Identity Adapter

[armin-zd]

I am exploring a personalisation model that has one adapter per user. When there's enough data for a given user, an adapter is trained for that user, otherwise the default model weights are used (no adapters).

The problem is inference with Parallel. A batch contains samples from multiple users, some have adapters and some don't. However, it's not possible to extract default model weights with Parallel, eg. by using None adapter.

What is the best way to achieve this? I'm thinking of making a PR with an Identity Adapter.

Thanks!

[adapter-hub-bert]

This issue has been automatically marked as stale because it has been without activity for 90 days. This issue will be closed in 14 days unless you comment or remove the stale label.

[adapter-hub-bert]

This issue was closed because it was stale for 14 days without any activity.

[calpt]

Sorry for not responding to this issue! Passing None for forward pass without adapters sounds like a sensible addition especially for Parallel and BatchSplit blocks (functionality would have to be added here).

Re-opening as feature requests. Open for PR, otherwise we'll look into it :)