I added some support for loading peft adapters without merging with the base model, instead the merging happens in the harness. the imports are dynamic, so people won't need to install peft==0.3 unless they want to use this. It adds the --peft_model argument, which points to the peft adapter, while the --model argument will be used as the base model.
I also cleaned up the if statement in main.py, and fixed a small bug with loading 8-bit models due to a typo: it checked if args.load_in_9-bit instead of if args.load_in_8bit.
I tested to see if it works with a LoRA adapter for starcoder. Runs just fine.
I added some support for loading peft adapters without merging with the base model, instead the merging happens in the harness. the imports are dynamic, so people won't need to install
peft==0.3
unless they want to use this. It adds the--peft_model
argument, which points to the peft adapter, while the--model
argument will be used as the base model.I also cleaned up the if statement in
main.py
, and fixed a small bug with loading 8-bit models due to a typo: it checkedif args.load_in_9-bit
instead ofif args.load_in_8bit
.I tested to see if it works with a LoRA adapter for starcoder. Runs just fine.