提一个小建议。不知道对不对，在解决导出模型时：”RuntimeError: "addmm_impl_cpu_" not implemented for 'Half'“

[ganchun1130]

Reminder

• I have read the README and searched the existing issues.

Reproduction

这个报错是因为cpu不支持bf32或者是fp16，原因在于没使用gpu，
而原文在src/llmtuner/hparams/model_args.py 大约132行 代码是：
export_device: str = field(
default="cpu",
metadata={"help": "The device used in model export."},
)
默认是cpu，所以会报错，改成cuda，就可以顺利导出模型！

Expected behavior

No response

System Info

No response

Others

No response

[hiyouga]

可以提供下训练的命令吗？是否用了除了 lora 外别的微调方法

[ganchun1130]

这是我保存的训练命令
{
"top.model_name": "Qwen-7B-Chat",
"top.finetuning_type": "lora",
"top.adapter_path": [],
"top.quantization_bit": "none",
"top.template": "qwen",
"top.rope_scaling": "none",
"top.booster": "none",
"train.training_stage": "Supervised Fine-Tuning",
"train.dataset_dir": "/usr/local/TFBOYS/gc/NLP/LLaMA-Factory/data",
"train.dataset": [
"ai_augmented_classification_data_zh_1"
],
"train.learning_rate": "2e-5",
"train.num_train_epochs": "5.0",
"train.max_grad_norm": "1.0",
"train.max_samples": "100000",
"train.compute_type": "fp32",
"train.cutoff_len": 2048,
"train.batch_size": 4,
"train.gradient_accumulation_steps": 4,
"train.val_size": 0,
"train.lr_scheduler_type": "cosine",
"train.logging_steps": 5,
"train.save_steps": 100,
"train.warmup_steps": 0,
"train.neftune_alpha": 0,
"train.optim": "adamw_torch",
"train.resize_vocab": false,
"train.packing": false,
"train.upcast_layernorm": false,
"train.use_llama_pro": false,
"train.shift_attn": false,
"train.report_to": false,
"train.num_layer_trainable": 3,
"train.name_module_trainable": "all",
"train.lora_rank": 8,
"train.lora_alpha": 16,
"train.lora_dropout": 0.1,
"train.loraplus_lr_ratio": 0,
"train.create_new_adapter": false,
"train.use_rslora": false,
"train.use_dora": true,
"train.lora_target": "all",
"train.additional_target": "",
"train.dpo_beta": 0.1,
"train.dpo_ftx": 0,
"train.orpo_beta": 0.1,
"train.reward_model": null,
"train.use_galore": false,
"train.galore_rank": 16,
"train.galore_update_interval": 200,
"train.galore_scale": 0.25,
"train.galore_target": "all"
}

[ganchun1130]

我也用过bf16的精度，也是会在导出模型时出现上述错误，后来改了代码之后就没有出现这个错误了

[lvsijian8]

#3434
一样的问题，测试在 A100 没有问题，v100 有问题

[hiyouga]

在 webui 上加入了这个选项