KV-planner sizes LLM GPU deployments

This is a useful infrastructure tool because it attacks a real deployment pain point: the gap between datasheet math and runtime behavior.

• –Self-calibration is the strongest feature, since MBU/MFU defaults vary sharply across vLLM, Ollama, llama.cpp, consumer GPUs, and datacenter GPUs.
• –The MoE kernel-launch correction is a good reminder that active-parameter math alone misses scheduling overhead in modern inference stacks.
• –The validation story is promising but still thin; four GPUs and one MoE coefficient are not enough for production trust across MI300X, L40S, TensorRT-LLM, SGLang, or multi-node setups.
• –The broad interface surface, including CLI, GUI, REST, TUI, and MCP, makes it more than a spreadsheet replacement, but the project needs broader benchmark contributions to become a sizing reference.