executorch

Why we included this project

Modeling teams that need to ship inference on a phone or a microcontroller usually discover a painful gap: the model trains fine on a server but runs far too slowly or bloated on the actual device. ExecuTorch is PyTorch's answer to that problem. It takes the model graph you already export from PyTorch, compiles it ahead of time with quantization and hardware partitioning, and emits a compact binary that a lightweight C++ runtime loads directly on targets from Android phones to embedded boards. You keep working with the usual PyTorch export APIs, can swap between CPU and vendor accelerators such as Qualcomm or Apple backends with a single line change, and never deal with intermediate formats or manual C++ rewrites. Because inference runs locally, it suits privacy-sensitive products where data should not leave the device, and the same pipeline covers LLMs, vision, speech, and multimodal models, so a team can standardize on one on-device stack instead of maintaining separate toolchains for each platform.