Add diffusion reward loop

[chenyingshu]

What does this PR do?

Initialize diffusion reward loop pipeline.
Add & pass unit test for DiffusionRewardLoopManager

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.

[Copilot]

Pull request overview

This PR adds a diffusion reward loop pipeline to support image-based reward computation in reinforcement learning workflows. It introduces a specialized reward loop manager (DiffusionRewardLoopManager) and reward manager (DiffusionRewardManager) designed to handle image outputs from diffusion models, along with tests and reward computation functions for OCR-based evaluation.

• Adds DiffusionRewardLoopManager and DiffusionRewardLoopWorker for distributed reward computation on image data
• Implements DiffusionRewardManager with support for async reward scoring of image outputs
• Provides OCR reward computation function with generative reward model (GRM) support using image-to-text extraction

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 11 comments.

Show a summary per file

File	Description
verl/experimental/reward_loop/reward_manager/diffusion.py	Implements DiffusionRewardManager that extends RewardManagerBase for image-based reward computation
verl/experimental/reward_loop/diffusion_reward_loop.py	Core implementation of DiffusionRewardLoopWorker and DiffusionRewardLoopManager for distributed image reward processing
verl/experimental/reward_loop/reward_manager/__init__.py	Registers DiffusionRewardManager in module exports
verl/experimental/reward_loop/__init__.py	Exports DiffusionRewardLoopManager for external use
tests/experimental/reward_loop/test_diffusion_reward_model_genrm.py	Adds unit test for DiffusionRewardLoopManager with OCR-based image evaluation
tests/experimental/reward_loop/reward_fn.py	Adds compute_score_ocr function for OCR reward computation using GRM with Levenshtein distance scoring

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The parameters chat and prompt are not used in the prepare_query method body. Consider removing them if they are not needed, or include them in the query if they were intended to be used.

[chenyingshu]

fixed

[zhtmike]

let me know if it is ok for merge

[zhtmike]

# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved. 

[chenyingshu]

fixed

[zhtmike]

need Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved. , CI check the whole sentence

[chenyingshu]

fixed