Fairchem and cuEquivariance

[husakm]

I was benchmarking MACE and Fairchem for aprox. the same model complexity. Speeds are similar. MACE speed can be significantly improved by the cuEquivariance library. Any planes of something similar for Fairchem ?

[rayg1234]

the UMA/Esen models doesn't use e3nn tensor products which is what cuEquivariance is based on, instead we use a variant of the simple so2 convs described in the original ESCN paper which is already fairly efficient compared to e3nn tensor produts. however, if someone wants to try rewriting UMA/Esen using e3nn/cuEquivariance, we'd be interested in the results :)

[husakm]

Unfortunately this is outside my know how :-) . I am an X-ray crystallographer trying to utilize the Fairchem engine for experimental crystals structures verification. I am just processing 250 000 entrees = speed is welcome ...