pytorch / torchtune

PyTorch native finetuning library
https://pytorch.org/torchtune/main/
BSD 3-Clause "New" or "Revised" License
4.35k stars 440 forks source link

PPO slow start #2044

Open felipemello1 opened 1 day ago

felipemello1 commented 1 day ago

1) I tried using this https://github.com/pytorch/torchtune/blob/main/recipes/configs/mistral/7B_full_ppo_low_memory.yaml and gave up after 10min no first step, even after turning compile=False

2) we shouldn't be compiling the entire model. Instead, we should use the utility that compiles per layer, like here: https://github.com/pytorch/torchtune/blob/e9fd56a812cf0ba151fa164a45eb04056d099726/recipes/lora_dpo_single_device.py#L296

3) DPO distributed doesn't compile the model. Should it?

cc: @SalmanMohammadi