when TensorRT-LLM clearly wins: early impressions

[vectorport]

Thanks, that helps. I don't think your conclusion is far off. The biggest change for me was simplifying the stack instead of adding one more optimization layer.