SFT训练模型的命令
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
--stage sft \
--model_name_or_path models/chatglm3-6b \
--do_train \
--dataset self_cognition \
--template chatglm3 \
--finetuning_type lora \
--lora_target query_key_value \
--output_dir output/chatglm3_sft_lora_self/ \
--overwrite_cache \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 200 \
--learning_rate 5e-5 \
--num_train_epochs 100 \
--plot_loss \
--fp16
模型基座是下载正确的,torch.cuda.is_available()也是True
训练报错:
11/21/2023 09:11:23 - INFO - llmtuner.data.loader - Loading dataset self_cognition.json...
Using custom data configuration default-aaabbbccc
Loading Dataset Infos from /usr/local/lib/python3.10/site-packages/datasets/packaged_modules/json
Generating dataset json (/root/.cache/huggingface/datasets/json/default-aaabbbccc/0.0.0/34bc96c741b2e8a1f18598ffdd8bb11242116d54740a1d4f2a2872c7a28b6900)
Downloading and preparing dataset json/default to /root/.cache/huggingface/datasets/json/default-aaabbbccc/0.0.0/34bc96c741b2e8a1f18598ffdd8bb11242116d54740a1d4f2a2872c7a28b6900...
Downloading data files: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 6364.65it/s]
Downloading took 0.0 min
Checksum Computation took 0.0 min
Extracting data files: 0%| | 0/1 [00:00<?, ?it/s]train.sft.1gpu.lora.fp16.self.sh: line 19: 2551 Segmentation fault (core dumped) CUDA_VISIBLE_DEVICES=0 python src/train_bash.py --stage sft --model_name_or_path models/chatglm3-6b --do_train --dataset self_cognition --template chatglm3 --finetuning_type lora --lora_target query_key_value --output_dir output/chatglm3_sft_lora_self/ --overwrite_cache --per_device_train_batch_size 4 --gradient_accumulation_steps 4 --lr_scheduler_type cosine --logging_steps 10 --save_steps 200 --learning_rate 5e-5 --num_train_epochs 100 --plot_loss --fp16
调试
- 用pycharm,跟踪入口,从src/train_bash.py进去,一步一步看
- 根据报错log,找到"Loading dataset {}..."
- 然后加断点,单步调试
调试后定位到:Segmentation fault (core dumped)是在 运行 datasets的 load_dataset() 函数发生的,单独调用这个函数也复现问题了: https://github.com/hiyouga/LLaMA-Factory/blob/main/src/llmtuner/data/loader.py#L56
试了好几个和datasets的Segmentation fault (core dumped)报错相关的修复方案,都也没有解决,包括网上搜到要更新一些相关库的版本,也都不行。
最终从报错的ubuntu18.04切换到20.04,换了个操作系统,就没有这个报错了