https://github.com/tttapa/cross-pytorch
Instructions for cross-compiling PyTorch with CUDA support on Linux.
https://github.com/tttapa/cross-pytorch
cmake cross-compilation pytorch
Last synced: 8 months ago
JSON representation
Instructions for cross-compiling PyTorch with CUDA support on Linux.
- Host: GitHub
- URL: https://github.com/tttapa/cross-pytorch
- Owner: tttapa
- Created: 2024-05-12T14:27:20.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2024-05-18T16:56:03.000Z (over 1 year ago)
- Last Synced: 2025-01-01T07:09:17.044Z (9 months ago)
- Topics: cmake, cross-compilation, pytorch
- Language: Shell
- Homepage:
- Size: 6.84 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# cross-pytorch
Scripts for cross-compiling PyTorch with CUDA support on Linux.
## Why?
Recent PyTorch binaries no longer support glibc 2.28 and earlier, and can no
longer be compiled using older compilers.
## How?
This repository has scripts that
1. Download a suitable cross-compiler to make sure that the resulting binaries are compatible with Ubuntu 18.04 Bionic, Debian 10 Buster, Rocky 8 and later.
2. Download a suitable version of CUDA.
3. Cross-compile PyTorch and (some of) its dependencies with the rigth compiler and CUDA version.
## Instructions
- In `scripts/download-cuda.sh`, select the appropriate `cuda_version` and `cudnn_version` (the former should be compatible with your driver version, use `nvidia-smi` to verify the version you have installed).
- In `scripts/install-torch.sh`, select the appropriate `CMAKE_CUDA_ARCHITECTURES` and `TORCH_CUDA_ARCH_LIST` options (see https://developer.nvidia.com/cuda-gpus).
- In `scripts/install-torch.sh`, select the version of PyTorch you need (`main` by default).
- You'll need a native C compiler, Python 3 and some standard tools like `wget` and `git`. Ccache is recommended and enabled by default, but can be commented out in `scripts/install-torch.sh`
- Finally, run the following commands:
```sh
./install-toolchain.sh # ~120 MiB download
./download-cuda.sh # ~2 GiB download
./install-torch.sh # May take a couple of hours, dependening on your machine's performance
```
Everything is downloaded and built locally in the current folder. No root privileges are required.
The resulting PyTorch install tree can be found in `toolchains/x86_64-bionic-linux-gnu/pytorch`.
Add it to your `CMAKE_FIND_ROOT_PATH`, and use the `scripts/toolchains/x86_64-bionic-linux-gnu.cmake` toolchain file so CMake can locate CUDA correctly when building your own project. A CMake project and a `CMakePresets.json` file are provided as an example:
```sh
cmake --preset develop
cmake --build build -j --config Release
export LD_LIBRARY_PATH="$PWD/toolchains/x-tools/x86_64-bionic-linux-gnu/x86_64-bionic-linux-gnu/lib64"
./build/Release/example-torch
```
## Further reading
- https://pytorch.org/cppdocs/
- https://pytorch.org/cppdocs/installing.html
- https://pytorch.org/tutorials/advanced/cpp_frontend.html