Update rl.py

This commit is contained in:
Daniel Han 2025-02-05 06:58:04 -08:00
commit aebeeb4901

View file

@ -39,15 +39,13 @@ def PatchRL(FastLanguageModel):
@contextmanager
def unsloth_unwrap_model_for_generation(model, accelerator):
# Must use for_inference to allow inference in Unsloth
with torch.inference_mode():
with unwrap_model_for_generation(model, accelerator) as unwrapped_model:
FastLanguageModel.for_inference(unwrapped_model)
try:
yield unwrapped_model
finally:
# Finally return back training
FastLanguageModel.for_training(model)
pass
with unwrap_model_for_generation(model, accelerator) as unwrapped_model:
FastLanguageModel.for_inference(unwrapped_model)
try:
yield unwrapped_model.eval()
finally:
# Finally return back training
FastLanguageModel.for_training(model)
pass
pass
pass