Ray OOM causes the process to be killed

[PKU-Fgx]

I found that as the training progressed, the System Memory Utilization (%) skyrocketed, and after a fixed point ray would report an out of memory error that would crash the training process.

• Error

• System Memory Utilization (%)

• Script

set -x
MODEL_PATH=<local_path>
export VLLM_ATTENTION_BACKEND=XFORMERS
python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=<local_path> \
    data.val_files=<local_path> \
    data.train_batch_size=64 \
    data.val_batch_size=64 \
    data.max_prompt_length=768 \
    data.max_response_length=3328 \
    actor_rollout_ref.model.path=$MODEL_PATH \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=12288 \
    actor_rollout_ref.actor.use_kl_loss=False \
    actor_rollout_ref.actor.kl_loss_coef=0. \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.use_dynamic_bsz=True \
    actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
    actor_rollout_ref.model.use_remove_padding=True \
    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.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.temperature=1.0 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.72 \
    actor_rollout_ref.rollout.n=32 \
    actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=12288 \
    actor_rollout_ref.rollout.enforce_eager=False \
    actor_rollout_ref.rollout.free_cache_engine=False \
    actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=12288 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.kl_ctrl.kl_coef=0. \
    trainer.critic_warmup=0 \
    trainer.logger=['wandb'] \
    trainer.project_name=<wandb> \
    trainer.experiment_name=<wandb> \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.default_local_dir=<local_path> \
    trainer.default_hdfs_dir=null \
    +trainer.val_before_train=False \
    trainer.save_freq=200 \
    trainer.test_freq=200 \
    trainer.total_epochs=3

Or is there some parameter I haven't configured correctly that is causing the memory usage to keep increasing?

[LinyeLi60]

same issue

[vermouth1992]

I suspect that this is memory leak when saving checkpoint.

[PKU-Fgx]

I suspect that this is memory leak when saving checkpoint.

But I only saved the checkpoint once or twice during the training process, so if it's a checkpoint saving memory leak, then there shouldn't be a linear increase in memory usage before saving.

[vermouth1992]

Then, I highly doubt the reward function that you use have memory leak... Try to switch to dummy reward function to see if there is memory leak

[yushuiwx]

same issue,

[kevin85421]

@PeterSH6 would you mind adding the label ray to the issue? Thanks!

[kevin85421]

@PKU-Fgx, this could be either a veRL issue, a Ray issue, or both. You can use jemalloc to profile it (see ray-project/ray#51031). If it turns out to be a Ray issue after you profile it, I'll take a look.

[PKU-Fgx]

@PeterSH6 would you mind adding the label ray to the issue? Thanks!

Okay, I'll look into that in the next few days.

[hiyouga]

Did you use vLLM 0.7+? It seems that the recent versions of vLLM are the cause of memory leak. You can try vllm 0.6.3

[NIL-zhuang]

same issue for vllm 0.7.3, i didn't save any checkpoint.

[LUMO666]

Did you use vLLM 0.7+? It seems that the recent versions of vLLM are the cause of memory leak. You can try vllm 0.6.3

same issue on 32B with 16 nodes. use vllm 0.6.3

[Skywuuuu]

In my case, I set actor_rollout_ref.rollout.free_cache_engine=False to solve this problem. My vllm version is 0.6.3.

[PKU-Fgx]

In my case, I set actor_rollout_ref.rollout.free_cache_engine=False to solve this problem. My vllm version is 0.6.3.

it does not work well for me, my version is 0.7.2

[yxliu0903]

In my case, I set actor_rollout_ref.rollout.free_cache_engine=False to solve this problem. My vllm version is 0.6.3.

it does not work well for me, my version is 0.6.3