I'm working on a personalisation problem that involves fine-tuning a model for each individual user. I'm experimenting with adapters, creating a unique adapter for every user.
I noticed that subsequent calls to add_adapter get increasingly slow and it practically halts after a few hundred iterations (Bert, various adapter types). I understand this might not be the intended usage, but are there any workarounds to allow scaling the number of adapters to thousands/millions?
I'm working on a personalisation problem that involves fine-tuning a model for each individual user. I'm experimenting with adapters, creating a unique adapter for every user.
I noticed that subsequent calls to
add_adapter
get increasingly slow and it practically halts after a few hundred iterations (Bert, various adapter types). I understand this might not be the intended usage, but are there any workarounds to allow scaling the number of adapters to thousands/millions?Many thanks!