Skip to content

[HACK] Turn off inference mode in RL trainer for now #3440

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

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
+5 −1
Draft

[HACK] Turn off inference mode in RL trainer for now #3440

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: 5 additions & 1 deletion unsloth/models/rl_replacements.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -431,7 +431,11 @@ def _get_per_token_logps_and_entropies(self, model, input_ids, attention_mask, l
unwrapped_model = self.accelerator.unwrap_model(model, keep_fp32_wrapper=False)

with torch.amp.autocast(device_type = 'cuda', dtype = self._autocast_dtype):
with torch.inference_mode():
# aten.t() for tensor subclasses raises RuntimeError:
# Cannot set version_counter for inference tensor
# Turning off inference mode for now
# See https://github.com/pytorch/pytorch/issues/164872 for more details
with torch.inference_mode(False):
if pixel_values is None:
attention_mask = input_ids != self.processing_class.pad_token_id
attention_mask = attention_mask.to(attention_mask.dtype)
Expand Down