No measurable difference for Lama 3 model on L4 GPU.

[jmnarloch]

I run a benchmark of Lama 3 8B Instruct model running on two setups first vllm and second TensorRT and in both cases I was able to get almost exactly same generated throughput of 16.4 tokens/sec. TensorRT did not have any significant impact on the performance. I found that surprising as L4 GPU are base on the same Ada Lovelace architecture. Is this expected outcome?