[trainer, fsdp, vllm, recipe] feat: one step off async training recipe

[imh966]

What does this PR do?

This PR provides a simple implementation of one step off async training with fsdp and vllm backend.

We conducted three different experiments with qwen2.5_3b model on 8 A100 GPUs:

1. baseline: all models are colocated
2. standalone rollout: rollout model runs on 4 GPUs and other models run on remaining 4GPUs
3. one step off: the same model placement as the second experiment, but with one step off async training

The pictures below demonstrate the results of these experiments:

In these experiments, baseline has the highest throughput, but we think it is just because we didn't find the best configure for one step off async training.

The exciting point is that our nccl based weights updating for rollout model has great performance. The latency is showed below:

At most of time, the latency is under 300ms, which is negligible for RLHF. Although it is only implemented with fsdp and vllm now, we think it is not complex to extend it to the other backend.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

To use this feature, hybrid_engine option must be disabled to separate actor model and rollout model into difference GPU cluster. rollout.n_gpus option has been added to configure file to indicate how many GPUs rollout model would be occupied. The script below is an example to train qwen2.5_3b with 8 GPUs.

python3 -m recipe.async.async_main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=$HOME/data/gsm8k/train.parquet \
    data.val_files=$HOME/data/gsm8k/test.parquet \
    data.train_batch_size=1024 \
    data.max_prompt_length=512 \
    data.max_response_length=1024 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    data.shuffle=False \
    actor_rollout_ref.model.path=Qwen/Qwen2.5-3B-Instruct \
    actor_rollout_ref.actor.optim.lr=3e-6 \
    actor_rollout_ref.hybrid_engine=False \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=256 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=40 \
    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=40 \
    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.rollout.n_gpus=4 \
    actor_rollout_ref.rollout.load_format=safetensors \
    actor_rollout_ref.rollout.layered_summon=True \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=40 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.val_before_train=True \
    trainer.logger=['console','wandb'] \
    trainer.project_name='verl_grpo_example_gsm8k' \
    trainer.experiment_name='qwen2.5_3b_grpo_async_one_step_off' \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.save_freq=-1 \
    trainer.test_freq=-1 \
    trainer.total_epochs=15 $@

High-Level Design

Demonstrate the high-level design if this PR is complex.

Specific Changes

1. nccl based weights updating for rollout model.
2. one step off async trainer.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace.

[PeterSH6]

Nice work

[PeterSH6]

Can you simplify the code in this file? There's too much redundancy

[imh966]

Can you simplify the code in this file? There's too much redundancy

OK, I will try.

[imh966]

@PeterSH6 I've removed some redundant code, but I'm not sure whether it's enought.

[ccclyu]

great work! Have you tried the testing on multiple nodes and observed some throughput delta?

[eric-haibin-lin]

thanks for the contribution! please add a README.md for the scope of this recipe, for instance, indicating the support status for features available in the original ray trainer such as vlm/multi-turn.
Please also make a copy of the doc to section to docs/advance/ for documentation. Please include the convergence curve in these docs

[eric-haibin-lin]

since we're deprecating fsdp, we can limit this recipe to fsdp2 only, and make sure it is tested with fsdp2

[eric-haibin-lin]

pls remove unused code

[CLAassistant]

All committers have signed the CLA.

[lzxdjb]

Hello! Thank you so much for implementing asynchronous RLHF framework. I have little questions:
1, Have you tested the time consumption between your awesome implementation with the standard verl?
2, I find that every time before generation, it needs synchronize the weight from actor. I think it is a little redundant. It actually just needs updating parameter once updating the actor? (Although it might mix some staleness batch, but the AReaL prove that using decoupled PPO algorithm can almost keep the training quality)
Looking forward to your reply~

[BounharAbdelaziz]

@lzxdjb I agree. Mistral AI does the same without even recomputing kv cache (in Magistral paper https://arxiv.org/pdf/2506.10910).

[eric-haibin-lin]

In the next PR, could you reuse the function _start_profiling and _stop_profiling from the parent class? thanks.
https://github.com/volcengine/verl/blob/main/verl/trainer/ppo/ray_trainer.py#L1042-L1063

[eric-haibin-lin]

The snapshot of this recipe development branch is pushed to https://github.com/volcengine/verl/tree/recipe/one_step_off_async. Thanks team for the great work!