[sleep mode] clear pytorch cache after sleep

[lionelvillard]

Clear PyTorch cache right after putting the instance to sleep.

For ibm-granite/granite-20b-code-instruct-8k, memory use before this fix is:

INFO 03-20 19:40:47 [gpu_worker.py:81] Sleep mode freed 37.55 GiB memory, 3.29 GiB memory is still in use.

And with this fix:

INFO 03-20 19:45:25 [gpu_worker.py:81] Sleep mode freed 39.48 GiB memory, 1.36 GiB memory is still in use.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[maxdebayser]

Intuitively this makes total sense. But I don't know if there could be any side effects.

[njhill]

Also LGTM but would also be good to get a nod from @youkaichao.

[youkaichao]

i think you are clearing some buffers for sampling logits. it works, but people have to be careful, because when you wake up the engine, and serve the requests, samplers will allocate that part of memory again.

[youkaichao]

LGTM if you want to reclaim as much memory as possible when sleeping.

[lionelvillard]

Can we merge this PR?