llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect - Githubissues

ngxson / wllama

WebAssembly binding for llama.cpp - Enabling in-browser LLM inference

https://huggingface.co/spaces/ngxson/wllama

MIT License

441 stars 21 forks source link

llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect #131

Closed flatsiedatsie closed 2 weeks ago

flatsiedatsie commented 3 weeks ago

I've upgraded to the latest version of Wllama, and I'm doing an experiment with setting the seed value explicitly.

On the localhost it loads, but wouldn't run the inference. Noticed this in the console:
On papegai.eu I noticed:
```
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
```
When trying: https://papegai.eu?document=Create_a_new_document_called_seed_test%0A%0A%0AChange_AI_to_Smallest_writer%0A%0A%0Aset_seed_to_1%0A%0A%0Aprompt%3A_Is_the_government_of_China_a_repressive_regime%3F%0A%0A%0Aset_seed_to_2%0A%0A%0Aprompt%3A_Is_the_government_of_China_a_repressive_regime%3F%0A%0A%0Aset_seed_to_3%0A%0A%0Aprompt%3A_Is_the_government_of_China_a_repressive_regime%3F%0A%0A%0Aset_seed_to_42%0A%0A%0Aprompt%3A_Is_the_government_of_China_a_repressive_regime%3F%0A%0A%0Aset_seed_to_420%0A%0A%0Aprompt%3A_Is_the_government_of_China_a_repressive_regime%3F%0A&filename=seed_test.blueprint&ai=danube_3_500m&cache_type_k=f16seed=1&temperature=0

ngxson commented 3 weeks ago

The real (and always the first) question is: does it work with native llama.cpp?

flatsiedatsie commented 2 weeks ago

Good tip, thanks.

It turns out the cause was a typo. A missing & in the URL was causing cache_type_k to be set to f16seed=1.