[rollout, vllm, sglang] fix: allow user customization of repetition_penalty to avoid watchdog timeout during GRPO rollout (volcengine#3309)

Allow user customization of `repetition_penalty` to avoid watchdog
timeout during GRPO rollout

### What does this PR do?

This PR adds an interface for users to specify `repetition_penalty`,
which helps avoid repetition in LLM generation and prevents watchdog
timeouts during GRPO rollout. If not specified, `repetition_penalty`
will remain at its default value of `1.0`.


### Checklist Before Starting

- [X] Search for similar PRs. No similar PRs found.
- [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

This PR can be vetted by existing CI test cases.

### API and Usage Example

Previously, users could not specify `repetition_penalty`, but this PR
adds support for it.

For example, users can now start GRPO training with a command like:

```bash
python -m verl.trainer.main_ppo \
    +actor_rollout_ref.rollout.repetition_penalty=1.05 \
    # other params here...
```

### Design & Code Changes

This PR adds an interface allowing users to specify the
`repetition_penalty` (e.g., `1.05`), while maintaining backward
compatibility with the default value of `1.0`.

### 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).)

Author: Mighten

verl/workers/rollout/sglang_rollout/async_sglang_server.py

@@ -82,6 +82,7 @@ async def generate(
         request_id: str,
         image_data: Optional[list[Any]] = None,
     ) -> TokenOutput:
+        sampling_params.setdefault("repetition_penalty", self.config.rollout.get("repetition_penalty", 1.0))
         return await self.master_worker.generate.remote(prompt_ids, sampling_params, request_id, image_data=image_data)
 
     async def wake_up(self):

verl/workers/rollout/sglang_rollout/sglang_rollout.py

@@ -481,7 +481,7 @@ def _init_sampling_params(self, **kwargs):
             max_new_tokens=self.config.response_length,
             presence_penalty=0.0,
             frequency_penalty=0.0,
-            repetition_penalty=1.0,
+            repetition_penalty=self.config.get("repetition_penalty", 1.0),
         )
         # supporting adding any sampling params from the config file
         for k in self.config.keys():

verl/workers/rollout/vllm_rollout/vllm_async_server.py

@@ -350,6 +350,7 @@ async def generate(
     ) -> TokenOutput:
         max_tokens = self.max_model_len - len(prompt_ids)
         sampling_params["logprobs"] = 0 if sampling_params.pop("logprobs", False) else None
+        sampling_params.setdefault("repetition_penalty", self.config.rollout.get("repetition_penalty", 1.0))
         sampling_params = SamplingParams(max_tokens=max_tokens, **sampling_params)
         prompt_ids = _qwen2_5_vl_dedup_image_tokens(prompt_ids, self.processor)
         prompt = TokensPrompt(

verl/workers/rollout/vllm_rollout/vllm_rollout_spmd.py

@@ -206,6 +206,7 @@ def __init__(
             n=1,
             logprobs=0,  # can be set to 0 and let actor to recompute
             max_tokens=config.response_length,
+            repetition_penalty=config.get("repetition_penalty", 1.0),
         )
 
         kwargs["detokenize"] = False