google

/

gemma-4-12B-it-qat-w4a16-ct

like 45

Follow

Google 61.9k

The vision_config section is missing num_soft_tokens and model_patch_size fields that are present in the base google/gemma-4-12B-it checkpoint.

This causes vLLM to crash with:
AttributeError: Gemma4UnifiedVisionConfig object has no attribute num_soft_tokens

Fix: add num_soft_tokens=280 and model_patch_size=48 to match the base model.

Ref: https://github.com/vllm-project/vllm/issues/45039