About model size

[DBpackage]

Hi, During train your gpcr model with BindingDB dataset, I got too many CUDA OOM Error.

Even I set the batch size as minum value(1), still the model needs more than 24GBs CUDA memories.

I could only run the model when I'm reduce default model n_heads parameter values dramatically.
(--n_heads :8 --> 2, then the model uses 20GBs CUDA memories)

Is that abnormal situation or not?

Best regards,