google

/

gemma-4-31B-it

like 2.15k

Follow

Google 52.2k

The chat_template field is missing from tokenizer_config.json. The template exists as a separate chat_template.jinja file, but AutoTokenizer.from_pretrained() only reads from tokenizer_config.json. This causes apply_chat_template() to fail in transformers.js and other non-Python tooling.

Gemma 2 and Gemma 3 models include this field correctly. This PR embeds the existing chat_template.jinja content into tokenizer_config.json.

Same fix as:

• https://huggingface.co/google/gemma-4-E4B-it/discussions/21
• https://huggingface.co/google/gemma-4-E2B-it/discussions/8