ExecuTorch adds Arm Cortex-M CPU support, running PyTorch models directly on microcontrollers

PyTorch's ExecuTorch now runs models directly on Arm Cortex-M CPUs through a backend that uses Arm's CMSIS-NN kernels, removing the need to convert models to other formats.

The backend covers Scalar, DSP and MVE/Helium tiers, and on the Alif Ensemble E8 four MLPerf Tiny models each run in 21.5 ms or less using under 90 KiB of RAM on one Cortex-M55 core. The ExecuTorch core runtime is under 20 KiB of flash.