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 2025-02-18 00:17:07 -08:00
commit 606fc099f3
1 changed files with 0 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

2
unsloth/models/rl_replacements.py
View file

@ -177,7 +177,6 @@ def grpo_trainer__get_per_token_logps(function_name, function):
if function_name != "_get_per_token_logps": return function
def _get_per_token_logps(self, model, input_ids, attention_mask, logits_to_keep):
print(self.args.unsloth_num_chunks)
if self.args.unsloth_num_chunks != 1: return None
if not hasattr(self, '_autocast_dtype'):
self._autocast_dtype = torch.float16 if os.environ.get('ACCELERATE_MIXED_PRECISION', 'fp16') == 'fp16' else torch.bfloat16
@ -240,7 +239,6 @@ def grpo_trainer_compute_loss(function_name, function):
ref_per_token_logps, per_token_logps, input_ids, completion_mask, self.beta, advantages, bsz,
)
else:
print(int(self.args.unsloth_num_chunks), end = ",")
loss, completion_length, mean_kl = grpo_accumulated_loss(
self, _input_ids, logits_to_keep, completion_mask, advantages,
n_chunks = self.args.unsloth_num_chunks,