[veomni, trainer] feat: add rl support for veomni backend

[ji-huazhong]

What does this PR do?

Add rl support for veomni backend.

Collaborated with @A1waysBeenHere

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, cfg, reward
  • 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.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

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

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.

[gemini-code-assist]

Code Review

This pull request adds support for the veomni backend for RL training. It introduces new configuration files for veomni actor, critic, reference policy, and reward models, along with a main PPO trainer configuration. The changes also update the worker selection logic to handle the veomni strategy.

My review has identified a few critical issues in the new configuration files that would lead to runtime errors. Specifically, a misconfigured legacy worker flag, a non-portable model path, and a placeholder value that needs to be replaced. There is also a minor redundancy in one of the configuration files. Addressing these points will improve the correctness and usability of the new veomni backend configuration.

[wuxibin89]

We're not going to support reward model with model engine, reward models include GRM/DistRM should use rollout engine(vllm/sglang/trt-llm). cc @yyDing1

[ji-huazhong]

Here’s the relevant PR: #5194. I’ll delete the reward model code later today. Cheers for pointing this out!

[ji-huazhong]

fixed in 22ee2f4. cc @wuxibin89

[A1waysBeenHere]

We conducted comparative experiments on Qwen3-8B using the GRPO. The results show that the reward trends of the FSDPEngine and the VeOmniEngine are consistent.

cc @ji-huazhong

[ji-huazhong]

The failed test case is unrelated to this PR.

[ji-huazhong]

Due to current resource constraints, I tested the following script on an A3:

set -x
ENGINE=${1:-vllm}

export MULTI_STREAM_MEMORY_REUSE=2
export VLLM_USE_V1=1
export TASK_QUEUE_ENABLE=1
export CPU_AFFINITY_CONF=1


python3 -m verl.trainer.main_ppo --config-path=config\
    --config-name="ppo_veomni_trainer.yaml" \
    algorithm.adv_estimator=grpo \
    data.train_files=/home/data/geo3k/train.parquet \
    data.val_files=/home/data/geo3k/test.parquet \
    data.train_batch_size=256 \
    data.max_prompt_length=2048 \
    data.max_response_length=2048 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    data.image_key=images \
    actor_rollout_ref.model.path=/home/model/Qwen3-VL-30B-A3B-Instruct \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.actor.optim.optimizer="anyprecision_adamw" \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=128 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.01 \
    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.veomni.param_offload=True \
    actor_rollout_ref.actor.veomni.optimizer_offload=True \
    actor_rollout_ref.actor.veomni.data_parallel_size=16 \
    actor_rollout_ref.actor.veomni.expert_parallel_size=16 \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=4 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=8 \
    actor_rollout_ref.rollout.name=$ENGINE \
    +actor_rollout_ref.rollout.engine_kwargs.vllm.disable_mm_preprocessor_cache=True \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.7 \
    actor_rollout_ref.rollout.enable_chunked_prefill=False \
    actor_rollout_ref.rollout.enforce_eager=False \
    actor_rollout_ref.rollout.free_cache_engine=True \
    actor_rollout_ref.rollout.n=5 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=4 \
    actor_rollout_ref.ref.veomni.param_offload=True \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.val_before_train=True \
    trainer.use_legacy_worker_impl=disable \
    trainer.logger='["console","wandb"]' \
    trainer.project_name='verl_grpo_example_geo3k' \
    trainer.experiment_name='qwen3_vl_30b_function_rm' \
    trainer.n_gpus_per_node=16 \
    trainer.nnodes=1 \
    trainer.save_freq=-1 \
    trainer.test_freq=-1 \
    trainer.total_epochs=15 $@

and the partial results obtained are as follows:

cc @wuxibin89

[wuxibin89]

Is it possible to unify ppo_veomni_trainer.yaml with ppo_trainer.yaml? Eventually, we will unify all trainer(fsdp/megatron/veomni/torchtitan) into one config file.

[ji-huazhong]

set -x
ENGINE=${1:-vllm}

export WANDB_MODE=offline
export MULTI_STREAM_MEMORY_REUSE=2
export VLLM_USE_V1=1
export TASK_QUEUE_ENABLE=1
export CPU_AFFINITY_CONF=1


python3 -m verl.trainer.main_ppo \
    model_engine=veomni \
    algorithm.adv_estimator=grpo \
    data.train_files=/home/lynn/data/geo3k/train.parquet \
    data.val_files=/home/lynn/data/geo3k/test.parquet \
    data.train_batch_size=256 \
    data.max_prompt_length=2048 \
    data.max_response_length=2048 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    data.image_key=images \
    actor_rollout_ref.model.path=/home/lynn/models/Qwen3-VL-30B-A3B-Instruct \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.actor.optim.optimizer="anyprecision_adamw" \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=128 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.01 \
    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.veomni.param_offload=True \
    actor_rollout_ref.actor.veomni.optimizer_offload=True \
    actor_rollout_ref.actor.veomni.fsdp_size=16 \
    actor_rollout_ref.actor.veomni.expert_parallel_size=16 \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=4 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=8 \
    actor_rollout_ref.rollout.name=$ENGINE \
    +actor_rollout_ref.rollout.engine_kwargs.vllm.disable_mm_preprocessor_cache=True \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.7 \
    actor_rollout_ref.rollout.enable_chunked_prefill=False \
    actor_rollout_ref.rollout.enforce_eager=False \
    actor_rollout_ref.rollout.free_cache_engine=True \
    actor_rollout_ref.rollout.n=5 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=4 \
    actor_rollout_ref.ref.veomni.param_offload=True \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.val_before_train=True \
    trainer.use_legacy_worker_impl=disable \
    trainer.logger='["console","wandb"]' \
    trainer.project_name='verl_grpo_example_geo3k' \
    trainer.experiment_name='qwen3_vl_30b_function_rm_baseline' \
    trainer.n_gpus_per_node=16 \
    trainer.nnodes=1 \
    trainer.save_freq=-1 \
    trainer.test_freq=-1 \
    trainer.total_epochs=15 $@