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, 06:00 UTC
[Fix] Fall back to Gloo when NCCL backend is unavailable (#19641)
Signed-off-by: conroy-cheers <conroy@corncheese.org>
Conroy Cheers committed
Jun 17, 2025 at 10:42 UTC
0860087affa8eb0ba0490e02bb965d3aec421952
1 file changed
+7
vllm/distributed/parallel_state.py
+7
@@ -938,6 +938,13 @@ def init_distributed_environment(
938
assert distributed_init_method is not None, (
939
"distributed_init_method must be provided when initializing "
940
"distributed environment")
941
+ if not torch.distributed.is_backend_available(backend):
942
+ logger.warning(
943
+ "Distributed backend %s is not available; "
944
+ "falling back to gloo.", backend)
945
+ assert torch.distributed.is_gloo_available(), (
946
+ "Fallback Gloo backend is not available.")
947
+ backend = "gloo"
948
# this backend is used for WORLD
949
torch.distributed.init_process_group(
950
backend=backend,