LoRA fine-tuning running on Apple's Neural Engine

The full story

Apple's Neural Engine, the machine-learning chip in every Apple Silicon Mac, has no public training API. A self-described non-ML engineer used Claude Code with Opus 4.6 to get LoRA fine-tuning gradients running on it anyway: 192 gradient dispatches, no GPU fallback, loss converging, at about 2.8 watts.

Iterating on real hardware turned up three surprises. The chip's matrix-multiply op compiles but never runs, so everything had to be rewritten as 1×1 convolutions. Spatial dimensions must be multiples of 16. And the compiler leaks handles and silently fails after about 119 compiles. It builds on earlier community reverse engineering of the Neural Engine.