[rollout] fix: delete problematic assert for max_tokens <= response_length in multi-turn scenario

[PeterSH6]

What does this PR do?

In multi-turn conversations, the length of prompts_id can exceed the original max_prompt_length, causing the current assertion to fail. This occurs because the accumulated prompt tokens from multiple turns grow beyond the initially configured limit.

This PR addresses the immediate assertion failure. Note that the max_tokens and max_model_len assignments should be revisited in a future PR to properly handle dynamic prompt length requirements in multi-turn scenarios.

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).)

[gemini-code-assist]

Code Review

This pull request fixes an assertion failure in multi-turn scenarios by removing the assertion that checks if max_tokens is within the calculated response_length. While this solves the immediate crash, it introduces a risk of passing invalid max_tokens values to the backend generation engines, which could lead to other failures. My review suggests a more robust fix by replacing the removed assertions with logic that caps the max_tokens value to a safe, non-negative range. This ensures stability while still addressing the original problem.

I am having trouble creating individual review comments. Click here to see my feedback.

verl/workers/rollout/sglang_rollout/async_sglang_server.py (274-276)

Simply removing the assertion is risky. If a user provides max_new_tokens larger than the available response_length, it could cause the backend generation engine to fail, especially since response_length can be negative if the prompt is too long.

A safer approach is to cap max_new_tokens to the valid range (i.e., between 0 and response_length). This handles both cases: when max_new_tokens is user-provided and might be too large, and when response_length itself is negative due to a long prompt.

        max_new_tokens = max(0, min(max_new_tokens, response_length))

verl/workers/rollout/vllm_rollout/vllm_async_server.py (473-475)

Simply removing the assertion is risky. If a user provides max_tokens larger than the available response_length, it could cause the backend vLLM engine to fail, especially since response_length can be negative if the prompt is too long.

A safer approach is to cap max_tokens to the valid range (i.e., between 0 and response_length). This handles both cases: when max_tokens is user-provided and might be too large, and when response_length itself is negative due to a long prompt.

        max_tokens = max(0, min(max_tokens, response_length))