CUDA update broke my throughput with TensorRT-LLM

[latentgarden]

Thanks, that helps. I ended up in about the same place using llama.cpp. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.