Porting/PyTorch
Binaries
Debian
PyTorch is packaged in Debian Bookworm and works fine, except for one bug. Debian uses PyTorch 1.13 and Python 3.11; these are not mutually compatible. To patch Debian's PyTorch to work properly with Python 3.11, apply this patch to /usr/lib/python3/dist-packages/torch/distributed/_shard/sharded_tensor/metadata.py:
- tensor_properties: TensorProperties = field(default=TensorProperties()) + tensor_properties: TensorProperties = field(default_factory=TensorProperties)
The above patch is derived from this Lightning comment , with a bugfix applied. The code affected by the patch may or may not be triggered depending on which application you're running; Real-ESRGAN doesn't need the patch while Stable Diffusion WebUI does.
Example workflow to install PyTorch and Real-ESRGAN on Debian Bookworm ppc64le:
sudo apt install python3-torch python3-torchvision python3-opencv python3-llvmlite python3-grpcio python3-pip python3-skimage python3-numba GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=true pip3 install --user --break-system-packages basicsr facexlib gfpgan # Log in and log out again to update $PATH to include ~/.local/bin git clone https://github.com/xinntao/Real-ESRGAN.git cd Real-ESRGAN pip3 install --user --break-system-packages -r requirements.txt pip3 install --user --break-system-packages . # Real-ESRGAN is now installed.
Stable Diffusion WebUI works fine with Debian's packaged PyTorch and Python, subject to the following caveats:
- You need an older version of Lightning; checking out this WebUI commit (predates v1.1.0) works fine.
- You need to patch
repositories/CodeFormer/facelib/detection/yolov5face/face_detector.pyas follows:
-IS_HIGH_VERSION = tuple(map(int, torch.__version__.split('+')[0].split('.'))) >= (1, 9, 0)
+IS_HIGH_VERSION = tuple(map(int, torch.__version__.split('+')[0].split('a')[0].split('.'))) >= (1, 9, 0)
It looks like a similar patch is already upstreamed to CodeFormer, but simply using that commit via export CODEFORMER_COMMIT_HASH=07c8cc6f6d9b5aee87046177a8d429ec041da54a doesn't work because that commit makes changes to CodeFormer's vendored BasicSR, which aren't compatible with Stable Diffusion WebUI's usage of upstream BasicSR.
Once you've done this, install the python3-sentencepiece package, and then the following script will launch the WebUI:
export TORCH_COMMAND=true export COMMANDLINE_ARGS='--skip-version-check --skip-torch-cuda-test --no-half' export PIP_USER=1 export PIP_BREAK_SYSTEM_PACKAGES=1 python3 launch.py
Conda
Open Cognitive Environment (Open-CE) provides distro-independent ppc64le binaries of PyTorch and related packages. Example workflow to install PyTorch and Real-ESRGAN on Fedora 38 ppc64le:
sudo dnf install python3.10 sudo dnf install conda conda create --name pytorch python=3.10 # Close and re-open terminal conda activate pytorch conda install -c https://ftp.osuosl.org/pub/open-ce/current/ pytorch-cpu torchvision-cpu py-opencv conda install -c numba llvmlite conda install -c conda-forge grpcio conda install -c conda-forge libstdcxx-ng conda install pip GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=true pip install basicsr facexlib gfpgan git clone https://github.com/xinntao/Real-ESRGAN.git cd Real-ESRGAN pip install -r requirements.txt python setup.py develop # Real-ESRGAN is now installed.
Example script to run Stable Diffusion WebUI (assuming you've cloned from Git) on Fedora 38 using the above installed PyTorch:
export TORCH_COMMAND=true export COMMANDLINE_ARGS='--skip-version-check --skip-torch-cuda-test --no-half' python launch.py
Docker
MarcusC maintains a Dockerfile that sets up PyTorch 2.0.1 and Stable Diffusion WebUI 1.5.2 on ppc64le.
Finished
- PyTorch Build From Source Error: ppc64le + gcc 7.3.1 + cuda 11.8 + python 3.9
- fallback to cpu_kernel for VSX
- Add missing conversion functions between half and float for ppc64le
In progress
- PR #88607 breaks build for POWER9 CPU
- PPC64le: vsx_helpers.h errors
- Wrong vector shift results on PowerPC
- Fix CPU bitwise shifts for out-of-limit values in VSX-vec
