working config for 32k context for Llama

[rocmcrate26]

I hit almost the same issue last night.

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