OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Revert "Fixing NCCL abort hang issue when a ProcessGroupNCCL manages multiple ncclComms (#150690)"

Browse Source 
This reverts commit 91173ff89a.

Reverted https://github.com/pytorch/pytorch/pull/150690 on behalf of https://github.com/atalman due to failing internal test ([comment](https://github.com/pytorch/pytorch/pull/150690#issuecomment-2787905966))
This commit is contained in:
PyTorch MergeBot 2025-04-09 00:06:30 +00:00
parent 27ded359a5
commit d9f47c75de
1 changed files with 0 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp
View File

@ -1371,9 +1371,6 @@ void ProcessGroupNCCL::abortCommsFromMap(
const std::optional<std::string>& abortReason) {
// The process may control multiple devices, loop through the communicators on
// each device
// NCCL expects Group abort when there are multiple communicators created in a
// device.
groupStart();
for (auto& it : ncclCommsMap) {
auto& devName = it.first;
auto& ncclComm = it.second;
@ -1394,7 +1391,6 @@ void ProcessGroupNCCL::abortCommsFromMap(
VLOG(2) << logPrefix() << "ProcessGroupNCCL destroyed "
<< " communicator on CUDA device: " << devName;
}
groupEnd();
}
// Abort all communicators on this rank