[Attention] Integrate FlashAttention 4 SM100 headdim 256 support (#42669)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
Matthew Bonanni
2026-07-27 18:25:50 +00:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
parent 8112b6c997
commit bf2b45b5d6
7 changed files with 66 additions and 34 deletions
+3 -1
View File
@@ -153,7 +153,9 @@ hardware and configuration.
> **‡** Automatic selection tries FlashAttention first. On Blackwell
> (SM100), the fallback order is TRT-LLM Ragged, FlashInfer, then
> TokenSpeed MLA. On other GPUs, only FlashAttention is considered.
> TokenSpeed MLA; for (qk_nope_head_dim=192, qk_rope_head_dim=64,
> v_head_dim=256) TRT-LLM Ragged is tried before FlashAttention. On other
> GPUs, only FlashAttention is considered.
### Decode Backends