> the (meta provided) code to load the initial weights is completely unnecessarily making two copies of the weights in RAM simultaneously This is the kind of t

> the (meta provided) code to load the initial weights is completely unnecessarily making two copies of the weights in RAM simultaneously This is the kind of thing that the stable diffusion community optimized the shit out