[Gemma 4] Gemma4UnifiedForConditionalGeneration text-only inference produces degenerate output (token repetition collapse)
![[Gemma 4] Gemma4UnifiedForConditionalGeneration text-only inference produces degenerate output (token repetition collapse)](https://www.chat-gpts.plus/wp-content/uploads/2026/06/46531-3f809be1-768x403.jpg)
用户在使用 Hugging Face Transformers 库( transformers )加载 Google Gemma 4 指令微调模型(如 google/gemma-4-12B-it )进行纯文本推理时,模型输出退化为重复单一 token(例如 "111111111111")。该问题在








