[worker] fix: respect free_cache_engine flag

[HollowMan6]

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.

Continuation of #1464

Now, recent changes have broken the free_cache_engine option again.

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.

Unit test cases might not be feasible as the sleep/wake_up call can happen anywhere in the codebase. An end-to-end test might be resource-consuming.

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 aims to correctly respect the free_cache_engine flag by adding conditional logic to sleep and wake_up calls for the inference engine. The changes are applied across several worker types and the vLLM rollout implementation. While the overall direction is correct, I've identified a potential issue in vLLMRollout.release where reset_prefix_cache() is called unconditionally. This could lead to performance degradation when free_cache_engine is false, as it would clear the cache that is intended to be preserved. My review includes a specific suggestion to address this.

[vermouth1992]

What's the downside of always turn on free_cache_engine=True?

[HollowMan6]

Please refer to #1464

While always alllowing inference engine to free cache can be ideal, it's unfortunately not supported on some devices, such as AMD Mi250x, since it doesn't support virtual memory management: vllm-project/vllm#12695 (comment)

So we would need to be able to turn it off so that verl can run on those devices.

You can also find more discussion about this starting from this comment: #1464 (comment)

[vermouth1992]

Is there any way to add a test so that it won't break in the future?

[HollowMan6]

Honestly, as mentioned in this PR description, unit test cases might not be feasible as the sleep/wake_up call can be added anywhere in the codebase. An end-to-end test might be resource-consuming and I'm not sure if it's worthwhile doing that, as I assume possible tests can work by patching sleep/wake_up with raising exceptions, and we will need to duplicate and run all the end-to-end test cases with free_cache_engine=False so that we can iterate through all the possible branches to make sure sleep/wake_up calls are controlled by free_cache_engine=False.