kentwait/cuda-mamba repository overview

⁠CUDA-Mamba

Nvidia CUDA container with Mamba package manager installed. This container is a useful base for containers that need Python package management and access to CUDA libraries.

⁠Downloading from Dockerhub

CUDA-Mamba can be downloaded from Dockerhub using the following

docker pull kentwait/cuda-mamba:latest

This downloads a prebuilt image with the default configuration⁠. Other supported image tags are listed in https://hub.docker.com/r/kentwait/cuda-mamba/tags⁠.

⁠Building from Dockerfile

For more flexibility and control over CUDA and Python versions, clone this repository⁠ and build using the included Dockerfile.

⁠Build arguments

• CUDA_VERSION: "11.7.1"
• CUDA_IMAGE_TYPE: "base"
• CUDA_OS: "ubuntu22.04"
• CONDA_DIR: "/opt/conda"
• PYTHON_VERSION: "3.9.15"

The build can be customized by specifying build argument values. Here is an example of building the container image using the default values.

docker build -t <image_name> \
  --build-arg CUDA_VERSION="11.7.1" \
  --build-arg CUDA_IMAGE_TYPE="base" \
  --build-arg CUDA_OS="ubuntu22.04" \
  --build-arg CONDA_DIR="/opt/conda" \
  --build-arg PYTHON_VERSION="3.9.15" .

This command will build a container image following the default configuration⁠.

Arguments staring with "CUDA_" are used to build the container tag in the format <CUDA_VERSION>-<CUDA_IMAGE_TYPE>-<CUDA_OS>. For example, the default arguments builds a container tag 11.7.1-base-ubuntu22.04.

A full list of supported container tags for the CUDA base container can be found here⁠. For CUDNN containers, append cudnn8- to "runtime" or "devel" in CUDA_IMAGE_TYPE. For example, CUDA_IMAGE_TYPE=cudnn8-runtime instead of CUDA_IMAGE_TYPE=runtime.

⁠Default configuration

• Uses nvidia/cuda:11.7.1-base-ubuntu22.04 base container
• Installs the latest Mamba with Python 3.9.15
• Install cuda-python <= 11.7.1 and cudatoolkit-dev <= 11.7.1