conda-lock not solving a solvable environment

[Siniara]

This env couldn't be solved by conda-lock but was easily be solved by mamba.

channels:
- conda-forge
dependencies:
- python=3.11
- cuda-version=12.6
- pytorch-gpu
- pandas
- numpy
- scikit-learn
- matplotlib
- pyyaml
platforms:
- win-64

I lost a lot of time struggling with this as I assumed that conda-lock solves it the same as mamba, so thought that there are missing packages on conda-forge or what not. But then almost by accident I installed it with mamba and it just worked.

[maresb]

Hi @Siniara, thanks for using conda-lock, and I'm terribly sorry about your lost time.

I was able to get conda-lock to work for your environment by specifying --with-cuda 12.6.

The challenge here is that installation of this environment will succeed on systems where CUDA >=12.6 is installed at the system level, and will fail otherwise. Since conda-lock is designed to (mostly) work the same regardless of the system it's installed on, we must by default make conservative options about the system requirements. This is where the virtual packages come into play, and the --with-cuda argument injects the specified CUDA into that.

I realize there are still lots of rough edges here. And probably it's safe to infer the --with-cuda value in this case from the cuda-version dependency, and we should do that.

I really appreciate you taking the time to provide this feedback, and I hope my explanation makes some sense.

[Siniara]

Thanks for getting back to me! It worked for me too with the flag! So this might have been a skill issue! So the default then is as thought the flag --without-cuda was passed? Maybe at least some warning might be nice, because I was just getting an infinite solve without any messages.

I did some more testing with the env.yaml file specification and the --with-cuda flag.

1. env.yaml cuda-version=12.6; --with-cuda not specified -> env doesn't solve, the case in the post
2. env.yaml cuda-version=12.6; --with-cuda 12.6 -> env solved with cuda 12.6
3. env.yaml cuda-version not specified; --with-cuda 12.6 -> env solved with cuda 12.9
4. env.yaml cuda-version=12.6; --with-cuda 12.9 -> env solved with cuda 12.6

It seems that the flag doesn't enforce a version? In which case it should maybe be treated as boolean instead?