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 2025-02-14 04:45:48 -08:00
commit 9b773201f2
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

@ -195,7 +195,7 @@ RL_FUNCTIONS["grpo_trainer"].append(grpo_trainer__get_per_token_logps)
# Custom compiled GRPO loss - creates 3 Triton kernels
@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
# @torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
def _grpo_compute_loss(old_logits, new_logits, input_ids, mask, beta):
old_logits = old_logits.to(torch.float32)
new_logits = new_logits.to(torch.float32)