Is there plan to support PEFT?

[matrixbot123]

Full fine-tuning of a foundational model on local, specific data defeats the purpose. The primary benefit of a pre-trained model is its extensive general knowledge; PEFT methods, by saving the original weights, are essential for efficient adaptation without risking catastrophic forgetting or losing that foundation.

[abdulfatir]

@matrixbot123 Indeed support for PEFT would be great but it requires significant experimentation to find a setup which works well in most cases. Although this is not on our immediate roadmap, we welcome community contributions in this regard.