The latest release of pytorch/executorch is v1.5.1. ExecuTorch is an open-source stack for running PyTorch models locally on phones, wearables, laptops, browsers, embedded systems, and microcontrollers. Models are exported with torch.export, optimized for the target hardware, and then run through APIs in several languages. However, the provided information lists no specific changes in v1.5.1, so the contents of this release cannot be confirmed.
The purpose of this stack is to run PyTorch models across many kinds of devices. The provided description mentions backends for CPU, NPU, DSP, and MCU. According to the README summary, it is used for on-device features on Instagram, WhatsApp, Facebook, Messenger, and Meta Quest and Ray-Ban Meta devices. The repository's primary language is Python, with 5105 stars. This is information about use cases and supported targets, not figures that show per-device performance.
No performance benchmarks or comparisons with other runtimes are included in the provided FACTS. It is therefore not possible to judge here what speed or resource usage it delivers on real devices. The license is listed as NOASSERTION, so the terms of use cannot be settled from this information alone. The fact that there are deployment use cases and the fact that the changes or performance of v1.5.1 have been confirmed should be kept distinct.