pytorch-riscv64

PyTorch wheels (whl) for riscv64 / RISC-V 64 bit devices

Install

pip

pip install torch --prefer-binary --extra-index-url https://ext.kmtea.eu/simple

• preferred index: https://ext.kmtea.eu/simple
• manually pick whl files: releases

Build

1. Install Docker Desktop on Windows

You can use Dokcer for Linux, macOS or any other platforms you like. And Docker on Windows is likely to have more performance loss. Here I use it just because it configures cross compilation environment automatically.

2. Create Ubuntu 22.04 Image

While the earliest version of Ubuntu that provided images for RISC-V is 20.04, it comes with cmake 3.16 but PyTorch requires 3.18.

Dockerfile

FROM --platform=linux/riscv64 buildpack-deps:jammy

3. Install Python

From the image built above, install Python: build-python.sh

You can also use my prebuilt binaries:

wget https://github.com/KumaTea/pytorch-riscv64/releases/download/2.4/python3.12_3.12.6+ubuntu2204_riscv64.deb
wget https://github.com/KumaTea/pytorch-riscv64/releases/download/2.4/checkinstall_1.6.2+riscv64patched-ubuntu2204_riscv64.deb
dpkg -i python3.12_3.12.6+ubuntu2204_riscv64.deb
dpkg -i checkinstall_1.6.2+riscv64patched-ubuntu2204_riscv64.deb

I must mention that checkinstall takes a horrible 6.5 hours to simply pack a Python installation.

4. Prepare Building Environment

install-deps.sh

python3 -m pip config set global.extra-index-url https://ext.kmtea.eu/simple

python3 -m pip install -U pip setuptools wheel auditwheel
python3 -m pip install -U cffi dataclasses future oldest-supported-numpy pillow pyyaml requests six typing_extensions tqdm

5. Build PyTorch

torch.sh

python3 setup.py build
python3 setup.py develop
python3 setup.py bdist_wheel

6. Audit Wheel (Optional)

LD_LIBRARY_PATH=/path/to/torch/build/lib auditwheel repair dist/*.whl --plat manylinux_2_35_riscv64