GitHub Codespaces: no space left on device — what takes up the extra space?

Question

I have a GitHub repository with a .devcontainer file to customize the Codespace image, that fails to start. The repository has the following structure:

.
├── .devcontainer
│   └── devcontainer.json
├── .github
│   └── workflows
│       └── build_docker_image.yml
└── ingredients
    └── .devcontainer
        └── devcontainer.json

The workflow in build_docker_image.yml simply builds a Docker container (based on the containers.dev tutorial on prebuilt devcontainers) and pushes the image to Docker Hub:

devcontainer build --workspace-folder ingredients --image-name ${{ github.repository }}
docker push ${{ github.repository }}

This is the image on Docker Hub: https://hub.docker.com/layers/kostrykin/mobi-devcontainer-python/latest/images/sha256-8158e3aaf271db4d2401563b1996ca9f1cf538c8cbdb3b3d372a697d10ebb199?context=repo

The contents of the ingredients/.devcontainer/devcontainer.json file which is used for the image:

{
  "name": "xxxx xxxx",
  "image": "mcr.microsoft.com/devcontainers/universal:2"
}

The contents of the .devcontainer/devcontainer.json file which is used to run the Codespace:

{
  "image": "docker.io/kostrykin/mobi-devcontainer-python:latest"
}

Shouldn't this be equivalent to running a Codespace with the default image, i.e. the mcr.microsoft.com/devcontainers/universal:2 image? However, when running the codespace with the custom docker.io/kostrykin/mobi-devcontainer-python:latest image, the Codespace fails to start:

[2B2023-11-22 11:16:32.738Z: docker: failed to register layer: ApplyLayer exit status 1 stdout: stderr: write /home/codespace/.local/lib/python3.10/site-packages/nvidia/cudnn/lib/libcudnn_adv_infer.so.8: no space left on device.

The full log is very long, so I posted it here.

My actual question is: What takes up the extra space, when using the custom image in comparison to using the original image? I see that there should be an extra layer, but shouldn't this be close to zero in size? Also, using docker inspect I can confirm that both images are actually equal in size.

And the natural follow-up question is: Is there any elegant way to solve this issue, aside from raising the host requirements from 32GB to 64GB, or using docker system prune --all --force as the initializeCommand of the devcontainer?

---

In case it matters, here is the output of dh -h ran inside of the Codespace after creating it with the initializeCommand mentioned above:

/workspaces/mobi-devcontainer-python (main) $ df -h
Filesystem      Size  Used Avail Use% Mounted on
overlay          32G   13G   18G  43% /
tmpfs            64M     0   64M   0% /dev
shm              64M     0   64M   0% /dev/shm
/dev/root        29G   22G  7.2G  76% /usr/sbin/docker-init
/dev/loop3       32G   13G   18G  43% /workspaces
/dev/sda1        44G  148K   42G   1% /tmp