prefill/decode separation in practice with Ollama

[deltapoint17]

I'd approach it in a boring order. If you're on TensorRT-LLM, I'd confirm the version first and then reduce variables until the behavior is obvious.