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

Update rl_replacements.py

Browse source
This commit is contained in:
Daniel Han 2025-03-18 02:05:49 -07:00
commit 442d3170e8
1 changed files with 1 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
unsloth/models/rl_replacements.py
View file

@ -213,7 +213,7 @@ def grpo_trainer__get_per_token_logps(function_name, function):
if not hasattr(self, '_autocast_dtype'):
self._autocast_dtype = torch.float16 if os.environ.get('ACCELERATE_MIXED_PRECISION', 'fp16') == 'fp16' else torch.bfloat16
if os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1': self._autocast_dtype = torch.float16
print(self._autocast_dtype)
print("GRPO", self._autocast_dtype)
with torch.amp.autocast(device_type = 'cuda', dtype = self._autocast_dtype):
# We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded
logits = model(input_ids=input_ids, attention_mask=attention_mask, logits_to_keep=logits_to_keep + 1).logits