[PP] Correct cache size check

[zhengy001]

Single long prompt gets preempted(as cannot allocate slots) during decode and leads to recomput and incorrect output length.

Basically, it is because CacheEngine splits kv cache into separate pipeline by self.num_gpu_blocks //= parallel_config.pipeline_parallel_size. Long prompt does not get blocked even though it exceeds kv cache capacity. There should be the same logic for kv cache size check.

[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.

🚀

[youkaichao]

@andoorve can you help take a look?

[andoorve]

This makes sense, thanks for taking a look!

[zhengy001]

Hi @DarkLight1337,
I ran into this CI error. And your PR ce26b16 made change prompt_tokens from 202 to 201 and it is 202 again. I dont think it is related to this PR. Could you please take a look?

[DarkLight1337]

Can you pull from latest main again? It might be fixed by #13890

[zhengy001]

Can you pull from latest main again? It might be fixed by #13890

Sure.

[DarkLight1337]

Doesn't look like it's fixed. Let me open another PR.

Updated: Opened #13895