EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

set default attention backend for deterministic inference (#11801)

Browse Source
This commit is contained in:
Minglei Zhu
2025-10-18 00:01:24 -07:00
committed by GitHub
parent e68a2b5b2f
commit f4488e9dd9
2 changed files with 26 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
python/sglang/srt/utils/common.py
View File

@@ -174,6 +174,15 @@ def is_blackwell():
return torch.cuda.get_device_capability()[0] == 10
@lru_cache(maxsize=1)
def is_sm120_supported(device=None) -> bool:
if not is_cuda_alike():
return False
return (torch.cuda.get_device_capability(device)[0] == 12) and (
torch.version.cuda >= "12.8"
)
@lru_cache(maxsize=1)
def is_sm100_supported(device=None) -> bool:
if not is_cuda_alike():