When using Gemma 4's thinking mode (enable_thinking=True) with a max_tokens budget in the range of 512–1024, the model sometimes returns a response containing only the <channel|> delimiter and n
python gemma thinking-mode inference token-budget 102 tokens