ANEForge
A Python library that compiles neural-network graphs directly into programs for Apple's Neural Engine, bypassing Core ML so that training, not just inference, can run on the chip.
ANEForge lets a developer train and run neural networks directly on the Apple Neural Engine from Python, without going through Core ML. Apple's Neural Engine is normally reachable only through Core ML for inference, and Core ML decides on its own whether a model actually lands on the engine or silently falls back to the CPU or GPU; ANEForge instead compiles a tensor graph into a single ANE program dispatched through the same private stack that Core ML, MPSGraph and Espresso use internally.
Because it controls the compiled program directly, ANEForge can run the forward pass, backward pass and Adam optimizer update on the engine, train a small model from scratch, and reach hardware layers that Core ML's compiler never emits, such as a fused-attention operation. It also runs pretrained models, language-model decoding with key-value caching, and Whisper speech-to-text end to end on the engine, and can cross-compile a graph for any of twenty-eight ANE hardware targets from one machine. The project describes itself as a research project on Apple Silicon.