This milestone tracks the progress for MagiCompiler v1.2.0. For this release, we are focusing on two main areas:

Acceleration: Breaking framework limitations with native FSDP full-graph capture, maximizing VRAM utilization via refactored AutoRecompute, and exploring extreme performance with deterministic inference and computation-communication fusions (MegaKernel).

Ecosystem: Upgrading our core compatibility to PyTorch 2.12, providing richer out-of-the-box training/inference examples, and launching our official technical blog series to share our design insights with the community.

Feel free to pick up any unassigned issues if you'd like to contribute!