Closed mikemhenry closed 1 year ago
Done!
This is a duplicate of #13
So sorry for the dupe! Feel free to close this one or keep it (does github let you merge issues :thinking: ) @peastman can you make a tag or a release? That way github will make an archive that conda-forge can use to build the package.
Thanks!
Once I get a tag or release I can update the PR: https://github.com/conda-forge/staged-recipes/pull/18353 everything passes so this should be a quick one
@peastman : Can you cut a prerelease tag so that we can start building versions of this on conda-forge
for testing?
You can create a build directly from the git hash without needing a release. Until we figure out the problems in #25 though, I don't think it's suitable for building even a prerelease.
If it is premature to distribute on conda-forge, that's fine. I didn't know that this works: https://github.com/openmm/openmm-ml/archive/28b4fadd761221fbe085570faac3e95a2d0be4e0.tar.gz I thought that the ref
had to be a tag or release and not a commit hash!
I've uploaded the build artifacts here https://anaconda.org/mmh/openmm-ml in case anyone wants to test/use them before we get the package on conda-forge. conda install -c mmh openmm-ml
@peastman I've got a PR staged to get this on conda-forge, but I would like your approval first.
EDIT: Ah looks like you want #25 resolved first!
There's a working for #25 in https://github.com/openmm/openmm/pull/3533. It also looks like the root problem may be fixed in PyTorch 1.11, though we still need to verify that.
Thanks for all of your feedback on the PR feedstock @peastman :open_hands: I think that because this is a pure python package (so easy for people to install/use it without it being on conda-forge), that getting it on conda-forge isn't as big of a priority and it can wait until openmm 7.7.1 comes out, thoughts?
That sounds reasonable. The plan is that when all the pieces are ready, we'll do a big release of "OpenMM for ML" which will involve coordinated packages for OpenMM, NNPOps, OpenMM-Torch, and OpenMM-ML.
That sounds good to me, for now I will close the PR then and re-open it when we are ready for the release!
@peastman is it time to get this package onto conda-forge?
@dominicrufa : Everything look good from your end too?
Absolutely! It and https://github.com/conda-forge/openmm-torch-feedstock/pull/23 are the last pieces we need to get the 8.0 beta out.
Working on that here: https://github.com/conda-forge/staged-recipes/pull/19223
Right now it is linux only, windows needs openmm-torch
and nnpops
.
OSX needs nnpops
Do we want to try and get OSX support?
Thanks! Yes, Mac support is important. Unfortunately, it's impossible for us to build Windows packages, because the conda-forge
version of PyTorch doesn't include Windows. The version on the pytorch
channel does, but of course we're not allowed to use it for packages distributed through CondaForge.
https://github.com/conda-forge/pytorch-cpu-feedstock/pull/63 We will see what we can do!
@raimis could you add OSX testing to the CI matrix for nnpops? I'd like to make sure upstream has 1st class support for OSX before trying to distribute it on conda-forge.
@peastman is CUDA 10.2 support required for this push? I'm having a hard time getting 10.2 to build for nnpops, if we need it that is fine but if we don't then I can invest more effort in OSX support.
It's nice to have but not essential, especially for the beta. Mac is more important.
:+1: There isn't cuda support for osx it looks like (or at least pytorch on osx is CPU only) so I will need to play around with the nnpops feedstock to get a osx cpu only version built.
That's because there's no CUDA for Macs. Apple hasn't used NVIDIA GPUs in many years. Starting in PyTorch 1.12 they added support for Apple GPUs through Metal. I don't know if that requires a special build.
Any progress on this?
Thanks!
I don't have access on this repo to assign this issue to myself, but please assignee it to me :)