[Perf] Improve/Fix-regression for FA3 in High QPS regimes (vllm-project#19463)

LucasWilkinson · web-flow · commit a045b7e89a24 · 2025-06-24T13:09:01.000-04:00
Signed-off-by: Lucas Wilkinson &lt;lwilkinson@neuralmagic.com&gt;
diff --git a/cmake/external_projects/vllm_flash_attn.cmake b/cmake/external_projects/vllm_flash_attn.cmake
@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 763ad155a1c826f71ff318f41edb1e4e5e376ddb
+          GIT_TAG 2c6bcfc0feb3d9d4a57b243fc159a68aa9933f5b
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
diff --git a/test-qwen b/test-qwen
@@ -0,0 +1 @@
+Subproject commit 34c31c0af8fc975140b8c85548fefa1eb7f523e4

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	`+Subproject commit 34c31c0af8fc975140b8c85548fefa1eb7f523e4`