[data] feat: TransferQueue - Unify the return of reward by walterchenchn · Pull Request #4902 · verl-project/verl

walterchenchn · 2026-01-13T03:59:43Z

What does this PR do?

Unify the return values of the reward function to make the logic clearer

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.

CLAassistant · 2026-01-13T03:59:50Z

All committers have signed the CLA.

gemini-code-assist

Code Review

This pull request aims to unify the return values of the reward function, which is a good refactoring goal. However, the current implementation introduces a critical bug related to the REMAX baseline calculation and also makes the function's type hint and docstring incorrect and misleading. I've provided two review comments with high and critical severity to address these issues. Please review them carefully.

Updated return type annotation for the reward computation method.

Copilot

Pull request overview

This PR refactors the _compute_or_extract_reward method in the PPO ray trainer to simplify and unify its return value logic. The parameter return_dict is renamed to reward_for_val to better reflect its purpose, and the return value handling is streamlined.

Changes:

Renamed parameter return_dict to reward_for_val to clarify that it controls validation vs training reward computation paths
Simplified return logic: now returns either a tensor (for REMAX with sum_reward) or a tuple (all other cases), removing the dictionary return type
Updated all call sites to match the new return signature (tuple unpacking instead of dictionary access)

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

0oshowero0 · 2026-01-13T09:46:00Z

-            If return_dict=True: dict with "reward_tensor" and "reward_extra_info"
-            If return_dict=False and sum_reward=True: summed reward_tensor (1D tensor)
-            If return_dict=False and sum_reward=False: reward_tensor (2D tensor)
+            If reward_for_val=True: reward_tensor (2D tensor)


Add descriptions for reward_extra_infos

0oshowero0 · 2026-01-13T09:46:37Z

                reward_tensor = reward_tensor.sum(dim=-1)
            reward_extra_info = result.get("reward_extra_info", {})
-            return {"reward_tensor": reward_tensor, "reward_extra_info": reward_extra_info}
+            return reward_tensor, reward_extra_info


unify with reward_extra_info_dict?

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

ji-huazhong · 2026-01-14T08:35:37Z

please run pre-commit run first

walterchenchn · 2026-01-14T10:47:32Z

please run pre-commit run first

done

…#4902) ### What does this PR do? Unify the return values of the reward function to make the logic clearer ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] 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. ```python # 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. - [x] Read the [Contribute Guide](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [x] Apply [pre-commit checks](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md#code-linting-and-formatting): `pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always` - [x] Add / Update [the documentation](https://github.com/volcengine/verl/tree/main/docs). - [x] Add unit or end-to-end test(s) to [the CI workflow](https://github.com/volcengine/verl/tree/main/.github/workflows) to cover all the code. If not feasible, explain why: ... - [x] Once your PR is ready for CI, send a message in [the `ci-request` channel](https://verl-project.slack.com/archives/C091TCESWB1) in [the `verl` Slack workspace](https://join.slack.com/t/verl-project/shared_invite/zt-3855yhg8g-CTkqXu~hKojPCmo7k_yXTQ). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [x] 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`. --------- Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Signed-off-by: 0oshowero0 <o0shower0o@outlook.com>

…#4902) ### What does this PR do? Unify the return values of the reward function to make the logic clearer ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] 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. ```python # 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. - [x] Read the [Contribute Guide](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [x] Apply [pre-commit checks](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md#code-linting-and-formatting): `pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always` - [x] Add / Update [the documentation](https://github.com/volcengine/verl/tree/main/docs). - [x] Add unit or end-to-end test(s) to [the CI workflow](https://github.com/volcengine/verl/tree/main/.github/workflows) to cover all the code. If not feasible, explain why: ... - [x] Once your PR is ready for CI, send a message in [the `ci-request` channel](https://verl-project.slack.com/archives/C091TCESWB1) in [the `verl` Slack workspace](https://join.slack.com/t/verl-project/shared_invite/zt-3855yhg8g-CTkqXu~hKojPCmo7k_yXTQ). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).) - [x] 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`. --------- Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Unify the return reward

51b236d

walterchenchn requested review from PeterSH6, eric-haibin-lin, tongyx361 and vermouth1992 as code owners January 13, 2026 03:59

gemini-code-assist bot reviewed Jan 13, 2026

View reviewed changes

Refine return type annotation in ray_trainer.py

42e16e0

Updated return type annotation for the reward computation method.

walterchenchn changed the title ~~[data] feat: TransferQueue - Unify the return reward~~ [data] feat: TransferQueue - Unify the return of reward Jan 13, 2026

0oshowero0 requested a review from Copilot January 13, 2026 08:57

Copilot started reviewing on behalf of 0oshowero0 January 13, 2026 08:57 View session

Copilot AI reviewed Jan 13, 2026

View reviewed changes

Comment thread verl/trainer/ppo/ray_trainer.py Outdated

Comment thread verl/trainer/ppo/ray_trainer.py Outdated

0oshowero0 reviewed Jan 13, 2026

View reviewed changes

walterchenchn and others added 2 commits January 13, 2026 17:49

Update ray_trainer.py

8ca8ba9

Apply suggestions from code review

775dfd8

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

after pre-commit

f514c38

yyDing1 approved these changes Jan 15, 2026

View reviewed changes

yyDing1 merged commit e204cd8 into verl-project:main Jan 15, 2026
69 of 74 checks passed

dodatboii pushed a commit to TransferQueue/verl that referenced this pull request Jan 30, 2026

adapt verl-project#4902

6c686d9

Signed-off-by: 0oshowero0 <o0shower0o@outlook.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[data] feat: TransferQueue - Unify the return of reward#4902

[data] feat: TransferQueue - Unify the return of reward#4902
yyDing1 merged 5 commits intoverl-project:mainfrom
walterchenchn:tq_dev

walterchenchn commented Jan 13, 2026 •

edited

Loading

Uh oh!

CLAassistant commented Jan 13, 2026 •

edited

Loading

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

0oshowero0 Jan 13, 2026

Uh oh!

walterchenchn Jan 14, 2026

Uh oh!

0oshowero0 Jan 13, 2026

Uh oh!

walterchenchn Jan 14, 2026

Uh oh!

ji-huazhong commented Jan 14, 2026

Uh oh!

walterchenchn commented Jan 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Conversation

walterchenchn commented Jan 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Checklist Before Starting

Test

API and Usage Example

Design & Code Changes

Checklist Before Submitting

Uh oh!

CLAassistant commented Jan 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

0oshowero0 Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

walterchenchn Jan 14, 2026

Choose a reason for hiding this comment

Uh oh!

0oshowero0 Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

walterchenchn Jan 14, 2026

Choose a reason for hiding this comment

Uh oh!

ji-huazhong commented Jan 14, 2026

Uh oh!

walterchenchn commented Jan 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

walterchenchn commented Jan 13, 2026 •

edited

Loading

CLAassistant commented Jan 13, 2026 •

edited

Loading