Skip to content

Use torch.zeros_like instead of empty_like to prevent accuracy drop in CAR #1387

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
valarLip merged 1 commit into from
Nov 12, 2025
Merged

Use torch.zeros_like instead of empty_like to prevent accuracy drop in CAR #1387

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions aiter/dist/device_communicators/custom_all_reduce.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -300,7 +300,7 @@ def custom_all_reduce(
else:
# if warm up, mimic the allocation pattern
# since custom allreduce is out-of-place
return torch.empty_like(input)
return torch.zeros_like(input)
else:
# note: outside of cuda graph context,
# custom allreduce incurs a cost of cudaMemcpy, which should
Expand Down Expand Up @@ -332,7 +332,7 @@ def custom_all_gather(self, inp: torch.Tensor) -> Optional[torch.Tensor]:
return self.all_gather_reg(inp)
else:
print("allgather capture hipgraph error")
return torch.empty_like(inp)
return torch.zeros_like(inp)
else:
return self.all_gather_unreg(inp)

Expand Down Expand Up @@ -371,7 +371,7 @@ def custom_fused_ar_rms(
if torch.cuda.is_current_stream_capturing():
return self.fused_ar_rms(input, w=weight, eps=eps, registered=True)
else:
return torch.empty_like(input), torch.empty_like(input)
return torch.zeros_like(input), torch.zeros_like(input)
else:
return self.fused_ar_rms(input, w=weight, eps=eps, registered=False)

Expand Down