Qwen2-7B-Instruct with TensorRT-LLM: consistently high tokens/SEC