[Misc] Add FA2 support to ViT MHA layer #12355

Isotr0py · 2025-01-23T13:14:26Z

Add FA2 support to ViT MHA layer.
I leave Qwen2-VL's ViT MHA layer consolidation to be done in a separate PR, because it has head_size=80 and doesn't work with vllm_flash_attn (head_size should be a multiple of 32).

Signed-off-by: Isotr0py <[email protected]>

github-actions · 2025-01-23T13:14:42Z

👋 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.

🚀

DarkLight1337 · 2025-01-24T06:04:38Z

QQ: Do we have tests for MHA layer?

Isotr0py · 2025-01-24T06:59:27Z

There is no test for MHA layer yet, we can add one to kernel tests in this PR later.

Signed-off-by: Isotr0py <[email protected]>

DarkLight1337

LGTM once you get the tests to pass

WoosukKwon · 2025-01-26T01:17:15Z

Unfortunately, this PR breaks the CI because the concurrent FA3 PR #12093 removed flash_attn_func.

This reverts commit f1fc051.

* [Misc] Use VisionArena Dataset for VLM Benchmarking (vllm-project#12389) Signed-off-by: Roger Wang <[email protected]> * [ci/build] fix wheel size check (vllm-project#12396) Signed-off-by: youkaichao <[email protected]> * [Hardware][Gaudi][Doc] Add missing step in setup instructions (vllm-project#12382) * [ci/build] sync default value for wheel size (vllm-project#12398) Signed-off-by: youkaichao <[email protected]> * [Misc] Enable proxy support in benchmark script (vllm-project#12356) Signed-off-by: Junichi Sato <[email protected]> * [Bugfix][Kernel] Fix CUDA 11.8 being broken by FA3 build (vllm-project#12375) Signed-off-by: Lucas Wilkinson <[email protected]> * [Misc] Remove deprecated code (vllm-project#12383) Signed-off-by: DarkLight1337 <[email protected]> * [Bugfix][Kernel] FA3 Fix - RuntimeError: This flash attention build only supports pack_gqa (for build size reasons). (vllm-project#12405) Signed-off-by: Lucas Wilkinson <[email protected]> * [Bugfix][Kernel] Fix moe align block issue for mixtral (vllm-project#12413) * [Bugfix] Fix BLIP-2 processing (vllm-project#12412) Signed-off-by: DarkLight1337 <[email protected]> * [ROCm][MoE] MI300 tuned configs Mixtral-8x(7B,22B) | fp16, fp8 (vllm-project#12408) Signed-off-by: Divakar Verma <[email protected]> * [Misc] Add FA2 support to ViT MHA layer (vllm-project#12355) Signed-off-by: Isotr0py <[email protected]> * [TPU][CI] Update torchxla version in requirement-tpu.txt (vllm-project#12422) Signed-off-by: Siyuan Liu <[email protected]> * [Misc][Bugfix] FA3 support to ViT MHA layer (vllm-project#12435) Signed-off-by: Roger Wang <[email protected]> Signed-off-by: Isotr0py <[email protected]> Co-authored-by: Isotr0py <[email protected]> * [V1][Perf] Reduce scheduling overhead in model runner after cuda sync (vllm-project#12094) Signed-off-by: Keyun Tong <[email protected]> * [V1][Bugfix] Fix assertion when mm hashing is turned off (vllm-project#12439) Signed-off-by: Roger Wang <[email protected]> * [Misc] Revert FA on ViT vllm-project#12355 and vllm-project#12435 (vllm-project#12445) * [Frontend] generation_config.json for maximum tokens(vllm-project#12242) Signed-off-by: Matthew Hendrey <[email protected]> Signed-off-by: Shangming Cai <[email protected]> Signed-off-by: youkaichao <[email protected]> Signed-off-by: Harry Mellor <[email protected]> Signed-off-by: Yuan Tang <[email protected]> Signed-off-by: Isotr0py <[email protected]> Signed-off-by: DarkLight1337 <[email protected]> Signed-off-by: Chen Zhang <[email protected]> Signed-off-by: wangxiyuan <[email protected]> Co-authored-by: shangmingc <[email protected]> Co-authored-by: youkaichao <[email protected]> Co-authored-by: Harry Mellor <[email protected]> Co-authored-by: Yuan Tang <[email protected]> Co-authored-by: Isotr0py <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Chen Zhang <[email protected]> Co-authored-by: wangxiyuan <[email protected]> * [Bugfix] Disable w16a16 2of4 sparse CompressedTensors24 (vllm-project#12417) Signed-off-by: Tyler Michael Smith <[email protected]> Co-authored-by: mgoin <[email protected]> * [Bugfix/CI] Fix broken kernels/test_mha.py (vllm-project#12450) * [Bugfix][Kernel] Fix perf regression caused by PR vllm-project#12405 (vllm-project#12434) Signed-off-by: Lucas Wilkinson <[email protected]> * [Build/CI] Fix libcuda.so linkage (vllm-project#12424) Signed-off-by: Tyler Michael Smith <[email protected]> * [Frontend] Rerank API (Jina- and Cohere-compatible API) (vllm-project#12376) Signed-off-by: Kyle Mistele <[email protected]> * [DOC] Add link to vLLM blog (vllm-project#12460) Signed-off-by: Yuan Tang <[email protected]> * [V1] Avoid list creation in input preparation (vllm-project#12457) Signed-off-by: Woosuk Kwon <[email protected]> * [Frontend] Support scores endpoint in run_batch (vllm-project#12430) Signed-off-by: Pooya Davoodi <[email protected]> * [Bugfix] Fix Granite 3.0 MoE model loading (vllm-project#12446) Signed-off-by: DarkLight1337 <[email protected]> * [Bugfix] Fix missing seq_start_loc in xformers prefill metadata (vllm-project#12464) Signed-off-by: Isotr0py <[email protected]> * [V1][Minor] Minor optimizations for update_from_output (vllm-project#12454) Signed-off-by: Woosuk Kwon <[email protected]> * [Bugfix] Fix gpt2 GGUF inference (vllm-project#12467) Signed-off-by: Isotr0py <[email protected]> --------- Signed-off-by: Roger Wang <[email protected]> Signed-off-by: youkaichao <[email protected]> Signed-off-by: Junichi Sato <[email protected]> Signed-off-by: Lucas Wilkinson <[email protected]> Signed-off-by: DarkLight1337 <[email protected]> Signed-off-by: Divakar Verma <[email protected]> Signed-off-by: Isotr0py <[email protected]> Signed-off-by: Siyuan Liu <[email protected]> Signed-off-by: Keyun Tong <[email protected]> Signed-off-by: Matthew Hendrey <[email protected]> Signed-off-by: Shangming Cai <[email protected]> Signed-off-by: Harry Mellor <[email protected]> Signed-off-by: Yuan Tang <[email protected]> Signed-off-by: Chen Zhang <[email protected]> Signed-off-by: wangxiyuan <[email protected]> Signed-off-by: Tyler Michael Smith <[email protected]> Signed-off-by: Kyle Mistele <[email protected]> Signed-off-by: Woosuk Kwon <[email protected]> Signed-off-by: Pooya Davoodi <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: youkaichao <[email protected]> Co-authored-by: Mohit Deopujari <[email protected]> Co-authored-by: Junichi Sato <[email protected]> Co-authored-by: Lucas Wilkinson <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: ElizaWszola <[email protected]> Co-authored-by: Divakar Verma <[email protected]> Co-authored-by: Isotr0py <[email protected]> Co-authored-by: Siyuan Liu <[email protected]> Co-authored-by: Isotr0py <[email protected]> Co-authored-by: Keyun Tong <[email protected]> Co-authored-by: Matthew Hendrey <[email protected]> Co-authored-by: shangmingc <[email protected]> Co-authored-by: Harry Mellor <[email protected]> Co-authored-by: Yuan Tang <[email protected]> Co-authored-by: Chen Zhang <[email protected]> Co-authored-by: wangxiyuan <[email protected]> Co-authored-by: Tyler Michael Smith <[email protected]> Co-authored-by: mgoin <[email protected]> Co-authored-by: Kyle Mistele <[email protected]> Co-authored-by: Woosuk Kwon <[email protected]> Co-authored-by: Pooya Davoodi <[email protected]>

Signed-off-by: Isotr0py <[email protected]>

…lm-project#12445)

Signed-off-by: Isotr0py <[email protected]>

…lm-project#12445)

Signed-off-by: Isotr0py <[email protected]>

…lm-project#12445) Signed-off-by: Isotr0py <[email protected]>

add FA2 support to MHA layer

6aba458

Signed-off-by: Isotr0py <[email protected]>

add mha kernel test

e8676c8

Signed-off-by: Isotr0py <[email protected]>

Isotr0py requested review from tlrmchlsmth and WoosukKwon as code owners January 24, 2025 14:47

fix dtype

9da906e

Signed-off-by: Isotr0py <[email protected]>

DarkLight1337 approved these changes Jan 24, 2025

View reviewed changes

Isotr0py added the ready ONLY add when PR is ready to merge/full CI is needed label Jan 24, 2025

DarkLight1337 merged commit f1fc051 into vllm-project:main Jan 25, 2025
61 checks passed

Isotr0py deleted the vit-attn-backend branch January 25, 2025 11:09

WoosukKwon added a commit that referenced this pull request Jan 26, 2025

Revert "[Misc] Add FA2 support to ViT MHA layer (#12355)"

1be4877

This reverts commit f1fc051.

LucasWilkinson mentioned this pull request Jan 26, 2025

Add back flash_attn_func api (and support FA3) [Don't Merge Yet] vllm-project/flash-attention#40

Open

ywang96 mentioned this pull request Jan 26, 2025

[Misc][Bugfix] FA3 support to ViT MHA layer #12435

Merged

WoosukKwon pushed a commit that referenced this pull request Jan 26, 2025

[Misc] Revert FA on ViT #12355 and #12435 (#12445)

a525527

tjtanaa pushed a commit to EmbeddedLLM/vllm that referenced this pull request Jan 28, 2025

[Misc] Add FA2 support to ViT MHA layer (vllm-project#12355)

e516889

Signed-off-by: Isotr0py <[email protected]>

tjtanaa pushed a commit to EmbeddedLLM/vllm that referenced this pull request Jan 28, 2025

[Misc] Revert FA on ViT vllm-project#12355 and vllm-project#12435 (vl…

2192644

…lm-project#12445)

rasmith pushed a commit to rasmith/vllm that referenced this pull request Jan 30, 2025

[Misc] Add FA2 support to ViT MHA layer (vllm-project#12355)

a62c186

Signed-off-by: Isotr0py <[email protected]>

rasmith pushed a commit to rasmith/vllm that referenced this pull request Jan 30, 2025

[Misc] Revert FA on ViT vllm-project#12355 and vllm-project#12435 (vl…

bec24a2

…lm-project#12445)

Isotr0py added a commit to Isotr0py/vllm that referenced this pull request Feb 2, 2025

[Misc] Add FA2 support to ViT MHA layer (vllm-project#12355)

5e64358

Signed-off-by: Isotr0py <[email protected]>

Isotr0py pushed a commit to Isotr0py/vllm that referenced this pull request Feb 2, 2025

[Misc] Revert FA on ViT vllm-project#12355 and vllm-project#12435 (vl…

44b8833

…lm-project#12445) Signed-off-by: Isotr0py <[email protected]>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Misc] Add FA2 support to ViT MHA layer #12355

[Misc] Add FA2 support to ViT MHA layer #12355

Isotr0py commented Jan 23, 2025 •

edited by github-actions bot

Loading

github-actions bot commented Jan 23, 2025

DarkLight1337 commented Jan 24, 2025 •

edited

Loading

Isotr0py commented Jan 24, 2025

DarkLight1337 left a comment

WoosukKwon commented Jan 26, 2025 •

edited

Loading

[Misc] Add FA2 support to ViT MHA layer #12355

[Misc] Add FA2 support to ViT MHA layer #12355

Conversation

Isotr0py commented Jan 23, 2025 • edited by github-actions bot Loading

github-actions bot commented Jan 23, 2025

DarkLight1337 commented Jan 24, 2025 • edited Loading

Isotr0py commented Jan 24, 2025

DarkLight1337 left a comment

Choose a reason for hiding this comment

WoosukKwon commented Jan 26, 2025 • edited Loading

Isotr0py commented Jan 23, 2025 •

edited by github-actions bot

Loading

DarkLight1337 commented Jan 24, 2025 •

edited

Loading

WoosukKwon commented Jan 26, 2025 •

edited

Loading