preprocessor_config.json is missing?

#19
by matthewgaddis-streamline - opened

when I try run

import { AutoProcessor } from "@huggingface/transformers";

export const processor = await AutoProcessor.from_pretrained(
  "google/functiongemma-270m-it"
);

I receive:

Error: Could not locate file: "https://huggingface.co/google/functiongemma-270m-it/resolve/main/preprocessor_config.json".

Hi @matthewgaddis-streamline ,

This is expected because functiongemma is specifically optimised exclusively for text-to-text function calling and reasoning, meaning the vision stack artifacts, specifically preprocessor_config.json, were intentionally omitted from the Hugging Face repository. AutoProcessor in Transformers.js fails here because it strictly enforces the presence of that configuration to initialise a full multimodal pipeline. you can try switch to AutoTokenizer, this will correctly latch onto the existing tokenizer.json and tokenizer_config.json allowing you to tokenise your text prompts and schema definitions without triggering the request for the nonexistent image processing configuration.

Thank you!

Sign up or log in to comment