Question

#1
by WesPro - opened

Hi, I was just wondering what the point is behind slerping the same model with itself? Is getting noticeably better/different or is the only difference the dtype?

Hi, I was just wondering what the point is behind slerping the same model with itself? Is getting noticeably better/different or is the only difference the dtype?

Probably to change the dtype of the model to F32. Which for some People yields "better" outputs compared to BF/FP16. VRAM Usage is higher when using F32 though, which is why people prefer BF/FP16.

CUDA optimised for 32 floating numbers
https://m.youtube.com/watch?v=h9Z4oGN89MU&t=9m23s

Sign up or log in to comment