You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Summary:
## Context
Implement some small miscellaneous fixes to enable installing ExecuTorch on Windows. Currently, the optimized kernels cannot be built successfully but the installation works in custom ops are not built.
## Installation
```powershell
cd executorch
# Need to disable building custom kernels; there are a bunch of build errors
$env:EXECUTORCH_BUILD_KERNELS_CUSTOM_AOT=0
# For some reason need to explicitly enable tensor extension otherwise I see a linker error
$env:CMAKE_ARGS="-DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON"
python install_executorch.py
# Unset if you want
Remove-Item Env:\EXECUTORCH_BUILD_KERNELS_CUSTOM_AOT
Remove-Item Env:\CMAKE_ARGS
```
## Testing
I am able to follow the [Getting Started tutorial](https://pytorch.org/executorch/stable/getting-started-setup.html#run-your-program) successfully.
## Outstanding Build Issues
Currently, with CMake I can successfully build with the following settings:
```
del -Recurse -Force cmake-out; `
cmake . `
-DCMAKE_INSTALL_PREFIX=cmake-out `
-DPYTHON_EXECUTABLE=C:\\Users\\ssjia\\AppData\\Local\\miniconda3\\python.exe `
-DCMAKE_PREFIX_PATH=C:\\Users\\ssjia\\AppData\\Local\\miniconda3\\Lib\\site-packages `
-DCMAKE_BUILD_TYPE=Release `
-DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON `
-DEXECUTORCH_BUILD_PYBIND=ON `
-DEXECUTORCH_BUILD_XNNPACK=ON `
-DEXECUTORCH_BUILD_KERNELS_QUANTIZED_AOT=ON `
-DEXECUTORCH_BUILD_KERNELS_CUSTOM_AOT=ON `
-DEXECUTORCH_BUILD_KERNELS_CUSTOM=OFF `
-T ClangCL `
-Bcmake-out; `
cmake --build cmake-out -j64 --target install
```
If I switch `EXECUTORCH_BUILD_KERNELS_CUSTOM` to `ON`, then the build fails. The primary offenders appear to be the `gelu`. The implementation includes a header from ATen which is causing the majority of the errors.
0 commit comments