mym2358

Managed to fix it. It should be dim=1. The norm should be calculated along the "embedding" axis. Problem comes from embedding_i = embedding_b[seg_mask_i] which break the dims to get a...