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 14, 05:30 UTC
fix pynccl reduce_scatter (#23648)
Co-authored-by: hongchao <hongchao@msh.team>
yzds committed
Aug 27, 2025 at 09:21 UTC
c7c80af084e4d87c4e73148cb71ee990970281ff
1 file changed
+3
-3
vllm/distributed/device_communicators/cuda_communicator.py
+3
-3
@@ -152,7 +152,7 @@ class CudaCommunicator(DeviceCommunicatorBase):
152
dtype=input_tensor.dtype,
153
device=input_tensor.device)
154
155
- pynccl_comm.reduce_scatter(output, input_)
155
+ pynccl_comm.reduce_scatter(output, input_tensor)
156
157
# Reshape before returning
158
return output.movedim(0, dim).contiguous()
@@ -186,9 +186,9 @@ class CudaCommunicator(DeviceCommunicatorBase):
186
device=input_tensor.device)
187
188
if sizes is not None:
189
- pynccl_comm.reduce_scatterv(output, input_, sizes=sizes)
189
+ pynccl_comm.reduce_scatterv(output, input_tensor, sizes=sizes)
190
else:
191
- pynccl_comm.reduce_scatter(output, input_)
191
+ pynccl_comm.reduce_scatter(output, input_tensor)
192
193
# Reshape before returning
194
return output.movedim(0, dim).contiguous()