Downstream Issues Due to Inference-time Optimization in Model Forward

#2
by theyorubayesian - opened

There's an optimization in the model forward that limits the logits returned to the last timestep (https://huggingface.co/BeardedMonster/SabiYarn-125M/blob/main/pretrained_model.py#L196). This causes issues during likelihood-based evaluations where the logits need to be processed.

I will work sort it out as soon as i can. Thank you.

BeardedMonster changed discussion status to closed

It has been fixed. Logits for all time steps should be returned now.

BeardedMonster changed discussion status to open
BeardedMonster changed discussion status to closed

Sign up or log in to comment