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 4, 03:45 UTC
[XPU] Implement out-of-place all-reduce functionality (#41808)
Signed-off-by: Chaojun,Zhang <chaojun.zhang@intel.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Chaojun Zhang committed
May 7, 2026 at 20:58 UTC
2a84da3b17a9d0b6b10f7c0347fe9529307596ec
1 file changed
+1
-1
vllm/distributed/device_communicators/xpu_communicator.py
+1
-1
@@ -42,7 +42,7 @@ class XpuCommunicator(DeviceCommunicatorBase):
42
logger.info("Using AgRs manager on XPU device.")
43
44
def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
45
- output = input_.clone() if torch.compiler.is_compiling() else input_
45
+ output = input_.clone()
46
dist.all_reduce(output, group=self.device_group)
47
return output
48