[vllm] fix: reset all caches after weight updates

[s-isaev]

What does this PR do?

This PR resets all available vLLM rollout caches after updating model weights.

Previously, verl only reset the prefix/KV cache after update_weights. For multimodal rollouts, vLLM may also cache multimodal inputs and encoder outputs. Reusing those cached values after a weight update can lead to stale features being used during rollout.

This change adds a clear_all_caches helper on the vLLM HTTP server and calls it after weight updates.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: https://github.com/verl-project/verl/pulls?q=is%3Apr+is%3Aopen+vllm+cache
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, 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, fully_async, one_step_off
  • 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.

[gemini-code-assist]

Code Review

This pull request introduces a clear_all_caches method in vllm_async_server.py to clear KV, multi-modal, and encoder caches, and updates vllm_rollout.py to call this new method instead of only clearing the KV cache after updating weights. Feedback was provided to replace fragile hardcoded vLLM version checks with dynamic hasattr checks to verify the presence of the cache-clearing methods on the engine.

[gemini-code-assist]

Using hardcoded version checks like _VLLM_VERSION >= version.parse('0.16.0') is fragile and can lead to bugs. For instance, current vLLM releases are in the 0.7.x range, meaning the check for 0.16.0 will evaluate to False and prevent reset_encoder_cache() from being called even if the installed vLLM version supports it.

A much more robust and idiomatic approach is to dynamically check for the presence of these methods on the engine using hasattr(). This ensures compatibility across different vLLM versions, custom forks, or backported features without relying on hardcoded version strings.

Suggested change

	if self.node_rank == 0:
	if _VLLM_VERSION >= version.parse("0.9.0"):
	await self.engine.reset_mm_cache()
	if _VLLM_VERSION >= version.parse("0.16.0"):
	await self.engine.reset_encoder_cache()
	if self.node_rank == 0:
	if hasattr(self.engine, 'reset_mm_cache'):
	await self.engine.reset_mm_cache()
	if hasattr(self.engine, 'reset_encoder_cache'):
	await self.engine.reset_encoder_cache()