A developer has found a way to train artificial intelligence models locally on MacBook Air M4 chips, bypassing Apple's software restrictions. The workaround unlocks up to 15.8 TFLOPS of compute power for neural engine operations that normally remain locked behind official frameworks.
Custom MIL framework bypasses Apple restrictions to enable full backpropagation workloads
The project targets the Apple Neural Engine (ANE) directly by building a custom Model Intermediate Language (MIL). This approach avoids Core ML and Metal APIs, allowing full training workloads like backpropagation to run on Apple Silicon hardware.
Early tests show transformer model steps completing in milliseconds on M4 chips using this method. The workaround also includes an exec() command that restarts the process when training stalls, enabling checkpoint saves instead of losing progress.
The developer reverse-engineered Apple's neural engine architecture to enable these capabilities without official support. The technique demonstrates how local AI workloads can run efficiently on consumer MacBooks despite software limitations.
Source: GIZMOCHINA
