A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 14:45 UTC
[FA4] Update flash-attention to latest upstream FA4 (#38690)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Apr 2, 2026 at 13:02 UTC
cb3935a8fc1968cdd2d7fa2dfd963900419820a0
2 files changed
+3
-3
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 29210221863736a08f71a866459e368ad1ac4a95
42
+ GIT_TAG c0ec424fd8a546d0cbbf4bf050bbcfe837c55afb
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
requirements/cuda.txt
+2
-2
@@ -16,5 +16,5 @@ flashinfer-cubin==0.6.7
16
nvidia-cudnn-frontend>=1.13.0,<1.19.0
17
18
# QuACK and Cutlass DSL for FA4 (cute-DSL implementation)
19
-nvidia-cutlass-dsl>=4.4.0.dev1
20
-quack-kernels>=0.2.7
19
+nvidia-cutlass-dsl>=4.4.2
20
+quack-kernels>=0.3.3