some problem about speed of torch_sim compare with ase

[hongmoxian]

hello, I conducted some test comparing the optimization speed of ASE and Torch_sim (without autobatch) for a single POSCAR structure and observed similar performance. However, when enabling autobatch to optimize two structures simultaneously, I encountered a GPU memory error despite using an RTX 4090 Ti (24GB VRAM).

Environment:
GPU: NVIDIA RTX 4090 Ti (24GB VRAM)

ML model: Fairchem model

Input files: [Attach bellow]

Hope for your reply. Thank you!

[orionarcher]

Could you try manually defining a autobatcher as described in Understanding Autobatching? Set the memory scaler to n_atoms and the max memory scaler to something that you know will fit in memory.

I believe the problem here is that, in practice, max system size fluctuates slightly. When the autobatcher determines max system size it may be finding something very close to the actual maximum such that a later fluctuation forces OOM.

I am going to add a fix shortly that adds a bit of buffer to the autobatcher.

[hongmoxian]

Your suggestion works well, thank you!