While Mesa is predominantly used for GPU drivers, open-source developer Tomeu Vizoso continues advancing the capabilities of NPUs within the Mesa driver stack.
Tomeu today merged support for YOLOX within Mesa's Torx code. Torx is the hardware backend in Mesa for serving as a compilation and execution path for machine learning models with PyTorch ExecuTorch.
With the now merged code, users can run YOLOX with PyTorch using NPU drivers in Mesa. This has been tested so far with the Arm EthosU NPU using the NXP i.MX93 SoC. The Etnaviv driver for Vivante NPUs should have all operations supported but still lacking infrastructure work for use with Torx.
Those wishing to learn more about this YOLOX support for Mesa NPU drivers via Torx can learn more via this merge request now part of Mesa 26.3.