微调 qwen2的时候为啥默认开启的是多卡，我明明用的是单卡训练，而且我也用webui试了单卡，但是它默认的还是多卡

[yxl23]

Reminder

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

System Info

Reproduction

llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml

Expected behavior

No response

Others

No response

[frozenarctic]

在cli.py的第80行，改成
if (not disable_torchrun) and (get_device_count() > 1):
然后到根目录下重新运行
pip install -e '.[torch,metrics]'
默认就是用本地单卡了

[hiyouga]

修复了

[yxl23]

我用llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml 8bit量化微调qwen2出现
{'loss': 0.0, 'grad_norm': nan, 'learning_rate': 0.0, 'epoch': 0.51}
{'loss': 0.0, 'grad_norm': nan, 'learning_rate': 0.0, 'epoch': 1.02}
{'loss': 0.0, 'grad_norm': nan, 'learning_rate': 0.0, 'epoch': 1.53}

[hiyouga]

用 bf16

[yxl23]

在哪添加啊

model

model_name_or_path: E:\LLaMA-Factory\qwen\Qwen2-7B-Instruct
quantization_bit: 8

method

stage: sft
do_train: true
finetuning_type: lora
lora_target: all

dataset

dataset: xunlian
template: qwen
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: saves/qwen2-7b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true

train

per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 100.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
fp16: true

eval

val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

[hiyouga]

fp16 换成 bf16

[yxl23]

好的谢谢