avolkov1
commented
2 years ago Refer to: https://github.com/NVIDIA/fsi-samples/blob/main/gQuant/docker/build.sh
That will setup a docker container. But if you want to manually set up conda with gquant plugin run something like this:
export RAPIDS_VERSION=21.06
export RAPIDS_CUDA_VER=11.2
conda install -y mamba -n base -c conda-forge
mamba install -y -c rapidsai -c nvidia -c conda-forge -c defaults \
rapids=$RAPIDS_VERSION cudatoolkit=$RAPIDS_CUDA_VER python=3.8 && \
conda clean --all -y
mamba install -y -c conda-forge -c defaults \
jupyterlab'>=3.0.0' jupyter-packaging'>=0.9.2' jupyterlab-system-monitor \
nodejs=12.4.0 python-graphviz pydot ruamel.yaml && \
conda clean --all -y && \
jlpm cache clean && \
jupyter lab clean
pip install bqplot==0.12.21 && \
jlpm cache clean && \
jupyter lab clean
# either git checkout the repo and then pip install
cd greenflowlab && pip install .
# or install from PyPi https://pypi.org/project/greenflowlab/
pip install greenflowlab
# greenflowlab will install greenflow as well.
# plugins are here https://github.com/NVIDIA/fsi-samples/tree/main/gQuant/plugins
# cd to each directory and install
cd gQuant/plugins/gquant_plugin
pip install .
# or from PyPi: https://pypi.org/project/greenflow-gquant-plugin/
pip install greenflow-gquant-plugin
Above I extracted parts from the dockerfile build script. Modify that to use a named conda environment if you prefer instead of installing everything into base conda environment.
We use mamba instead of conda to resolve dependencies in conda. Conda dependency resolution is very slow or broken. Also, we have not tested with Cuda 11.5 and the very latest Rapids. It might work with latest versions, but we haven't verified. Please use Cuda 11.2 and Rapids 21.06. Those were the latest versions we tested with.
complyue
At time of writing,
gquant_pluginassumesrapids=0.19:https://github.com/NVIDIA/fsi-samples/blob/main/gQuant/plugins/gquant_plugin/README.md#install-the-external-example-plugin
I tried to create a virtual env for
rapids=0.19withconda create -n fsi -c rapidsai -c nvidia -c conda-forge rapids=0.19 cudatoolkit, but conda just can't solve it even after several hours of running.Blindly tried with CUDA 11.5 and rapids 21.12, jupyter-lab just won't start.