[matthewolfe]

To echo the other replies, the tokenizer is definitely not the bottleneck. It just happens to be the first step in inference, so it's what I did first.