[recipe, rollout] feat: enable gpt-oss training for tool agent add gpt-oss for retool recipe by HJSang · Pull Request #3837 · verl-project/verl

HJSang · 2025-10-20T21:17:40Z

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.

add the tool response parsing logic for gpt-oss models
add training recipe for retool

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
- 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 support for gpt-oss models in the tool agent, including a new training recipe and logic for handling tool responses. The refactoring of tool response formatting into a shared utility is a good step towards code reuse. However, this change has led to a critical runtime bug in two files due to incorrect argument types being passed to the new utility function. I've provided specific comments and code suggestions to address these critical issues. Once these fixes are applied, the changes look solid.

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

wuxibin89 · 2025-10-21T02:51:06Z

@HJSang Do you have any gpt-oss performance metrics on retool?

HJSang · 2025-10-21T03:58:29Z

@HJSang Do you have any gpt-oss performance metrics on retool?

I attached details to this issue: #3794.
Overall, we still see an issue for grad norm. It may be due to some inconsistency between inference and training or kernel issue. We will continue investigate.

mirrorboat · 2025-10-22T05:49:57Z

-                lambda: self.tokenizer.apply_chat_template(add_messages, add_generation_prompt=True, tokenize=True),
-            )
-        response_ids = response_ids[len(self.system_prompt) :]
+            if self.tool_parser == "gpt-oss":


@HJSang self.tool_parser is NOT a string, thus the condition of the if statement here is always false.

HJSang · 2025-10-22T06:23:05Z

Good catch. Let me fix it.

…

On Tue, Oct 21, 2025 at 10:50 PM mirrorboat ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In verl/experimental/agent_loop/tool_agent_loop.py <#3837 (comment)>: > @@ -341,11 +344,25 @@ async def _handle_processing_tools_state(self, agent_data: AgentData) -> AgentSt model_inputs = self.processor(text=[raw_tool_response], images=current_images, return_tensors="pt") response_ids = model_inputs.pop("input_ids").squeeze(0).tolist() else: - response_ids = await self.loop.run_in_executor( - None, - lambda: self.tokenizer.apply_chat_template(add_messages, add_generation_prompt=True, tokenize=True), - ) - response_ids = response_ids[len(self.system_prompt) :] + if self.tool_parser == "gpt-oss": @HJSang <https://github.com/HJSang> self.tool_parser is NOT a string, thus the condition of the if statement here is always false. — Reply to this email directly, view it on GitHub <#3837 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AI3YZXRQ5APOQNBJQFIKK5L3Y4LJXAVCNFSM6AAAAACJW7OPBCVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZTGNRTG44TSMBRHA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

…t-oss for retool recipe (verl-project#3837) ### 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. * add the tool response parsing logic for gpt-oss models * add training recipe for retool ### 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` - 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` - [ ] Add / Update [the documentation](https://github.com/volcengine/verl/tree/main/docs). - [ ] 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: ... - [ ] 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).) --------- Co-authored-by: Hejian Sang <hsang@linkedin.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

add gpt-oss for retool recipe

482e14b

gemini-code-assist Bot reviewed Oct 20, 2025

View reviewed changes

Comment thread recipe/langgraph_agent/chat_model.py Outdated

Comment thread verl/experimental/agent_loop/tool_agent_loop.py Outdated

HJSang changed the title ~~[recipe, rollout]: feat, enable gpt-oss training for tool agent add gpt-oss for retool recipe~~ [recipe, rollout] feat: enable gpt-oss training for tool agent add gpt-oss for retool recipe Oct 20, 2025

HJSang and others added 2 commits October 20, 2025 14:19

Update recipe/langgraph_agent/chat_model.py

38061a6

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

Update verl/experimental/agent_loop/tool_agent_loop.py

d684f8f

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

HJSang mentioned this pull request Oct 21, 2025

Agentic RL Support in GPT-OSS #3794

Open

wuxibin89 approved these changes Oct 21, 2025

View reviewed changes

wuxibin89 merged commit 6244d36 into verl-project:main Oct 21, 2025
69 of 70 checks passed

mirrorboat reviewed Oct 22, 2025

View reviewed changes

Navigation Menu

Repository navigation

[recipe, rollout] feat: enable gpt-oss training for tool agent add gpt-oss for retool recipe - #3837

[recipe, rollout] feat: enable gpt-oss training for tool agent add gpt-oss for retool recipe#3837
wuxibin89 merged 3 commits into
verl-project:mainfrom
HJSang:hejian/retool

HJSang commented Oct 20, 2025 •

edited

Loading

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

wuxibin89 commented Oct 21, 2025

Uh oh!

HJSang commented Oct 21, 2025 •

edited

Loading

Uh oh!

Uh oh!

mirrorboat Oct 22, 2025

Uh oh!

HJSang commented Oct 22, 2025 via email

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

HJSang commented Oct 20, 2025 • 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!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

wuxibin89 commented Oct 21, 2025

Uh oh!

HJSang commented Oct 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

mirrorboat Oct 22, 2025

Choose a reason for hiding this comment

Uh oh!

HJSang commented Oct 22, 2025 via email

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

HJSang commented Oct 20, 2025 •

edited

Loading

HJSang commented Oct 21, 2025 •

edited

Loading