[fsdp] feat: upcast MoE routing to FP32 for better accuracy

[Shangwei-Li]

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.

Upcast MoE routing to FP32 for better accuracy. In Megatron training, MoE router will usually be upcast to fp32 or fp64, but original Transformers implementation uses only bf16, which hurts accuracy in certain cases like fully async training. Upcasting router to FP32 helps with accuracy and stability in such cases.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: https://github.com/verl-project/verl/pulls?q=fp32+moe
• 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
  • 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.

Fully async DAPO aime 2024 validation score. The red line represents vanilla implementation and the pink line represents FP32 implementation.

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 introduces a custom torch.autograd.Function to upcast Mixture-of-Experts (MoE) routing computations to FP32, aiming to improve accuracy. The changes correctly replace the standard nn.Linear gate with this new function in MoE blocks. My review found a potential bug in the new RouterGatingLinearFunction where the bias parameter is handled in the backward pass but not applied in the forward pass. Although this doesn't affect the current implementation as no bias is used, I've provided a suggestion to fix this for future robustness.