Tie queue to GPU (if there are multiple), or stop multiple instances of the webui from loading models twice.

[gxmikvid]

First of all, this is a HUGE request. So low priority by default.

I'd like to see multiple instances of the webui, started with different cuda device arguments, share the loaded models. (Or just a built-in way of starting multiple instances to make memory-sharing easier.)
Or if there are multiple devices available, a split queue system.

This idea is raw, and god knows I don't know how any of this works. But if someone who knows more than me can reply, telling me if it's even doable, I'd appreciate that.

Keep up the good work.