PrimeIntellect‑ai has published a GitHub repository called dion that contains efficient implementations of the Dion and Muon optimizers for distributed machine‑learning training. The code works with modern PyTorch (v2.7 or newer) and supports DTensor‑based parallelism such as DDP, FSDP2 and tensor parallelism. Users can install the package directly from the repo with pip. The README includes sample scripts for training a GPT‑small model on eight GPUs.
Why it matters
It gives researchers a ready‑to‑use, communication‑efficient optimizer that can speed up large‑scale model training.
