Apple Neural Engine experiments on Linux
An engineering log and Python test scripts that verify Apple Neural Engine primitives such as gemm, softmax, attention and transformer blocks on Linux, comparing ANE output against NumPy on real M1 and M1 Max Asahi Linux hardware.
This repository tests the Apple Neural Engine on Linux, targeting an M1 MacBook Pro running Omarchy with the Asahi kernel. Python scripts drive the ANE directly and compare its output against a NumPy reference implementation for primitives such as add-multiply, small matrix multiplication, 1x1 convolution and linear layers of specific shapes, with results and status captured as JSON receipts under a receipts directory that every reported number links back to.
Current status notes include a staged decode benchmark for a Qwen model that outperforms equivalent macOS timings, a bit-exact contract check against a Parakeet speech model on the installed ANE kernel module, and root-caused history of past converter defects, such as an over-fetching task-record bug and a channel-polarity mismatch between an exported task family and libane's expected layout, each tied to a specific fixing commit. A third host, an M2 MacBook, is not yet qualified for live inference.