[v1][attention] Support Hybrid Allocator + FlashInfer

[heheda12345]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

Support hybrid allocator + flashinfer backend. Achieved by letting the attention backend know the set of layers that use this backend, and only performs plan for these layers.

Limitation:
For a model with both sliding window attention and full attention, when hybrid allocator is disabled, both the two types of layer use the same attention metadata builder, and flashinfer cannot handle this case. As a temporary solution, I add an error message to tell user to set disable_sliding_window manually. (This config was set automatically when using flashinfer backend before this PR).

Test Plan

Run basic.py with "google/gemma-3-1b-it"

Test Result

Not cap the max length to the sliding window size, and generate meaningful result

VLLM_ATTENTION_BACKEND=FLASHINFER python3 basic.py

output
------------------------------------------------------------
Prompt:    'Hello, my name is'
Output:    " Alex. I'm a software engineer.\n\nI'm passionate about creating"
------------------------------------------------------------
Prompt:    'The president of the United States is'
Output:    ' currently Joe Biden. He was elected in November 2020.\n\n'
------------------------------------------------------------
Prompt:    'The capital of France is'
Output:    ' Paris.\nThe largest city in the world is Tokyo.\nThe most populous'
------------------------------------------------------------
Prompt:    'The future of AI is'
Output:    ' being shaped by a critical question: can we build AI that is aligned with human'
------------------------------------------------------------

(Optional) Documentation Update

[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.

🚀

[gemini-code-assist]

Warning

Gemini is unable to generate a review due to a potential policy violation.

[heheda12345]

CC @LucasWilkinson @WoosukKwon

[heheda12345]

@LucasWilkinson seems that we met the same problem when hybrid allocator is disabled as here. #21093 (comment)

[WoosukKwon]

For models like Gemma 3, is it using flash-attn for global attn and flashinfer for sliding window attn?

[heheda12345]

@WoosukKwon No, all layers are using flashinfer. Mixing flash-attn + flashinfer should be easy with the current design but prefer to put it into next pr if you think it is needed.

[WoosukKwon]

@heheda12345 Great. I think we shouldn't mix the two.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @heheda12345.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork