Managing Missing Adapters in Parallel

[armin-zd]

I am exploring a personalization model that utilizes one adapter for each user.

When performing inference, I process a batch containing samples from various users. For users with insufficient training data, no adapters are available, and the None active adapter should be used to return the base model's outputs.

To reduce latency, my ideal solution would involve a single forward pass using Parallel composition. However, this method does not support None adapters. Are there any alternative approaches you are aware of?

I attempted to create an "identity" adapter using PfeifferConfig(non_linearity="linear", reduction_factor=1, ...), but the resulting outputs differ from those of the base model.

Thanks!