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 15, 18:30 UTC
[Misc][V1] Skip device checking if not available (#15061)
Signed-off-by: Cody Yu <hao.yu.cody@gmail.com>
Cody Yu committed
Mar 18, 2025 at 19:33 UTC
4f065f12f5d33217bfade1d1f88d539e4abf5431
1 file changed
+4
vllm/engine/arg_utils.py
+4
@@ -1469,8 +1469,12 @@ class EngineArgs:
1469
return False
1470
1471
# Need at least Ampere for now (FA support required).
1472
+ # Skip this check if we are running on a non-GPU platform,
1473
+ # or if the device capability is not available
1474
+ # (e.g. in a Ray actor without GPUs).
1475
from vllm.platforms import current_platform
1476
if (current_platform.is_cuda()
1477
+ and current_platform.get_device_capability()
1478
and current_platform.get_device_capability().major < 8):
1479
_raise_or_fallback(feature_name="Compute Capability < 8.0",
1480
recommend_to_remove=False)