Remove unused kwargs from model definitions

[hmellor]

Follow up for #11967 which removes kv_cache and attn_metadata from all model definitions.

Summary of changes:

• Remove kv_cache and attn_metadata from Attention.forward() args
• Use forward context instead of attn_metadata arg in MambaMixer and MambaMixer2
• Remove kv_caches, kv_cache and attn_metadata from forward() args of all model modules
• Remove kv_caches and attn_metadata from new model docs
• Leave kv_caches arg (but try not to use it) in all child classes of ModelRunnerBase.execute_model() to avoid further complication

[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.

🚀

[DarkLight1337]

Can you fix the failure in helm chart CI?

[hmellor]

Can you fix the failure in helm chart CI?

This failure is likely due to the CPU runner not being updated correctly. If I can fix the CPU model runner in buildkite then this workflow should also pass.

[hmellor]

The above commit should stop errors coming from Attention.forward.

I'm not sure what to do about CustomModelForCausalLM.forward. I could add some code to inspect the function signature in all the model runners and pass None to kv_caches and attn_metadata if needed?

[youkaichao]

@youkaichao will this break out of tree models?

I think it's fine to ask out-of-tree models to update to new vLLM versions. We cannot let out-of-tree models to slow down the development. In fact we can remove these code several months ago, and I think we have waited for long enough.

[youkaichao]

agree on the direction, thanks for cleaning up the code!

[hmellor]

Reverting the deprecation of kv_cache and attn_metadata from Attention.forward because it was only a half measure that does not fully solve the problem of breaking out of tree models.

This way, it is unambiguous that we do not expect these arguments in Model.forward or Attention.forward.

[heheda12345]

I think it won't be difficult for out-of-tree models to remove these kwargs. One thing that is already broken is models with kv cache sharing (multiple attention layers using the same kv cache). It exists in open source models but not in vllm repo. We need to design the interface for it after we get the first pr for kv cache sharing model.

I think we need to tell the users how to debug kv cache related problems. These two variables are accessed frequently during debugging the models.

[heheda12345]

Not necessary in this pr, but can we update these notes to help people print kv cache & attn_metadata to debug kv-cache related problems?