Skip to content

Retry partial.to to fix accelerate invalid argument error for first moe layer for >4 GPU setups #2163

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
merged 2 commits into from
Nov 2, 2025
Merged

Retry partial.to to fix accelerate invalid argument error for first moe layer for >4 GPU setups #2163

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9d6ee27
retry partial.to to fix accelerate invalid argument error for first m…
avtc Nov 1, 2025
c8f2a64
remove second loop, reduce delay time
Qubitium Nov 2, 2025
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
16 changes: 13 additions & 3 deletions gptqmodel/quantization/gptq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -500,9 +500,19 @@ def _materialize_global_hessian(self, target_device: Optional[torch.device] = No

for partial_device, partial in self._device_hessian_partials.items():
if partial.device != result_accum.device or partial.dtype != torch.float32:
tmp = partial.to(device=result_accum.device, dtype=torch.float32)
result_accum.add_(tmp)
del tmp
# TODO FIXME multi-3090 using P2P is revaling an issue where result_accum and/or partial is not ready for consolidation on the main thread
# when parials are calculated on the individual
try:
tmp = partial.to(device=result_accum.device, dtype=torch.float32)
result_accum.add_(tmp)
del tmp
except:
log.warn(f"Quantization: Module `{self.name}` -> Retry 1/2 partial.to in 0.5s")
time.sleep(0.25)
tmp = partial.to(device=result_accum.device, dtype=torch.float32)
result_accum.add_(tmp)
del tmp

else:
result_accum.add_(partial)

Expand Down