Kv cache quantization: ignorance, or malice?

I run Qwen-3.6 27B FP8 on vllm for long-horizon agentic coding harness workloads with high context window and concurrent sub-agents. On two 3090s that aren’t used for anything else, it seems reasonable to expect a good balance between speed and reliability. I want to bring up a particular point of contention regarding this optimization process. I have extensive software engineering background but am relatively new to this so feel free to correct me if I’m not on the right track. It seems like c