conda-forge / conda-forge/tensorflow-feedstock
Help mamba decide on CUDA variant
- Dominant language
- Shell
- Stars
- 93
- Forks
- 93
- PR merge metrics
- No merged PRs in 30d
Description
Mamba currently can't properly decide which `tensorflow` cuda package to prefer since we don't directly depend on `cudatoolkit` in the `tensorflow` package.
Only in the `tensorflow-base` package we have a `cudatoolkit` dependency.
However, `tensorflow-base` is connected to `tensorflow` via an exact dependency.
If we'd also add the `cudatoolkit` to the `tensorflow` package, mamba would be able to select the best version.
OR we could remove the tensorflow variant completely and just have a single metapackage that depends on `tensorflow-base {{ VERSION }}` (but not exact). Then mamba could decide the tensorflow-base version more freely.
Contributor guide
Assessment
This issue has not been assessed yet.