But for BLOOM inference -
which is a very large model - dynamic batching is essential to provide a decent experience for everyone.