load model, compute a 1k token response (ie, do a thousand forward passes in sequence, one per token), load a different model, compute a response, I would expe

load model, compute a 1k token response (ie, do a thousand forward passes in sequence, one per token), load a different model, compute a response, I would expect the model loading to take basically zero percent of the time in the above workflow