ML-stuffing CUDA scripts for Ubuntu-server with PyTorch, Apex, Deepspeed
After sh ./install-nvidia-docker.sh done, please run reboot
When everything established you may build and run image with:
# set DOCKER_BUILDKIT=1 as environment variable to boost docker
docker image build --tag 'your-image-name'
# you'll get path to your image, so:
docker container run --gpus all path/to/your-image-name <blah-blah>
Or use make to build and run with Jupyter Notebook
make build
make jupyter
Also minimal test, if cuda works, looks like this:
docker run --gpus all nvidia/cuda:10.0-base nvidia-smi