CUDA update broke my throughput with TensorRT-LLM

[thetapath]

What worked for me was pretty simple. If you're on vLLM, I'd confirm the version first and then reduce variables until the behavior is obvious.