Have you done preliminary numbers on replacing tokenizer on, say, llama-server?
Added numbers here: https://news.ycombinator.com/item?id=49015014
Running the numbers now
Added numbers here: https://news.ycombinator.com/item?id=49015014