How do you indicate CUDA_VERSION=12.4.1?

[f0nzie]

I am trying to build the CUDA r-base containers but I am receiving errors during the build: " requirement error: unsatisfied condition: cuda>=12.5".

Host OS: Linux Fedora Silverblue 40 Host CUDA: 12.4.1 GPU: NVIDIA RTX 3080 Build: Local IDE: VS Code Method: "Reopen in Container" Extension: DevContainer

It seems that something changed during the past days because last week I was able to build cuda-r-base, and yesterday and today, the containers do not build anymore.

Last week, I made a build status table to check which images built fine.

But as of today, none of them run.

Any clues?

[benz0li]

Any clues?

My fault.

I forgot to update the README: CUDA-based images are now derived from nvidia/cuda:12.5.0-runtime-ubuntu22.04.

Updated right now: https://github.com/b-data/data-science-devcontainers/commit/bc3af9f8821c736fc1d3268772fdfb5a227b5ab0

[benz0li]

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

[benz0li]

In order to run the current Data Science Dev Containers I recommend a regular NVIDIA Linux driver version ≥ 555.42.02.

https://www.nvidia.com/download/driverResults.aspx/224751/en-us/

[benz0li]

The latest versions of the CUDA Julia/Python/R Data Science Dev Containers are [currently] based on the latest version of nvidia/cuda.

Any other version than latest is frozen (see Security), i.e. does not get updated [any more].
ℹ️ E.g. the last Dev Containers that are based on nvidia/cuda:11.8.0-cudnn8-devel-ubuntu22.04 are

• R 4.3.3
• Python 3.11[.9]
• Julia 1.10.1

👉 Pass R_VERSION=4.3.3, PYTHON_VERSION=3.11[.9] or JULIA_VERSION=1.10.1 as build.args when building the Data Science Dev Container to get an image that runs with any regular NVIDIA Linux driver version ≥ 520.61.05.

[f0nzie]

Oh, I see. I have the Driver Version: 550.78 for CUDA Version: 12.4.

"Pass R_VERSION=4.3.3, PYTHON_VERSION=3.11[.9] or JULIA_VERSION=1.10.1 as build.args to create the Data Science Dev Container to get an image that runs with any NVIDIA Linux driver version ≥ 520.61.05."

Do you mean, I could be able to build an image with a lesser CUDA version (12.4) if I choose R-4.3.3 instead of latest?

[benz0li]

Do you mean, I could be able to build an image with a lesser CUDA version (12.4) if I choose R-4.3.3 instead of latest?

Yes (R_VERSION=4.3.3 instead of R_VERSION=4.4.0).

[benz0li]

See also

• Julia: CUDA Version Matrix
• Python: CUDA Version Matrix
• R: CUDA Version Matrix

[benz0li]

Another possibility is to use a Long Term Support Branch (LTSB) driver:

• 470: CUDA 11.0+
  • All Data Science Dev Containers: Works due to the CUDA forward compat package[^1]
• 535: CUDA 12.2+
  • 'CUDA 12'-based Data Science Dev Containers: Works due to the CUDA forward compat package[^1]
  • 'CUDA 11'-based Data Science Dev Containers: Supported due to backward compatibility

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

Cross references:

• https://docs.nvidia.com/deploy/cuda-compatibility/index.html#id3
• https://gitlab.com/nvidia/container-images/cuda/-/blob/master/doc/container_tags.pdf

[^1]: The Data Science Dev Containers, i.e. the nvidia/cuda images, already have the right CUDA forward compat package installed.

[benz0li]

I have updated the R CUDA Version Matrix accordingly.

@f0nzie Please close this issue if your problem is resolved.

[f0nzie]

Yes (R_VERSION=4.3.3 instead of R_VERSION=4.4.0).

Awesome! Tried downgrading to one version previous to latest last night and all CUDA images work!

Thank you so much!

[f0nzie]

• 470: CUDA 11.0+
• 535: CUDA 12.2+

That is a good alternative. I have the 550.78 driver in my machine. That is maybe the reason the image build failed.

[benz0li]

@f0nzie The last versions running CUDA 11.8.0 are

• R 4.3.3
• Python 3.12.1
  • Python 3.11.9+ ℹ️ I am going to keep CUDA 11.8.0 for all future Python 3.11 releases
• Julia 1.10.1

[benz0li]

• 470: CUDA 11.0+
• 535: CUDA 12.2+

That is a good alternative. I have the 550.78 driver in my machine. That is maybe the reason the image build failed.

If you install[^1] Linux driver version 535, you can build and run all available Data Science Dev Containers for the next two years; this driver does not reach EOL until June 2026.

[^1]: and update from time to time