[Misc][V1] Enhance performance of KVCacheManager._get_cached_block

[imkero]

Enhance KVCacheManager cached block dict retrieve by:

• fusing key in dict check and dict getitem operation
• cache dict.get fn (cache the fn itself, rather than access dict.get every time) will get better performance
• use for _ in dict.values() for better performance than next(iter(dict.values())) because for loop use native GET_ITER and FOR_ITER instruction with better performance

A slight time cost grow in case of "unknown block", but we may expect one or more "cached block" hit when enable_prefix_cache=True, and it would act like cache_hit, cache_hit, cache_hit, ..., cache_hit, cache_miss. So I think it's worth to choose this impl?

Benchmark

repeat 10000 times

code	unknown block	cached block
main	0.00107s	0.00427s
ver 0 (worse performance when cache miss)	0.00139s (+29.9%)	0.00225s (-47.3%)
ver 1 (this PR)	0.00114s (+6.5%)	0.00179s (-58.1%)

Reference

10+ years performance problem of dict.get

https://www.reddit.com/r/Python/comments/1atp5s/comment/c90onqq/

Iteration instruction compare

def v0(dict):
    return next(iter(dict.values()), None)
---
  1           0 RESUME                   0

  2           2 LOAD_GLOBAL              1 (NULL + next)
             12 LOAD_GLOBAL              3 (NULL + iter)
             22 LOAD_FAST                0 (dict)
             24 LOAD_ATTR                5 (NULL|self + values)
             44 CALL                     0
             52 CALL                     1
             60 LOAD_CONST               0 (None)
             62 CALL                     2
             70 RETURN_VALUE

def v1(dict):
    for v in dict.values():
        return v
    return None
---
  1           0 RESUME                   0

  2           2 LOAD_FAST                0 (dict)
              4 LOAD_ATTR                1 (NULL|self + values)
             24 CALL                     0
             32 GET_ITER
             34 FOR_ITER                 5 (to 48)
             38 STORE_FAST               1 (v)

  3          40 LOAD_FAST                1 (v)
             42 SWAP                     2
             44 POP_TOP
             46 RETURN_VALUE

  2     >>   48 END_FOR

  4          50 RETURN_CONST             0 (None)

[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.

🚀