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 13, 14:30 UTC
[flashinfer][fix] do not check nvcc availability when using pre-downloaded cubins (#27990)
Signed-off-by: Xiaozhu <mxz297@gmail.com> Co-authored-by: Lu Fang <30275821+houseroad@users.noreply.github.com>
Xiaozhu Meng committed
Nov 7, 2025 at 22:25 UTC
4a36681f854889f2d5279e189cc4dd0c8e6c85d8
1 file changed
+6
-2
vllm/utils/flashinfer.py
+6
-2
@@ -43,9 +43,13 @@ def has_flashinfer() -> bool:
43
if importlib.util.find_spec("flashinfer") is None:
44
logger.debug_once("FlashInfer unavailable since package was not found")
45
return False
46
+ # When not using flashinfer cubin,
47
# Also check if nvcc is available since it's required to JIT compile flashinfer
47
- if shutil.which("nvcc") is None:
48
- logger.debug_once("FlashInfer unavailable since nvcc was not found")
48
+ if not envs.VLLM_HAS_FLASHINFER_CUBIN and shutil.which("nvcc") is None:
49
+ logger.debug_once(
50
+ "FlashInfer unavailable since nvcc was not found "
51
+ "and not using pre-downloaded cubins"
52
+ )
53
return False
54
return True
55