30X SPEEDUP IN MODEL INIT FOR HF TRANSFORMERS (GITHUB PR)

Summary from the TLDR AI newsletter

You can speed up model initialization dramatically without much loss in tokens per second by moving some lazy loading to the model on the first pass.