Skip to content

PPO training CUDA out of memory #52

Description

@14H034160212

你好,
我现在在训练PPO的时候出现了CUDA out of memory的问题,我是用了8个A100 GPUs,每一个GPU有80GB显存。下面是我运行的命令。我是用的stanford-alpaca提供的代码用8个A100全参微调训练了llama2-13B的sft model,reward模型是用的LLM-tuning项目提供的训练reward的代码基于llama2-13B训练的。现在就是在运行下面的ppo的时候出现了爆显存的问题,请问有什么办法可以降低显存吗?谢谢

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python rl_training.py \
    --base_model_name /data/qbao775/Explanation-Generation/llama-2/llama-2-13B \
    --merged_sft_model_path /data/qbao775/Explanation-Generation/llama_2_13B_merged_all_generator_avg_3_lenexp_10 \
    --sft_model_lora_path /data/qbao775/Explanation-Generation/llama_2_13B_merged_all_generator_avg_3_lenexp_10 \
    --reward_model_lora_path ../weights/llama-2-13B_beyond_reward_chinese_5000_peft_last_checkpoint \
    --adafactor False \
    --save_freq 10 \
    --output_max_length 64 \
    --batch_size 1 \
    --gradient_accumulation_steps 1 \
    --batched_gen True \
    --ppo_epochs 4 \
    --seed 0 \
    --learning_rate 1e-5 \
    --early_stopping True \
    --output_dir weights/llama-13_rlhf_beyond_test_6 \
    --log_with wandb

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions