Update vllm to point to flash-attention commit that builds FA3 with torch stable API. (Retry) (#49599)
Signed-off-by: Chris Leonard <chleonar@redhat.com> Signed-off-by: Shengqi Chen <harry-chen@outlook.com> Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
Chris Leonard committed
Aug 6, 2026 at 11:37 UTC
41e7746b82b43dd3454cd842d1bcfc30665eddb2
2 files changed
+2
-5
.buildkite/check-torch-abi.py
+1
-4
@@ -12,10 +12,7 @@ from torch_abi_audit.report import ExtensionReport, PackageReport
12
13
# Temporary allowlist of extensions not yet on the stable ABI.
14
# Shrink and remove over time.
15
-ALLOWED_UNSTABLE_LIBRARIES: tuple[str, ...] = (
16
- "vllm_flash_attn/_vllm_fa3_C.abi3.so",
17
- "third_party/deep_gemm/_C*.so",
18
-)
15
+ALLOWED_UNSTABLE_LIBRARIES: tuple[str, ...] = ("third_party/deep_gemm/_C*.so",)
16
17
18
def _relative_path(lib: ExtensionReport, package_root: Path) -> str:
cmake/external_projects/vllm_flash_attn.cmake
+1
-1
@@ -39,7 +39,7 @@ else()
39
FetchContent_Declare(
40
vllm-flash-attn
41
GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
42
- GIT_TAG 28e862d21806bc3580207aa0ad4e2759151e9827
42
+ GIT_TAG f3e1a4f74c99145c0717709860bf765de1703779
43
GIT_PROGRESS TRUE
44
# Don't share the vllm-flash-attn build between build types
45
BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn