We are excited to announce the official release of MagiCompiler v1.0.0! 🎉
MagiCompiler is an advanced compiler and runtime augmentation framework built on top of torch.compile, designed specifically for large-scale Transformer-like models. Going beyond local operator optimization, it brings system-level optimization to both multi-modal inference and large-model training.
Release Highlights
- Whole-graph compilation for multi-modal inference across Transformer boundaries.
- FSDP-aware whole-layer compilation for large-model training with transparent parameter sharding.
- Plug-and-play integration with minimal code intrusion.
- Smart asynchronous offloading and heuristic activation recomputation to perfectly balance compute efficiency and memory footprint.
- Better interpretability through human-readable graph and kernel artifacts dumps.
About This Release
v1.0.0 marks our first official open-source release and the starting point of our Compiler as Manager vision: reimagining the compiler not just as a kernel generator, but as a system-level manager orchestrating execution, dataflow, and memory.
We welcome feedback, issues, and contributions from the community.