[ROCm] Auto-Select Attention Backend

[vllmellm]

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

The use of environment variables , especially for Aiter kernels on ROCm, has been a pain point for some users as mentioned in #21138.

This PR introduces:

• Selection of attention backends for ROCm based on priority(performance) and support instead of environment variables.
• Graceful handling of unsupported attention backends.

Additionally, the attention selection logic in this PR maintains the ability to force a backend thorough the VLLM_ATTENTION_BACKEND variable, allowing users to easily switch backends.

Although the selection is implemented for ROCm hardware only, it can be extended to other hardwares in future PRs.

Test Plan

Implement unit test for the backend selection function. To run, use the following command

pytest tests/attention/test_attention_selector.py

Test Result

tests/attention/test_attention_selector.py ................                                                               [100%]

====================================================== 16 passed in 2.85s =======================================================

[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]

Code Review

The pull request refactors the pa rocm code. The changes include adding new parameters to the _IsSupported dataclass and is_attn_backend_supported function in vllm/attention/selector.py, modifying environment variables in vllm/envs.py, and updating attention backend selection logic in vllm/platforms/rocm.py. Additionally, new files and modifications are introduced to handle attention backends in vllm/v1/attention/backends/.

[DarkLight1337]

This should be inside selector.py imo. Also it is rather confusing that not all existing backends are considered in this function.

[vllmellm]

For now, we are only considering attention backends used with rocm on V1. We are considering supporting all attention backends in a future PRs; however, we will add some comments to clarify for other developers.

[DarkLight1337]

I think that it is more straightforward to set VLLM_ATTENTION_BACKEND directly. If this is only used to help test the attention selector, we can directly patch global variables.

[tjtanaa]

One of the motivations is to follow kernels abstraction where developer can define which attention backend can run on which hardware based on the dependencies available in the environment. So, we can always pick the fastest backend for a hardware and not to always use Triton implementation as a default.

The abstraction is also to make it clear to developer and user where they should find these custom logic where defines the default behavior of vLLM's attention backend.

[mergify]

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

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

[mergify]

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

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

[vllmellm]

cc @LucasWilkinson @DarkLight1337 @tjtanaa

[DarkLight1337]

IMO this part should be handled by the platform

[DarkLight1337]

Or at least, it needs to accept the platform being used

[tjtanaa]

Or at least, it needs to accept the platform being used
@DarkLight1337 @vllmellm
I think the second approach is better, it delegates the responsibility to the Attention class itself. All the checks if a backend is supported should be centralized in the class itself.
platform should just a place to retrieve platform information, it should not determine if an attention backend can run or not.

This can improve readability and maintainability.

[mergify]

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

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