Anyone have the tokenizer?

#1
by jackboot - opened

Can't use HF samplers without it.

It's certainly possible to recreate it from the gguf, but it's not something that AFAIK has been done before.

Don't you have it saved from the conversion? I uploaded the 57b one but for some reason that GGUF didn't work. Didn't get to check if they are the same.

edit: I was able to adapt the 57b tokenizer and get it working. It had bad EOS token and no bos token. https://huggingface.co/jackboot/quill-57b-tokenizer/tree/main

Unfortunately I don't have a copy - the transformers model is converted immediately to an unquantized gguf after download and then deleted. The unquantized source gguf survives quite a bit longer.

mradermacher pinned discussion

So just to be clear, the 57b that got uploaded is broken or just not supported by llama.cpp?

llama complained about a missing tensor during imatrix generation, which is either a bug in llama, a shortcoming in the conversion script not supporting this model or a bug in the model. That's why I deleted it.

Sign up or log in to comment