Speed up batch size discovery hundreds of times (to fit into your GPU RAM)

2 posts were merged into an existing topic: The “BS<=32” paper