- [ ] Rewrite the async transfer engine to enable stall-free prefetching - [ ] Add a modular system for automatic model conversion - [ ] Support automatic conversion on more models - [x] Implement more efficient CPU optimizers - [ ] Compatibility fix on transformers v5 - [ ] Add LoRA fine-tuning tutorial