Deep learning images in Docker Hub: jinchaolove/dl developed from nvidia/cuda-cudnn-devel-ubuntu, with support of communications between containers and GPUs.
UCX, Open MPI, ADMA, OFED, SHARP are default installed to support distributed training with NVIDIA NCCL. All images are tested and passed the NVIDIA NCCL Tests.
conda3 is installed in a multi-user manner (run adduser $USER condaGroup for new users). Private envs are created in ~/.conda and common pkgs are shared in /use/local/conda by default. The python (python3.9) and pip (pip3) are soft linked from conda3 by default.
CUDA compatibility is enabled by adding ENTRYPOINT script in /entrypoint.sh. See Best practices for working with mismatched driver versions.
ubuntu:18.04(32 & 64-bit, python2 & 3),20.04(64-bit, python3)cuda:10.2,11.3,11.6(backward and minor version forward compatible)
- Build the image (recommend modify yours from the base image in jinchaolove/dl instead of re-built from start, see example in
Dockerfile.msra):
# docker build -t <name>:<tag> -f Dockerfile . # base image
cat Dockerfile.msra | docker build - -t <name>:<tag> # job-specified image- Run the container:
docker run -ti <name>:<tag> /bin/bash- Others: please see docker docs
Please give me a 🌟 if this repository helps you 🤗
If you have any questions, please feel free to issue or contact me (Jinchao).