Open felipemello1 opened 3 weeks ago
cc: @ebsmothers
I believe it happens because we only quantize base if it has LoRA. By using LoRA on less modules, we quantize less: https://github.com/pytorch/torchtune/blob/f9f75bb563ecae371492a9d49da4a9f514c081b3/torchtune/models/llama3_1/_component_builders.py#L337
Is that the expected behavior?
Running:
Gives mes
but running
gives me
this doesnt make sense, since the default of the config is to have more modules.
environment: