diff --git a/vllm/distributed/device_communicators/flashinfer_all_reduce.py b/vllm/distributed/device_communicators/flashinfer_all_reduce.py index 8d3e8170924..4e4635035cc 100644 --- a/vllm/distributed/device_communicators/flashinfer_all_reduce.py +++ b/vllm/distributed/device_communicators/flashinfer_all_reduce.py @@ -67,6 +67,12 @@ def _create_workspace( comm_backend=comm_backend, group=group, ) + if backend == "mnnvl" and not getattr(workspace, "mc_ptr", 0): + workspace.destroy() + logger.warning_once( + "FlashInfer MNNVL multicast is unavailable on the current topology." + ) + return None except Exception as e: if "multicast" in str(e).lower(): logger.warning_once(