distillation that keeps useful behavior with llama.cpp

[vramridge]

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