[REQUEST] Ability to switch out weights when using deepspeed inference

[jihan-yin]

Is your feature request related to a problem? Please describe.
If I have multiple fine-tuned versions of the same base model where only a small number of weights are different, I'd like to be able to change out the weights quickly without relaunching deepspeed. Not sure if this is already possible, so marking this issue as a feature request.

Describe the solution you'd like
Have a straightforward and fast way to reload weights to specific modules within a pytorch model, that has already been prepared by deepspeed for multi-gpu single-node inference, across multiple processes.

Describe alternatives you've considered
If not using deepspeed for inference, it is quite simple to reload weights for a small portion of the model.

Additional context
I'm wondering if this is possible for models for which deepspeed inference does not support custom kernels (T5), and models that deepspeed inference does support custom kernels for (GPT)