Closed dmitrysarov closed 3 years ago
Hey @dmitrysarov,
It looks like your installation is not using CUDA. Share some details on how you build the image and install pytorch_scatter. This will give us more information about what is going wrong.
Otherwise, pytorch_scatter may not be compatible with NVIDIA PyTorch 1.8.0... What do you thing about @rusty1s?
I can say that everything works fine with nvcr.io/nvidia/pytorch:20.03-py3
.
@mgomesborges thanks for answering
It actually does compatible because I can do build from a source inside an already running container.
In my Dockerfile, I tried simple pip install torch_scatter
as well as providing ENV variables like FORCE_CUDA=1 and CUDA version. Maybe I miss something obvious.
@mgomesborges can you share your Dockerfile with nvcr.io/nvidia/pytorch:20.03-py3
? Probably it will answer my question
This issue had no activity for 6 months. It will be closed in 2 weeks unless there is some new activity.
@mgomesborges What is the most recent version of nvcr.io/nvidia images that you get to run? How do you install pytorch scatter inside the container?
I tried 20.08 as it's the only cuda toolkit and torch version that should be supported. But I do not manage to install with conda, pip or from scratch. Always some libraries are not correctly linked. I tried to follow all instructions I found in the README and other issues.
Will now try 20.03 as you pointed out. Would really appreciate some pointers :)
Hey @jan-engelmann,
I keep using my dev env but it is a bit outdated:
I believe you need to update the CUDA version you are using. From my experience, the last version compatible with all 3D libraries was CUDA 11.2. I believe it is possible to use a more recent version, but you have to see compatibility with all the libraries you use.
Here is how I used to install the libraries FROM nvcr.io/nvidia/pytorch:21.02-py3
:
# Conda init
source ${CONDA_DIR}/etc/profile.d/conda.sh
conda activate ${CONDA_ENV}
# Install PyTorch
pip install --no-cache-dir \
torch==1.8.0 \
torchvision==0.9.0
pip install --no-cache-dir torch-points-kernels
pip install --no-cache-dir torchnet
# Install torch-geometric and dependencies
pip install --no-cache-dir torch-scatter
pip install --no-cache-dir torch-sparse
pip install --no-cache-dir torch-cluster
pip install --no-cache-dir torch-spline-conv
pip install --no-cache-dir torch-geometric
# Install MinkowskiEngine
apt-get install --yes --quiet --no-install-recommends libopenblas-dev
pip install --no-cache-dir --verbose --no-deps \
--install-option="--blas=openblas" \
MinkowskiEngine==v0.4.3
# Install torchsparse
apt-get install --yes --quiet --no-install-recommends libsparsehash-dev
pip install --no-cache-dir --upgrade git+https://github.com/mit-han-lab/torchsparse.git
pip install pycuda
# Install Torch Points 3D
pip install --no-cache-dir \
omegaconf wandb plyfile hydra-core==0.11.3 pytorch-metric-learning
# pip install --no-cache-dir torch-points3d==1.2.0
thanks a lot @mgomesborges !!!
I ended up using the pytorch image on dockerhub. Less other dependencies than the nvcr version.
"docker://pytorch/pytorch:2.0.0-cuda11.7-cudnn8-runtime" from that I can install pytorch-scatter with this:
pip3 install torch-scatter -f https://data.pyg.org/whl/torch-2.0.0+cu117.html
And everything works!
Thanks anyways. If I end up needing the other libraries I'll get back to your response :)
I struggle to build an image with pytorch_scatter based on the official nvcr.io/nvidia/pytorch:20.12-py3 image. But, interestingly, I am able to correctly install pytorch_scatter in an already running container.