This PR addresses issue #10 by adding support for a FSDP-compatible HF QLoRA baseline to the our benchmarks.
Feature
This will allow users to specify a no_peft_model field in the plugin config bnb.yaml. Specifying this field will bypass the plugin.augmentation function and allow SFTTrainer to manage the PEFT preparation of the model instead.
NOTE:
While the open-source approach to FSDP-compatible QLoRA removes the extraneous dtype casting in prepare_model_for_kbit_training, it only does so when the model is sharded. When on single device, it continues to use prepare_model_for_kbit_training and users will continue to experience a slowdown due to the extraneous casting.
This PR addresses issue #10 by adding support for a FSDP-compatible HF QLoRA baseline to the our benchmarks.
Feature
This will allow users to specify a
no_peft_model
field in the plugin configbnb.yaml
. Specifying this field will bypass theplugin.augmentation
function and allow SFTTrainer to manage the PEFT preparation of the model instead.NOTE:
prepare_model_for_kbit_training
, it only does so when the model is sharded. When on single device, it continues to useprepare_model_for_kbit_training
and users will continue to experience a slowdown due to the extraneous casting.