MagiCompiler

Run `npx skills add SandAI-org/MagiCompiler --all -g -y` in your terminal to install all skills in this repository globally.

It is a compiler framework built on torch.compile that speeds up large Transformer training and inference through full-graph capture, smart offloading, and automatic activation recomputation.

Decorate your core Transformer block once with @magi_compile to get FSDP-aware layer-wise compilation and up to 20% extra speedup with no model refactoring.

Ask your agent to review a PR or branch, and it fetches only the ahead-of-base diff, applies a strict P0-P4 severity framework, and saves a markdown report in the repo.

Almost none. Two decorators deliver the acceleration out of the box, and it integrates into existing multi-modality frameworks with minimal intrusion.