Hi!
I probably can’t help too much, but I found this post, maybe you can try it this way (creating a new test dataloader).
I ran the collab filtering notebook from the book, learn.get_preds() worked fine there, learn.predict(df.iloc[0]) gave me the same error as you experienced. This is with the default flattened MSE loss. I could not even get it to train with BCELoss, did that work for you?
Regarding this error:
RuntimeError: CUDA error: device-side assert triggered
I have tried in the past to switch to a non GPU environment. This way, the error messages are sometimes clearer than the generic CUDA error.