nvidia

/

diffusiongemma-26B-A4B-it-NVFP4

like 109

Follow

NVIDIA 63.9k

Syncs chat_template.jinja with upstream google/diffusiongemma-26B-A4B-it @ f7f5b7f ("fix: chat template - null handling, reasoning preservation, turn-tag balance, input validation"), which this repo has not picked up. The file is upstream's current version verbatim.

Why

Rendering six fixtures through both templates (via AutoTokenizer.apply_chat_template from this repo):

The headline: standard multi-turn tool calling crashes on the current template (assistant tool_calls messages carry content: None, which the old template concatenates). Upstream #23's null handling fixes it. The image_url/input_audio type aliases also fix silent multimodal content loss.

Notes for maintainers

• Your PR #2 change is preserved. This repo removed the empty-thought prefill for non-thinking mode; upstream's rewrite also removed it (the only remaining enable_thinking | default(false) is variable setup, and the one thought prefill left is gated on enable_thinking after a tool response). Syncing does not reintroduce what #2 removed.
• String tool-arguments become an explicit error. The old tolerant elif function['arguments'] is string branch is gone upstream — but it was unreachable anyway (the null-content crash fired first), so nothing that works today stops working; the failure just becomes actionable. If you want to accept OpenAI-style string arguments as-is, that branch would need to be carried as a deliberate downstream delta.
• Rendered prompts change for consecutive-assistant and trailing-tool-response cases (turn-tag balancing, and a new thinking-gated generation prompt after tool responses). If the README's published accuracy tables were measured with the old template, you may want to re-verify.
• Scope: only chat_template.jinja. tokenizer_config.json (2.74 kB) has no chat_template key, so this file is authoritative. NVFP4 quantization touches weights only — no interaction.

…<|turn>user
Weather in Oslo?<turn|>
<|turn>model
<|tool_call>call:get_weather{city:<|"|>Oslo<|"|>}<tool_call|><|tool_response>response:unknown{value:<|"|>12C, rain<|"|>}<tool_response|>

 <bos><|turn>user
 Two answers please<turn|>
 <|turn>model
-First.<turn|>
-Second.<turn|>
+First.Second.<turn|>
 <|turn>user
 Thanks<turn|>

 <bos><|turn>user
-Describe:<turn|>
+Describe:<|image|><turn|>
 <|turn>model