[trainer] feat: support one step off policy for diffusion model

[chenyingshu]

What does this PR do?

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

Support one-step-off async policy training for diffusion model with FlowGRPO algorithm (FSDP2).

Main changes:

• Created new trainer for it.
• Use nccl backend, change some logics when updating weights.
• Support async reward loop

Progress:

• Passed UT (w/ FSDP2)
• e2e training validation, check performance: examples/flowgrpo_trainer/run_flowgrpo_one_step_off.sh
• Code cleaning and refactoring
• Next PR for fully-async policy

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, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • 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.

See unit test in tests/special_e2e/run_one_step_off_policy.sh. Run the test by:

CUDA_VISIBLE_DEVICES=0,1,2 NUM_GPUS=2 MODEL_ID=tiny-random/Qwen-Image adv_estimator=flow_grpo n_gpu_rollout=1 bash tests/special_e2e/run_one_step_off_policy.sh

API and Usage Example

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

1. Change configs:
   Major changes include using 'nccl' backend for separate trainer and actor, #gpus for trainer and rollout. E.g.,

actor_rollou t_ref.hybrid_engine=False
actor_rollout_ref.rollout.checkpoint_engine.backend=nccl

trainer.nnodes=1
trainer.n_gpus_per_node=${n_gpus_training}
rollout.nnodes=1
rollout.n_gpus_per_node=${n_gpus_rollout}

2. Run one-step-off training:

python3 -m verl.experimental.one_step_off_policy.main_ppo \
    --config-path=config \
    --config-name='one_step_off_ppo_diffusion_trainer.yaml' \
    algorithm.adv_estimator=flow_grpo \
    actor_rollout_ref.actor.strategy=fsdp2 \
    actor_rollout_ref.rollout.checkpoint_engine.backend=nccl \
    # actor and rollout are placed separately
    actor_rollout_ref.hybrid_engine=False \
    # actor and rollout resource
    trainer.nnodes=1 \
    trainer.n_gpus_per_node=4 \
    rollout.nnodes=1 \
    rollout.n_gpus_per_node=4

or run example script:

bash examples/flowgrpo_trainer/run_flowgrpo_one_step_off.sh

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

• Trainer:
  • New class SeparateRayFlowGRPOTrainer in verl/experimental/separation/ray_diffusion_trainer.py.
    Note that it is expected to be reused for fully-async policy, which will be implemented in next PR.
  • New class OneStepOffRayFlowGRPOTrainer in verl/experimental/one_step_off_policy/ray_diffusion_trainer.py
  • Added condition to use OneStepOffRayFlowGRPOTrainer and standalone reward loop in verl/experimental/one_step_off_policy/main_ppo.py
• Config: New config verl/experimental/one_step_off_policy/config/one_step_off_ppo_diffusion_trainer.yaml
• Weight Update: Some minor changes were made to apply lora for updating weights of the separate rollout. The changes mainly occurred due to nccl backend usage:
  • Return peft_config and base_sync_done in ActorRolloutRefWorker.update_weights() in verl/workers/engine_workers.py
  • Pass peft_config and base_sync_done to CheckpointEngineWorker.update_weights() in verl/checkpoint_engine/base.py.
  • Modified collect_lora_params in verl/utils/fsdp_utils.py since when using nccl backend, neither layered_summon nor base_sync_done is passed.
    Note: (susan) I am not sure here.
• UT: Added unit test in .github/workflows/e2e_one_step_off_policy.yml, and tests/special_e2e/run_one_step_off_policy.sh.
• Doc: Updated doc in docs/advance/one_step_off.md and verl/experimental/one_step_off_policy/README.md

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. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.