ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

[pre-commit.ci] auto fixes from pre-commit.com hooks

Browse source 
for more information, see https://pre-commit.ci
This commit is contained in:
pre-commit-ci[bot] 2026-01-01 05:13:11 +00:00
commit 5200e67cd2
1 changed files with 8 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

10
unsloth/models/rl_replacements.py
View file

@ -861,13 +861,19 @@ def grpo_trainer_compute_loss(function_name, function):
else torch.tensor(0.0, device = self.model.device)
)
self._metrics[mode]["sampling/importance_sampling_ratio/min"].append(
self.accelerator.gather(min_importance_sampling_ratio).nan_to_num(nan=float('inf')).min().item()
self.accelerator.gather(min_importance_sampling_ratio)
.nan_to_num(nan = float("inf"))
.min()
.item()
)
self._metrics[mode]["sampling/importance_sampling_ratio/mean"].append(
self.accelerator.gather(mean_importance_sampling_ratio).nanmean().item()
)
self._metrics[mode]["sampling/importance_sampling_ratio/max"].append(
self.accelerator.gather(max_importance_sampling_ratio).nan_to_num(nan=float('-inf')).max().item()
self.accelerator.gather(max_importance_sampling_ratio)
.nan_to_num(nan = float("-inf"))
.max()
.item()
)
return loss