Skip to content

CUDA.md

Latest commit

 

History

History
196 lines (142 loc) · 6.07 KB

CUDA.md

File metadata and controls

196 lines (142 loc) · 6.07 KB

CUDA-enabled JupyterLab Python docker stack

GPU accelerated, multi-arch (linux/amd64, linux/arm64/v8) docker images:

Images available for Python versions ≥ 3.11.1.

🔬 Check out jupyterlab/cuda/python/scipy at https://demo.cuda.jupyter.b-data.ch.
👉 You can ask b-data to whitelist your GitHub account for access.

Build chain

The same as the JupyterLab Python docker stack.

Features

The same as the JupyterLab Python docker stack plus

  • CUDA runtime, CUDA math libraries, NCCL and cuDNN
    • including development libraries and headers
  • TensortRT and TensorRT plugin libraries
    • including development libraries and headers

👉 See the CUDA Version Matrix for detailed information.

Subtags

The same as the JupyterLab Python docker stack.

Table of Contents

Prerequisites

The same as the JupyterLab Python docker stack plus

  • NVIDIA GPU
  • NVIDIA Linux driver
  • NVIDIA Container Toolkit

ℹ️ The host running the GPU accelerated images only requires the NVIDIA driver, the CUDA toolkit does not have to be installed.

Install

To install the NVIDIA Container Toolkit, follow the instructions for your platform:

Usage

Build image (base)

latest:

cd base && docker build \
  --build-arg BASE_IMAGE=ubuntu \
  --build-arg BASE_IMAGE_TAG=22.04 \
  --build-arg BUILD_ON_IMAGE=glcr.b-data.ch/cuda/python/ver \
  --build-arg PYTHON_VERSION=3.11.3 \
  --build-arg CUDA_IMAGE_FLAVOR=devel \
  -t jupyterlab/cuda/python/base \
  -f latest.Dockerfile .

version:

cd base && docker build \
  --build-arg BASE_IMAGE=ubuntu \
  --build-arg BASE_IMAGE_TAG=22.04 \
  --build-arg BUILD_ON_IMAGE=glcr.b-data.ch/cuda/python/ver \
  --build-arg CUDA_IMAGE_FLAVOR=devel \
  -t jupyterlab/cuda/python/base:MAJOR.MINOR.PATCH \
  -f MAJOR.MINOR.PATCH.Dockerfile .

For MAJOR.MINOR.PATCH ≥ 3.11.1.

Create home directory

Create an empty directory:

mkdir jupyterlab-jovyan
sudo chown 1000:100 jupyterlab-jovyan

It will be bind mounted as the JupyterLab user's home directory and automatically populated on first run.

Run container

❗ Always mount the user's entire home directory.
Mounting a subfolder prevents the container from starting.1

self built:

docker run -it --rm \
  --gpus '"device=all"' \
  -p 8888:8888 \
  -u root \
  -v "${PWD}/jupyterlab-jovyan":/home/jovyan \
  -e NB_UID=$(id -u) \
  -e NB_GID=$(id -g) \
  -e CHOWN_HOME=yes \
  -e CHOWN_HOME_OPTS='-R' \
  jupyterlab/cuda/python/base[:MAJOR.MINOR.PATCH]

from the project's GitLab Container Registries:

docker run -it --rm \
  --gpus '"device=all"' \
  -p 8888:8888 \
  -u root \
  -v "${PWD}/jupyterlab-jovyan":/home/jovyan \
  -e NB_UID=$(id -u) \
  -e NB_GID=$(id -g) \
  -e CHOWN_HOME=yes \
  -e CHOWN_HOME_OPTS='-R' \
  IMAGE[:MAJOR[.MINOR[.PATCH]]]

IMAGE being one of

The use of the -v flag in the command mounts the empty directory on the host (${PWD}/jupyterlab-jovyan in the command) as /home/jovyan in the container.

-e NB_UID=$(id -u) -e NB_GID=$(id -g) instructs the startup script to switch the user ID and the primary group ID of ${NB_USER} to the user and group ID of the one executing the command.

-e CHOWN_HOME=yes -e CHOWN_HOME_OPTS='-R' instructs the startup script to recursively change the ${NB_USER} home directory owner and group to the current value of ${NB_UID} and ${NB_GID}.
ℹ️ This is only required for the first run.

The server logs appear in the terminal.

Using Docker Desktop

sudo chown 1000:100 jupyterlab-jovyan might not be required. Also

docker run -it --rm \
  --gpus '"device=all"' \
  -p 8888:8888 \
  -v "${PWD}/jupyterlab-jovyan":/home/jovyan \
  IMAGE[:MAJOR[.MINOR[.PATCH]]]

might be sufficient.

Similar projects

What makes this project different:

  1. Multi-arch: linux/amd64, linux/arm64/v8
  2. Derived from nvidia/cuda:11.8.0-cudnn8-devel-ubuntu22.04
    • including development libraries and headers
  3. TensortRT and TensorRT plugin libraries
    • including development libraries and headers
  4. IDE: code-server next to JupyterLab
  5. Just Python – no Conda / Mamba

See Notes for tweaks, settings, etc.

Footnotes

  1. The only exception is the use case described at Jupyter Docker Stacks > Quick Start > Example 2. ↩