nvidia

/

Gemma-4-31B-IT-NVFP4

Like 588

Follow

NVIDIA 74.3k

Syncs tokenizer_config.json with upstream google/gemma-4-31B-it.

What

Adds the response_template block introduced upstream in
google/gemma-4-31B-it@842da37 ("Add response_template to
tokenizer_config.json"). It declares structured
output parsing for content, thinking and tool_calls, replacing
the need to rely on the x-regex fallback.

Why

This repo's tokenizer_config.json has not been updated since April
and predates the upstream change. Downstream tooling that reads
response_template (transformers parse_response, and serving stacks
that key off it for tool-call extraction) currently falls back to
regex-based parsing, which is more brittle for multi-turn tool calls.

Scope

• Only tokenizer_config.json is touched.
• Purely additive: the response_template key is new, no existing
  keys are modified or removed.
• No change to tokenizer.json, vocabulary, added tokens or special
  tokens — quantized weights and the embedding matrix are unaffected.
• Upstream's README changes in the same commit are not included here,
  as they document the transformers code path rather than serving.