Repository navigation

[Kernel] Integrate FA4 SM90 FP8 MMA for Gemma-4 full attention - #60372

Merged

vllm-bot merged 26 commits into

vllm-project:mainfrom

katec846:fa4-fp8mma-draft-20261007

Nov 5, 2026

Merged

[Kernel] Integrate FA4 SM90 FP8 MMA for Gemma-4 full attention#60372
vllm-bot merged 26 commits into
vllm-project:mainfrom
katec846:fa4-fp8mma-draft-20261007

Commits on Oct 27, 2026
- [Attention] Integrate SM90 d512 FP8 MMA with vLLM routing
  
  authored and
  katec846
  committed
  a962321
- [Test] Supply parallel config for SM90 attention fixtures
  
  authored and
  katec846
  committed
  0530c91
- [Test] Narrow FP8 rejection reason for mypy
  
  katec846
  and
  codex
  committed
  d20b103
- [Attention] Consolidate SM90 FP8 wrapper dispatch flags
  
  katec846
  and
  codex
  committed
  5fc01d3
- [Attention] Guard unvalidated SM90 FP8 noncausal full attention
  
  katec846
  and
  codex
  committed
  7ea8800
- [Review] Trim SM90 FP8 selector tests and comments
  
  katec846
  and
  codex
  committed
  b283eac
- [Review] Remove redundant Diff-KV probe cache check
  
  katec846
  and
  codex
  committed
  e9344ed
- [Review] Remove remaining added attention tests
  
  katec846
  and
  codex
  committed
  fc937f5
- [Review] Clarify FP8 MMA comments and noncausal error
  
  katec846
  and
  codex
  committed
  847cee6
- [Review] Simplify SM90 FP8 guards and error messages
  
  katec846
  and
  codex
  committed
  ac73cba
- [Attention] Remove the unreachable FA4 -> FA3 FP8 fallback on SM90
  
  simon-veitner-redhat
  authored and
  katec846
  committed
  6c178bc
- [Attention] Reject DCP with FA4 SM90 FP8 KV at impl init
  
  simon-veitner-redhat
  authored and
  katec846
  committed
  f116077
- [Test] Cover FA4 SM90 FP8 MMA routing in a CPU selector test
  
  simon-veitner-redhat
  authored and
  katec846
  committed
  d0d39b0
- [CI] Add Hopper GSM8K eval for Gemma-4 with FP8 KV (FA4 SM90 FP8 MMA)
  
  simon-veitner-redhat
  authored and
  katec846
  committed
  9fe5c82
- Update SM90 FP8 MMA FlashAttention review fixes
  
  katec846
  and
  codex
  committed
  237043e
Commits on Oct 29, 2026
- Pin merged FlashAttention SM90 FP8 MMA commit
  
  katec846
  and
  codex
  committed
  3a01ead
Commits on Nov 4, 2026
Commits on Nov 5, 2026
- Merge branch 'main' into fa4-fp8mma-draft-20261007
  DarkLight1337
  authored
  78e295c
- [Test] Remove obsolete input dtype from Zen MoE conversion test
  
  katec846
  and
  codex
  committed
  cf3b814
- Merge branch 'main' into fa4-fp8mma-draft-20261007
  DarkLight1337
  authored
  feba607
- Move Gemma-4 FP8 KV eval into the H100 large-model nightly
  
  MatthewBonanni
  and
  claude
  committed
  419626d
- Set Gemma-4 FP8 KV GSM8K reference to the measured 0.937
  
  MatthewBonanni
  and
  claude
  committed
  7894b61
- Merge branch 'main' into fa4-fp8mma-draft-20261007
  MatthewBonanni
  committed
  f2e9e8f
- Merge branch 'main' into fa4-fp8mma-draft-20261007
  MatthewBonanni
  authored
  7928f5f

Navigation Menu

Repository navigation

[Kernel] Integrate FA4 SM90 FP8 MMA for Gemma-4 full attention - #60372

[Kernel] Integrate FA4 SM90 FP8 MMA for Gemma-4 full attention#60372
vllm-bot merged 26 commits into
vllm-project:mainfrom
katec846:fa4-fp8mma-draft-20261007

Commits on Oct 27, 2026

[Attention] Integrate SM90 d512 FP8 MMA with vLLM routing

[Test] Supply parallel config for SM90 attention fixtures

[Test] Narrow FP8 rejection reason for mypy

[Attention] Consolidate SM90 FP8 wrapper dispatch flags

[Attention] Guard unvalidated SM90 FP8 noncausal full attention

[Review] Trim SM90 FP8 selector tests and comments

[Review] Remove redundant Diff-KV probe cache check

[Review] Remove remaining added attention tests

[Review] Clarify FP8 MMA comments and noncausal error

[Review] Simplify SM90 FP8 guards and error messages

[Attention] Remove the unreachable FA4 -> FA3 FP8 fallback on SM90

[Attention] Reject DCP with FA4 SM90 FP8 KV at impl init

[Test] Cover FA4 SM90 FP8 MMA routing in a CPU selector test

[CI] Add Hopper GSM8K eval for Gemma-4 with FP8 KV (FA4 SM90 FP8 MMA)

Update SM90 FP8 MMA FlashAttention review fixes

Commits on Oct 29, 2026

Pin merged FlashAttention SM90 FP8 MMA commit

Commits on Nov 4, 2026

Merge branch 'main' into fa4-fp8mma-draft-20261007

Skip SM90 FP8 MMA selection test without CUDA FA extensions

Merge remote-tracking branch 'upstream-main-now' into fix-60372-ci

Commits on Nov 5, 2026

Merge branch 'main' into fa4-fp8mma-draft-20261007

[Test] Remove obsolete input dtype from Zen MoE conversion test

Merge branch 'main' into fa4-fp8mma-draft-20261007

Move Gemma-4 FP8 KV eval into the H100 large-model nightly

Set Gemma-4 FP8 KV GSM8K reference to the measured 0.937

Merge branch 'main' into fa4-fp8mma-draft-20261007

Merge branch 'main' into fa4-fp8mma-draft-20261007