[v1] fix compilation cache

[youkaichao]

compilation cache is not working due to a bug, self.to_be_compiled_sizes should be self.compile_sizes.copy() rather than self.compile_sizes.union(self.capture_sizes) .

this should be merged after #11596 . our compilation cache does not consider the code change in the model's code.

before this pr:

Compiling a graph for general shape takes 12.69 s

After this pr:

Compiling a graph for general shape takes 2.27 s

cc @tlrmchlsmth how can we take all these code into consideration? like rope / activation etc.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[tlrmchlsmth]

our compilation cache does not consider the code change in the model's code.

I thought this would be handled by adding the version number here. Maybe we need to explicitly add the git hash as well?

vllm/vllm/config.py

Line 3000 in faef77c

vllm_factors.append(__version__)

[tlrmchlsmth]

Why do we need to append "None" now?

[youkaichao]

To be safer, to distinguish:

config 1: None
config 2: hash abc

config 1: hash abc
config 2: None

[youkaichao]

I thought this would be handled by adding the version number here. Maybe we need to explicitly add the git hash as well?

that would be kind of overkill, for the developers (like me), the git hash changes from time to time, and it means I cannot reuse any compilation cache.