ExecuTorch 1.0 RC1

[GregoryComer]

The first release candidate build for ExecuTorch 1.0 is now available. The 1.0 release is targeted for October 17th, though this date is subject to change. The release is focused on stability and maturity of the core framework. Draft release notes will be available closer to the release date.

Instructions for installing the release candidate build for Python, Android, and iOS are included below. Any bug reports are greatly appreciated. Issues can be filed against the ExecuTorch GitHub repo. If you believe that an issue is critical and not being addressed, feel free to tag myself (@GregoryComer) or one of the other release managers (@abhinaykukkadapu or @psiddh) to triage or escalate it.

Installing the Release Candidate

Platform-specific instructions are included below.

Python Packages

Python packages are available from the pytorch test channel and can be installed with the following pip command:

pip install executorch==1.0.0 torch==2.9.0 torchvision==0.24.0 torchao==0.13.0  torchaudio==2.9.0 --extra-index-url https://download.pytorch.org/whl/test/cpu

Android

The release candidate binary for ARMv8 can be downloaded from https://ossci-android.s3.amazonaws.com/executorch/release/1.0.0-rc1/executorch.aar. See Using ExecuTorch on Android for information on how to use the AAR.

iOS

The release candidate frameworks are available through SwiftPM. Specify swiftpm-1.0.0 as the target branch when adding the ExecuTorch dependencies from XCode. See Using ExecuTorch on iOS for more information.

Source

To build from source, following the Building From Source documentation using the release/1.0 git branch.

Known Issues

There are several known issues in RC1, including several critical issues noted below. The current status for all release blocking issues is tracked in https://github.com/orgs/pytorch/projects/142/views/1.

• Python runtime bindings (pybindings) can't be imported when installing from a wheel (Pybindings can't load on Windows when installing wheel #14443, Fix RPATH issues on MacOS #14421).
• Models with quantized linear ops on XNNPACK may fail to load when weight cache is enabled (Use weight cache for quantized tensor scale data #14448).