[FA/Chore] Bump FA version

[LucasWilkinson]

Bump vLLM Flash Attention to pickup:

vllm-project/flash-attention#96
vllm-project/flash-attention#95

[gemini-code-assist]

Code Review

This pull request updates the vllm-flash-attn dependency to a newer commit to incorporate fixes for building with recent CUDA versions. The change is correct and necessary. My review includes one suggestion to improve the long-term maintainability of the build configuration by adding a comment to explain the purpose of the new commit hash, as using raw commit hashes without context can make future updates and debugging more difficult.

[gemini-code-assist]

Using a raw commit hash for a dependency makes it difficult to understand what version is being used and why it was chosen. This can complicate future maintenance and updates. It is a best practice to use version tags for dependencies. If a tag is not available for this commit, please add a comment explaining what this commit hash corresponds to (e.g., what features or fixes it includes) and why it was chosen. This will improve the maintainability of the build system.

          # Pick up fixes for CUDA 12.4 (GH#95) and 12.5 (GH#96)
          GIT_TAG 45b5dac5497e848d788af686bb68cbe5cf2e56bc

[johnnynunez]

super! thank you guys!

[johnnynunez]

@LucasWilkinson @ProExpertProg @mgoin could you review this? #26098

[johnnynunez]

@ProExpertProg i don't know why is it failing, because im using it on my gh200/thor and compiles and works well

[jmkuebler]

@LucasWilkinson can you make sense of the logs? I can't relate it to the changes that are being picked up.

[ProExpertProg]

Can someone try to reproduce the CI failures? It might just be flaky tests

[jmkuebler]

@ProExpertProg I installed the fork from source (to really make it pick up the FA3 changes) and reran the two tests that failed:

• ✅ pytest tests/entrypoints/openai/test_completion_with_function_calling::test_tool_id_kimi_k2[required-True-Qwen/Qwen3-0.6B] --> 1 passed, 1 warning in 100.42s (failed here https://buildkite.com/vllm/ci/builds/33751/steps/canvas?jid=0199bb64-e3eb-4a9c-b08f-df0f064bb67c#0199bb64-e3eb-4a9c-b08f-df0f064bb67c/156-2707)
• ✅ pytest tests/v1/distributed/test_async_llm_dp.py --> 8 passed, 17 warnings in 257.25s (0:04:17) (failed all 8 here https://buildkite.com/vllm/ci/builds/33628/steps/canvas?jid=0199b666-16af-49aa-b2bd-5e5c5f188375#0199b666-16af-49aa-b2bd-5e5c5f188375/209-2296)

So I think this PR is not the issue here. Can we restart the test or overwrite?

[mgoin]

Actually the Distributed Tests (2 GPUs) failure looks different from the ones on main

PR: https://buildkite.com/vllm/ci/builds/33628/steps/canvas?jid=0199b666-16af-49aa-b2bd-5e5c5f188375
Nightly: https://buildkite.com/vllm/ci/builds/34132/steps/canvas?sid=0199c721-28cd-4a02-b8bd-3864658542ef

And this landed commit on main should have fixed it https://buildkite.com/vllm/ci/builds/34222/steps/canvas?jid=0199c900-b19d-4a2c-ab37-15327e7d63b0

[mergify]

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

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

[LucasWilkinson]

superseded by: #25049