[Fix] Fall back to triton MOE for GPT-OSS on Blackwell with driver >= 595 (#21780)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Baizhou Zhang
2026-03-31 15:52:10 -07:00
committed by GitHub
co-authored by Claude Opus 4.6
parent 9191b02eda
commit f60f2ccc10
4 changed files with 61 additions and 32 deletions
+5 -13
View File
@@ -77,19 +77,11 @@ def _run_smi(query, query_type="gpu"):
def _get_smi_version():
"""Return nvidia-smi driver version and CUDA version, or None on failure."""
try:
out = subprocess.check_output(
[
"nvidia-smi",
"--query-gpu=driver_version",
"--format=csv,noheader,nounits",
],
text=True,
timeout=10,
)
driver = out.strip().splitlines()[0].strip() if out.strip() else "unknown"
except (subprocess.SubprocessError, FileNotFoundError, IndexError):
"""Return nvidia-smi driver version and GPU name, or None on failure."""
from sglang.srt.utils.common import get_nvidia_driver_version_str
driver = get_nvidia_driver_version_str()
if driver is None:
return None
try:
out = subprocess.check_output(