下载模型

python
展开代码
curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh |  bash
apt-get install git-lfs
git clone https://huggingface.co/Qwen/Qwen2.5-3B-Instruct

模型路径：

bash
展开代码
/mnt/jfs6/model/Qwen2.5-3B-Instruct

bash
展开代码
MODEL_NAME_OR_PATH="/mnt/jfs6/model/Qwen2.5-3B-Instruct"
JOBNAME=shai1014-11
CACHE_DIR=/mnt/jfs/cache/$JOBNAME
OUTPUT_DIR="/mnt/jfs/output/$JOBNAME"
SWANLAB_NAME=$JOBNAME

export SWANLAB_API_KEY=pM7Xvs5OS2EeXPO5gKXfJ
export SWANLAB_LOG_DIR=/swanlab_log
export SWANLAB_MODE=cloud


llamafactory-cli train \
      --model_name_or_path $MODEL_NAME_OR_PATH \
      --dataset_dir /app/data \
      --cache_dir ${CACHE_DIR}/cache_dir \
      --tokenized_path ${CACHE_DIR}/tokenized_cache \
      --overwrite_cache false  \
      --dataset alpaca_en_demo  \
      --template qwen \
      --output_dir $OUTPUT_DIR \
      --per_device_train_batch_size 4 \
      --gradient_accumulation_steps 1 \
      --image_max_pixels 451584 \
      --video_max_pixels 16384 \
      --trust_remote_code true \
      --stage sft \
      --do_train true \
      --finetuning_type full \
      --freeze_vision_tower false \
      --freeze_multi_modal_projector false \
      --freeze_language_model false \
      --cutoff_len 2048 \
      --preprocessing_num_workers 32 \
      --preprocessing_batch_size 32 \
      --dataloader_num_workers 32 \
      --logging_steps 10 \
      --plot_loss True  \
      --overwrite_output_dir false  \
      --save_only_model false \
      --learning_rate 1.0e-6 \
      --num_train_epochs 1000.0 \
      --save_steps 500 \
      --lr_scheduler_type cosine \
      --bf16 True  \
      --flash_attn auto \
      --report_to none \
      --use_swanlab True \
      --swanlab_project $SWANLAB_NAME \
      --swanlab_mode cloud \
      --save_strategy steps \
      --data_shared_file_system true \
      --ddp_timeout 180000000