Open alexbarghi-nv opened 3 months ago
Got it. We'll look into it.
This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you
(not stale)
This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you
(not stale)
This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you
(not stale)
Sorry, we're currently short of resources to work on this work item. So there's no ETA yet.
Yeah no worries. Just trying to keep the bot at bay 😉
Edit: Should add am more than happy to discuss & review with someone from your team when they are available
🚀 Feature
Currently, DGL packages are separated by CUDA version using labels. i.e. to install DGL for CUDA 12.1, you use the
dglteam/label/cu121
channel and for CUDA 11.8 you use thedglteam/label/cu118
channel. This is causing issues with RAPIDS CI where we test against several different CUDA versions, and are running into issues where we cannot resolve the right version of DGL because our dependency generator doesn't support setting a channel by CUDA version. Most other packages have a dependency oncuda-version
that ensures the correct version of the package gets installed, so this is only an issue with DGL right now.Motivation
The RAPIDS Graph team has been overhauling our testing process for GNN packages (cuGraph-DGL, cuGraph-PyG, WholeGraph, etc.) and we want to expand CUDA 12 testing of our packages to align with the rest of RAPIDS. We have run into issues in CI where our automated process picks the wrong channel or package, and have resolved all of these problems with the exception of this issue for DGL.
Alternatives
We might be able to make some changes to our dependency file generator to support this in the short-term, but in the long term we would like to rely on
cuda-version
which we do for all our other dependencies in conda.Pitch
We would like DGL to adopt dependency on
cuda-version
and remove the separate labels per CUDA version. So you could install for CUDA 12.1 by runningconda install -c conda-forge -c dglteam dgl cuda-version=12.1
Additional context
We're encountering this issue here: https://github.com/rapidsai/cugraph/pull/4346
@VibhuJawa @jakirkham