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 17, 04:15 UTC
[Bugfix] [Hotfix] fix nccl library name (#3661)
youkaichao committed
Mar 27, 2024 at 10:23 UTC
d18f4e73f39a72bc2abe74bc6b02c61805e1c3d1
1 file changed
+2
-2
vllm/model_executor/parallel_utils/pynccl.py
+2
-2
@@ -39,9 +39,9 @@ if so_file:
39
f"Loading nccl from environment variable VLLM_NCCL_SO_PATH={so_file}")
40
else:
41
if torch.version.cuda is not None:
42
- so_file = "libnccl.so"
42
+ so_file = "libnccl.so.2"
43
elif torch.version.hip is not None:
44
- so_file = "librccl.so"
44
+ so_file = "librccl.so.2"
45
else:
46
raise ValueError("NCCL only supports CUDA and ROCm backends.")
47
logger.debug(f"Loading nccl from library {so_file}")