[szopa]

How I wish this post had appeared a few days earlier... I am writing on my own library for some agent experiments (in go, to make my life more interesting I guess), and knowing the number of tokens is important to implement a token buffer memory (as you approach the model's context window size, you prune enough messages from the beginning of the conversation that the whole thing keeps some given size, in tokens). While there's a nice native library in go for OpenAI models (https://github.com/tiktoken-go/tokenizer), the only library I found for Hugging Face models (and Claude, they published their tokenizer spec in the same JSON format) calls into HF's Rust implementation, which makes it challenging as a dependency in Go. What is more, any tokenizer needs to keep some representation of its vocabulary in memory. So, in the end I removed the true tokenizers, and ended up using an approximate version (just split it in on spaces and multiply by a factor I determined experimentally for the models I use using the real tokenizer, with a little extra for safety). If it turns out someone needs the real thing they can always provide their own token counter). I was actually rather happy with this result: I have less dependencies, and use less memory. But to get there I needed to do a deep dive too understand BPE tokenizers :)

(The library, if anyone is interested: https://github.com/ryszard/agency.)

[Ambix]

You might reuse simple LLaMA tokenizer right in your Go code, look there:

https://github.com/gotzmann/llama.go/blob/8cc54ca81e6bfbce25...