[perf] feat: add remote reward manager and fix math verify issue

[yyDing1]

What does this PR do?

This PR introduces a remote reward manager to reward loop, which launches multiple separate processes to handle reward computation requests, as

Solve the following issues:

1. Certain reward computation methods (e.g., Math-Verify) cannot run in the current async + thread–based execution and therefore require execution in a separate process. math_verify reward return 0 when using async #3407
2. For CPU-intensive workloads such as Math-verify and maybe code reasoning, a process-based execution can be more appropriate than threading.

The proposed reward manager launches remote Ray reward-compute actors across Ray nodes, with each actor running in an independent process, thereby enabling effective load balancing across the cluster.

@ray.remote
class RewardComputeWorker:
    """
    WARNING: This class cannot have async methods.
    """

    def __init__(self, compute_score_fn):
        # since the reward function may not be pickleable, we need to init it in the worker
        self.compute_score_fn = compute_score_fn

    def compute_score(self, **kwargs) -> dict:
        return self.compute_score_fn(**kwargs)


num_reward_workers = config.reward_model.num_workers
# in the rollout & reward parallel mode
# the sum of final reward workers will be agent_loop_workers * num_reward_workers
self.reward_worker = [
    # register the reward worker in the same node
    RewardComputeWorker.options(
        scheduling_strategy=ray.util.scheduling_strategies.NodeAffinitySchedulingStrategy(
            node_id=ray.get_runtime_context().get_node_id(),
            soft=True,
        ),
    ).remote(self.compute_score)
    for _ in range(num_reward_workers)
]

Credit to @gxy-gxy with the solution using a separate process.

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).)

[gemini-code-assist]

Code Review

This pull request introduces a RemoteRewardManager to handle reward computations in separate processes using Ray actors, which is a solid approach for CPU-intensive tasks like math-verify. The implementation is mostly well-done. However, I've identified a potential race condition in the reward worker selection logic that could lead to imbalanced loads. Additionally, the new test file includes hardcoded paths, which compromises its portability and will likely cause issues in CI environments. I've provided specific feedback and suggestions to address these high-severity issues.