ExecuTorch on Microcontrollers: Deploying PyTorch To...
summary
ExecuTorch brings PyTorch models to microcontrollers and tiny NPUs with an ahead‑of‑time pipeline: torch.xport graph tracing, node‑level quantization from torch.ao, EXIR graph optimizations, and a selectable kernel+dtype build so binaries only include what a model uses. Runtime is OS‑agnostic and kept under 50 kilobytes, backends include Cadence DSPs, ARM Ethos‑U, Cortex‑M and NXP Neutron with over 12 backends supported. Constraints: <10 MB RAM, milliwatt power; int8 gives 4x model size reduction, defaults allow 2,000 kernels but pruning saves ~20 KB RAM; deployed in Ray‑Ban Meta glasses and Meta Vanguard.
AI-written summary · about
chapters (7)
- 0:00Introduction to ExecuTorch on microcontrollers
- 1:07Deploy PyTorch on MCUs
- 2:52Hardware for TinyML
- 3:33ExecuTorch for constrained devices
- 6:25MCU deployment tips
- 7:51Real-world models
- 9:13Partners and integrations
related talks
22mIntroducing ExecuTorch from PyTorch Edge: On-Device AI...
13mAI inference for the smallest processors
33mExecutorch: Enabling and Accelerating PyTorch Models on Edge Devices
56mHow to: Bringing up ExecuTorch on a RISC-V Microcontroller | Corrected Audio
25mExecuTorch 1.0: General Availability Status for Mobile and Embedded...
56mHow to: Bringing up ExecuTorch on a RISC-V Microcontroller
20mPyTorch Edge: Developer Journey for Deploying AI Models Onto Edge Devices
11mLightning Talk: Accelerating Experimentation...