By default, the ANI models are loaded with gradient required on parameters. This slows down ANI in inference mode. Should be a good idea to disable it in the Calculator class. Alternatively, this could be done in BuiltinModel.
My quick test of ASE LBFGS optimization shows ~25% speed improvement on CPU and 10% on GPU (tested on Google Colab).
By default, the ANI models are loaded with gradient required on parameters. This slows down ANI in inference mode. Should be a good idea to disable it in the Calculator class. Alternatively, this could be done in BuiltinModel. My quick test of ASE LBFGS optimization shows ~25% speed improvement on CPU and 10% on GPU (tested on Google Colab).