A teammate wants to swap LayerNorm for BatchNorm in your decoder-only language model to share code with a vision project.

Training batches mix sequences of very different lengths, and inference serves one request at a time.