This is a port of my pure rust image/video trainer along with its inference and different image/vid model trainers into mojo. It is a autograd and tensor lib, it has also has training for many models. ZImage and Klein 9b confirmed working as others are being worked on. It has no Max support for a simple reason, I am Gpu poor and only have a NVIDIA 3090 ti so i could not use it. Inference support is already done for SDXL, Flux, Zimage, Qwen and many more models. I used Claude Code to port it from Rust and speed wise, with optimizations is just a tad, tad bit slower. It has no pytorch/python.
1 Like