verl Low GPU Utilization, help on training parameters. 训练时显卡利用率很低，如何提升训练效率？

Low GPU Utilization, help on training parameters. 训练时显卡利用率很低，如何提升训练效率？

Open lezhang7 opened this issue 7 months ago • 2 comments

trafficstars

I can train the model. However, I find the GPU utilization extremely low with following training script. Considering long prompt length and response lenght, I set the batch size to be small. As shown in the screenshot, for more than half time the GPU is not working. I'm wondering how to better ultilize GPUs.

如图所示，训练时大部分时间显卡基本空置，怀疑是训练参数设置有问题，但是自己调了一下也没有很好的work，请问这种情况下应该如何调整训练参数来最大化显卡利用率呢？

    data.max_prompt_length=3072 \
    data.max_response_length=2048 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    actor_rollout_ref.model.path=Qwen/Qwen2.5-7B-Instruct \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=16 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.entropy_coeff=0 \
    actor_rollout_ref.model.enable_gradient_checkpointing=True \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.n=5 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.val_before_train=false \
    trainer.logger=['console','wandb'] \
    trainer.project_name='qwen2' \
    trainer.experiment_name='qwen2_7b' \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.save_freq=150 \
    trainer.test_freq=-1 \
    trainer.total_epochs=15 $@```