Lambda Stack Dockerfiles

Dockerfiles with rolling-release Lambda Stack, designed for use with nvidia-container-toolkit

Installing nvidia-container-toolkit

1. Ensure that you have a docker version > 19.03. On Ubuntu, you can simply run sudo apt-get install docker.io. On a different OS, or if you prefer to use upstream docker, follow Docker's installation instructions

2. If using Lambda Stack on your host machine, install nvidia-container-toolkit with sudo apt-get install nvidia-container-toolkit. Otherwise, follow NVIDIA's installation instructions

Building images

Build the image with the appropriate command for the distribution you wish to use.

sudo docker build -t lambda-stack:18.04 -f Dockerfile.bionic .
sudo docker build -t lambda-stack:20.04 -f Dockerfile.focal .
sudo docker build -t lambda-stack:22.04 -f Dockerfile.jammy .
sudo docker build -t lambda-stack:22.04 -f Dockerfile.jammy.llmSolar .

Note that building these docker images requires acceptance of the cuDNN license agreement

Testing images

Here's a simple PyTorch test to make sure that your GPUs are usable in your docker images

$ sudo docker run --gpus 2 lambda-stack:22.04 python3 -c "import torch; print(torch.cuda.device_count())"
2

Run this command from the root of the repo that you want to use your code inside the lambda container with access to 4 GPUs.

docker run -it --gpus 4 --mount type=bind,source="$(pwd)",target=/app  lambda-stack:22.04 

Run this command from the root of the repo that you want to use your code inside the lambda container with access to all GPUs.

docker run --gpus all --mount type=bind,source="$(pwd)",target=/app  lambda-stack:22.04 

Command to check GPUs mounted in your machine.

sudo lshw -C display