llamafactory SFT 从断点恢复训练

背景

我使用llamafactory sft 微调模型的时候。gpu停止运行了。日志文件没有任何的报错信息。

显存还是占用状态。

查看llamafactory的进程是下述信息：

151312 151306 91 17:42 ? 03:58:10 [llamafactory-cl]

既然如此，那就只能从断点恢复训练了。

这里是 GitHub 仓库对应的 issues: https://github.com/hiyouga/LLaMA-Factory/issues/1713

断点重启

如果只是恢复一个模型的训练，手动修改 resume_from_checkpoint 参数：

bash 复制代码

llamafactory-cli train \
        --resume_from_checkpoint ${save_path}/${resume_from_checkpoint} \
        --stage sft \
        --do_train True \
        --model_name_or_path ${model_name_or_path} \
        --preprocessing_num_workers 16 \
        --finetuning_type lora \
        --template $template \
        --flash_attn auto \
        --dataset_dir $dataset_dir \
        --dataset $alpaca_dataset_name \
        --cutoff_len 2048 \
        --learning_rate 2e-05 \
        --num_train_epochs 2.0 \
        --per_device_train_batch_size 2 \
        --gradient_accumulation_steps 4 \
        --lr_scheduler_type cosine \
        --max_grad_norm 1.0 \
        --logging_steps 5 \
        --save_steps 100 \
        --warmup_steps 100 \
        --packing False \
        --report_to none \
        --output_dir ${save_path} \
        --fp16 True \
        --plot_loss True \
        --trust_remote_code True \
        --ddp_timeout 180000000 \
        --include_num_input_tokens_seen True \
        --optim adamw_torch \
        --lora_rank 16 \
        --lora_alpha 32 \
        --lora_dropout 0 \
        --lora_target all \
        --val_size 0.1 \
        --eval_strategy steps \
        --eval_steps 100 \
        --per_device_eval_batch_size 1 >${log_dir}/sft.log 2>&1

但我想使用 shell 脚本，自动地从断点恢复训练，减轻手动负担：

下述脚本可以找到当前文件夹，最大的checkpoint文件夹:

bash 复制代码

# 检查 all_results.json 是否不存在，同时 trainer_log.jsonl 是否存在
if [ ! -f "all_results.json" ] && [ -f "trainer_log.jsonl" ]; then
    # 找出所有 checkpoint-数字 的文件夹，并提取最大的编号
    max_folder=$(ls -d checkpoint-*/ 2>/dev/null | awk -F '-' '{print $2}' | sort -n | tail -1)

    if [ -n "$max_folder" ]; then
        echo "最大的 checkpoint 文件夹是: checkpoint-$max_folder"
    else
        echo "没有找到任何 checkpoint 文件夹"
    fi
else
    echo "条件不满足：all_results.json 存在 或 trainer_log.jsonl 不存在"
fi

下述是一个自动加载checkpoint，恢复训练的shell脚本：

bash 复制代码

resume_from_checkpoint=None

# 检查 all_results.json 不存在，同时 trainer_log.jsonl 存在，代表模型没有训练完成就退出了
if [ ! -f "${save_path}/all_results.json" ] && [ -f "${save_path}/trainer_log.jsonl" ]; then

    # 找出所有 checkpoint-数字 的文件夹，并提取最大的编号
    max_folder=$(ls -d ${save_path}/checkpoint-*/ 2>/dev/null | awk -F '-' '{print $2}' | sort -n | tail -1)
    if [ -n "${max_folder}" ]; then
        echo "最大的 checkpoint 文件夹是: checkpoint-${max_folder}"
        resume_from_checkpoint=checkpoint-${max_folder}
    fi
else
    echo "模型可能微调完成，退出微调"
fi

echo "resume_from_checkpoint: ${resume_from_checkpoint}"

if [ -f "${save_path}/all_results.json" ]; then
    echo "${save_path}/trainer_log.jsonl, just skip sft"
else
    llamafactory-cli train \
        --resume_from_checkpoint ${save_path}/${resume_from_checkpoint} \
        --stage sft \
        --do_train True \
        --model_name_or_path ${model_name_or_path} \
        --preprocessing_num_workers 16 \
        --finetuning_type lora \
        --template $template \
        --flash_attn auto \
        --dataset_dir $dataset_dir \
        --dataset $alpaca_dataset_name \
        --cutoff_len 2048 \
        --learning_rate 2e-05 \
        --num_train_epochs 2.0 \
        --per_device_train_batch_size 2 \
        --gradient_accumulation_steps 4 \
        --lr_scheduler_type cosine \
        --max_grad_norm 1.0 \
        --logging_steps 5 \
        --save_steps 100 \
        --warmup_steps 100 \
        --packing False \
        --report_to none \
        --output_dir ${save_path} \
        --fp16 True \
        --plot_loss True \
        --trust_remote_code True \
        --ddp_timeout 180000000 \
        --include_num_input_tokens_seen True \
        --optim adamw_torch \
        --lora_rank 16 \
        --lora_alpha 32 \
        --lora_dropout 0 \
        --lora_target all \
        --val_size 0.1 \
        --eval_strategy steps \
        --eval_steps 100 \
        --per_device_eval_batch_size 1 >${log_dir}/sft.log 2>&1

    echo "lora finish save at ${save_path}"
fi

运行脚本后恢复训练，日志文件信息如下：

复制代码

[INFO|trainer.py:748] 2025-05-16 21:51:23,017 >> Using auto half precision backend
[INFO|trainer.py:2813] 2025-05-16 21:51:23,018 >> Loading model from /mnt/mydisk/github/papers/plan_agent_rag/answer_learn/many_datasets/wiki_mqa/llama3/output/saves/lora//checkpoint-300/.
[INFO|trainer.py:2414] 2025-05-16 21:51:28,423 >> ***** Running training *****
[INFO|trainer.py:2415] 2025-05-16 21:51:28,423 >>   Num examples = 3,591
[INFO|trainer.py:2416] 2025-05-16 21:51:28,423 >>   Num Epochs = 2
[INFO|trainer.py:2417] 2025-05-16 21:51:28,423 >>   Instantaneous batch size per device = 2
[INFO|trainer.py:2420] 2025-05-16 21:51:28,423 >>   Total train batch size (w. parallel, distributed & accumulation) = 8
[INFO|trainer.py:2421] 2025-05-16 21:51:28,423 >>   Gradient Accumulation steps = 4
[INFO|trainer.py:2422] 2025-05-16 21:51:28,423 >>   Total optimization steps = 898
[INFO|trainer.py:2423] 2025-05-16 21:51:28,425 >>   Number of trainable parameters = 41,943,040
[INFO|trainer.py:2445] 2025-05-16 21:51:28,425 >>   Continuing training from checkpoint, will skip to saved global_step
[INFO|trainer.py:2446] 2025-05-16 21:51:28,425 >>   Continuing training from epoch 0
[INFO|trainer.py:2447] 2025-05-16 21:51:28,425 >>   Continuing training from global step 300
[INFO|trainer.py:2449] 2025-05-16 21:51:28,425 >>   Will skip the first 0 epochs then the first 1200 batches in the first epoch.

  0%|          | 0/898 [00:00<?, ?it/s]
 34%|███▎      | 301/898 [00:08<00:17, 34.10it/s]
 34%|███▎      | 302/898 [00:22<00:17, 34.10it/s]
 34%|███▎      | 303/898 [00:24<01:00,  9.90it/s]
 34%|███▍      | 304/898 [00:31<01:28,  6.69it/s]

脚本中可能有一些本地文件、日志保存的变量，请大家忽略掉！