[core] gemma2 full context length support

[youkaichao]

the scheduler treats it as a model without sliding window, and sliding window is only used for computation.

FIX #6220
FIX #8580

[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.

🚀

[youkaichao]

fixes #9517

[WoosukKwon]

Just a heads up: the paged attention kernel we use for the xFormers backend doesn't support sliding window attention. This PR will introduce a slight correctness bug in the xformers backend.

[youkaichao]

Just a heads up: the paged attention kernel we use for the xFormers backend doesn't support sliding window attention. This PR will introduce a slight correctness bug in the xformers backend.

for xformers, we keep the original behavior of capping the max-model-length.

[noamgat]

This looks great! Which attention backend do you recommend for gemma 2 now?

[youkaichao]

@noamgat the default one (flash attention) should work.

[azsh1725]

Hi! Thanks for this fix.

Can you, please, tell me when the release with this fix is ​​planned?

[azsh1725]

Hi! Thanks for this fix.

Can you, please, tell me when the release with this fix is ​​planned?

For those interested, the release with this fix is ​​version 0.6.5