Executorch: Enabling and Accelerating PyTorch Models on Edge Devices
summary
ExecuTorch is a PyTorch 2–native stack that captures models with PyTorch 2 export IR, applies quantization, delegation and memory-planning passes, and AOT-compiles into a small C++ instruction bundle executed by a portable runtime. The runtime targets <50KB runtime base size, links only needed kernels or hardware delegates, and provides profiling/debug handles that map runtime nodes back to exported graph lines. MVP is open-sourced (Oct 2023); delegate development is currently via partner program with alpha and beta releases planned in 2024.
AI-written summary · about
chapters (10)
- 0:00ExecuTorch for edge AI
- 1:52AI on edge devices
- 5:50Deploying PyTorch on devices
- 10:36ExecuTorch as PyTorch solution
- 13:29Transforming PyTorch programs
- 17:43ExecuTorch runtime design
- 20:11ExecuTorch timeline and community
- 21:35ExecuTorch resource needs
- 23:02Performance vs battery on edge
- 28:50Q&A
related talks
22mIntroducing ExecuTorch from PyTorch Edge: On-Device AI...
23mBringing ExecuTorch To the Next Frontiers of Edge AI
20mPyTorch Edge: Developer Journey for Deploying AI Models Onto Edge Devices
10mExecuTorch on Microcontrollers: Deploying PyTorch To...
25mExecuTorch 1.0: General Availability Status for Mobile and Embedded...
25mLeveraging PyTorch for Generative AI in Distributed Edge Clouds
14mFlexible Deployment of PyTorch Models on MCU-Class...
16mLightning Talk: Efficient Inference at the Edge: Performance You Need at the Lowest...