Speeding up PyTorch inference on Apple devices with AI-generated Metal kernels