Batch inference

[qpessemesse]

I am wondering if batch inference can be used with MACE to efficiently handle large datasets of small molecules, it seems wasteful to use a full gpu node for molecules with ~50-100 atoms, I would like to optimize e.g. 10 molecules at the same time. Has this been implemented/done ?
Thx
Q.

[ilyes319]

hey @qpessemesse, torchsim handles that with mace here: https://github.com/TorchSim/torch-sim, so you can batch the md or geometry optimisation.

[ilyes319]

also alchemy does it https://github.com/NVIDIA/nvalchemi-toolkit

[qpessemesse]

Thanks a lot! Am I correct that neither of them currently have saddle-point optimizers ?

[ilyes319]

I dont think so, might be worth opening an issue or a PR! One hack you can do in mace is create a big system where you seperate each molecule by a lot of vacuum.

[ilyes319]

Depending on the optimiser it might influence the results or not.

[qpessemesse]

Will try these options and open issues as needed, many thanks!Le 21 mai 2026 à 19:13, Ilyes Batatia ***@***.***> a écrit : Depending on the optimiser it might influence the results or not. —Reply to this email directly, view it on GitHub, or unsubscribe.Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you were mentioned.Message ID: ***@***.***>