Fix Qwen3 thinking mode + increase max_new_tokens: inference.py 4afdc43 verified piyush-mk commited on Apr 25