On-Device Diffusion Kernels
This project turns Modulated Diffusion (MoDiff) from an operation-count story into a hardware result. The manuscript reports fused low-bit kernels and a cache-update fusion strategy that cuts extra memory traffic during iterative denoising.
On the evaluated setup, the implementation reaches up to 1.8× runtime speedup over FP32 and up to 42.2% lower memory I/O. The public tree lives at github.com/xiaruize0911/MoDiff, forked from the official ICML 2025 MoDiff codebase and used as the systems implementation path for the paper.
