cuda-version>12 tries to pull cudatoolkit>12

[RaulPPelaez]

Solution to issue cannot be found in the documentation.

• [X] I checked the documentation.

Issue

I am trying to install CUDA 12 in an environment which has cudatoolkit (11.8) installed. In particular, I am expecting this sequence to work:

$ mamba create -n test cudatoolkit 

Looking for: ['cudatoolkit']

conda-forge/linux-64                                        Using cache
conda-forge/noarch                                          Using cache
Transaction

  Prefix: /home/raul/miniforge3/envs/test

  Updating specs:

   - cudatoolkit

  Package          Version  Build        Channel           Size
─────────────────────────────────────────────────────────────────
  Install:
─────────────────────────────────────────────────────────────────

  + _libgcc_mutex      0.1  conda_forge  conda-forge     Cached
  + libstdcxx-ng    13.2.0  h7e041cc_3   conda-forge     Cached
  + libgomp         13.2.0  h807b86a_3   conda-forge     Cached
  + _openmp_mutex      4.5  2_gnu        conda-forge     Cached
  + libgcc-ng       13.2.0  h807b86a_3   conda-forge     Cached
  + cudatoolkit     11.8.0  h4ba93d1_12  conda-forge     Cached

  Summary:

  Install: 6 packages

  Total download: 0 B

─────────────────────────────────────────────────────────────────

Confirm changes: [Y/n] y

Downloading and Extracting Packages:

Preparing transaction: done
Verifying transaction: done
Executing transaction: / By downloading and using the CUDA Toolkit conda packages, you accept the terms and conditions of the CUDA End User License Agreement (EULA): https://docs.nvidia.com/cuda/eula/index.html

done

To activate this environment, use

     $ mamba activate test

To deactivate an active environment, use

     $ mamba deactivate

$ mamba activate test
$ mamba install "cuda-version>12"

Looking for: ["cuda-version[version='>12']"]

conda-forge/linux-64                                        Using cache
conda-forge/noarch                                          Using cache
Could not solve for environment specs
The following packages are incompatible
├─ cuda-version >12  is installable with the potential options
│  ├─ cuda-version 12.1 would require
│  │  └─ cudatoolkit 12.1|12.1.* , which can be installed;
│  └─ cuda-version 12.2 would require
│     └─ cudatoolkit 12.2|12.2.* , which can be installed;
└─ cudatoolkit is not installable because it conflicts with any installable versions previously reported.

I would expect cudatoolkit>11.8 to be a metapackage depending on, for instance, the "cuda" package. Maybe this is intended behavior?

Installed packages

N/A

Environment info

N/A

[leofang]

For CUDA >= 12, there's no cudatoolkit. It was the mega package (containing only runtime libraries) that we used for CUDA <= 11. The meta package that you seek is now named as cuda-toolkit.

[jakirkham]

Would recommend not explicitly installing the cudatoolkit package for the reason Leo mentioned

Instead would explicitly install the cuda-version package with the intended version and any other packages that you would like. For example

conda create -n cupy python=3.10 cuda-version=11.8 cupy -y
conda activate cupy
conda install cuda-version=12.2 -y

This way Conda understands cudatoolkit is an implicit dependency (pulled in by other packages) and won't try to include cudatoolkit to satisfy the solve

Also as cuda-version is built for all CUDA versions conda-forge had supported. It can be used across the CUDA 11/12 divide easily

Hope that helps

[RaulPPelaez]

Thanks guys. Your commands do not work for me. I am trying to install a package that depends on cudatoolkit in its cuda 11.8 version, for instance:

mamba create -n test cupy cuda-version=11.8

Then, when I try to update cuda-version:

$ mamba install "cuda-version>12"

Looking for: ["cuda-version[version='>12']"]

conda-forge/linux-64                                        Using cache
conda-forge/noarch                                          Using cache

Pinned packages:
  - python 3.12.*

Could not solve for environment specs
The following packages are incompatible
├─ cuda-version >12  is installable with the potential options
│  ├─ cuda-version 12.1 would require
│  │  └─ cudatoolkit 12.1|12.1.* , which can be installed;
│  └─ cuda-version 12.2 would require
│     └─ cudatoolkit 12.2|12.2.* , which can be installed;
└─ cudatoolkit is not installable because it conflicts with any installable versions previously reported.

[RaulPPelaez]

Actually your commands work when using conda, the culprit is mamba.

[RaulPPelaez]

I opened an issue at mamba to address this https://github.com/mamba-org/mamba/issues/3120

[jakirkham]

Yep had tested these with Conda. Interesting to know Mamba has a bug. Let's follow up there

Will close this out as answered. We can track the upstream issue