Hotswap speaker weights during inference with same base model?

[michaellin99999]

Hi,
I have finetuned two versions of ESPNet myself. one with a mans voice and one with a womans, the base model is the same based on VITS.

Is there a way where I can hotswap weights during runtime?
the experience I am looking for is I only need to host 1 model. and can simulate a man, woman talking with each other jsut by hotswapping the weights during inference? preferably less than 2 second latency. I am currently running on a 4090 GPU.

[sw005320]

How about loading two models and preparing two instances of Text2Speech with these different models instead of swapping the weights?

[michaellin99999]

I'm looking into hot swap during inference considering scale. hosting one model for every user is resource intensive.

[sw005320]

OK.
We don't prepare such a function, and I recommend implementing it yourself.