Tag: diffusion

From Python + CUDA to Pure Rust + Metal 4: How zml-flow-rs Stops Us Writing a Kernel Per Model

Every new diffusion model ships as PyTorch with CUDA assumptions baked in, and every port to Apple Silicon has meant hand-writing its kernels again. zml-flow-rs reads only the weights, discovers the semantics from one reference trace, synthesizes its fused kernels from data, and compiles the result into a standalone Rust binary. Two different architectures, zero model-specific code, and a Krea-2 PNG that is bit-identical to the reference.

Porting ERNIE-Image-Turbo to Zig and Metal: An 8B Diffusion Transformer From Scratch on Apple Silicon

I rewrote the hot path of Baidu's 8B-parameter ERNIE-Image-Turbo (DiT + scheduler) from scratch in Zig with hand-written Metal kernels. 13.06s/step at 1024×1024 on M2 Max. The bug archaeology behind getting 36 layers numerically correct, plus the honest accounting of what still has Python in it.