Raimondas Galvelis

@peastman I have just finished integrating NNPOps (https://github.com/torchmd/torchmd-net/pull/50). The performance (https://github.com/torchmd/torchmd-net/blob/main/benchmarks/graph_network.ipynb) is just 2-3 time better for the small molecules (10-100 atoms) and no significant improvement for the larger ones....

Just before going into NNPOps, I checked how much CUDA Graphs (https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/) can help. CUDA Graphs don't work with TorchMD_GN due to https://github.com/rusty1s/pytorch_cluster/issues/123. To circumvent this, I have implemented a...

The current implementation doesn't support batching, but it could be implemented.

**Optimization: round 2** I have wrote optimized kernels for the neighbor search (#61) and message passing (#69). The kernels are drop-in replacement for the generic kernels from PyTorch Geometric and...

Regarding the variable names, my preference is `d/dy`, but `energy/forces` is fine. Could they be made consistent in `models.model.TorchMD_Net` too.

Another options: `models.model.TorchMD_Net.forward` does not only the computation of `y` (the forward pass), but also the computation of `dy` (the backward pass). We can factor out the gradient computation into...

``` (openmm) mateoc@mateoc-IdeaPad-5-15ITL05:~/Documentos/practiques/tleap_propka$ python -c "import openmm" Traceback (most recent call last): File "", line 1, in File "/home/mateoc/Documentos/practiques/tleap_propka/openmm.py", line 1, in from openmm.app import * ModuleNotFoundError: No module named...

Do you install the same version of `torch_cluser` with `conda` and `pip`?

Key differences with #134: - The atom- and molecular-wise operations are implement with `pre_reduce` and `post_reduce` methods. This is consistent with the output model. - The priors are executed after...

@stefdoerr @PhilippThoelke what do you think?