Is there a way to accelerate IS2RS task optimization speed

Hi OCP team,

I am evaluating the GemNet-OC optimisation for the IS2RS task using different test sets. It took me 3 days to finish one test set (test-id) with one 3090 GPU. I saw that structures are optimised one by one, and I am curious is there a way to accelerate the optimisation process (e.g. in-batch or multi-GPU acceleration)?

Thanks

Hi -

If you’re running relaxations via main.py and specifying a config-yml then it will batch things with a batch size defined by ocp/gemnet-oc.yml at 7d1040bddc590dc70790493b802f0ef6dd968d95 · Open-Catalyst-Project/ocp · GitHub. You can modify this as necessary depending on the memory of your card.

The repo also supports multi-GPU acceleration via the following command -

python -u -m torch.distributed.launch --nproc_per_node=8 main.py --distributed --num-gpus 8 [...]

More details can be found here.

Thanks, Muhammed. I will have a try and hope everyone in the Meta OCP team is okay.