Dear OCP team or Prof. or Dr. Mshuaibi:
OC20/22 datasets and pre-trianed models are valueable for catalyst study.
Just want to let you know the results of my trials of different S2EFS pre-trained models.
I focus on metal oxide so I only care about AI model trained on OC22 (+OC20)
When using fairchem.core pkg, I found 6 different pre-trianed model (5 GemNet** + 1 Equivformer-v2(31M/121M) ) from Pretrained FAIRChem models
Only [Equivformer-v2] works
[Equivformer-v2-31M] work in CPU and GPU mode
[Equivformer-v2-121M] work in CPU + ~120GB Mem, no matter on personal PC or HPC
About using [Equivformer-v2-121M] in GPU mode, I have tried (all use py3.11):
win-server2012-py3.11-Tesla K80 12GB
ubuntu20- GTX 3080Ti 12GB
HPC (compute canada) V100l 32 GB
HPC (compute canada) A100 40 GB
All failed with [out of mem in GPU],
From my tests results, I assume that the problem is in fairchem-core pkg.
I have attached the test file in pack.rar (a cif file, jupyter notebook running in CPU/GPU) https://drive.google.com/file/d/1sB24BaNVOgPEbKZ12y8s4wppTM0xe-ij/view?usp=drive_link
Thanks in advance for your assitance and advice.
Best,
Zhaoyang