NNPOps 0.6

[raimis]

Let's start the new release.

New features:

• Particle Mesh Ewald #103

Testing improvements:

• Add Self-Hosted AWS GPU Runner #100
• Fix CMake not adding some tests #104

[raimis]

@RaulPPelaez any change to finish #105?

[raimis]

Does anybody want to implement/finish something more?

Ping: @peastman

[RaulPPelaez]

@RaulPPelaez any change to finish #105?

It defeated me tbh. I do not think it is worth stopping the release for it. I will give it another spin.

[raimis]

The test on the GPU don't pass: https://github.com/openmm/NNPOps/actions/runs/5645655178

[RaulPPelaez]

I think both failures could be attributed to the inaccuracy introduced by the non-deterministic order of operations in the CUDA versions.
The first test yields an error of 0.0343 (tolerance is 0.025) after doing a sum operation.
The second has a relative error tolerance of 1e-7, which is already a bit too much for single precision

[peastman]

Sounds good to me.

[mikemhenry]

Do we want to update the tests to make them less strict?

[RaulPPelaez]

Maybe this is enough?
https://pytorch.org/docs/stable/generated/torch.use_deterministic_algorithms.html

[RaulPPelaez]

Both tests pass locally in a 4090 at the current master 5e2438d , perhaps the CI was just a fluke?
EDIT: Also tried on a GPU with arch 75 (as the one in the CI) just in case and one of the BatchedNN tests failed also because some tolerance.

        assert energy_error < 5e-7
        if molFile == '3o99':
            assert grad_error < 0.05 # Some numerical instability
        else:
>           assert grad_error < 5e-3
E           AssertionError: assert tensor(0.0071, device='cuda:0') < 0.005

I do not think this is due to a bug, I would blame whatever kernel torch uses for sm_75 being a tad more inaccurate.
Using deterministic algorithms does indeed make the error go away. I opened #108 to solve try in the CI.

[mikemhenry]

Okay so now that #108 is merged are we ready for a new release? There have been some breaking but very good changes in conda-forge packaging for CUDA conda-forge/conda-forge.github.io#1963 so it might take a bit to get that sorted. I will get the release on conda-forge started, but will wait to actually merge it in until we decide if the release is ready.

[RaulPPelaez]

The changes should not affect CUDA<12 builds, right?
AFAIK the new packages are only for CUDA >=12.
Anyhow I say we are ready to release v0.6.
cc @raimis, could you please?.

[raimis]

OK! I'm making the release

[raimis]

We have a release: https://github.com/openmm/NNPOps/releases/tag/v0.6

[raimis]

@mikemhenry you can proceed with conda-forge/nnpops-feedstock#24. If CUDA 12 is an issue, we can build just for CUDA 11. And then sort out CUDA 12.

[raimis]

NNPOps 0.6 is out!

@jchodera could you make a tweet (or whatever is it called now)?