides = current_value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_value_31_cast_fp16")]; + tensor var_3501_cast_fp16 = mul(x = var_87_cast_fp16_15, y = var_207_cast_fp16)[name = tensor("op_3501_cast_fp16")]; + tensor var_3502_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3502_cast_fp16")]; + tensor key_61_cast_fp16 = add(x = var_3501_cast_fp16, y = var_3502_cast_fp16)[name = tensor("key_61_cast_fp16")]; + tensor var_3505_cast_fp16 = mul(x = var_114_cast_fp16_15, y = var_207_cast_fp16)[name = tensor("op_3505_cast_fp16")]; + tensor var_3506_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3506_cast_fp16")]; + tensor value_61_cast_fp16 = add(x = var_3505_cast_fp16, y = var_3506_cast_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_3510 = const()[name = tensor("op_3510"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_61_cast_fp16 = reshape(shape = var_3510, x = query_61_cast_fp16)[name = tensor("mh_q_61_cast_fp16")]; + tensor var_3512_to_fp16 = const()[name = tensor("op_3512_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3513_cast_fp16 = mul(x = mh_q_61_cast_fp16, y = var_3512_to_fp16)[name = tensor("op_3513_cast_fp16")]; + tensor var_3516 = const()[name = tensor("op_3516"), val = tensor([1, 16, 64, 448])]; + tensor var_3517_cast_fp16 = reshape(shape = var_3516, x = key_61_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor mh_w_91_transpose_x_0 = const()[name = tensor("mh_w_91_transpose_x_0"), val = tensor(true)]; + tensor mh_w_91_transpose_y_0 = const()[name = tensor("mh_w_91_transpose_y_0"), val = tensor(false)]; + tensor mh_w_91_cast_fp16 = matmul(transpose_x = mh_w_91_transpose_x_0, transpose_y = mh_w_91_transpose_y_0, x = var_3513_cast_fp16, y = var_3517_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; + tensor mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; + tensor var_3525_cast_fp16 = softmax(axis = var_3437, x = mh_w_93_cast_fp16)[name = tensor("op_3525_cast_fp16")]; + tensor var_3526 = const()[name = tensor("op_3526"), val = tensor([1, 16, 64, 448])]; + tensor var_3527_cast_fp16 = reshape(shape = var_3526, x = value_61_cast_fp16)[name = tensor("op_3527_cast_fp16")]; + tensor attn_61_transpose_x_0 = const()[name = tensor("attn_61_transpose_x_0"), val = tensor(false)]; + tensor attn_61_transpose_y_0 = const()[name = tensor("attn_61_transpose_y_0"), val = tensor(true)]; + tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3527_cast_fp16, y = var_3525_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_3530 = const()[name = tensor("op_3530"), val = tensor([1, 1024, 1, 1])]; + tensor input_151_cast_fp16 = reshape(shape = var_3530, x = attn_61_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor obj_217_pad_type_0 = const()[name = tensor("obj_217_pad_type_0"), val = tensor("valid")]; + tensor obj_217_strides_0 = const()[name = tensor("obj_217_strides_0"), val = tensor([1, 1])]; + tensor obj_217_pad_0 = const()[name = tensor("obj_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_217_dilations_0 = const()[name = tensor("obj_217_dilations_0"), val = tensor([1, 1])]; + tensor obj_217_groups_0 = const()[name = tensor("obj_217_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617301248)))]; + tensor layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619398464)))]; + tensor obj_217_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = obj_217_dilations_0, groups = obj_217_groups_0, pad = obj_217_pad_0, pad_type = obj_217_pad_type_0, strides = obj_217_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("obj_217_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_217_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor out_93_axes_0 = const()[name = tensor("out_93_axes_0"), val = tensor([1])]; + tensor var_3552_to_fp16 = const()[name = tensor("op_3552_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_3552_to_fp16, x = inputs_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_219_gamma_0_to_fp16 = const()[name = tensor("obj_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619400576)))]; + tensor obj_219_beta_0_to_fp16 = const()[name = tensor("obj_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619402688)))]; + tensor obj_219_epsilon_0_to_fp16 = const()[name = tensor("obj_219_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_219_cast_fp16 = batch_norm(beta = obj_219_beta_0_to_fp16, epsilon = obj_219_epsilon_0_to_fp16, gamma = obj_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_219_cast_fp16")]; + tensor query_63_pad_type_0 = const()[name = tensor("query_63_pad_type_0"), val = tensor("valid")]; + tensor query_63_strides_0 = const()[name = tensor("query_63_strides_0"), val = tensor([1, 1])]; + tensor query_63_pad_0 = const()[name = tensor("query_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_63_dilations_0 = const()[name = tensor("query_63_dilations_0"), val = tensor([1, 1])]; + tensor query_63_groups_0 = const()[name = tensor("query_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619404800)))]; + tensor layers_15_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(621502016)))]; + tensor query_63_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_bias_to_fp16, dilations = query_63_dilations_0, groups = query_63_groups_0, pad = query_63_pad_0, pad_type = query_63_pad_type_0, strides = query_63_strides_0, weight = layers_15_encoder_attn_q_proj_weight_to_fp16, x = obj_219_cast_fp16)[name = tensor("query_63_cast_fp16")]; + tensor key_63_pad_type_0 = const()[name = tensor("key_63_pad_type_0"), val = tensor("valid")]; + tensor key_63_strides_0 = const()[name = tensor("key_63_strides_0"), val = tensor([1, 1])]; + tensor key_63_pad_0 = const()[name = tensor("key_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_63_dilations_0 = const()[name = tensor("key_63_dilations_0"), val = tensor([1, 1])]; + tensor key_63_groups_0 = const()[name = tensor("key_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(621504128)))]; + tensor key_63_cast_fp16 = conv(dilations = key_63_dilations_0, groups = key_63_groups_0, pad = key_63_pad_0, pad_type = key_63_pad_type_0, strides = key_63_strides_0, weight = layers_15_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_63_cast_fp16")]; + tensor value_63_pad_type_0 = const()[name = tensor("value_63_pad_type_0"), val = tensor("valid")]; + tensor value_63_strides_0 = const()[name = tensor("value_63_strides_0"), val = tensor([1, 1])]; + tensor value_63_pad_0 = const()[name = tensor("value_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_63_dilations_0 = const()[name = tensor("value_63_dilations_0"), val = tensor([1, 1])]; + tensor value_63_groups_0 = const()[name = tensor("value_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(623601344)))]; + tensor layers_15_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(625698560)))]; + tensor value_63_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_bias_to_fp16, dilations = value_63_dilations_0, groups = value_63_groups_0, pad = value_63_pad_0, pad_type = value_63_pad_type_0, strides = value_63_strides_0, weight = layers_15_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_63_cast_fp16")]; + tensor var_3588 = const()[name = tensor("op_3588"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_63_cast_fp16 = reshape(shape = var_3588, x = query_63_cast_fp16)[name = tensor("mh_q_63_cast_fp16")]; + tensor var_3590_to_fp16 = const()[name = tensor("op_3590_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3591_cast_fp16 = mul(x = mh_q_63_cast_fp16, y = var_3590_to_fp16)[name = tensor("op_3591_cast_fp16")]; + tensor var_3594 = const()[name = tensor("op_3594"), val = tensor([1, 16, 64, 1500])]; + tensor var_3595_cast_fp16 = reshape(shape = var_3594, x = key_63_cast_fp16)[name = tensor("op_3595_cast_fp16")]; + tensor mh_w_95_transpose_x_0 = const()[name = tensor("mh_w_95_transpose_x_0"), val = tensor(true)]; + tensor mh_w_95_transpose_y_0 = const()[name = tensor("mh_w_95_transpose_y_0"), val = tensor(false)]; + tensor mh_w_95_cast_fp16 = matmul(transpose_x = mh_w_95_transpose_x_0, transpose_y = mh_w_95_transpose_y_0, x = var_3591_cast_fp16, y = var_3595_cast_fp16)[name = tensor("mh_w_95_cast_fp16")]; + tensor obj_223_cast_fp16 = softmax(axis = var_3437, x = mh_w_95_cast_fp16)[name = tensor("obj_223_cast_fp16")]; + tensor var_3599 = const()[name = tensor("op_3599"), val = tensor([1, 16, 64, 1500])]; + tensor var_3600_cast_fp16 = reshape(shape = var_3599, x = value_63_cast_fp16)[name = tensor("op_3600_cast_fp16")]; + tensor attn_63_transpose_x_0 = const()[name = tensor("attn_63_transpose_x_0"), val = tensor(false)]; + tensor attn_63_transpose_y_0 = const()[name = tensor("attn_63_transpose_y_0"), val = tensor(true)]; + tensor attn_63_cast_fp16 = matmul(transpose_x = attn_63_transpose_x_0, transpose_y = attn_63_transpose_y_0, x = var_3600_cast_fp16, y = obj_223_cast_fp16)[name = tensor("attn_63_cast_fp16")]; + tensor var_3603 = const()[name = tensor("op_3603"), val = tensor([1, 1024, 1, 1])]; + tensor input_153_cast_fp16 = reshape(shape = var_3603, x = attn_63_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor obj_221_pad_type_0 = const()[name = tensor("obj_221_pad_type_0"), val = tensor("valid")]; + tensor obj_221_strides_0 = const()[name = tensor("obj_221_strides_0"), val = tensor([1, 1])]; + tensor obj_221_pad_0 = const()[name = tensor("obj_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_221_dilations_0 = const()[name = tensor("obj_221_dilations_0"), val = tensor([1, 1])]; + tensor obj_221_groups_0 = const()[name = tensor("obj_221_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(625700672)))]; + tensor layers_15_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627797888)))]; + tensor obj_221_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_bias_to_fp16, dilations = obj_221_dilations_0, groups = obj_221_groups_0, pad = obj_221_pad_0, pad_type = obj_221_pad_type_0, strides = obj_221_strides_0, weight = layers_15_encoder_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("obj_221_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_221_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor out_95_axes_0 = const()[name = tensor("out_95_axes_0"), val = tensor([1])]; + tensor var_3624_to_fp16 = const()[name = tensor("op_3624_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_3624_to_fp16, x = inputs_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_155_gamma_0_to_fp16 = const()[name = tensor("input_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627800000)))]; + tensor input_155_beta_0_to_fp16 = const()[name = tensor("input_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627802112)))]; + tensor input_155_epsilon_0_to_fp16 = const()[name = tensor("input_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor input_157_pad_type_0 = const()[name = tensor("input_157_pad_type_0"), val = tensor("valid")]; + tensor input_157_strides_0 = const()[name = tensor("input_157_strides_0"), val = tensor([1, 1])]; + tensor input_157_pad_0 = const()[name = tensor("input_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_157_dilations_0 = const()[name = tensor("input_157_dilations_0"), val = tensor([1, 1])]; + tensor input_157_groups_0 = const()[name = tensor("input_157_groups_0"), val = tensor(1)]; + tensor layers_15_fc1_weight_to_fp16 = const()[name = tensor("layers_15_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627804224)))]; + tensor layers_15_fc1_bias_to_fp16 = const()[name = tensor("layers_15_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(636192896)))]; + tensor input_157_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = layers_15_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_mode_0 = const()[name = tensor("input_159_mode_0"), val = tensor("EXACT")]; + tensor input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor hidden_states_33_pad_type_0 = const()[name = tensor("hidden_states_33_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_33_strides_0 = const()[name = tensor("hidden_states_33_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_33_pad_0 = const()[name = tensor("hidden_states_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_33_dilations_0 = const()[name = tensor("hidden_states_33_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_33_groups_0 = const()[name = tensor("hidden_states_33_groups_0"), val = tensor(1)]; + tensor layers_15_fc2_weight_to_fp16 = const()[name = tensor("layers_15_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(636201152)))]; + tensor layers_15_fc2_bias_to_fp16 = const()[name = tensor("layers_15_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644589824)))]; + tensor hidden_states_33_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_15_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("hidden_states_33_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_3660 = const()[name = tensor("op_3660"), val = tensor(3)]; + tensor out_97_axes_0 = const()[name = tensor("out_97_axes_0"), val = tensor([1])]; + tensor var_3685_to_fp16 = const()[name = tensor("op_3685_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_3685_to_fp16, x = inputs_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor obj_225_gamma_0_to_fp16 = const()[name = tensor("obj_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644591936)))]; + tensor obj_225_beta_0_to_fp16 = const()[name = tensor("obj_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644594048)))]; + tensor obj_225_epsilon_0_to_fp16 = const()[name = tensor("obj_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_225_cast_fp16 = batch_norm(beta = obj_225_beta_0_to_fp16, epsilon = obj_225_epsilon_0_to_fp16, gamma = obj_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("obj_225_cast_fp16")]; + tensor query_65_pad_type_0 = const()[name = tensor("query_65_pad_type_0"), val = tensor("valid")]; + tensor query_65_strides_0 = const()[name = tensor("query_65_strides_0"), val = tensor([1, 1])]; + tensor query_65_pad_0 = const()[name = tensor("query_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_65_dilations_0 = const()[name = tensor("query_65_dilations_0"), val = tensor([1, 1])]; + tensor query_65_groups_0 = const()[name = tensor("query_65_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644596160)))]; + tensor layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(646693376)))]; + tensor query_65_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = query_65_dilations_0, groups = query_65_groups_0, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = query_65_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("query_65_cast_fp16")]; + tensor current_key_33_pad_type_0 = const()[name = tensor("current_key_33_pad_type_0"), val = tensor("valid")]; + tensor current_key_33_strides_0 = const()[name = tensor("current_key_33_strides_0"), val = tensor([1, 1])]; + tensor current_key_33_pad_0 = const()[name = tensor("current_key_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_33_dilations_0 = const()[name = tensor("current_key_33_dilations_0"), val = tensor([1, 1])]; + tensor current_key_33_groups_0 = const()[name = tensor("current_key_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(646695488)))]; + tensor current_key_33_cast_fp16 = conv(dilations = current_key_33_dilations_0, groups = current_key_33_groups_0, pad = current_key_33_pad_0, pad_type = current_key_33_pad_type_0, strides = current_key_33_strides_0, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_key_33_cast_fp16")]; + tensor current_value_33_pad_type_0 = const()[name = tensor("current_value_33_pad_type_0"), val = tensor("valid")]; + tensor current_value_33_strides_0 = const()[name = tensor("current_value_33_strides_0"), val = tensor([1, 1])]; + tensor current_value_33_pad_0 = const()[name = tensor("current_value_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_33_dilations_0 = const()[name = tensor("current_value_33_dilations_0"), val = tensor([1, 1])]; + tensor current_value_33_groups_0 = const()[name = tensor("current_value_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(648792704)))]; + tensor layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650889920)))]; + tensor current_value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = current_value_33_dilations_0, groups = current_value_33_groups_0, pad = current_value_33_pad_0, pad_type = current_value_33_pad_type_0, strides = current_value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_value_33_cast_fp16")]; + tensor var_3724_cast_fp16 = mul(x = var_87_cast_fp16_16, y = var_207_cast_fp16)[name = tensor("op_3724_cast_fp16")]; + tensor var_3725_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3725_cast_fp16")]; + tensor key_65_cast_fp16 = add(x = var_3724_cast_fp16, y = var_3725_cast_fp16)[name = tensor("key_65_cast_fp16")]; + tensor var_3728_cast_fp16 = mul(x = var_114_cast_fp16_16, y = var_207_cast_fp16)[name = tensor("op_3728_cast_fp16")]; + tensor var_3729_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3729_cast_fp16")]; + tensor value_65_cast_fp16 = add(x = var_3728_cast_fp16, y = var_3729_cast_fp16)[name = tensor("value_65_cast_fp16")]; + tensor var_3733 = const()[name = tensor("op_3733"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_65_cast_fp16 = reshape(shape = var_3733, x = query_65_cast_fp16)[name = tensor("mh_q_65_cast_fp16")]; + tensor var_3735_to_fp16 = const()[name = tensor("op_3735_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3736_cast_fp16 = mul(x = mh_q_65_cast_fp16, y = var_3735_to_fp16)[name = tensor("op_3736_cast_fp16")]; + tensor var_3739 = const()[name = tensor("op_3739"), val = tensor([1, 16, 64, 448])]; + tensor var_3740_cast_fp16 = reshape(shape = var_3739, x = key_65_cast_fp16)[name = tensor("op_3740_cast_fp16")]; + tensor mh_w_97_transpose_x_0 = const()[name = tensor("mh_w_97_transpose_x_0"), val = tensor(true)]; + tensor mh_w_97_transpose_y_0 = const()[name = tensor("mh_w_97_transpose_y_0"), val = tensor(false)]; + tensor mh_w_97_cast_fp16 = matmul(transpose_x = mh_w_97_transpose_x_0, transpose_y = mh_w_97_transpose_y_0, x = var_3736_cast_fp16, y = var_3740_cast_fp16)[name = tensor("mh_w_97_cast_fp16")]; + tensor mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_99_cast_fp16")]; + tensor var_3748_cast_fp16 = softmax(axis = var_3660, x = mh_w_99_cast_fp16)[name = tensor("op_3748_cast_fp16")]; + tensor var_3749 = const()[name = tensor("op_3749"), val = tensor([1, 16, 64, 448])]; + tensor var_3750_cast_fp16 = reshape(shape = var_3749, x = value_65_cast_fp16)[name = tensor("op_3750_cast_fp16")]; + tensor attn_65_transpose_x_0 = const()[name = tensor("attn_65_transpose_x_0"), val = tensor(false)]; + tensor attn_65_transpose_y_0 = const()[name = tensor("attn_65_transpose_y_0"), val = tensor(true)]; + tensor attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_0, transpose_y = attn_65_transpose_y_0, x = var_3750_cast_fp16, y = var_3748_cast_fp16)[name = tensor("attn_65_cast_fp16")]; + tensor var_3753 = const()[name = tensor("op_3753"), val = tensor([1, 1024, 1, 1])]; + tensor input_161_cast_fp16 = reshape(shape = var_3753, x = attn_65_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor obj_231_pad_type_0 = const()[name = tensor("obj_231_pad_type_0"), val = tensor("valid")]; + tensor obj_231_strides_0 = const()[name = tensor("obj_231_strides_0"), val = tensor([1, 1])]; + tensor obj_231_pad_0 = const()[name = tensor("obj_231_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_231_dilations_0 = const()[name = tensor("obj_231_dilations_0"), val = tensor([1, 1])]; + tensor obj_231_groups_0 = const()[name = tensor("obj_231_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650892032)))]; + tensor layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652989248)))]; + tensor obj_231_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = obj_231_dilations_0, groups = obj_231_groups_0, pad = obj_231_pad_0, pad_type = obj_231_pad_type_0, strides = obj_231_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("obj_231_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_231_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor out_99_axes_0 = const()[name = tensor("out_99_axes_0"), val = tensor([1])]; + tensor var_3775_to_fp16 = const()[name = tensor("op_3775_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_3775_to_fp16, x = inputs_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor obj_233_gamma_0_to_fp16 = const()[name = tensor("obj_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652991360)))]; + tensor obj_233_beta_0_to_fp16 = const()[name = tensor("obj_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652993472)))]; + tensor obj_233_epsilon_0_to_fp16 = const()[name = tensor("obj_233_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_233_cast_fp16 = batch_norm(beta = obj_233_beta_0_to_fp16, epsilon = obj_233_epsilon_0_to_fp16, gamma = obj_233_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("obj_233_cast_fp16")]; + tensor query_67_pad_type_0 = const()[name = tensor("query_67_pad_type_0"), val = tensor("valid")]; + tensor query_67_strides_0 = const()[name = tensor("query_67_strides_0"), val = tensor([1, 1])]; + tensor query_67_pad_0 = const()[name = tensor("query_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_67_dilations_0 = const()[name = tensor("query_67_dilations_0"), val = tensor([1, 1])]; + tensor query_67_groups_0 = const()[name = tensor("query_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652995584)))]; + tensor layers_16_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(655092800)))]; + tensor query_67_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_bias_to_fp16, dilations = query_67_dilations_0, groups = query_67_groups_0, pad = query_67_pad_0, pad_type = query_67_pad_type_0, strides = query_67_strides_0, weight = layers_16_encoder_attn_q_proj_weight_to_fp16, x = obj_233_cast_fp16)[name = tensor("query_67_cast_fp16")]; + tensor key_67_pad_type_0 = const()[name = tensor("key_67_pad_type_0"), val = tensor("valid")]; + tensor key_67_strides_0 = const()[name = tensor("key_67_strides_0"), val = tensor([1, 1])]; + tensor key_67_pad_0 = const()[name = tensor("key_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_67_dilations_0 = const()[name = tensor("key_67_dilations_0"), val = tensor([1, 1])]; + tensor key_67_groups_0 = const()[name = tensor("key_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(655094912)))]; + tensor key_67_cast_fp16 = conv(dilations = key_67_dilations_0, groups = key_67_groups_0, pad = key_67_pad_0, pad_type = key_67_pad_type_0, strides = key_67_strides_0, weight = layers_16_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_67_cast_fp16")]; + tensor value_67_pad_type_0 = const()[name = tensor("value_67_pad_type_0"), val = tensor("valid")]; + tensor value_67_strides_0 = const()[name = tensor("value_67_strides_0"), val = tensor([1, 1])]; + tensor value_67_pad_0 = const()[name = tensor("value_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_67_dilations_0 = const()[name = tensor("value_67_dilations_0"), val = tensor([1, 1])]; + tensor value_67_groups_0 = const()[name = tensor("value_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657192128)))]; + tensor layers_16_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(659289344)))]; + tensor value_67_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_bias_to_fp16, dilations = value_67_dilations_0, groups = value_67_groups_0, pad = value_67_pad_0, pad_type = value_67_pad_type_0, strides = value_67_strides_0, weight = layers_16_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_67_cast_fp16")]; + tensor var_3811 = const()[name = tensor("op_3811"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_67_cast_fp16 = reshape(shape = var_3811, x = query_67_cast_fp16)[name = tensor("mh_q_67_cast_fp16")]; + tensor var_3813_to_fp16 = const()[name = tensor("op_3813_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3814_cast_fp16 = mul(x = mh_q_67_cast_fp16, y = var_3813_to_fp16)[name = tensor("op_3814_cast_fp16")]; + tensor var_3817 = const()[name = tensor("op_3817"), val = tensor([1, 16, 64, 1500])]; + tensor var_3818_cast_fp16 = reshape(shape = var_3817, x = key_67_cast_fp16)[name = tensor("op_3818_cast_fp16")]; + tensor mh_w_101_transpose_x_0 = const()[name = tensor("mh_w_101_transpose_x_0"), val = tensor(true)]; + tensor mh_w_101_transpose_y_0 = const()[name = tensor("mh_w_101_transpose_y_0"), val = tensor(false)]; + tensor mh_w_101_cast_fp16 = matmul(transpose_x = mh_w_101_transpose_x_0, transpose_y = mh_w_101_transpose_y_0, x = var_3814_cast_fp16, y = var_3818_cast_fp16)[name = tensor("mh_w_101_cast_fp16")]; + tensor obj_237_cast_fp16 = softmax(axis = var_3660, x = mh_w_101_cast_fp16)[name = tensor("obj_237_cast_fp16")]; + tensor var_3822 = const()[name = tensor("op_3822"), val = tensor([1, 16, 64, 1500])]; + tensor var_3823_cast_fp16 = reshape(shape = var_3822, x = value_67_cast_fp16)[name = tensor("op_3823_cast_fp16")]; + tensor attn_67_transpose_x_0 = const()[name = tensor("attn_67_transpose_x_0"), val = tensor(false)]; + tensor attn_67_transpose_y_0 = const()[name = tensor("attn_67_transpose_y_0"), val = tensor(true)]; + tensor attn_67_cast_fp16 = matmul(transpose_x = attn_67_transpose_x_0, transpose_y = attn_67_transpose_y_0, x = var_3823_cast_fp16, y = obj_237_cast_fp16)[name = tensor("attn_67_cast_fp16")]; + tensor var_3826 = const()[name = tensor("op_3826"), val = tensor([1, 1024, 1, 1])]; + tensor input_163_cast_fp16 = reshape(shape = var_3826, x = attn_67_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor obj_235_pad_type_0 = const()[name = tensor("obj_235_pad_type_0"), val = tensor("valid")]; + tensor obj_235_strides_0 = const()[name = tensor("obj_235_strides_0"), val = tensor([1, 1])]; + tensor obj_235_pad_0 = const()[name = tensor("obj_235_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_235_dilations_0 = const()[name = tensor("obj_235_dilations_0"), val = tensor([1, 1])]; + tensor obj_235_groups_0 = const()[name = tensor("obj_235_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(659291456)))]; + tensor layers_16_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661388672)))]; + tensor obj_235_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_bias_to_fp16, dilations = obj_235_dilations_0, groups = obj_235_groups_0, pad = obj_235_pad_0, pad_type = obj_235_pad_type_0, strides = obj_235_strides_0, weight = layers_16_encoder_attn_o_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("obj_235_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_235_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor out_101_axes_0 = const()[name = tensor("out_101_axes_0"), val = tensor([1])]; + tensor var_3847_to_fp16 = const()[name = tensor("op_3847_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_3847_to_fp16, x = inputs_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor input_165_gamma_0_to_fp16 = const()[name = tensor("input_165_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661390784)))]; + tensor input_165_beta_0_to_fp16 = const()[name = tensor("input_165_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661392896)))]; + tensor input_165_epsilon_0_to_fp16 = const()[name = tensor("input_165_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = batch_norm(beta = input_165_beta_0_to_fp16, epsilon = input_165_epsilon_0_to_fp16, gamma = input_165_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_pad_type_0 = const()[name = tensor("input_167_pad_type_0"), val = tensor("valid")]; + tensor input_167_strides_0 = const()[name = tensor("input_167_strides_0"), val = tensor([1, 1])]; + tensor input_167_pad_0 = const()[name = tensor("input_167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_167_dilations_0 = const()[name = tensor("input_167_dilations_0"), val = tensor([1, 1])]; + tensor input_167_groups_0 = const()[name = tensor("input_167_groups_0"), val = tensor(1)]; + tensor layers_16_fc1_weight_to_fp16 = const()[name = tensor("layers_16_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661395008)))]; + tensor layers_16_fc1_bias_to_fp16 = const()[name = tensor("layers_16_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(669783680)))]; + tensor input_167_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = input_167_dilations_0, groups = input_167_groups_0, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = input_167_strides_0, weight = layers_16_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor hidden_states_35_pad_type_0 = const()[name = tensor("hidden_states_35_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_35_strides_0 = const()[name = tensor("hidden_states_35_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_35_pad_0 = const()[name = tensor("hidden_states_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_35_dilations_0 = const()[name = tensor("hidden_states_35_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_35_groups_0 = const()[name = tensor("hidden_states_35_groups_0"), val = tensor(1)]; + tensor layers_16_fc2_weight_to_fp16 = const()[name = tensor("layers_16_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(669791936)))]; + tensor layers_16_fc2_bias_to_fp16 = const()[name = tensor("layers_16_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678180608)))]; + tensor hidden_states_35_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = layers_16_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("hidden_states_35_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor var_3883 = const()[name = tensor("op_3883"), val = tensor(3)]; + tensor out_103_axes_0 = const()[name = tensor("out_103_axes_0"), val = tensor([1])]; + tensor var_3908_to_fp16 = const()[name = tensor("op_3908_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3908_to_fp16, x = inputs_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor obj_239_gamma_0_to_fp16 = const()[name = tensor("obj_239_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678182720)))]; + tensor obj_239_beta_0_to_fp16 = const()[name = tensor("obj_239_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678184832)))]; + tensor obj_239_epsilon_0_to_fp16 = const()[name = tensor("obj_239_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_239_cast_fp16 = batch_norm(beta = obj_239_beta_0_to_fp16, epsilon = obj_239_epsilon_0_to_fp16, gamma = obj_239_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("obj_239_cast_fp16")]; + tensor query_69_pad_type_0 = const()[name = tensor("query_69_pad_type_0"), val = tensor("valid")]; + tensor query_69_strides_0 = const()[name = tensor("query_69_strides_0"), val = tensor([1, 1])]; + tensor query_69_pad_0 = const()[name = tensor("query_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_69_dilations_0 = const()[name = tensor("query_69_dilations_0"), val = tensor([1, 1])]; + tensor query_69_groups_0 = const()[name = tensor("query_69_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678186944)))]; + tensor layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(680284160)))]; + tensor query_69_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = query_69_dilations_0, groups = query_69_groups_0, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = query_69_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("query_69_cast_fp16")]; + tensor current_key_35_pad_type_0 = const()[name = tensor("current_key_35_pad_type_0"), val = tensor("valid")]; + tensor current_key_35_strides_0 = const()[name = tensor("current_key_35_strides_0"), val = tensor([1, 1])]; + tensor current_key_35_pad_0 = const()[name = tensor("current_key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_35_dilations_0 = const()[name = tensor("current_key_35_dilations_0"), val = tensor([1, 1])]; + tensor current_key_35_groups_0 = const()[name = tensor("current_key_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(680286272)))]; + tensor current_key_35_cast_fp16 = conv(dilations = current_key_35_dilations_0, groups = current_key_35_groups_0, pad = current_key_35_pad_0, pad_type = current_key_35_pad_type_0, strides = current_key_35_strides_0, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_key_35_cast_fp16")]; + tensor current_value_35_pad_type_0 = const()[name = tensor("current_value_35_pad_type_0"), val = tensor("valid")]; + tensor current_value_35_strides_0 = const()[name = tensor("current_value_35_strides_0"), val = tensor([1, 1])]; + tensor current_value_35_pad_0 = const()[name = tensor("current_value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_35_dilations_0 = const()[name = tensor("current_value_35_dilations_0"), val = tensor([1, 1])]; + tensor current_value_35_groups_0 = const()[name = tensor("current_value_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(682383488)))]; + tensor layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684480704)))]; + tensor current_value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = current_value_35_dilations_0, groups = current_value_35_groups_0, pad = current_value_35_pad_0, pad_type = current_value_35_pad_type_0, strides = current_value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_value_35_cast_fp16")]; + tensor var_3947_cast_fp16 = mul(x = var_87_cast_fp16_17, y = var_207_cast_fp16)[name = tensor("op_3947_cast_fp16")]; + tensor var_3948_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3948_cast_fp16")]; + tensor key_69_cast_fp16 = add(x = var_3947_cast_fp16, y = var_3948_cast_fp16)[name = tensor("key_69_cast_fp16")]; + tensor var_3951_cast_fp16 = mul(x = var_114_cast_fp16_17, y = var_207_cast_fp16)[name = tensor("op_3951_cast_fp16")]; + tensor var_3952_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3952_cast_fp16")]; + tensor value_69_cast_fp16 = add(x = var_3951_cast_fp16, y = var_3952_cast_fp16)[name = tensor("value_69_cast_fp16")]; + tensor var_3956 = const()[name = tensor("op_3956"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_69_cast_fp16 = reshape(shape = var_3956, x = query_69_cast_fp16)[name = tensor("mh_q_69_cast_fp16")]; + tensor var_3958_to_fp16 = const()[name = tensor("op_3958_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3959_cast_fp16 = mul(x = mh_q_69_cast_fp16, y = var_3958_to_fp16)[name = tensor("op_3959_cast_fp16")]; + tensor var_3962 = const()[name = tensor("op_3962"), val = tensor([1, 16, 64, 448])]; + tensor var_3963_cast_fp16 = reshape(shape = var_3962, x = key_69_cast_fp16)[name = tensor("op_3963_cast_fp16")]; + tensor mh_w_103_transpose_x_0 = const()[name = tensor("mh_w_103_transpose_x_0"), val = tensor(true)]; + tensor mh_w_103_transpose_y_0 = const()[name = tensor("mh_w_103_transpose_y_0"), val = tensor(false)]; + tensor mh_w_103_cast_fp16 = matmul(transpose_x = mh_w_103_transpose_x_0, transpose_y = mh_w_103_transpose_y_0, x = var_3959_cast_fp16, y = var_3963_cast_fp16)[name = tensor("mh_w_103_cast_fp16")]; + tensor mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_105_cast_fp16")]; + tensor var_3971_cast_fp16 = softmax(axis = var_3883, x = mh_w_105_cast_fp16)[name = tensor("op_3971_cast_fp16")]; + tensor var_3972 = const()[name = tensor("op_3972"), val = tensor([1, 16, 64, 448])]; + tensor var_3973_cast_fp16 = reshape(shape = var_3972, x = value_69_cast_fp16)[name = tensor("op_3973_cast_fp16")]; + tensor attn_69_transpose_x_0 = const()[name = tensor("attn_69_transpose_x_0"), val = tensor(false)]; + tensor attn_69_transpose_y_0 = const()[name = tensor("attn_69_transpose_y_0"), val = tensor(true)]; + tensor attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_0, transpose_y = attn_69_transpose_y_0, x = var_3973_cast_fp16, y = var_3971_cast_fp16)[name = tensor("attn_69_cast_fp16")]; + tensor var_3976 = const()[name = tensor("op_3976"), val = tensor([1, 1024, 1, 1])]; + tensor input_171_cast_fp16 = reshape(shape = var_3976, x = attn_69_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor obj_245_pad_type_0 = const()[name = tensor("obj_245_pad_type_0"), val = tensor("valid")]; + tensor obj_245_strides_0 = const()[name = tensor("obj_245_strides_0"), val = tensor([1, 1])]; + tensor obj_245_pad_0 = const()[name = tensor("obj_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_245_dilations_0 = const()[name = tensor("obj_245_dilations_0"), val = tensor([1, 1])]; + tensor obj_245_groups_0 = const()[name = tensor("obj_245_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684482816)))]; + tensor layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686580032)))]; + tensor obj_245_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = obj_245_dilations_0, groups = obj_245_groups_0, pad = obj_245_pad_0, pad_type = obj_245_pad_type_0, strides = obj_245_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("obj_245_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_245_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor out_105_axes_0 = const()[name = tensor("out_105_axes_0"), val = tensor([1])]; + tensor var_3998_to_fp16 = const()[name = tensor("op_3998_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3998_to_fp16, x = inputs_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor obj_247_gamma_0_to_fp16 = const()[name = tensor("obj_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686582144)))]; + tensor obj_247_beta_0_to_fp16 = const()[name = tensor("obj_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686584256)))]; + tensor obj_247_epsilon_0_to_fp16 = const()[name = tensor("obj_247_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_247_cast_fp16 = batch_norm(beta = obj_247_beta_0_to_fp16, epsilon = obj_247_epsilon_0_to_fp16, gamma = obj_247_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("obj_247_cast_fp16")]; + tensor query_71_pad_type_0 = const()[name = tensor("query_71_pad_type_0"), val = tensor("valid")]; + tensor query_71_strides_0 = const()[name = tensor("query_71_strides_0"), val = tensor([1, 1])]; + tensor query_71_pad_0 = const()[name = tensor("query_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_71_dilations_0 = const()[name = tensor("query_71_dilations_0"), val = tensor([1, 1])]; + tensor query_71_groups_0 = const()[name = tensor("query_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686586368)))]; + tensor layers_17_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(688683584)))]; + tensor query_71_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_bias_to_fp16, dilations = query_71_dilations_0, groups = query_71_groups_0, pad = query_71_pad_0, pad_type = query_71_pad_type_0, strides = query_71_strides_0, weight = layers_17_encoder_attn_q_proj_weight_to_fp16, x = obj_247_cast_fp16)[name = tensor("query_71_cast_fp16")]; + tensor key_71_pad_type_0 = const()[name = tensor("key_71_pad_type_0"), val = tensor("valid")]; + tensor key_71_strides_0 = const()[name = tensor("key_71_strides_0"), val = tensor([1, 1])]; + tensor key_71_pad_0 = const()[name = tensor("key_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_71_dilations_0 = const()[name = tensor("key_71_dilations_0"), val = tensor([1, 1])]; + tensor key_71_groups_0 = const()[name = tensor("key_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(688685696)))]; + tensor key_71_cast_fp16 = conv(dilations = key_71_dilations_0, groups = key_71_groups_0, pad = key_71_pad_0, pad_type = key_71_pad_type_0, strides = key_71_strides_0, weight = layers_17_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_71_cast_fp16")]; + tensor value_71_pad_type_0 = const()[name = tensor("value_71_pad_type_0"), val = tensor("valid")]; + tensor value_71_strides_0 = const()[name = tensor("value_71_strides_0"), val = tensor([1, 1])]; + tensor value_71_pad_0 = const()[name = tensor("value_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_71_dilations_0 = const()[name = tensor("value_71_dilations_0"), val = tensor([1, 1])]; + tensor value_71_groups_0 = const()[name = tensor("value_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(690782912)))]; + tensor layers_17_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(692880128)))]; + tensor value_71_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_bias_to_fp16, dilations = value_71_dilations_0, groups = value_71_groups_0, pad = value_71_pad_0, pad_type = value_71_pad_type_0, strides = value_71_strides_0, weight = layers_17_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_71_cast_fp16")]; + tensor var_4034 = const()[name = tensor("op_4034"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_71_cast_fp16 = reshape(shape = var_4034, x = query_71_cast_fp16)[name = tensor("mh_q_71_cast_fp16")]; + tensor var_4036_to_fp16 = const()[name = tensor("op_4036_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4037_cast_fp16 = mul(x = mh_q_71_cast_fp16, y = var_4036_to_fp16)[name = tensor("op_4037_cast_fp16")]; + tensor var_4040 = const()[name = tensor("op_4040"), val = tensor([1, 16, 64, 1500])]; + tensor var_4041_cast_fp16 = reshape(shape = var_4040, x = key_71_cast_fp16)[name = tensor("op_4041_cast_fp16")]; + tensor mh_w_107_transpose_x_0 = const()[name = tensor("mh_w_107_transpose_x_0"), val = tensor(true)]; + tensor mh_w_107_transpose_y_0 = const()[name = tensor("mh_w_107_transpose_y_0"), val = tensor(false)]; + tensor mh_w_107_cast_fp16 = matmul(transpose_x = mh_w_107_transpose_x_0, transpose_y = mh_w_107_transpose_y_0, x = var_4037_cast_fp16, y = var_4041_cast_fp16)[name = tensor("mh_w_107_cast_fp16")]; + tensor obj_251_cast_fp16 = softmax(axis = var_3883, x = mh_w_107_cast_fp16)[name = tensor("obj_251_cast_fp16")]; + tensor var_4045 = const()[name = tensor("op_4045"), val = tensor([1, 16, 64, 1500])]; + tensor var_4046_cast_fp16 = reshape(shape = var_4045, x = value_71_cast_fp16)[name = tensor("op_4046_cast_fp16")]; + tensor attn_71_transpose_x_0 = const()[name = tensor("attn_71_transpose_x_0"), val = tensor(false)]; + tensor attn_71_transpose_y_0 = const()[name = tensor("attn_71_transpose_y_0"), val = tensor(true)]; + tensor attn_71_cast_fp16 = matmul(transpose_x = attn_71_transpose_x_0, transpose_y = attn_71_transpose_y_0, x = var_4046_cast_fp16, y = obj_251_cast_fp16)[name = tensor("attn_71_cast_fp16")]; + tensor var_4049 = const()[name = tensor("op_4049"), val = tensor([1, 1024, 1, 1])]; + tensor input_173_cast_fp16 = reshape(shape = var_4049, x = attn_71_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor obj_249_pad_type_0 = const()[name = tensor("obj_249_pad_type_0"), val = tensor("valid")]; + tensor obj_249_strides_0 = const()[name = tensor("obj_249_strides_0"), val = tensor([1, 1])]; + tensor obj_249_pad_0 = const()[name = tensor("obj_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_249_dilations_0 = const()[name = tensor("obj_249_dilations_0"), val = tensor([1, 1])]; + tensor obj_249_groups_0 = const()[name = tensor("obj_249_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(692882240)))]; + tensor layers_17_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(694979456)))]; + tensor obj_249_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_bias_to_fp16, dilations = obj_249_dilations_0, groups = obj_249_groups_0, pad = obj_249_pad_0, pad_type = obj_249_pad_type_0, strides = obj_249_strides_0, weight = layers_17_encoder_attn_o_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("obj_249_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_249_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor out_107_axes_0 = const()[name = tensor("out_107_axes_0"), val = tensor([1])]; + tensor var_4070_to_fp16 = const()[name = tensor("op_4070_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_4070_to_fp16, x = inputs_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_175_gamma_0_to_fp16 = const()[name = tensor("input_175_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(694981568)))]; + tensor input_175_beta_0_to_fp16 = const()[name = tensor("input_175_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(694983680)))]; + tensor input_175_epsilon_0_to_fp16 = const()[name = tensor("input_175_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_175_cast_fp16 = batch_norm(beta = input_175_beta_0_to_fp16, epsilon = input_175_epsilon_0_to_fp16, gamma = input_175_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_pad_type_0 = const()[name = tensor("input_177_pad_type_0"), val = tensor("valid")]; + tensor input_177_strides_0 = const()[name = tensor("input_177_strides_0"), val = tensor([1, 1])]; + tensor input_177_pad_0 = const()[name = tensor("input_177_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_177_dilations_0 = const()[name = tensor("input_177_dilations_0"), val = tensor([1, 1])]; + tensor input_177_groups_0 = const()[name = tensor("input_177_groups_0"), val = tensor(1)]; + tensor layers_17_fc1_weight_to_fp16 = const()[name = tensor("layers_17_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(694985792)))]; + tensor layers_17_fc1_bias_to_fp16 = const()[name = tensor("layers_17_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(703374464)))]; + tensor input_177_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = input_177_dilations_0, groups = input_177_groups_0, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = input_177_strides_0, weight = layers_17_fc1_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor hidden_states_37_pad_type_0 = const()[name = tensor("hidden_states_37_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_37_strides_0 = const()[name = tensor("hidden_states_37_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_37_pad_0 = const()[name = tensor("hidden_states_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_37_dilations_0 = const()[name = tensor("hidden_states_37_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_37_groups_0 = const()[name = tensor("hidden_states_37_groups_0"), val = tensor(1)]; + tensor layers_17_fc2_weight_to_fp16 = const()[name = tensor("layers_17_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(703382720)))]; + tensor layers_17_fc2_bias_to_fp16 = const()[name = tensor("layers_17_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711771392)))]; + tensor hidden_states_37_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_17_fc2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_4106 = const()[name = tensor("op_4106"), val = tensor(3)]; + tensor out_109_axes_0 = const()[name = tensor("out_109_axes_0"), val = tensor([1])]; + tensor var_4131_to_fp16 = const()[name = tensor("op_4131_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_4131_to_fp16, x = inputs_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor obj_253_gamma_0_to_fp16 = const()[name = tensor("obj_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711773504)))]; + tensor obj_253_beta_0_to_fp16 = const()[name = tensor("obj_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711775616)))]; + tensor obj_253_epsilon_0_to_fp16 = const()[name = tensor("obj_253_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_253_cast_fp16 = batch_norm(beta = obj_253_beta_0_to_fp16, epsilon = obj_253_epsilon_0_to_fp16, gamma = obj_253_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("obj_253_cast_fp16")]; + tensor query_73_pad_type_0 = const()[name = tensor("query_73_pad_type_0"), val = tensor("valid")]; + tensor query_73_strides_0 = const()[name = tensor("query_73_strides_0"), val = tensor([1, 1])]; + tensor query_73_pad_0 = const()[name = tensor("query_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_73_dilations_0 = const()[name = tensor("query_73_dilations_0"), val = tensor([1, 1])]; + tensor query_73_groups_0 = const()[name = tensor("query_73_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711777728)))]; + tensor layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(713874944)))]; + tensor query_73_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = query_73_dilations_0, groups = query_73_groups_0, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = query_73_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("query_73_cast_fp16")]; + tensor current_key_37_pad_type_0 = const()[name = tensor("current_key_37_pad_type_0"), val = tensor("valid")]; + tensor current_key_37_strides_0 = const()[name = tensor("current_key_37_strides_0"), val = tensor([1, 1])]; + tensor current_key_37_pad_0 = const()[name = tensor("current_key_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_37_dilations_0 = const()[name = tensor("current_key_37_dilations_0"), val = tensor([1, 1])]; + tensor current_key_37_groups_0 = const()[name = tensor("current_key_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(713877056)))]; + tensor current_key_37_cast_fp16 = conv(dilations = current_key_37_dilations_0, groups = current_key_37_groups_0, pad = current_key_37_pad_0, pad_type = current_key_37_pad_type_0, strides = current_key_37_strides_0, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_key_37_cast_fp16")]; + tensor current_value_37_pad_type_0 = const()[name = tensor("current_value_37_pad_type_0"), val = tensor("valid")]; + tensor current_value_37_strides_0 = const()[name = tensor("current_value_37_strides_0"), val = tensor([1, 1])]; + tensor current_value_37_pad_0 = const()[name = tensor("current_value_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_37_dilations_0 = const()[name = tensor("current_value_37_dilations_0"), val = tensor([1, 1])]; + tensor current_value_37_groups_0 = const()[name = tensor("current_value_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(715974272)))]; + tensor layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(718071488)))]; + tensor current_value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = current_value_37_dilations_0, groups = current_value_37_groups_0, pad = current_value_37_pad_0, pad_type = current_value_37_pad_type_0, strides = current_value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_value_37_cast_fp16")]; + tensor var_4170_cast_fp16 = mul(x = var_87_cast_fp16_18, y = var_207_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor var_4171_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4171_cast_fp16")]; + tensor key_73_cast_fp16 = add(x = var_4170_cast_fp16, y = var_4171_cast_fp16)[name = tensor("key_73_cast_fp16")]; + tensor var_4174_cast_fp16 = mul(x = var_114_cast_fp16_18, y = var_207_cast_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor var_4175_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4175_cast_fp16")]; + tensor value_73_cast_fp16 = add(x = var_4174_cast_fp16, y = var_4175_cast_fp16)[name = tensor("value_73_cast_fp16")]; + tensor var_4179 = const()[name = tensor("op_4179"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_73_cast_fp16 = reshape(shape = var_4179, x = query_73_cast_fp16)[name = tensor("mh_q_73_cast_fp16")]; + tensor var_4181_to_fp16 = const()[name = tensor("op_4181_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4182_cast_fp16 = mul(x = mh_q_73_cast_fp16, y = var_4181_to_fp16)[name = tensor("op_4182_cast_fp16")]; + tensor var_4185 = const()[name = tensor("op_4185"), val = tensor([1, 16, 64, 448])]; + tensor var_4186_cast_fp16 = reshape(shape = var_4185, x = key_73_cast_fp16)[name = tensor("op_4186_cast_fp16")]; + tensor mh_w_109_transpose_x_0 = const()[name = tensor("mh_w_109_transpose_x_0"), val = tensor(true)]; + tensor mh_w_109_transpose_y_0 = const()[name = tensor("mh_w_109_transpose_y_0"), val = tensor(false)]; + tensor mh_w_109_cast_fp16 = matmul(transpose_x = mh_w_109_transpose_x_0, transpose_y = mh_w_109_transpose_y_0, x = var_4182_cast_fp16, y = var_4186_cast_fp16)[name = tensor("mh_w_109_cast_fp16")]; + tensor mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_111_cast_fp16")]; + tensor var_4194_cast_fp16 = softmax(axis = var_4106, x = mh_w_111_cast_fp16)[name = tensor("op_4194_cast_fp16")]; + tensor var_4195 = const()[name = tensor("op_4195"), val = tensor([1, 16, 64, 448])]; + tensor var_4196_cast_fp16 = reshape(shape = var_4195, x = value_73_cast_fp16)[name = tensor("op_4196_cast_fp16")]; + tensor attn_73_transpose_x_0 = const()[name = tensor("attn_73_transpose_x_0"), val = tensor(false)]; + tensor attn_73_transpose_y_0 = const()[name = tensor("attn_73_transpose_y_0"), val = tensor(true)]; + tensor attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_0, transpose_y = attn_73_transpose_y_0, x = var_4196_cast_fp16, y = var_4194_cast_fp16)[name = tensor("attn_73_cast_fp16")]; + tensor var_4199 = const()[name = tensor("op_4199"), val = tensor([1, 1024, 1, 1])]; + tensor input_181_cast_fp16 = reshape(shape = var_4199, x = attn_73_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor obj_259_pad_type_0 = const()[name = tensor("obj_259_pad_type_0"), val = tensor("valid")]; + tensor obj_259_strides_0 = const()[name = tensor("obj_259_strides_0"), val = tensor([1, 1])]; + tensor obj_259_pad_0 = const()[name = tensor("obj_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_259_dilations_0 = const()[name = tensor("obj_259_dilations_0"), val = tensor([1, 1])]; + tensor obj_259_groups_0 = const()[name = tensor("obj_259_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(718073600)))]; + tensor layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720170816)))]; + tensor obj_259_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = obj_259_dilations_0, groups = obj_259_groups_0, pad = obj_259_pad_0, pad_type = obj_259_pad_type_0, strides = obj_259_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("obj_259_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_259_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor out_111_axes_0 = const()[name = tensor("out_111_axes_0"), val = tensor([1])]; + tensor var_4221_to_fp16 = const()[name = tensor("op_4221_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_4221_to_fp16, x = inputs_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor obj_261_gamma_0_to_fp16 = const()[name = tensor("obj_261_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720172928)))]; + tensor obj_261_beta_0_to_fp16 = const()[name = tensor("obj_261_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720175040)))]; + tensor obj_261_epsilon_0_to_fp16 = const()[name = tensor("obj_261_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_261_cast_fp16 = batch_norm(beta = obj_261_beta_0_to_fp16, epsilon = obj_261_epsilon_0_to_fp16, gamma = obj_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("obj_261_cast_fp16")]; + tensor query_75_pad_type_0 = const()[name = tensor("query_75_pad_type_0"), val = tensor("valid")]; + tensor query_75_strides_0 = const()[name = tensor("query_75_strides_0"), val = tensor([1, 1])]; + tensor query_75_pad_0 = const()[name = tensor("query_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_75_dilations_0 = const()[name = tensor("query_75_dilations_0"), val = tensor([1, 1])]; + tensor query_75_groups_0 = const()[name = tensor("query_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720177152)))]; + tensor layers_18_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722274368)))]; + tensor query_75_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_bias_to_fp16, dilations = query_75_dilations_0, groups = query_75_groups_0, pad = query_75_pad_0, pad_type = query_75_pad_type_0, strides = query_75_strides_0, weight = layers_18_encoder_attn_q_proj_weight_to_fp16, x = obj_261_cast_fp16)[name = tensor("query_75_cast_fp16")]; + tensor key_75_pad_type_0 = const()[name = tensor("key_75_pad_type_0"), val = tensor("valid")]; + tensor key_75_strides_0 = const()[name = tensor("key_75_strides_0"), val = tensor([1, 1])]; + tensor key_75_pad_0 = const()[name = tensor("key_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_75_dilations_0 = const()[name = tensor("key_75_dilations_0"), val = tensor([1, 1])]; + tensor key_75_groups_0 = const()[name = tensor("key_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722276480)))]; + tensor key_75_cast_fp16 = conv(dilations = key_75_dilations_0, groups = key_75_groups_0, pad = key_75_pad_0, pad_type = key_75_pad_type_0, strides = key_75_strides_0, weight = layers_18_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_75_cast_fp16")]; + tensor value_75_pad_type_0 = const()[name = tensor("value_75_pad_type_0"), val = tensor("valid")]; + tensor value_75_strides_0 = const()[name = tensor("value_75_strides_0"), val = tensor([1, 1])]; + tensor value_75_pad_0 = const()[name = tensor("value_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_75_dilations_0 = const()[name = tensor("value_75_dilations_0"), val = tensor([1, 1])]; + tensor value_75_groups_0 = const()[name = tensor("value_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724373696)))]; + tensor layers_18_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726470912)))]; + tensor value_75_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_bias_to_fp16, dilations = value_75_dilations_0, groups = value_75_groups_0, pad = value_75_pad_0, pad_type = value_75_pad_type_0, strides = value_75_strides_0, weight = layers_18_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_75_cast_fp16")]; + tensor var_4257 = const()[name = tensor("op_4257"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_75_cast_fp16 = reshape(shape = var_4257, x = query_75_cast_fp16)[name = tensor("mh_q_75_cast_fp16")]; + tensor var_4259_to_fp16 = const()[name = tensor("op_4259_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4260_cast_fp16 = mul(x = mh_q_75_cast_fp16, y = var_4259_to_fp16)[name = tensor("op_4260_cast_fp16")]; + tensor var_4263 = const()[name = tensor("op_4263"), val = tensor([1, 16, 64, 1500])]; + tensor var_4264_cast_fp16 = reshape(shape = var_4263, x = key_75_cast_fp16)[name = tensor("op_4264_cast_fp16")]; + tensor mh_w_113_transpose_x_0 = const()[name = tensor("mh_w_113_transpose_x_0"), val = tensor(true)]; + tensor mh_w_113_transpose_y_0 = const()[name = tensor("mh_w_113_transpose_y_0"), val = tensor(false)]; + tensor mh_w_113_cast_fp16 = matmul(transpose_x = mh_w_113_transpose_x_0, transpose_y = mh_w_113_transpose_y_0, x = var_4260_cast_fp16, y = var_4264_cast_fp16)[name = tensor("mh_w_113_cast_fp16")]; + tensor obj_265_cast_fp16 = softmax(axis = var_4106, x = mh_w_113_cast_fp16)[name = tensor("obj_265_cast_fp16")]; + tensor var_4268 = const()[name = tensor("op_4268"), val = tensor([1, 16, 64, 1500])]; + tensor var_4269_cast_fp16 = reshape(shape = var_4268, x = value_75_cast_fp16)[name = tensor("op_4269_cast_fp16")]; + tensor attn_75_transpose_x_0 = const()[name = tensor("attn_75_transpose_x_0"), val = tensor(false)]; + tensor attn_75_transpose_y_0 = const()[name = tensor("attn_75_transpose_y_0"), val = tensor(true)]; + tensor attn_75_cast_fp16 = matmul(transpose_x = attn_75_transpose_x_0, transpose_y = attn_75_transpose_y_0, x = var_4269_cast_fp16, y = obj_265_cast_fp16)[name = tensor("attn_75_cast_fp16")]; + tensor var_4272 = const()[name = tensor("op_4272"), val = tensor([1, 1024, 1, 1])]; + tensor input_183_cast_fp16 = reshape(shape = var_4272, x = attn_75_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor obj_263_pad_type_0 = const()[name = tensor("obj_263_pad_type_0"), val = tensor("valid")]; + tensor obj_263_strides_0 = const()[name = tensor("obj_263_strides_0"), val = tensor([1, 1])]; + tensor obj_263_pad_0 = const()[name = tensor("obj_263_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_263_dilations_0 = const()[name = tensor("obj_263_dilations_0"), val = tensor([1, 1])]; + tensor obj_263_groups_0 = const()[name = tensor("obj_263_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726473024)))]; + tensor layers_18_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(728570240)))]; + tensor obj_263_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_bias_to_fp16, dilations = obj_263_dilations_0, groups = obj_263_groups_0, pad = obj_263_pad_0, pad_type = obj_263_pad_type_0, strides = obj_263_strides_0, weight = layers_18_encoder_attn_o_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("obj_263_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_263_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor out_113_axes_0 = const()[name = tensor("out_113_axes_0"), val = tensor([1])]; + tensor var_4293_to_fp16 = const()[name = tensor("op_4293_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_4293_to_fp16, x = inputs_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor input_185_gamma_0_to_fp16 = const()[name = tensor("input_185_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(728572352)))]; + tensor input_185_beta_0_to_fp16 = const()[name = tensor("input_185_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(728574464)))]; + tensor input_185_epsilon_0_to_fp16 = const()[name = tensor("input_185_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = batch_norm(beta = input_185_beta_0_to_fp16, epsilon = input_185_epsilon_0_to_fp16, gamma = input_185_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor input_187_pad_type_0 = const()[name = tensor("input_187_pad_type_0"), val = tensor("valid")]; + tensor input_187_strides_0 = const()[name = tensor("input_187_strides_0"), val = tensor([1, 1])]; + tensor input_187_pad_0 = const()[name = tensor("input_187_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_187_dilations_0 = const()[name = tensor("input_187_dilations_0"), val = tensor([1, 1])]; + tensor input_187_groups_0 = const()[name = tensor("input_187_groups_0"), val = tensor(1)]; + tensor layers_18_fc1_weight_to_fp16 = const()[name = tensor("layers_18_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(728576576)))]; + tensor layers_18_fc1_bias_to_fp16 = const()[name = tensor("layers_18_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736965248)))]; + tensor input_187_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = input_187_dilations_0, groups = input_187_groups_0, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = input_187_strides_0, weight = layers_18_fc1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_mode_0 = const()[name = tensor("input_189_mode_0"), val = tensor("EXACT")]; + tensor input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor hidden_states_39_pad_type_0 = const()[name = tensor("hidden_states_39_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_39_strides_0 = const()[name = tensor("hidden_states_39_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_39_pad_0 = const()[name = tensor("hidden_states_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_39_dilations_0 = const()[name = tensor("hidden_states_39_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_39_groups_0 = const()[name = tensor("hidden_states_39_groups_0"), val = tensor(1)]; + tensor layers_18_fc2_weight_to_fp16 = const()[name = tensor("layers_18_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736973504)))]; + tensor layers_18_fc2_bias_to_fp16 = const()[name = tensor("layers_18_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(745362176)))]; + tensor hidden_states_39_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = layers_18_fc2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("hidden_states_39_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor var_4329 = const()[name = tensor("op_4329"), val = tensor(3)]; + tensor out_115_axes_0 = const()[name = tensor("out_115_axes_0"), val = tensor([1])]; + tensor var_4354_to_fp16 = const()[name = tensor("op_4354_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_4354_to_fp16, x = inputs_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor obj_267_gamma_0_to_fp16 = const()[name = tensor("obj_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(745364288)))]; + tensor obj_267_beta_0_to_fp16 = const()[name = tensor("obj_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(745366400)))]; + tensor obj_267_epsilon_0_to_fp16 = const()[name = tensor("obj_267_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_267_cast_fp16 = batch_norm(beta = obj_267_beta_0_to_fp16, epsilon = obj_267_epsilon_0_to_fp16, gamma = obj_267_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("obj_267_cast_fp16")]; + tensor query_77_pad_type_0 = const()[name = tensor("query_77_pad_type_0"), val = tensor("valid")]; + tensor query_77_strides_0 = const()[name = tensor("query_77_strides_0"), val = tensor([1, 1])]; + tensor query_77_pad_0 = const()[name = tensor("query_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_77_dilations_0 = const()[name = tensor("query_77_dilations_0"), val = tensor([1, 1])]; + tensor query_77_groups_0 = const()[name = tensor("query_77_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(745368512)))]; + tensor layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(747465728)))]; + tensor query_77_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = query_77_dilations_0, groups = query_77_groups_0, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = query_77_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("query_77_cast_fp16")]; + tensor current_key_39_pad_type_0 = const()[name = tensor("current_key_39_pad_type_0"), val = tensor("valid")]; + tensor current_key_39_strides_0 = const()[name = tensor("current_key_39_strides_0"), val = tensor([1, 1])]; + tensor current_key_39_pad_0 = const()[name = tensor("current_key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_39_dilations_0 = const()[name = tensor("current_key_39_dilations_0"), val = tensor([1, 1])]; + tensor current_key_39_groups_0 = const()[name = tensor("current_key_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(747467840)))]; + tensor current_key_39_cast_fp16 = conv(dilations = current_key_39_dilations_0, groups = current_key_39_groups_0, pad = current_key_39_pad_0, pad_type = current_key_39_pad_type_0, strides = current_key_39_strides_0, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_key_39_cast_fp16")]; + tensor current_value_39_pad_type_0 = const()[name = tensor("current_value_39_pad_type_0"), val = tensor("valid")]; + tensor current_value_39_strides_0 = const()[name = tensor("current_value_39_strides_0"), val = tensor([1, 1])]; + tensor current_value_39_pad_0 = const()[name = tensor("current_value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_39_dilations_0 = const()[name = tensor("current_value_39_dilations_0"), val = tensor([1, 1])]; + tensor current_value_39_groups_0 = const()[name = tensor("current_value_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749565056)))]; + tensor layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(751662272)))]; + tensor current_value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = current_value_39_dilations_0, groups = current_value_39_groups_0, pad = current_value_39_pad_0, pad_type = current_value_39_pad_type_0, strides = current_value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_value_39_cast_fp16")]; + tensor var_4393_cast_fp16 = mul(x = var_87_cast_fp16_19, y = var_207_cast_fp16)[name = tensor("op_4393_cast_fp16")]; + tensor var_4394_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4394_cast_fp16")]; + tensor key_77_cast_fp16 = add(x = var_4393_cast_fp16, y = var_4394_cast_fp16)[name = tensor("key_77_cast_fp16")]; + tensor var_4397_cast_fp16 = mul(x = var_114_cast_fp16_19, y = var_207_cast_fp16)[name = tensor("op_4397_cast_fp16")]; + tensor var_4398_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4398_cast_fp16")]; + tensor value_77_cast_fp16 = add(x = var_4397_cast_fp16, y = var_4398_cast_fp16)[name = tensor("value_77_cast_fp16")]; + tensor var_4402 = const()[name = tensor("op_4402"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_77_cast_fp16 = reshape(shape = var_4402, x = query_77_cast_fp16)[name = tensor("mh_q_77_cast_fp16")]; + tensor var_4404_to_fp16 = const()[name = tensor("op_4404_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4405_cast_fp16 = mul(x = mh_q_77_cast_fp16, y = var_4404_to_fp16)[name = tensor("op_4405_cast_fp16")]; + tensor var_4408 = const()[name = tensor("op_4408"), val = tensor([1, 16, 64, 448])]; + tensor var_4409_cast_fp16 = reshape(shape = var_4408, x = key_77_cast_fp16)[name = tensor("op_4409_cast_fp16")]; + tensor mh_w_115_transpose_x_0 = const()[name = tensor("mh_w_115_transpose_x_0"), val = tensor(true)]; + tensor mh_w_115_transpose_y_0 = const()[name = tensor("mh_w_115_transpose_y_0"), val = tensor(false)]; + tensor mh_w_115_cast_fp16 = matmul(transpose_x = mh_w_115_transpose_x_0, transpose_y = mh_w_115_transpose_y_0, x = var_4405_cast_fp16, y = var_4409_cast_fp16)[name = tensor("mh_w_115_cast_fp16")]; + tensor mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_117_cast_fp16")]; + tensor var_4417_cast_fp16 = softmax(axis = var_4329, x = mh_w_117_cast_fp16)[name = tensor("op_4417_cast_fp16")]; + tensor var_4418 = const()[name = tensor("op_4418"), val = tensor([1, 16, 64, 448])]; + tensor var_4419_cast_fp16 = reshape(shape = var_4418, x = value_77_cast_fp16)[name = tensor("op_4419_cast_fp16")]; + tensor attn_77_transpose_x_0 = const()[name = tensor("attn_77_transpose_x_0"), val = tensor(false)]; + tensor attn_77_transpose_y_0 = const()[name = tensor("attn_77_transpose_y_0"), val = tensor(true)]; + tensor attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_0, transpose_y = attn_77_transpose_y_0, x = var_4419_cast_fp16, y = var_4417_cast_fp16)[name = tensor("attn_77_cast_fp16")]; + tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1024, 1, 1])]; + tensor input_191_cast_fp16 = reshape(shape = var_4422, x = attn_77_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor obj_273_pad_type_0 = const()[name = tensor("obj_273_pad_type_0"), val = tensor("valid")]; + tensor obj_273_strides_0 = const()[name = tensor("obj_273_strides_0"), val = tensor([1, 1])]; + tensor obj_273_pad_0 = const()[name = tensor("obj_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_273_dilations_0 = const()[name = tensor("obj_273_dilations_0"), val = tensor([1, 1])]; + tensor obj_273_groups_0 = const()[name = tensor("obj_273_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(751664384)))]; + tensor layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(753761600)))]; + tensor obj_273_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = obj_273_dilations_0, groups = obj_273_groups_0, pad = obj_273_pad_0, pad_type = obj_273_pad_type_0, strides = obj_273_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("obj_273_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_273_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor out_117_axes_0 = const()[name = tensor("out_117_axes_0"), val = tensor([1])]; + tensor var_4444_to_fp16 = const()[name = tensor("op_4444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_4444_to_fp16, x = inputs_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor obj_275_gamma_0_to_fp16 = const()[name = tensor("obj_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(753763712)))]; + tensor obj_275_beta_0_to_fp16 = const()[name = tensor("obj_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(753765824)))]; + tensor obj_275_epsilon_0_to_fp16 = const()[name = tensor("obj_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_275_cast_fp16 = batch_norm(beta = obj_275_beta_0_to_fp16, epsilon = obj_275_epsilon_0_to_fp16, gamma = obj_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("obj_275_cast_fp16")]; + tensor query_79_pad_type_0 = const()[name = tensor("query_79_pad_type_0"), val = tensor("valid")]; + tensor query_79_strides_0 = const()[name = tensor("query_79_strides_0"), val = tensor([1, 1])]; + tensor query_79_pad_0 = const()[name = tensor("query_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_79_dilations_0 = const()[name = tensor("query_79_dilations_0"), val = tensor([1, 1])]; + tensor query_79_groups_0 = const()[name = tensor("query_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(753767936)))]; + tensor layers_19_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(755865152)))]; + tensor query_79_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_bias_to_fp16, dilations = query_79_dilations_0, groups = query_79_groups_0, pad = query_79_pad_0, pad_type = query_79_pad_type_0, strides = query_79_strides_0, weight = layers_19_encoder_attn_q_proj_weight_to_fp16, x = obj_275_cast_fp16)[name = tensor("query_79_cast_fp16")]; + tensor key_79_pad_type_0 = const()[name = tensor("key_79_pad_type_0"), val = tensor("valid")]; + tensor key_79_strides_0 = const()[name = tensor("key_79_strides_0"), val = tensor([1, 1])]; + tensor key_79_pad_0 = const()[name = tensor("key_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_79_dilations_0 = const()[name = tensor("key_79_dilations_0"), val = tensor([1, 1])]; + tensor key_79_groups_0 = const()[name = tensor("key_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(755867264)))]; + tensor key_79_cast_fp16 = conv(dilations = key_79_dilations_0, groups = key_79_groups_0, pad = key_79_pad_0, pad_type = key_79_pad_type_0, strides = key_79_strides_0, weight = layers_19_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_79_cast_fp16")]; + tensor value_79_pad_type_0 = const()[name = tensor("value_79_pad_type_0"), val = tensor("valid")]; + tensor value_79_strides_0 = const()[name = tensor("value_79_strides_0"), val = tensor([1, 1])]; + tensor value_79_pad_0 = const()[name = tensor("value_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_79_dilations_0 = const()[name = tensor("value_79_dilations_0"), val = tensor([1, 1])]; + tensor value_79_groups_0 = const()[name = tensor("value_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(757964480)))]; + tensor layers_19_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(760061696)))]; + tensor value_79_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_bias_to_fp16, dilations = value_79_dilations_0, groups = value_79_groups_0, pad = value_79_pad_0, pad_type = value_79_pad_type_0, strides = value_79_strides_0, weight = layers_19_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_79_cast_fp16")]; + tensor var_4480 = const()[name = tensor("op_4480"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_79_cast_fp16 = reshape(shape = var_4480, x = query_79_cast_fp16)[name = tensor("mh_q_79_cast_fp16")]; + tensor var_4482_to_fp16 = const()[name = tensor("op_4482_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4483_cast_fp16 = mul(x = mh_q_79_cast_fp16, y = var_4482_to_fp16)[name = tensor("op_4483_cast_fp16")]; + tensor var_4486 = const()[name = tensor("op_4486"), val = tensor([1, 16, 64, 1500])]; + tensor var_4487_cast_fp16 = reshape(shape = var_4486, x = key_79_cast_fp16)[name = tensor("op_4487_cast_fp16")]; + tensor mh_w_119_transpose_x_0 = const()[name = tensor("mh_w_119_transpose_x_0"), val = tensor(true)]; + tensor mh_w_119_transpose_y_0 = const()[name = tensor("mh_w_119_transpose_y_0"), val = tensor(false)]; + tensor mh_w_119_cast_fp16 = matmul(transpose_x = mh_w_119_transpose_x_0, transpose_y = mh_w_119_transpose_y_0, x = var_4483_cast_fp16, y = var_4487_cast_fp16)[name = tensor("mh_w_119_cast_fp16")]; + tensor obj_279_cast_fp16 = softmax(axis = var_4329, x = mh_w_119_cast_fp16)[name = tensor("obj_279_cast_fp16")]; + tensor var_4491 = const()[name = tensor("op_4491"), val = tensor([1, 16, 64, 1500])]; + tensor var_4492_cast_fp16 = reshape(shape = var_4491, x = value_79_cast_fp16)[name = tensor("op_4492_cast_fp16")]; + tensor attn_79_transpose_x_0 = const()[name = tensor("attn_79_transpose_x_0"), val = tensor(false)]; + tensor attn_79_transpose_y_0 = const()[name = tensor("attn_79_transpose_y_0"), val = tensor(true)]; + tensor attn_79_cast_fp16 = matmul(transpose_x = attn_79_transpose_x_0, transpose_y = attn_79_transpose_y_0, x = var_4492_cast_fp16, y = obj_279_cast_fp16)[name = tensor("attn_79_cast_fp16")]; + tensor var_4495 = const()[name = tensor("op_4495"), val = tensor([1, 1024, 1, 1])]; + tensor input_193_cast_fp16 = reshape(shape = var_4495, x = attn_79_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor obj_277_pad_type_0 = const()[name = tensor("obj_277_pad_type_0"), val = tensor("valid")]; + tensor obj_277_strides_0 = const()[name = tensor("obj_277_strides_0"), val = tensor([1, 1])]; + tensor obj_277_pad_0 = const()[name = tensor("obj_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_277_dilations_0 = const()[name = tensor("obj_277_dilations_0"), val = tensor([1, 1])]; + tensor obj_277_groups_0 = const()[name = tensor("obj_277_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(760063808)))]; + tensor layers_19_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762161024)))]; + tensor obj_277_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_bias_to_fp16, dilations = obj_277_dilations_0, groups = obj_277_groups_0, pad = obj_277_pad_0, pad_type = obj_277_pad_type_0, strides = obj_277_strides_0, weight = layers_19_encoder_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("obj_277_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_277_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor out_119_axes_0 = const()[name = tensor("out_119_axes_0"), val = tensor([1])]; + tensor var_4513_to_fp16 = const()[name = tensor("op_4513_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_4513_to_fp16, x = inputs_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor input_195_gamma_0_to_fp16 = const()[name = tensor("input_195_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762163136)))]; + tensor input_195_beta_0_to_fp16 = const()[name = tensor("input_195_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762165248)))]; + tensor input_195_epsilon_0_to_fp16 = const()[name = tensor("input_195_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_pad_type_0 = const()[name = tensor("input_197_pad_type_0"), val = tensor("valid")]; + tensor input_197_strides_0 = const()[name = tensor("input_197_strides_0"), val = tensor([1, 1])]; + tensor input_197_pad_0 = const()[name = tensor("input_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_197_dilations_0 = const()[name = tensor("input_197_dilations_0"), val = tensor([1, 1])]; + tensor input_197_groups_0 = const()[name = tensor("input_197_groups_0"), val = tensor(1)]; + tensor layers_19_fc1_weight_to_fp16 = const()[name = tensor("layers_19_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762167360)))]; + tensor layers_19_fc1_bias_to_fp16 = const()[name = tensor("layers_19_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(770556032)))]; + tensor input_197_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = layers_19_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_mode_0 = const()[name = tensor("input_199_mode_0"), val = tensor("EXACT")]; + tensor input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor hidden_states_41_pad_type_0 = const()[name = tensor("hidden_states_41_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_41_strides_0 = const()[name = tensor("hidden_states_41_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_41_pad_0 = const()[name = tensor("hidden_states_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_41_dilations_0 = const()[name = tensor("hidden_states_41_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_41_groups_0 = const()[name = tensor("hidden_states_41_groups_0"), val = tensor(1)]; + tensor layers_19_fc2_weight_to_fp16 = const()[name = tensor("layers_19_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(770564288)))]; + tensor layers_19_fc2_bias_to_fp16 = const()[name = tensor("layers_19_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(778952960)))]; + tensor hidden_states_41_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = hidden_states_41_dilations_0, groups = hidden_states_41_groups_0, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = hidden_states_41_strides_0, weight = layers_19_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("hidden_states_41_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_4548 = const()[name = tensor("op_4548"), val = tensor(3)]; + tensor out_121_axes_0 = const()[name = tensor("out_121_axes_0"), val = tensor([1])]; + tensor var_4573_to_fp16 = const()[name = tensor("op_4573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_4573_to_fp16, x = inputs_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor obj_281_gamma_0_to_fp16 = const()[name = tensor("obj_281_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(778955072)))]; + tensor obj_281_beta_0_to_fp16 = const()[name = tensor("obj_281_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(778957184)))]; + tensor obj_281_epsilon_0_to_fp16 = const()[name = tensor("obj_281_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_281_cast_fp16 = batch_norm(beta = obj_281_beta_0_to_fp16, epsilon = obj_281_epsilon_0_to_fp16, gamma = obj_281_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("obj_281_cast_fp16")]; + tensor query_81_pad_type_0 = const()[name = tensor("query_81_pad_type_0"), val = tensor("valid")]; + tensor query_81_strides_0 = const()[name = tensor("query_81_strides_0"), val = tensor([1, 1])]; + tensor query_81_pad_0 = const()[name = tensor("query_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_81_dilations_0 = const()[name = tensor("query_81_dilations_0"), val = tensor([1, 1])]; + tensor query_81_groups_0 = const()[name = tensor("query_81_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(778959296)))]; + tensor layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(781056512)))]; + tensor query_81_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = query_81_dilations_0, groups = query_81_groups_0, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = query_81_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("query_81_cast_fp16")]; + tensor current_key_41_pad_type_0 = const()[name = tensor("current_key_41_pad_type_0"), val = tensor("valid")]; + tensor current_key_41_strides_0 = const()[name = tensor("current_key_41_strides_0"), val = tensor([1, 1])]; + tensor current_key_41_pad_0 = const()[name = tensor("current_key_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_41_dilations_0 = const()[name = tensor("current_key_41_dilations_0"), val = tensor([1, 1])]; + tensor current_key_41_groups_0 = const()[name = tensor("current_key_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(781058624)))]; + tensor current_key_41_cast_fp16 = conv(dilations = current_key_41_dilations_0, groups = current_key_41_groups_0, pad = current_key_41_pad_0, pad_type = current_key_41_pad_type_0, strides = current_key_41_strides_0, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_key_41_cast_fp16")]; + tensor current_value_41_pad_type_0 = const()[name = tensor("current_value_41_pad_type_0"), val = tensor("valid")]; + tensor current_value_41_strides_0 = const()[name = tensor("current_value_41_strides_0"), val = tensor([1, 1])]; + tensor current_value_41_pad_0 = const()[name = tensor("current_value_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_41_dilations_0 = const()[name = tensor("current_value_41_dilations_0"), val = tensor([1, 1])]; + tensor current_value_41_groups_0 = const()[name = tensor("current_value_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(783155840)))]; + tensor layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(785253056)))]; + tensor current_value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = current_value_41_dilations_0, groups = current_value_41_groups_0, pad = current_value_41_pad_0, pad_type = current_value_41_pad_type_0, strides = current_value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_value_41_cast_fp16")]; + tensor var_4612_cast_fp16 = mul(x = var_87_cast_fp16_20, y = var_207_cast_fp16)[name = tensor("op_4612_cast_fp16")]; + tensor var_4613_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4613_cast_fp16")]; + tensor key_81_cast_fp16 = add(x = var_4612_cast_fp16, y = var_4613_cast_fp16)[name = tensor("key_81_cast_fp16")]; + tensor var_4616_cast_fp16 = mul(x = var_114_cast_fp16_20, y = var_207_cast_fp16)[name = tensor("op_4616_cast_fp16")]; + tensor var_4617_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4617_cast_fp16")]; + tensor value_81_cast_fp16 = add(x = var_4616_cast_fp16, y = var_4617_cast_fp16)[name = tensor("value_81_cast_fp16")]; + tensor var_4621 = const()[name = tensor("op_4621"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_81_cast_fp16 = reshape(shape = var_4621, x = query_81_cast_fp16)[name = tensor("mh_q_81_cast_fp16")]; + tensor var_4623_to_fp16 = const()[name = tensor("op_4623_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4624_cast_fp16 = mul(x = mh_q_81_cast_fp16, y = var_4623_to_fp16)[name = tensor("op_4624_cast_fp16")]; + tensor var_4627 = const()[name = tensor("op_4627"), val = tensor([1, 16, 64, 448])]; + tensor var_4628_cast_fp16 = reshape(shape = var_4627, x = key_81_cast_fp16)[name = tensor("op_4628_cast_fp16")]; + tensor mh_w_121_transpose_x_0 = const()[name = tensor("mh_w_121_transpose_x_0"), val = tensor(true)]; + tensor mh_w_121_transpose_y_0 = const()[name = tensor("mh_w_121_transpose_y_0"), val = tensor(false)]; + tensor mh_w_121_cast_fp16 = matmul(transpose_x = mh_w_121_transpose_x_0, transpose_y = mh_w_121_transpose_y_0, x = var_4624_cast_fp16, y = var_4628_cast_fp16)[name = tensor("mh_w_121_cast_fp16")]; + tensor mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_123_cast_fp16")]; + tensor var_4636_cast_fp16 = softmax(axis = var_4548, x = mh_w_123_cast_fp16)[name = tensor("op_4636_cast_fp16")]; + tensor var_4637 = const()[name = tensor("op_4637"), val = tensor([1, 16, 64, 448])]; + tensor var_4638_cast_fp16 = reshape(shape = var_4637, x = value_81_cast_fp16)[name = tensor("op_4638_cast_fp16")]; + tensor attn_81_transpose_x_0 = const()[name = tensor("attn_81_transpose_x_0"), val = tensor(false)]; + tensor attn_81_transpose_y_0 = const()[name = tensor("attn_81_transpose_y_0"), val = tensor(true)]; + tensor attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_0, transpose_y = attn_81_transpose_y_0, x = var_4638_cast_fp16, y = var_4636_cast_fp16)[name = tensor("attn_81_cast_fp16")]; + tensor var_4641 = const()[name = tensor("op_4641"), val = tensor([1, 1024, 1, 1])]; + tensor input_201_cast_fp16 = reshape(shape = var_4641, x = attn_81_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor obj_287_pad_type_0 = const()[name = tensor("obj_287_pad_type_0"), val = tensor("valid")]; + tensor obj_287_strides_0 = const()[name = tensor("obj_287_strides_0"), val = tensor([1, 1])]; + tensor obj_287_pad_0 = const()[name = tensor("obj_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_287_dilations_0 = const()[name = tensor("obj_287_dilations_0"), val = tensor([1, 1])]; + tensor obj_287_groups_0 = const()[name = tensor("obj_287_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(785255168)))]; + tensor layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(787352384)))]; + tensor obj_287_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = obj_287_dilations_0, groups = obj_287_groups_0, pad = obj_287_pad_0, pad_type = obj_287_pad_type_0, strides = obj_287_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("obj_287_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_287_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor out_123_axes_0 = const()[name = tensor("out_123_axes_0"), val = tensor([1])]; + tensor var_4663_to_fp16 = const()[name = tensor("op_4663_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_4663_to_fp16, x = inputs_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor obj_289_gamma_0_to_fp16 = const()[name = tensor("obj_289_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(787354496)))]; + tensor obj_289_beta_0_to_fp16 = const()[name = tensor("obj_289_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(787356608)))]; + tensor obj_289_epsilon_0_to_fp16 = const()[name = tensor("obj_289_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_289_cast_fp16 = batch_norm(beta = obj_289_beta_0_to_fp16, epsilon = obj_289_epsilon_0_to_fp16, gamma = obj_289_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("obj_289_cast_fp16")]; + tensor query_83_pad_type_0 = const()[name = tensor("query_83_pad_type_0"), val = tensor("valid")]; + tensor query_83_strides_0 = const()[name = tensor("query_83_strides_0"), val = tensor([1, 1])]; + tensor query_83_pad_0 = const()[name = tensor("query_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_83_dilations_0 = const()[name = tensor("query_83_dilations_0"), val = tensor([1, 1])]; + tensor query_83_groups_0 = const()[name = tensor("query_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(787358720)))]; + tensor layers_20_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789455936)))]; + tensor query_83_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_bias_to_fp16, dilations = query_83_dilations_0, groups = query_83_groups_0, pad = query_83_pad_0, pad_type = query_83_pad_type_0, strides = query_83_strides_0, weight = layers_20_encoder_attn_q_proj_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("query_83_cast_fp16")]; + tensor key_83_pad_type_0 = const()[name = tensor("key_83_pad_type_0"), val = tensor("valid")]; + tensor key_83_strides_0 = const()[name = tensor("key_83_strides_0"), val = tensor([1, 1])]; + tensor key_83_pad_0 = const()[name = tensor("key_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_83_dilations_0 = const()[name = tensor("key_83_dilations_0"), val = tensor([1, 1])]; + tensor key_83_groups_0 = const()[name = tensor("key_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789458048)))]; + tensor key_83_cast_fp16 = conv(dilations = key_83_dilations_0, groups = key_83_groups_0, pad = key_83_pad_0, pad_type = key_83_pad_type_0, strides = key_83_strides_0, weight = layers_20_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_83_cast_fp16")]; + tensor value_83_pad_type_0 = const()[name = tensor("value_83_pad_type_0"), val = tensor("valid")]; + tensor value_83_strides_0 = const()[name = tensor("value_83_strides_0"), val = tensor([1, 1])]; + tensor value_83_pad_0 = const()[name = tensor("value_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_83_dilations_0 = const()[name = tensor("value_83_dilations_0"), val = tensor([1, 1])]; + tensor value_83_groups_0 = const()[name = tensor("value_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(791555264)))]; + tensor layers_20_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(793652480)))]; + tensor value_83_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_bias_to_fp16, dilations = value_83_dilations_0, groups = value_83_groups_0, pad = value_83_pad_0, pad_type = value_83_pad_type_0, strides = value_83_strides_0, weight = layers_20_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_83_cast_fp16")]; + tensor var_4699 = const()[name = tensor("op_4699"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_83_cast_fp16 = reshape(shape = var_4699, x = query_83_cast_fp16)[name = tensor("mh_q_83_cast_fp16")]; + tensor var_4701_to_fp16 = const()[name = tensor("op_4701_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4702_cast_fp16 = mul(x = mh_q_83_cast_fp16, y = var_4701_to_fp16)[name = tensor("op_4702_cast_fp16")]; + tensor var_4705 = const()[name = tensor("op_4705"), val = tensor([1, 16, 64, 1500])]; + tensor var_4706_cast_fp16 = reshape(shape = var_4705, x = key_83_cast_fp16)[name = tensor("op_4706_cast_fp16")]; + tensor mh_w_125_transpose_x_0 = const()[name = tensor("mh_w_125_transpose_x_0"), val = tensor(true)]; + tensor mh_w_125_transpose_y_0 = const()[name = tensor("mh_w_125_transpose_y_0"), val = tensor(false)]; + tensor mh_w_125_cast_fp16 = matmul(transpose_x = mh_w_125_transpose_x_0, transpose_y = mh_w_125_transpose_y_0, x = var_4702_cast_fp16, y = var_4706_cast_fp16)[name = tensor("mh_w_125_cast_fp16")]; + tensor obj_293_cast_fp16 = softmax(axis = var_4548, x = mh_w_125_cast_fp16)[name = tensor("obj_293_cast_fp16")]; + tensor var_4710 = const()[name = tensor("op_4710"), val = tensor([1, 16, 64, 1500])]; + tensor var_4711_cast_fp16 = reshape(shape = var_4710, x = value_83_cast_fp16)[name = tensor("op_4711_cast_fp16")]; + tensor attn_83_transpose_x_0 = const()[name = tensor("attn_83_transpose_x_0"), val = tensor(false)]; + tensor attn_83_transpose_y_0 = const()[name = tensor("attn_83_transpose_y_0"), val = tensor(true)]; + tensor attn_83_cast_fp16 = matmul(transpose_x = attn_83_transpose_x_0, transpose_y = attn_83_transpose_y_0, x = var_4711_cast_fp16, y = obj_293_cast_fp16)[name = tensor("attn_83_cast_fp16")]; + tensor var_4714 = const()[name = tensor("op_4714"), val = tensor([1, 1024, 1, 1])]; + tensor input_203_cast_fp16 = reshape(shape = var_4714, x = attn_83_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor obj_291_pad_type_0 = const()[name = tensor("obj_291_pad_type_0"), val = tensor("valid")]; + tensor obj_291_strides_0 = const()[name = tensor("obj_291_strides_0"), val = tensor([1, 1])]; + tensor obj_291_pad_0 = const()[name = tensor("obj_291_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_291_dilations_0 = const()[name = tensor("obj_291_dilations_0"), val = tensor([1, 1])]; + tensor obj_291_groups_0 = const()[name = tensor("obj_291_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(793654592)))]; + tensor layers_20_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795751808)))]; + tensor obj_291_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_bias_to_fp16, dilations = obj_291_dilations_0, groups = obj_291_groups_0, pad = obj_291_pad_0, pad_type = obj_291_pad_type_0, strides = obj_291_strides_0, weight = layers_20_encoder_attn_o_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("obj_291_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_291_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor out_125_axes_0 = const()[name = tensor("out_125_axes_0"), val = tensor([1])]; + tensor var_4735_to_fp16 = const()[name = tensor("op_4735_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_4735_to_fp16, x = inputs_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor input_205_gamma_0_to_fp16 = const()[name = tensor("input_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795753920)))]; + tensor input_205_beta_0_to_fp16 = const()[name = tensor("input_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795756032)))]; + tensor input_205_epsilon_0_to_fp16 = const()[name = tensor("input_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_205_cast_fp16 = batch_norm(beta = input_205_beta_0_to_fp16, epsilon = input_205_epsilon_0_to_fp16, gamma = input_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_pad_type_0 = const()[name = tensor("input_207_pad_type_0"), val = tensor("valid")]; + tensor input_207_strides_0 = const()[name = tensor("input_207_strides_0"), val = tensor([1, 1])]; + tensor input_207_pad_0 = const()[name = tensor("input_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_207_dilations_0 = const()[name = tensor("input_207_dilations_0"), val = tensor([1, 1])]; + tensor input_207_groups_0 = const()[name = tensor("input_207_groups_0"), val = tensor(1)]; + tensor layers_20_fc1_weight_to_fp16 = const()[name = tensor("layers_20_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795758144)))]; + tensor layers_20_fc1_bias_to_fp16 = const()[name = tensor("layers_20_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804146816)))]; + tensor input_207_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = layers_20_fc1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_mode_0 = const()[name = tensor("input_209_mode_0"), val = tensor("EXACT")]; + tensor input_209_cast_fp16 = gelu(mode = input_209_mode_0, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor hidden_states_43_pad_type_0 = const()[name = tensor("hidden_states_43_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_43_strides_0 = const()[name = tensor("hidden_states_43_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_43_pad_0 = const()[name = tensor("hidden_states_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_43_dilations_0 = const()[name = tensor("hidden_states_43_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_43_groups_0 = const()[name = tensor("hidden_states_43_groups_0"), val = tensor(1)]; + tensor layers_20_fc2_weight_to_fp16 = const()[name = tensor("layers_20_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804155072)))]; + tensor layers_20_fc2_bias_to_fp16 = const()[name = tensor("layers_20_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(812543744)))]; + tensor hidden_states_43_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_20_fc2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor var_4771 = const()[name = tensor("op_4771"), val = tensor(3)]; + tensor out_127_axes_0 = const()[name = tensor("out_127_axes_0"), val = tensor([1])]; + tensor var_4796_to_fp16 = const()[name = tensor("op_4796_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_4796_to_fp16, x = inputs_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor obj_295_gamma_0_to_fp16 = const()[name = tensor("obj_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(812545856)))]; + tensor obj_295_beta_0_to_fp16 = const()[name = tensor("obj_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(812547968)))]; + tensor obj_295_epsilon_0_to_fp16 = const()[name = tensor("obj_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_295_cast_fp16 = batch_norm(beta = obj_295_beta_0_to_fp16, epsilon = obj_295_epsilon_0_to_fp16, gamma = obj_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("obj_295_cast_fp16")]; + tensor query_85_pad_type_0 = const()[name = tensor("query_85_pad_type_0"), val = tensor("valid")]; + tensor query_85_strides_0 = const()[name = tensor("query_85_strides_0"), val = tensor([1, 1])]; + tensor query_85_pad_0 = const()[name = tensor("query_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_85_dilations_0 = const()[name = tensor("query_85_dilations_0"), val = tensor([1, 1])]; + tensor query_85_groups_0 = const()[name = tensor("query_85_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(812550080)))]; + tensor layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814647296)))]; + tensor query_85_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = query_85_dilations_0, groups = query_85_groups_0, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = query_85_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("query_85_cast_fp16")]; + tensor current_key_43_pad_type_0 = const()[name = tensor("current_key_43_pad_type_0"), val = tensor("valid")]; + tensor current_key_43_strides_0 = const()[name = tensor("current_key_43_strides_0"), val = tensor([1, 1])]; + tensor current_key_43_pad_0 = const()[name = tensor("current_key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_43_dilations_0 = const()[name = tensor("current_key_43_dilations_0"), val = tensor([1, 1])]; + tensor current_key_43_groups_0 = const()[name = tensor("current_key_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814649408)))]; + tensor current_key_43_cast_fp16 = conv(dilations = current_key_43_dilations_0, groups = current_key_43_groups_0, pad = current_key_43_pad_0, pad_type = current_key_43_pad_type_0, strides = current_key_43_strides_0, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_key_43_cast_fp16")]; + tensor current_value_43_pad_type_0 = const()[name = tensor("current_value_43_pad_type_0"), val = tensor("valid")]; + tensor current_value_43_strides_0 = const()[name = tensor("current_value_43_strides_0"), val = tensor([1, 1])]; + tensor current_value_43_pad_0 = const()[name = tensor("current_value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_43_dilations_0 = const()[name = tensor("current_value_43_dilations_0"), val = tensor([1, 1])]; + tensor current_value_43_groups_0 = const()[name = tensor("current_value_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816746624)))]; + tensor layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(818843840)))]; + tensor current_value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = current_value_43_dilations_0, groups = current_value_43_groups_0, pad = current_value_43_pad_0, pad_type = current_value_43_pad_type_0, strides = current_value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_value_43_cast_fp16")]; + tensor var_4835_cast_fp16 = mul(x = var_87_cast_fp16_21, y = var_207_cast_fp16)[name = tensor("op_4835_cast_fp16")]; + tensor var_4836_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4836_cast_fp16")]; + tensor key_85_cast_fp16 = add(x = var_4835_cast_fp16, y = var_4836_cast_fp16)[name = tensor("key_85_cast_fp16")]; + tensor var_4839_cast_fp16 = mul(x = var_114_cast_fp16_21, y = var_207_cast_fp16)[name = tensor("op_4839_cast_fp16")]; + tensor var_4840_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4840_cast_fp16")]; + tensor value_85_cast_fp16 = add(x = var_4839_cast_fp16, y = var_4840_cast_fp16)[name = tensor("value_85_cast_fp16")]; + tensor var_4844 = const()[name = tensor("op_4844"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_85_cast_fp16 = reshape(shape = var_4844, x = query_85_cast_fp16)[name = tensor("mh_q_85_cast_fp16")]; + tensor var_4846_to_fp16 = const()[name = tensor("op_4846_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4847_cast_fp16 = mul(x = mh_q_85_cast_fp16, y = var_4846_to_fp16)[name = tensor("op_4847_cast_fp16")]; + tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 16, 64, 448])]; + tensor var_4851_cast_fp16 = reshape(shape = var_4850, x = key_85_cast_fp16)[name = tensor("op_4851_cast_fp16")]; + tensor mh_w_127_transpose_x_0 = const()[name = tensor("mh_w_127_transpose_x_0"), val = tensor(true)]; + tensor mh_w_127_transpose_y_0 = const()[name = tensor("mh_w_127_transpose_y_0"), val = tensor(false)]; + tensor mh_w_127_cast_fp16 = matmul(transpose_x = mh_w_127_transpose_x_0, transpose_y = mh_w_127_transpose_y_0, x = var_4847_cast_fp16, y = var_4851_cast_fp16)[name = tensor("mh_w_127_cast_fp16")]; + tensor mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_129_cast_fp16")]; + tensor var_4859_cast_fp16 = softmax(axis = var_4771, x = mh_w_129_cast_fp16)[name = tensor("op_4859_cast_fp16")]; + tensor var_4860 = const()[name = tensor("op_4860"), val = tensor([1, 16, 64, 448])]; + tensor var_4861_cast_fp16 = reshape(shape = var_4860, x = value_85_cast_fp16)[name = tensor("op_4861_cast_fp16")]; + tensor attn_85_transpose_x_0 = const()[name = tensor("attn_85_transpose_x_0"), val = tensor(false)]; + tensor attn_85_transpose_y_0 = const()[name = tensor("attn_85_transpose_y_0"), val = tensor(true)]; + tensor attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_0, transpose_y = attn_85_transpose_y_0, x = var_4861_cast_fp16, y = var_4859_cast_fp16)[name = tensor("attn_85_cast_fp16")]; + tensor var_4864 = const()[name = tensor("op_4864"), val = tensor([1, 1024, 1, 1])]; + tensor input_211_cast_fp16 = reshape(shape = var_4864, x = attn_85_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor obj_301_pad_type_0 = const()[name = tensor("obj_301_pad_type_0"), val = tensor("valid")]; + tensor obj_301_strides_0 = const()[name = tensor("obj_301_strides_0"), val = tensor([1, 1])]; + tensor obj_301_pad_0 = const()[name = tensor("obj_301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_301_dilations_0 = const()[name = tensor("obj_301_dilations_0"), val = tensor([1, 1])]; + tensor obj_301_groups_0 = const()[name = tensor("obj_301_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(818845952)))]; + tensor layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(820943168)))]; + tensor obj_301_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = obj_301_dilations_0, groups = obj_301_groups_0, pad = obj_301_pad_0, pad_type = obj_301_pad_type_0, strides = obj_301_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("obj_301_cast_fp16")]; + tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_301_cast_fp16)[name = tensor("inputs_129_cast_fp16")]; + tensor out_129_axes_0 = const()[name = tensor("out_129_axes_0"), val = tensor([1])]; + tensor var_4886_to_fp16 = const()[name = tensor("op_4886_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_129_cast_fp16 = layer_norm(axes = out_129_axes_0, epsilon = var_4886_to_fp16, x = inputs_129_cast_fp16)[name = tensor("out_129_cast_fp16")]; + tensor obj_303_gamma_0_to_fp16 = const()[name = tensor("obj_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(820945280)))]; + tensor obj_303_beta_0_to_fp16 = const()[name = tensor("obj_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(820947392)))]; + tensor obj_303_epsilon_0_to_fp16 = const()[name = tensor("obj_303_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_303_cast_fp16 = batch_norm(beta = obj_303_beta_0_to_fp16, epsilon = obj_303_epsilon_0_to_fp16, gamma = obj_303_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor("obj_303_cast_fp16")]; + tensor query_87_pad_type_0 = const()[name = tensor("query_87_pad_type_0"), val = tensor("valid")]; + tensor query_87_strides_0 = const()[name = tensor("query_87_strides_0"), val = tensor([1, 1])]; + tensor query_87_pad_0 = const()[name = tensor("query_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_87_dilations_0 = const()[name = tensor("query_87_dilations_0"), val = tensor([1, 1])]; + tensor query_87_groups_0 = const()[name = tensor("query_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(820949504)))]; + tensor layers_21_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(823046720)))]; + tensor query_87_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_bias_to_fp16, dilations = query_87_dilations_0, groups = query_87_groups_0, pad = query_87_pad_0, pad_type = query_87_pad_type_0, strides = query_87_strides_0, weight = layers_21_encoder_attn_q_proj_weight_to_fp16, x = obj_303_cast_fp16)[name = tensor("query_87_cast_fp16")]; + tensor key_87_pad_type_0 = const()[name = tensor("key_87_pad_type_0"), val = tensor("valid")]; + tensor key_87_strides_0 = const()[name = tensor("key_87_strides_0"), val = tensor([1, 1])]; + tensor key_87_pad_0 = const()[name = tensor("key_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_87_dilations_0 = const()[name = tensor("key_87_dilations_0"), val = tensor([1, 1])]; + tensor key_87_groups_0 = const()[name = tensor("key_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(823048832)))]; + tensor key_87_cast_fp16 = conv(dilations = key_87_dilations_0, groups = key_87_groups_0, pad = key_87_pad_0, pad_type = key_87_pad_type_0, strides = key_87_strides_0, weight = layers_21_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_87_cast_fp16")]; + tensor value_87_pad_type_0 = const()[name = tensor("value_87_pad_type_0"), val = tensor("valid")]; + tensor value_87_strides_0 = const()[name = tensor("value_87_strides_0"), val = tensor([1, 1])]; + tensor value_87_pad_0 = const()[name = tensor("value_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_87_dilations_0 = const()[name = tensor("value_87_dilations_0"), val = tensor([1, 1])]; + tensor value_87_groups_0 = const()[name = tensor("value_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(825146048)))]; + tensor layers_21_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827243264)))]; + tensor value_87_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_bias_to_fp16, dilations = value_87_dilations_0, groups = value_87_groups_0, pad = value_87_pad_0, pad_type = value_87_pad_type_0, strides = value_87_strides_0, weight = layers_21_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_87_cast_fp16")]; + tensor var_4922 = const()[name = tensor("op_4922"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_87_cast_fp16 = reshape(shape = var_4922, x = query_87_cast_fp16)[name = tensor("mh_q_87_cast_fp16")]; + tensor var_4924_to_fp16 = const()[name = tensor("op_4924_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4925_cast_fp16 = mul(x = mh_q_87_cast_fp16, y = var_4924_to_fp16)[name = tensor("op_4925_cast_fp16")]; + tensor var_4928 = const()[name = tensor("op_4928"), val = tensor([1, 16, 64, 1500])]; + tensor var_4929_cast_fp16 = reshape(shape = var_4928, x = key_87_cast_fp16)[name = tensor("op_4929_cast_fp16")]; + tensor mh_w_131_transpose_x_0 = const()[name = tensor("mh_w_131_transpose_x_0"), val = tensor(true)]; + tensor mh_w_131_transpose_y_0 = const()[name = tensor("mh_w_131_transpose_y_0"), val = tensor(false)]; + tensor mh_w_131_cast_fp16 = matmul(transpose_x = mh_w_131_transpose_x_0, transpose_y = mh_w_131_transpose_y_0, x = var_4925_cast_fp16, y = var_4929_cast_fp16)[name = tensor("mh_w_131_cast_fp16")]; + tensor obj_307_cast_fp16 = softmax(axis = var_4771, x = mh_w_131_cast_fp16)[name = tensor("obj_307_cast_fp16")]; + tensor var_4933 = const()[name = tensor("op_4933"), val = tensor([1, 16, 64, 1500])]; + tensor var_4934_cast_fp16 = reshape(shape = var_4933, x = value_87_cast_fp16)[name = tensor("op_4934_cast_fp16")]; + tensor attn_87_transpose_x_0 = const()[name = tensor("attn_87_transpose_x_0"), val = tensor(false)]; + tensor attn_87_transpose_y_0 = const()[name = tensor("attn_87_transpose_y_0"), val = tensor(true)]; + tensor attn_87_cast_fp16 = matmul(transpose_x = attn_87_transpose_x_0, transpose_y = attn_87_transpose_y_0, x = var_4934_cast_fp16, y = obj_307_cast_fp16)[name = tensor("attn_87_cast_fp16")]; + tensor var_4937 = const()[name = tensor("op_4937"), val = tensor([1, 1024, 1, 1])]; + tensor input_213_cast_fp16 = reshape(shape = var_4937, x = attn_87_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor obj_305_pad_type_0 = const()[name = tensor("obj_305_pad_type_0"), val = tensor("valid")]; + tensor obj_305_strides_0 = const()[name = tensor("obj_305_strides_0"), val = tensor([1, 1])]; + tensor obj_305_pad_0 = const()[name = tensor("obj_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_305_dilations_0 = const()[name = tensor("obj_305_dilations_0"), val = tensor([1, 1])]; + tensor obj_305_groups_0 = const()[name = tensor("obj_305_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827245376)))]; + tensor layers_21_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829342592)))]; + tensor obj_305_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_bias_to_fp16, dilations = obj_305_dilations_0, groups = obj_305_groups_0, pad = obj_305_pad_0, pad_type = obj_305_pad_type_0, strides = obj_305_strides_0, weight = layers_21_encoder_attn_o_proj_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("obj_305_cast_fp16")]; + tensor inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_305_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; + tensor out_131_axes_0 = const()[name = tensor("out_131_axes_0"), val = tensor([1])]; + tensor var_4958_to_fp16 = const()[name = tensor("op_4958_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_131_cast_fp16 = layer_norm(axes = out_131_axes_0, epsilon = var_4958_to_fp16, x = inputs_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; + tensor input_215_gamma_0_to_fp16 = const()[name = tensor("input_215_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829344704)))]; + tensor input_215_beta_0_to_fp16 = const()[name = tensor("input_215_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829346816)))]; + tensor input_215_epsilon_0_to_fp16 = const()[name = tensor("input_215_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = batch_norm(beta = input_215_beta_0_to_fp16, epsilon = input_215_epsilon_0_to_fp16, gamma = input_215_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor input_217_pad_type_0 = const()[name = tensor("input_217_pad_type_0"), val = tensor("valid")]; + tensor input_217_strides_0 = const()[name = tensor("input_217_strides_0"), val = tensor([1, 1])]; + tensor input_217_pad_0 = const()[name = tensor("input_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_217_dilations_0 = const()[name = tensor("input_217_dilations_0"), val = tensor([1, 1])]; + tensor input_217_groups_0 = const()[name = tensor("input_217_groups_0"), val = tensor(1)]; + tensor layers_21_fc1_weight_to_fp16 = const()[name = tensor("layers_21_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829348928)))]; + tensor layers_21_fc1_bias_to_fp16 = const()[name = tensor("layers_21_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(837737600)))]; + tensor input_217_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = input_217_dilations_0, groups = input_217_groups_0, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = input_217_strides_0, weight = layers_21_fc1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_mode_0 = const()[name = tensor("input_219_mode_0"), val = tensor("EXACT")]; + tensor input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor hidden_states_45_pad_type_0 = const()[name = tensor("hidden_states_45_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_45_strides_0 = const()[name = tensor("hidden_states_45_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_45_pad_0 = const()[name = tensor("hidden_states_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_45_dilations_0 = const()[name = tensor("hidden_states_45_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_45_groups_0 = const()[name = tensor("hidden_states_45_groups_0"), val = tensor(1)]; + tensor layers_21_fc2_weight_to_fp16 = const()[name = tensor("layers_21_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(837745856)))]; + tensor layers_21_fc2_bias_to_fp16 = const()[name = tensor("layers_21_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846134528)))]; + tensor hidden_states_45_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = layers_21_fc2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("hidden_states_45_cast_fp16")]; + tensor inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_133_cast_fp16")]; + tensor var_4994 = const()[name = tensor("op_4994"), val = tensor(3)]; + tensor out_133_axes_0 = const()[name = tensor("out_133_axes_0"), val = tensor([1])]; + tensor var_5019_to_fp16 = const()[name = tensor("op_5019_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_133_cast_fp16 = layer_norm(axes = out_133_axes_0, epsilon = var_5019_to_fp16, x = inputs_133_cast_fp16)[name = tensor("out_133_cast_fp16")]; + tensor obj_309_gamma_0_to_fp16 = const()[name = tensor("obj_309_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846136640)))]; + tensor obj_309_beta_0_to_fp16 = const()[name = tensor("obj_309_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846138752)))]; + tensor obj_309_epsilon_0_to_fp16 = const()[name = tensor("obj_309_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_309_cast_fp16 = batch_norm(beta = obj_309_beta_0_to_fp16, epsilon = obj_309_epsilon_0_to_fp16, gamma = obj_309_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor("obj_309_cast_fp16")]; + tensor query_89_pad_type_0 = const()[name = tensor("query_89_pad_type_0"), val = tensor("valid")]; + tensor query_89_strides_0 = const()[name = tensor("query_89_strides_0"), val = tensor([1, 1])]; + tensor query_89_pad_0 = const()[name = tensor("query_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_89_dilations_0 = const()[name = tensor("query_89_dilations_0"), val = tensor([1, 1])]; + tensor query_89_groups_0 = const()[name = tensor("query_89_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846140864)))]; + tensor layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(848238080)))]; + tensor query_89_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = query_89_dilations_0, groups = query_89_groups_0, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = query_89_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("query_89_cast_fp16")]; + tensor current_key_45_pad_type_0 = const()[name = tensor("current_key_45_pad_type_0"), val = tensor("valid")]; + tensor current_key_45_strides_0 = const()[name = tensor("current_key_45_strides_0"), val = tensor([1, 1])]; + tensor current_key_45_pad_0 = const()[name = tensor("current_key_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_45_dilations_0 = const()[name = tensor("current_key_45_dilations_0"), val = tensor([1, 1])]; + tensor current_key_45_groups_0 = const()[name = tensor("current_key_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(848240192)))]; + tensor current_key_45_cast_fp16 = conv(dilations = current_key_45_dilations_0, groups = current_key_45_groups_0, pad = current_key_45_pad_0, pad_type = current_key_45_pad_type_0, strides = current_key_45_strides_0, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_key_45_cast_fp16")]; + tensor current_value_45_pad_type_0 = const()[name = tensor("current_value_45_pad_type_0"), val = tensor("valid")]; + tensor current_value_45_strides_0 = const()[name = tensor("current_value_45_strides_0"), val = tensor([1, 1])]; + tensor current_value_45_pad_0 = const()[name = tensor("current_value_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_45_dilations_0 = const()[name = tensor("current_value_45_dilations_0"), val = tensor([1, 1])]; + tensor current_value_45_groups_0 = const()[name = tensor("current_value_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850337408)))]; + tensor layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(852434624)))]; + tensor current_value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = current_value_45_dilations_0, groups = current_value_45_groups_0, pad = current_value_45_pad_0, pad_type = current_value_45_pad_type_0, strides = current_value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_value_45_cast_fp16")]; + tensor var_5058_cast_fp16 = mul(x = var_87_cast_fp16_22, y = var_207_cast_fp16)[name = tensor("op_5058_cast_fp16")]; + tensor var_5059_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_5059_cast_fp16")]; + tensor key_89_cast_fp16 = add(x = var_5058_cast_fp16, y = var_5059_cast_fp16)[name = tensor("key_89_cast_fp16")]; + tensor var_5062_cast_fp16 = mul(x = var_114_cast_fp16_22, y = var_207_cast_fp16)[name = tensor("op_5062_cast_fp16")]; + tensor var_5063_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor value_89_cast_fp16 = add(x = var_5062_cast_fp16, y = var_5063_cast_fp16)[name = tensor("value_89_cast_fp16")]; + tensor var_5067 = const()[name = tensor("op_5067"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_89_cast_fp16 = reshape(shape = var_5067, x = query_89_cast_fp16)[name = tensor("mh_q_89_cast_fp16")]; + tensor var_5069_to_fp16 = const()[name = tensor("op_5069_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5070_cast_fp16 = mul(x = mh_q_89_cast_fp16, y = var_5069_to_fp16)[name = tensor("op_5070_cast_fp16")]; + tensor var_5073 = const()[name = tensor("op_5073"), val = tensor([1, 16, 64, 448])]; + tensor var_5074_cast_fp16 = reshape(shape = var_5073, x = key_89_cast_fp16)[name = tensor("op_5074_cast_fp16")]; + tensor mh_w_133_transpose_x_0 = const()[name = tensor("mh_w_133_transpose_x_0"), val = tensor(true)]; + tensor mh_w_133_transpose_y_0 = const()[name = tensor("mh_w_133_transpose_y_0"), val = tensor(false)]; + tensor mh_w_133_cast_fp16 = matmul(transpose_x = mh_w_133_transpose_x_0, transpose_y = mh_w_133_transpose_y_0, x = var_5070_cast_fp16, y = var_5074_cast_fp16)[name = tensor("mh_w_133_cast_fp16")]; + tensor mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_135_cast_fp16")]; + tensor var_5082_cast_fp16 = softmax(axis = var_4994, x = mh_w_135_cast_fp16)[name = tensor("op_5082_cast_fp16")]; + tensor var_5083 = const()[name = tensor("op_5083"), val = tensor([1, 16, 64, 448])]; + tensor var_5084_cast_fp16 = reshape(shape = var_5083, x = value_89_cast_fp16)[name = tensor("op_5084_cast_fp16")]; + tensor attn_89_transpose_x_0 = const()[name = tensor("attn_89_transpose_x_0"), val = tensor(false)]; + tensor attn_89_transpose_y_0 = const()[name = tensor("attn_89_transpose_y_0"), val = tensor(true)]; + tensor attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_0, transpose_y = attn_89_transpose_y_0, x = var_5084_cast_fp16, y = var_5082_cast_fp16)[name = tensor("attn_89_cast_fp16")]; + tensor var_5087 = const()[name = tensor("op_5087"), val = tensor([1, 1024, 1, 1])]; + tensor input_221_cast_fp16 = reshape(shape = var_5087, x = attn_89_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor obj_315_pad_type_0 = const()[name = tensor("obj_315_pad_type_0"), val = tensor("valid")]; + tensor obj_315_strides_0 = const()[name = tensor("obj_315_strides_0"), val = tensor([1, 1])]; + tensor obj_315_pad_0 = const()[name = tensor("obj_315_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_315_dilations_0 = const()[name = tensor("obj_315_dilations_0"), val = tensor([1, 1])]; + tensor obj_315_groups_0 = const()[name = tensor("obj_315_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(852436736)))]; + tensor layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854533952)))]; + tensor obj_315_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = obj_315_dilations_0, groups = obj_315_groups_0, pad = obj_315_pad_0, pad_type = obj_315_pad_type_0, strides = obj_315_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("obj_315_cast_fp16")]; + tensor inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_315_cast_fp16)[name = tensor("inputs_135_cast_fp16")]; + tensor out_135_axes_0 = const()[name = tensor("out_135_axes_0"), val = tensor([1])]; + tensor var_5109_to_fp16 = const()[name = tensor("op_5109_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_135_cast_fp16 = layer_norm(axes = out_135_axes_0, epsilon = var_5109_to_fp16, x = inputs_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; + tensor obj_317_gamma_0_to_fp16 = const()[name = tensor("obj_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854536064)))]; + tensor obj_317_beta_0_to_fp16 = const()[name = tensor("obj_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854538176)))]; + tensor obj_317_epsilon_0_to_fp16 = const()[name = tensor("obj_317_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_317_cast_fp16 = batch_norm(beta = obj_317_beta_0_to_fp16, epsilon = obj_317_epsilon_0_to_fp16, gamma = obj_317_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor("obj_317_cast_fp16")]; + tensor query_91_pad_type_0 = const()[name = tensor("query_91_pad_type_0"), val = tensor("valid")]; + tensor query_91_strides_0 = const()[name = tensor("query_91_strides_0"), val = tensor([1, 1])]; + tensor query_91_pad_0 = const()[name = tensor("query_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_91_dilations_0 = const()[name = tensor("query_91_dilations_0"), val = tensor([1, 1])]; + tensor query_91_groups_0 = const()[name = tensor("query_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854540288)))]; + tensor layers_22_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(856637504)))]; + tensor query_91_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_bias_to_fp16, dilations = query_91_dilations_0, groups = query_91_groups_0, pad = query_91_pad_0, pad_type = query_91_pad_type_0, strides = query_91_strides_0, weight = layers_22_encoder_attn_q_proj_weight_to_fp16, x = obj_317_cast_fp16)[name = tensor("query_91_cast_fp16")]; + tensor key_91_pad_type_0 = const()[name = tensor("key_91_pad_type_0"), val = tensor("valid")]; + tensor key_91_strides_0 = const()[name = tensor("key_91_strides_0"), val = tensor([1, 1])]; + tensor key_91_pad_0 = const()[name = tensor("key_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_91_dilations_0 = const()[name = tensor("key_91_dilations_0"), val = tensor([1, 1])]; + tensor key_91_groups_0 = const()[name = tensor("key_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(856639616)))]; + tensor key_91_cast_fp16 = conv(dilations = key_91_dilations_0, groups = key_91_groups_0, pad = key_91_pad_0, pad_type = key_91_pad_type_0, strides = key_91_strides_0, weight = layers_22_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_91_cast_fp16")]; + tensor value_91_pad_type_0 = const()[name = tensor("value_91_pad_type_0"), val = tensor("valid")]; + tensor value_91_strides_0 = const()[name = tensor("value_91_strides_0"), val = tensor([1, 1])]; + tensor value_91_pad_0 = const()[name = tensor("value_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_91_dilations_0 = const()[name = tensor("value_91_dilations_0"), val = tensor([1, 1])]; + tensor value_91_groups_0 = const()[name = tensor("value_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(858736832)))]; + tensor layers_22_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(860834048)))]; + tensor value_91_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_bias_to_fp16, dilations = value_91_dilations_0, groups = value_91_groups_0, pad = value_91_pad_0, pad_type = value_91_pad_type_0, strides = value_91_strides_0, weight = layers_22_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_91_cast_fp16")]; + tensor var_5145 = const()[name = tensor("op_5145"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_91_cast_fp16 = reshape(shape = var_5145, x = query_91_cast_fp16)[name = tensor("mh_q_91_cast_fp16")]; + tensor var_5147_to_fp16 = const()[name = tensor("op_5147_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5148_cast_fp16 = mul(x = mh_q_91_cast_fp16, y = var_5147_to_fp16)[name = tensor("op_5148_cast_fp16")]; + tensor var_5151 = const()[name = tensor("op_5151"), val = tensor([1, 16, 64, 1500])]; + tensor var_5152_cast_fp16 = reshape(shape = var_5151, x = key_91_cast_fp16)[name = tensor("op_5152_cast_fp16")]; + tensor mh_w_137_transpose_x_0 = const()[name = tensor("mh_w_137_transpose_x_0"), val = tensor(true)]; + tensor mh_w_137_transpose_y_0 = const()[name = tensor("mh_w_137_transpose_y_0"), val = tensor(false)]; + tensor mh_w_137_cast_fp16 = matmul(transpose_x = mh_w_137_transpose_x_0, transpose_y = mh_w_137_transpose_y_0, x = var_5148_cast_fp16, y = var_5152_cast_fp16)[name = tensor("mh_w_137_cast_fp16")]; + tensor obj_321_cast_fp16 = softmax(axis = var_4994, x = mh_w_137_cast_fp16)[name = tensor("obj_321_cast_fp16")]; + tensor var_5156 = const()[name = tensor("op_5156"), val = tensor([1, 16, 64, 1500])]; + tensor var_5157_cast_fp16 = reshape(shape = var_5156, x = value_91_cast_fp16)[name = tensor("op_5157_cast_fp16")]; + tensor attn_91_transpose_x_0 = const()[name = tensor("attn_91_transpose_x_0"), val = tensor(false)]; + tensor attn_91_transpose_y_0 = const()[name = tensor("attn_91_transpose_y_0"), val = tensor(true)]; + tensor attn_91_cast_fp16 = matmul(transpose_x = attn_91_transpose_x_0, transpose_y = attn_91_transpose_y_0, x = var_5157_cast_fp16, y = obj_321_cast_fp16)[name = tensor("attn_91_cast_fp16")]; + tensor var_5160 = const()[name = tensor("op_5160"), val = tensor([1, 1024, 1, 1])]; + tensor input_223_cast_fp16 = reshape(shape = var_5160, x = attn_91_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor obj_319_pad_type_0 = const()[name = tensor("obj_319_pad_type_0"), val = tensor("valid")]; + tensor obj_319_strides_0 = const()[name = tensor("obj_319_strides_0"), val = tensor([1, 1])]; + tensor obj_319_pad_0 = const()[name = tensor("obj_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_319_dilations_0 = const()[name = tensor("obj_319_dilations_0"), val = tensor([1, 1])]; + tensor obj_319_groups_0 = const()[name = tensor("obj_319_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(860836160)))]; + tensor layers_22_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(862933376)))]; + tensor obj_319_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_bias_to_fp16, dilations = obj_319_dilations_0, groups = obj_319_groups_0, pad = obj_319_pad_0, pad_type = obj_319_pad_type_0, strides = obj_319_strides_0, weight = layers_22_encoder_attn_o_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("obj_319_cast_fp16")]; + tensor inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_319_cast_fp16)[name = tensor("inputs_137_cast_fp16")]; + tensor out_137_axes_0 = const()[name = tensor("out_137_axes_0"), val = tensor([1])]; + tensor var_5178_to_fp16 = const()[name = tensor("op_5178_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_137_cast_fp16 = layer_norm(axes = out_137_axes_0, epsilon = var_5178_to_fp16, x = inputs_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; + tensor input_225_gamma_0_to_fp16 = const()[name = tensor("input_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(862935488)))]; + tensor input_225_beta_0_to_fp16 = const()[name = tensor("input_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(862937600)))]; + tensor input_225_epsilon_0_to_fp16 = const()[name = tensor("input_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = batch_norm(beta = input_225_beta_0_to_fp16, epsilon = input_225_epsilon_0_to_fp16, gamma = input_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_pad_type_0 = const()[name = tensor("input_227_pad_type_0"), val = tensor("valid")]; + tensor input_227_strides_0 = const()[name = tensor("input_227_strides_0"), val = tensor([1, 1])]; + tensor input_227_pad_0 = const()[name = tensor("input_227_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_227_dilations_0 = const()[name = tensor("input_227_dilations_0"), val = tensor([1, 1])]; + tensor input_227_groups_0 = const()[name = tensor("input_227_groups_0"), val = tensor(1)]; + tensor layers_22_fc1_weight_to_fp16 = const()[name = tensor("layers_22_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(862939712)))]; + tensor layers_22_fc1_bias_to_fp16 = const()[name = tensor("layers_22_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(871328384)))]; + tensor input_227_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = input_227_dilations_0, groups = input_227_groups_0, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = input_227_strides_0, weight = layers_22_fc1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor hidden_states_47_pad_type_0 = const()[name = tensor("hidden_states_47_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_47_strides_0 = const()[name = tensor("hidden_states_47_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_47_pad_0 = const()[name = tensor("hidden_states_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_47_dilations_0 = const()[name = tensor("hidden_states_47_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_47_groups_0 = const()[name = tensor("hidden_states_47_groups_0"), val = tensor(1)]; + tensor layers_22_fc2_weight_to_fp16 = const()[name = tensor("layers_22_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(871336640)))]; + tensor layers_22_fc2_bias_to_fp16 = const()[name = tensor("layers_22_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(879725312)))]; + tensor hidden_states_47_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_22_fc2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("hidden_states_47_cast_fp16")]; + tensor inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_139_cast_fp16")]; + tensor var_5213 = const()[name = tensor("op_5213"), val = tensor(3)]; + tensor out_139_axes_0 = const()[name = tensor("out_139_axes_0"), val = tensor([1])]; + tensor var_5238_to_fp16 = const()[name = tensor("op_5238_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_139_cast_fp16 = layer_norm(axes = out_139_axes_0, epsilon = var_5238_to_fp16, x = inputs_139_cast_fp16)[name = tensor("out_139_cast_fp16")]; + tensor obj_323_gamma_0_to_fp16 = const()[name = tensor("obj_323_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(879727424)))]; + tensor obj_323_beta_0_to_fp16 = const()[name = tensor("obj_323_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(879729536)))]; + tensor obj_323_epsilon_0_to_fp16 = const()[name = tensor("obj_323_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_323_cast_fp16 = batch_norm(beta = obj_323_beta_0_to_fp16, epsilon = obj_323_epsilon_0_to_fp16, gamma = obj_323_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor("obj_323_cast_fp16")]; + tensor query_93_pad_type_0 = const()[name = tensor("query_93_pad_type_0"), val = tensor("valid")]; + tensor query_93_strides_0 = const()[name = tensor("query_93_strides_0"), val = tensor([1, 1])]; + tensor query_93_pad_0 = const()[name = tensor("query_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_93_dilations_0 = const()[name = tensor("query_93_dilations_0"), val = tensor([1, 1])]; + tensor query_93_groups_0 = const()[name = tensor("query_93_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(879731648)))]; + tensor layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881828864)))]; + tensor query_93_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = query_93_dilations_0, groups = query_93_groups_0, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = query_93_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("query_93_cast_fp16")]; + tensor current_key_pad_type_0 = const()[name = tensor("current_key_pad_type_0"), val = tensor("valid")]; + tensor current_key_strides_0 = const()[name = tensor("current_key_strides_0"), val = tensor([1, 1])]; + tensor current_key_pad_0 = const()[name = tensor("current_key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_dilations_0 = const()[name = tensor("current_key_dilations_0"), val = tensor([1, 1])]; + tensor current_key_groups_0 = const()[name = tensor("current_key_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881830976)))]; + tensor current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor current_value_pad_type_0 = const()[name = tensor("current_value_pad_type_0"), val = tensor("valid")]; + tensor current_value_strides_0 = const()[name = tensor("current_value_strides_0"), val = tensor([1, 1])]; + tensor current_value_pad_0 = const()[name = tensor("current_value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_dilations_0 = const()[name = tensor("current_value_dilations_0"), val = tensor([1, 1])]; + tensor current_value_groups_0 = const()[name = tensor("current_value_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883928192)))]; + tensor layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886025408)))]; + tensor current_value_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_5277_cast_fp16 = mul(x = var_87_cast_fp16_23, y = var_207_cast_fp16)[name = tensor("op_5277_cast_fp16")]; + tensor var_5278_cast_fp16 = mul(x = current_key_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_5278_cast_fp16")]; + tensor key_93_cast_fp16 = add(x = var_5277_cast_fp16, y = var_5278_cast_fp16)[name = tensor("key_93_cast_fp16")]; + tensor var_5281_cast_fp16 = mul(x = var_114_cast_fp16_23, y = var_207_cast_fp16)[name = tensor("op_5281_cast_fp16")]; + tensor var_5282_cast_fp16 = mul(x = current_value_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_5282_cast_fp16")]; + tensor value_93_cast_fp16 = add(x = var_5281_cast_fp16, y = var_5282_cast_fp16)[name = tensor("value_93_cast_fp16")]; + tensor var_5286 = const()[name = tensor("op_5286"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_93_cast_fp16 = reshape(shape = var_5286, x = query_93_cast_fp16)[name = tensor("mh_q_93_cast_fp16")]; + tensor var_5288_to_fp16 = const()[name = tensor("op_5288_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5289_cast_fp16 = mul(x = mh_q_93_cast_fp16, y = var_5288_to_fp16)[name = tensor("op_5289_cast_fp16")]; + tensor var_5292 = const()[name = tensor("op_5292"), val = tensor([1, 16, 64, 448])]; + tensor var_5293_cast_fp16 = reshape(shape = var_5292, x = key_93_cast_fp16)[name = tensor("op_5293_cast_fp16")]; + tensor mh_w_139_transpose_x_0 = const()[name = tensor("mh_w_139_transpose_x_0"), val = tensor(true)]; + tensor mh_w_139_transpose_y_0 = const()[name = tensor("mh_w_139_transpose_y_0"), val = tensor(false)]; + tensor mh_w_139_cast_fp16 = matmul(transpose_x = mh_w_139_transpose_x_0, transpose_y = mh_w_139_transpose_y_0, x = var_5289_cast_fp16, y = var_5293_cast_fp16)[name = tensor("mh_w_139_cast_fp16")]; + tensor mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_229_cast_fp16)[name = tensor("mh_w_141_cast_fp16")]; + tensor var_5301_cast_fp16 = softmax(axis = var_5213, x = mh_w_141_cast_fp16)[name = tensor("op_5301_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 16, 64, 448])]; + tensor var_5303_cast_fp16 = reshape(shape = var_5302, x = value_93_cast_fp16)[name = tensor("op_5303_cast_fp16")]; + tensor attn_93_transpose_x_0 = const()[name = tensor("attn_93_transpose_x_0"), val = tensor(false)]; + tensor attn_93_transpose_y_0 = const()[name = tensor("attn_93_transpose_y_0"), val = tensor(true)]; + tensor attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_0, transpose_y = attn_93_transpose_y_0, x = var_5303_cast_fp16, y = var_5301_cast_fp16)[name = tensor("attn_93_cast_fp16")]; + tensor var_5306 = const()[name = tensor("op_5306"), val = tensor([1, 1024, 1, 1])]; + tensor input_231_cast_fp16 = reshape(shape = var_5306, x = attn_93_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor obj_329_pad_type_0 = const()[name = tensor("obj_329_pad_type_0"), val = tensor("valid")]; + tensor obj_329_strides_0 = const()[name = tensor("obj_329_strides_0"), val = tensor([1, 1])]; + tensor obj_329_pad_0 = const()[name = tensor("obj_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_329_dilations_0 = const()[name = tensor("obj_329_dilations_0"), val = tensor([1, 1])]; + tensor obj_329_groups_0 = const()[name = tensor("obj_329_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886027520)))]; + tensor layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888124736)))]; + tensor obj_329_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = obj_329_dilations_0, groups = obj_329_groups_0, pad = obj_329_pad_0, pad_type = obj_329_pad_type_0, strides = obj_329_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("obj_329_cast_fp16")]; + tensor inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_329_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; + tensor out_141_axes_0 = const()[name = tensor("out_141_axes_0"), val = tensor([1])]; + tensor var_5328_to_fp16 = const()[name = tensor("op_5328_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_141_cast_fp16 = layer_norm(axes = out_141_axes_0, epsilon = var_5328_to_fp16, x = inputs_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; + tensor obj_331_gamma_0_to_fp16 = const()[name = tensor("obj_331_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888126848)))]; + tensor obj_331_beta_0_to_fp16 = const()[name = tensor("obj_331_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888128960)))]; + tensor obj_331_epsilon_0_to_fp16 = const()[name = tensor("obj_331_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_331_cast_fp16 = batch_norm(beta = obj_331_beta_0_to_fp16, epsilon = obj_331_epsilon_0_to_fp16, gamma = obj_331_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor("obj_331_cast_fp16")]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("valid")]; + tensor query_strides_0 = const()[name = tensor("query_strides_0"), val = tensor([1, 1])]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_dilations_0 = const()[name = tensor("query_dilations_0"), val = tensor([1, 1])]; + tensor query_groups_0 = const()[name = tensor("query_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888131072)))]; + tensor layers_23_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890228288)))]; + tensor query_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_23_encoder_attn_q_proj_weight_to_fp16, x = obj_331_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("valid")]; + tensor key_strides_0 = const()[name = tensor("key_strides_0"), val = tensor([1, 1])]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_dilations_0 = const()[name = tensor("key_dilations_0"), val = tensor([1, 1])]; + tensor key_groups_0 = const()[name = tensor("key_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890230400)))]; + tensor key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_23_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("valid")]; + tensor value_strides_0 = const()[name = tensor("value_strides_0"), val = tensor([1, 1])]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_dilations_0 = const()[name = tensor("value_dilations_0"), val = tensor([1, 1])]; + tensor value_groups_0 = const()[name = tensor("value_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(892327616)))]; + tensor layers_23_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894424832)))]; + tensor value_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_23_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_5364 = const()[name = tensor("op_5364"), val = tensor([1, 16, 64, 1])]; + tensor mh_q_cast_fp16 = reshape(shape = var_5364, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_5366_to_fp16 = const()[name = tensor("op_5366_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5367_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_5366_to_fp16)[name = tensor("op_5367_cast_fp16")]; + tensor var_5370 = const()[name = tensor("op_5370"), val = tensor([1, 16, 64, 1500])]; + tensor var_5371_cast_fp16 = reshape(shape = var_5370, x = key_cast_fp16)[name = tensor("op_5371_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_5367_cast_fp16, y = var_5371_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_335_cast_fp16 = softmax(axis = var_5213, x = mh_w_cast_fp16)[name = tensor("obj_335_cast_fp16")]; + tensor var_5375 = const()[name = tensor("op_5375"), val = tensor([1, 16, 64, 1500])]; + tensor var_5376_cast_fp16 = reshape(shape = var_5375, x = value_cast_fp16)[name = tensor("op_5376_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_5376_cast_fp16, y = obj_335_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_5379 = const()[name = tensor("op_5379"), val = tensor([1, 1024, 1, 1])]; + tensor input_233_cast_fp16 = reshape(shape = var_5379, x = attn_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor obj_333_pad_type_0 = const()[name = tensor("obj_333_pad_type_0"), val = tensor("valid")]; + tensor obj_333_strides_0 = const()[name = tensor("obj_333_strides_0"), val = tensor([1, 1])]; + tensor obj_333_pad_0 = const()[name = tensor("obj_333_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_333_dilations_0 = const()[name = tensor("obj_333_dilations_0"), val = tensor([1, 1])]; + tensor obj_333_groups_0 = const()[name = tensor("obj_333_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894426944)))]; + tensor layers_23_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(896524160)))]; + tensor obj_333_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_bias_to_fp16, dilations = obj_333_dilations_0, groups = obj_333_groups_0, pad = obj_333_pad_0, pad_type = obj_333_pad_type_0, strides = obj_333_strides_0, weight = layers_23_encoder_attn_o_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("obj_333_cast_fp16")]; + tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_333_cast_fp16)[name = tensor("inputs_143_cast_fp16")]; + tensor out_143_axes_0 = const()[name = tensor("out_143_axes_0"), val = tensor([1])]; + tensor var_5397_to_fp16 = const()[name = tensor("op_5397_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_143_cast_fp16 = layer_norm(axes = out_143_axes_0, epsilon = var_5397_to_fp16, x = inputs_143_cast_fp16)[name = tensor("out_143_cast_fp16")]; + tensor input_235_gamma_0_to_fp16 = const()[name = tensor("input_235_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(896526272)))]; + tensor input_235_beta_0_to_fp16 = const()[name = tensor("input_235_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(896528384)))]; + tensor input_235_epsilon_0_to_fp16 = const()[name = tensor("input_235_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_pad_type_0 = const()[name = tensor("input_237_pad_type_0"), val = tensor("valid")]; + tensor input_237_strides_0 = const()[name = tensor("input_237_strides_0"), val = tensor([1, 1])]; + tensor input_237_pad_0 = const()[name = tensor("input_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_237_dilations_0 = const()[name = tensor("input_237_dilations_0"), val = tensor([1, 1])]; + tensor input_237_groups_0 = const()[name = tensor("input_237_groups_0"), val = tensor(1)]; + tensor layers_23_fc1_weight_to_fp16 = const()[name = tensor("layers_23_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(896530496)))]; + tensor layers_23_fc1_bias_to_fp16 = const()[name = tensor("layers_23_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(904919168)))]; + tensor input_237_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = layers_23_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_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_237_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor hidden_states_49_pad_type_0 = const()[name = tensor("hidden_states_49_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_49_strides_0 = const()[name = tensor("hidden_states_49_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_49_pad_0 = const()[name = tensor("hidden_states_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_49_dilations_0 = const()[name = tensor("hidden_states_49_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_49_groups_0 = const()[name = tensor("hidden_states_49_groups_0"), val = tensor(1)]; + tensor layers_23_fc2_weight_to_fp16 = const()[name = tensor("layers_23_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(904927424)))]; + tensor layers_23_fc2_bias_to_fp16 = const()[name = tensor("layers_23_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(913316096)))]; + tensor hidden_states_49_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = hidden_states_49_dilations_0, groups = hidden_states_49_groups_0, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = hidden_states_49_strides_0, weight = layers_23_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_143_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; + tensor var_5439_to_fp16 = const()[name = tensor("op_5439_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5439_to_fp16, x = inputs_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(913318208)))]; + 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(913320320)))]; + 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_5450_axes_0 = const()[name = tensor("op_5450_axes_0"), val = tensor([2])]; + tensor var_5450_cast_fp16 = squeeze(axes = var_5450_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_5450_cast_fp16")]; + tensor var_5453_perm_0 = const()[name = tensor("op_5453_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(913322432)))]; + tensor var_5453_cast_fp16 = transpose(perm = var_5453_perm_0, x = var_5450_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_5453_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_5457 = const()[name = tensor("op_5457"), val = tensor(1)]; + tensor obj_339_interleave_0 = const()[name = tensor("obj_339_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_5457, interleave = obj_339_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_23_cast_fp16, current_key_25_cast_fp16, current_key_27_cast_fp16, current_key_29_cast_fp16, current_key_31_cast_fp16, current_key_33_cast_fp16, current_key_35_cast_fp16, current_key_37_cast_fp16, current_key_39_cast_fp16, current_key_41_cast_fp16, current_key_43_cast_fp16, current_key_45_cast_fp16, current_key_cast_fp16))[name = tensor("obj_339_cast_fp16")]; + tensor var_5460 = const()[name = tensor("op_5460"), val = tensor(1)]; + tensor obj_341_interleave_0 = const()[name = tensor("obj_341_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_5460, interleave = obj_341_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_23_cast_fp16, current_value_25_cast_fp16, current_value_27_cast_fp16, current_value_29_cast_fp16, current_value_31_cast_fp16, current_value_33_cast_fp16, current_value_35_cast_fp16, current_value_37_cast_fp16, current_value_39_cast_fp16, current_value_41_cast_fp16, current_value_43_cast_fp16, current_value_45_cast_fp16, current_value_cast_fp16))[name = tensor("obj_341_cast_fp16")]; + tensor var_5471_begin_0 = const()[name = tensor("op_5471_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5471_end_0 = const()[name = tensor("op_5471_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_5471_end_mask_0 = const()[name = tensor("op_5471_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5471_cast_fp16 = slice_by_index(begin = var_5471_begin_0, end = var_5471_end_0, end_mask = var_5471_end_mask_0, x = obj_167_cast_fp16)[name = tensor("op_5471_cast_fp16")]; + tensor var_5474_begin_0 = const()[name = tensor("op_5474_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5474_end_0 = const()[name = tensor("op_5474_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5474_end_mask_0 = const()[name = tensor("op_5474_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5474_squeeze_mask_0 = const()[name = tensor("op_5474_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5474_cast_fp16 = slice_by_index(begin = var_5474_begin_0, end = var_5474_end_0, end_mask = var_5474_end_mask_0, squeeze_mask = var_5474_squeeze_mask_0, x = var_5471_cast_fp16)[name = tensor("op_5474_cast_fp16")]; + tensor var_5489_begin_0 = const()[name = tensor("op_5489_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_5489_end_0 = const()[name = tensor("op_5489_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_5489_end_mask_0 = const()[name = tensor("op_5489_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5489_cast_fp16 = slice_by_index(begin = var_5489_begin_0, end = var_5489_end_0, end_mask = var_5489_end_mask_0, x = obj_209_cast_fp16)[name = tensor("op_5489_cast_fp16")]; + tensor var_5492_begin_0 = const()[name = tensor("op_5492_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5492_end_0 = const()[name = tensor("op_5492_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5492_end_mask_0 = const()[name = tensor("op_5492_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5492_squeeze_mask_0 = const()[name = tensor("op_5492_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5492_cast_fp16 = slice_by_index(begin = var_5492_begin_0, end = var_5492_end_0, end_mask = var_5492_end_mask_0, squeeze_mask = var_5492_squeeze_mask_0, x = var_5489_cast_fp16)[name = tensor("op_5492_cast_fp16")]; + tensor var_5507_begin_0 = const()[name = tensor("op_5507_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5507_end_0 = const()[name = tensor("op_5507_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_5507_end_mask_0 = const()[name = tensor("op_5507_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5507_cast_fp16 = slice_by_index(begin = var_5507_begin_0, end = var_5507_end_0, end_mask = var_5507_end_mask_0, x = obj_209_cast_fp16)[name = tensor("op_5507_cast_fp16")]; + tensor var_5510_begin_0 = const()[name = tensor("op_5510_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5510_end_0 = const()[name = tensor("op_5510_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5510_end_mask_0 = const()[name = tensor("op_5510_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5510_squeeze_mask_0 = const()[name = tensor("op_5510_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5510_cast_fp16 = slice_by_index(begin = var_5510_begin_0, end = var_5510_end_0, end_mask = var_5510_end_mask_0, squeeze_mask = var_5510_squeeze_mask_0, x = var_5507_cast_fp16)[name = tensor("op_5510_cast_fp16")]; + tensor var_5525_begin_0 = const()[name = tensor("op_5525_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5525_end_0 = const()[name = tensor("op_5525_end_0"), val = tensor([1, 15, 1, 1500])]; + tensor var_5525_end_mask_0 = const()[name = tensor("op_5525_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5525_cast_fp16 = slice_by_index(begin = var_5525_begin_0, end = var_5525_end_0, end_mask = var_5525_end_mask_0, x = obj_209_cast_fp16)[name = tensor("op_5525_cast_fp16")]; + tensor var_5528_begin_0 = const()[name = tensor("op_5528_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5528_end_0 = const()[name = tensor("op_5528_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5528_end_mask_0 = const()[name = tensor("op_5528_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5528_squeeze_mask_0 = const()[name = tensor("op_5528_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5528_cast_fp16 = slice_by_index(begin = var_5528_begin_0, end = var_5528_end_0, end_mask = var_5528_end_mask_0, squeeze_mask = var_5528_squeeze_mask_0, x = var_5525_cast_fp16)[name = tensor("op_5528_cast_fp16")]; + tensor var_5543_begin_0 = const()[name = tensor("op_5543_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5543_end_0 = const()[name = tensor("op_5543_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_5543_end_mask_0 = const()[name = tensor("op_5543_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5543_cast_fp16 = slice_by_index(begin = var_5543_begin_0, end = var_5543_end_0, end_mask = var_5543_end_mask_0, x = obj_223_cast_fp16)[name = tensor("op_5543_cast_fp16")]; + tensor var_5546_begin_0 = const()[name = tensor("op_5546_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5546_end_0 = const()[name = tensor("op_5546_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5546_end_mask_0 = const()[name = tensor("op_5546_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5546_squeeze_mask_0 = const()[name = tensor("op_5546_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5546_cast_fp16 = slice_by_index(begin = var_5546_begin_0, end = var_5546_end_0, end_mask = var_5546_end_mask_0, squeeze_mask = var_5546_squeeze_mask_0, x = var_5543_cast_fp16)[name = tensor("op_5546_cast_fp16")]; + tensor var_5561_begin_0 = const()[name = tensor("op_5561_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5561_end_0 = const()[name = tensor("op_5561_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5561_end_mask_0 = const()[name = tensor("op_5561_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5561_cast_fp16 = slice_by_index(begin = var_5561_begin_0, end = var_5561_end_0, end_mask = var_5561_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_5561_cast_fp16")]; + tensor var_5564_begin_0 = const()[name = tensor("op_5564_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5564_end_0 = const()[name = tensor("op_5564_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5564_end_mask_0 = const()[name = tensor("op_5564_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5564_squeeze_mask_0 = const()[name = tensor("op_5564_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5564_cast_fp16 = slice_by_index(begin = var_5564_begin_0, end = var_5564_end_0, end_mask = var_5564_end_mask_0, squeeze_mask = var_5564_squeeze_mask_0, x = var_5561_cast_fp16)[name = tensor("op_5564_cast_fp16")]; + tensor var_5579_begin_0 = const()[name = tensor("op_5579_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5579_end_0 = const()[name = tensor("op_5579_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_5579_end_mask_0 = const()[name = tensor("op_5579_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5579_cast_fp16 = slice_by_index(begin = var_5579_begin_0, end = var_5579_end_0, end_mask = var_5579_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_5579_cast_fp16")]; + tensor var_5582_begin_0 = const()[name = tensor("op_5582_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5582_end_0 = const()[name = tensor("op_5582_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5582_end_mask_0 = const()[name = tensor("op_5582_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5582_squeeze_mask_0 = const()[name = tensor("op_5582_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5582_cast_fp16 = slice_by_index(begin = var_5582_begin_0, end = var_5582_end_0, end_mask = var_5582_end_mask_0, squeeze_mask = var_5582_squeeze_mask_0, x = var_5579_cast_fp16)[name = tensor("op_5582_cast_fp16")]; + tensor var_5597_begin_0 = const()[name = tensor("op_5597_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_5597_end_0 = const()[name = tensor("op_5597_end_0"), val = tensor([1, 10, 1, 1500])]; + tensor var_5597_end_mask_0 = const()[name = tensor("op_5597_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5597_cast_fp16 = slice_by_index(begin = var_5597_begin_0, end = var_5597_end_0, end_mask = var_5597_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_5597_cast_fp16")]; + tensor var_5600_begin_0 = const()[name = tensor("op_5600_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5600_end_0 = const()[name = tensor("op_5600_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5600_end_mask_0 = const()[name = tensor("op_5600_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5600_squeeze_mask_0 = const()[name = tensor("op_5600_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5600_cast_fp16 = slice_by_index(begin = var_5600_begin_0, end = var_5600_end_0, end_mask = var_5600_end_mask_0, squeeze_mask = var_5600_squeeze_mask_0, x = var_5597_cast_fp16)[name = tensor("op_5600_cast_fp16")]; + tensor var_5615_begin_0 = const()[name = tensor("op_5615_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5615_end_0 = const()[name = tensor("op_5615_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_5615_end_mask_0 = const()[name = tensor("op_5615_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5615_cast_fp16 = slice_by_index(begin = var_5615_begin_0, end = var_5615_end_0, end_mask = var_5615_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_5615_cast_fp16")]; + tensor var_5618_begin_0 = const()[name = tensor("op_5618_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5618_end_0 = const()[name = tensor("op_5618_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5618_end_mask_0 = const()[name = tensor("op_5618_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5618_squeeze_mask_0 = const()[name = tensor("op_5618_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5618_cast_fp16 = slice_by_index(begin = var_5618_begin_0, end = var_5618_end_0, end_mask = var_5618_end_mask_0, squeeze_mask = var_5618_squeeze_mask_0, x = var_5615_cast_fp16)[name = tensor("op_5618_cast_fp16")]; + tensor var_5633_begin_0 = const()[name = tensor("op_5633_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5633_end_0 = const()[name = tensor("op_5633_end_0"), val = tensor([1, 15, 1, 1500])]; + tensor var_5633_end_mask_0 = const()[name = tensor("op_5633_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5633_cast_fp16 = slice_by_index(begin = var_5633_begin_0, end = var_5633_end_0, end_mask = var_5633_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_5633_cast_fp16")]; + tensor var_5636_begin_0 = const()[name = tensor("op_5636_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5636_end_0 = const()[name = tensor("op_5636_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5636_end_mask_0 = const()[name = tensor("op_5636_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5636_squeeze_mask_0 = const()[name = tensor("op_5636_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5636_cast_fp16 = slice_by_index(begin = var_5636_begin_0, end = var_5636_end_0, end_mask = var_5636_end_mask_0, squeeze_mask = var_5636_squeeze_mask_0, x = var_5633_cast_fp16)[name = tensor("op_5636_cast_fp16")]; + tensor var_5651_begin_0 = const()[name = tensor("op_5651_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_5651_end_0 = const()[name = tensor("op_5651_end_0"), val = tensor([1, 8, 1, 1500])]; + tensor var_5651_end_mask_0 = const()[name = tensor("op_5651_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5651_cast_fp16 = slice_by_index(begin = var_5651_begin_0, end = var_5651_end_0, end_mask = var_5651_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_5651_cast_fp16")]; + tensor var_5654_begin_0 = const()[name = tensor("op_5654_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5654_end_0 = const()[name = tensor("op_5654_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5654_end_mask_0 = const()[name = tensor("op_5654_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5654_squeeze_mask_0 = const()[name = tensor("op_5654_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5654_cast_fp16 = slice_by_index(begin = var_5654_begin_0, end = var_5654_end_0, end_mask = var_5654_end_mask_0, squeeze_mask = var_5654_squeeze_mask_0, x = var_5651_cast_fp16)[name = tensor("op_5654_cast_fp16")]; + tensor var_5669_begin_0 = const()[name = tensor("op_5669_begin_0"), val = tensor([0, 10, 0, 0])]; + tensor var_5669_end_0 = const()[name = tensor("op_5669_end_0"), val = tensor([1, 11, 1, 1500])]; + tensor var_5669_end_mask_0 = const()[name = tensor("op_5669_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5669_cast_fp16 = slice_by_index(begin = var_5669_begin_0, end = var_5669_end_0, end_mask = var_5669_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_5669_cast_fp16")]; + tensor var_5672_begin_0 = const()[name = tensor("op_5672_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5672_end_0 = const()[name = tensor("op_5672_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5672_end_mask_0 = const()[name = tensor("op_5672_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5672_squeeze_mask_0 = const()[name = tensor("op_5672_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5672_cast_fp16 = slice_by_index(begin = var_5672_begin_0, end = var_5672_end_0, end_mask = var_5672_end_mask_0, squeeze_mask = var_5672_squeeze_mask_0, x = var_5669_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor var_5687_begin_0 = const()[name = tensor("op_5687_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5687_end_0 = const()[name = tensor("op_5687_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5687_end_mask_0 = const()[name = tensor("op_5687_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5687_cast_fp16 = slice_by_index(begin = var_5687_begin_0, end = var_5687_end_0, end_mask = var_5687_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_5687_cast_fp16")]; + tensor var_5690_begin_0 = const()[name = tensor("op_5690_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5690_end_0 = const()[name = tensor("op_5690_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5690_end_mask_0 = const()[name = tensor("op_5690_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5690_squeeze_mask_0 = const()[name = tensor("op_5690_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5690_cast_fp16 = slice_by_index(begin = var_5690_begin_0, end = var_5690_end_0, end_mask = var_5690_end_mask_0, squeeze_mask = var_5690_squeeze_mask_0, x = var_5687_cast_fp16)[name = tensor("op_5690_cast_fp16")]; + tensor var_5705_begin_0 = const()[name = tensor("op_5705_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5705_end_0 = const()[name = tensor("op_5705_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5705_end_mask_0 = const()[name = tensor("op_5705_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5705_cast_fp16 = slice_by_index(begin = var_5705_begin_0, end = var_5705_end_0, end_mask = var_5705_end_mask_0, x = obj_293_cast_fp16)[name = tensor("op_5705_cast_fp16")]; + tensor var_5708_begin_0 = const()[name = tensor("op_5708_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5708_end_0 = const()[name = tensor("op_5708_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5708_end_mask_0 = const()[name = tensor("op_5708_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5708_squeeze_mask_0 = const()[name = tensor("op_5708_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5708_cast_fp16 = slice_by_index(begin = var_5708_begin_0, end = var_5708_end_0, end_mask = var_5708_end_mask_0, squeeze_mask = var_5708_squeeze_mask_0, x = var_5705_cast_fp16)[name = tensor("op_5708_cast_fp16")]; + tensor var_5723_begin_0 = const()[name = tensor("op_5723_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_5723_end_0 = const()[name = tensor("op_5723_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_5723_end_mask_0 = const()[name = tensor("op_5723_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5723_cast_fp16 = slice_by_index(begin = var_5723_begin_0, end = var_5723_end_0, end_mask = var_5723_end_mask_0, x = obj_293_cast_fp16)[name = tensor("op_5723_cast_fp16")]; + tensor var_5726_begin_0 = const()[name = tensor("op_5726_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5726_end_0 = const()[name = tensor("op_5726_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5726_end_mask_0 = const()[name = tensor("op_5726_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5726_squeeze_mask_0 = const()[name = tensor("op_5726_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5726_cast_fp16 = slice_by_index(begin = var_5726_begin_0, end = var_5726_end_0, end_mask = var_5726_end_mask_0, squeeze_mask = var_5726_squeeze_mask_0, x = var_5723_cast_fp16)[name = tensor("op_5726_cast_fp16")]; + tensor var_5741_begin_0 = const()[name = tensor("op_5741_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_5741_end_0 = const()[name = tensor("op_5741_end_0"), val = tensor([1, 10, 1, 1500])]; + tensor var_5741_end_mask_0 = const()[name = tensor("op_5741_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5741_cast_fp16 = slice_by_index(begin = var_5741_begin_0, end = var_5741_end_0, end_mask = var_5741_end_mask_0, x = obj_293_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5744_begin_0 = const()[name = tensor("op_5744_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5744_end_0 = const()[name = tensor("op_5744_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5744_end_mask_0 = const()[name = tensor("op_5744_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5744_squeeze_mask_0 = const()[name = tensor("op_5744_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5744_cast_fp16 = slice_by_index(begin = var_5744_begin_0, end = var_5744_end_0, end_mask = var_5744_end_mask_0, squeeze_mask = var_5744_squeeze_mask_0, x = var_5741_cast_fp16)[name = tensor("op_5744_cast_fp16")]; + tensor var_5759_begin_0 = const()[name = tensor("op_5759_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_5759_end_0 = const()[name = tensor("op_5759_end_0"), val = tensor([1, 15, 1, 1500])]; + tensor var_5759_end_mask_0 = const()[name = tensor("op_5759_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5759_cast_fp16 = slice_by_index(begin = var_5759_begin_0, end = var_5759_end_0, end_mask = var_5759_end_mask_0, x = obj_293_cast_fp16)[name = tensor("op_5759_cast_fp16")]; + tensor var_5762_begin_0 = const()[name = tensor("op_5762_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5762_end_0 = const()[name = tensor("op_5762_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5762_end_mask_0 = const()[name = tensor("op_5762_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5762_squeeze_mask_0 = const()[name = tensor("op_5762_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5762_cast_fp16 = slice_by_index(begin = var_5762_begin_0, end = var_5762_end_0, end_mask = var_5762_end_mask_0, squeeze_mask = var_5762_squeeze_mask_0, x = var_5759_cast_fp16)[name = tensor("op_5762_cast_fp16")]; + tensor var_5777_begin_0 = const()[name = tensor("op_5777_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_5777_end_0 = const()[name = tensor("op_5777_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_5777_end_mask_0 = const()[name = tensor("op_5777_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5777_cast_fp16 = slice_by_index(begin = var_5777_begin_0, end = var_5777_end_0, end_mask = var_5777_end_mask_0, x = obj_307_cast_fp16)[name = tensor("op_5777_cast_fp16")]; + tensor var_5780_begin_0 = const()[name = tensor("op_5780_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5780_end_0 = const()[name = tensor("op_5780_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5780_end_mask_0 = const()[name = tensor("op_5780_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5780_squeeze_mask_0 = const()[name = tensor("op_5780_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5780_cast_fp16 = slice_by_index(begin = var_5780_begin_0, end = var_5780_end_0, end_mask = var_5780_end_mask_0, squeeze_mask = var_5780_squeeze_mask_0, x = var_5777_cast_fp16)[name = tensor("op_5780_cast_fp16")]; + tensor var_5787 = const()[name = tensor("op_5787"), val = tensor(1)]; + tensor var_5788_interleave_0 = const()[name = tensor("op_5788_interleave_0"), val = tensor(false)]; + tensor var_5788_cast_fp16 = concat(axis = var_5787, interleave = var_5788_interleave_0, values = (var_5474_cast_fp16, var_5492_cast_fp16, var_5510_cast_fp16, var_5528_cast_fp16, var_5546_cast_fp16, var_5564_cast_fp16, var_5582_cast_fp16, var_5600_cast_fp16, var_5618_cast_fp16, var_5636_cast_fp16, var_5654_cast_fp16, var_5672_cast_fp16, var_5690_cast_fp16, var_5708_cast_fp16, var_5726_cast_fp16, var_5744_cast_fp16, var_5762_cast_fp16, var_5780_cast_fp16))[name = tensor("op_5788_cast_fp16")]; + tensor obj_axes_0 = const()[name = tensor("obj_axes_0"), val = tensor([1])]; + tensor obj_keep_dims_0 = const()[name = tensor("obj_keep_dims_0"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = obj_keep_dims_0, x = var_5788_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file