Can deepspeed support lora in peft, especially the future multi-adapter version?

[piekey1994]

A recent branch of peft is about to support multiple lora adapters. This implementation feels very suitable for the training in ppo stage. An sft model can be used as an actor, a value function and a reward model at the same time as long as three lora adapters are combined. This is also beneficial for ppo stage to support more than 6b models.

[yaozhewei]

Hi @piekey1994, thanks for letting us know this cool feature form the new PeFT release. We implemented our own LoRA becuase we want to make LoRA compatible with our Hybrid Engine. As you know, the PeFT library is still under development (not in a stable version) yet, we cannot simplt use that for our release.

Please let us know what specific features you are interested in can significantly benefit you and others. We can discuss that internally :)

Also, we are very welcome users to contribute :)

[yaozhewei]

closed due to no response for 2 weeks.