It seems that, in your situation, you do not have any metadata features for your users. If you don’t provide user features to LightFM, LightFM will create indicator variable features for them at fit time.
Given that your users have only indicator variables, recommendations can’t be predicted for your users unless their interaction data is present at fit time. Having the interactions present at fit-time allows LightFM to learn an embedding for each user’s indicator variable.
Interaction data is not consumed at predict time — it is consumed when fitting in order to learn the embeddings.
Hope that helps!