Think Before You Grid-Search: Floor-First Triage for LLM Serving

This paper proposes Floor First, a residual-driven workflow for optimizing LLM serving before launching expensive profiling or broad configuration searches. Each decode step is represented as a five-dimensional resource vector: HBM bytes, FLOPs, network bytes, network messages, and KV capacity. Summing each resource and taking the maximum yields optimistic and pessimistic floors; a measured result inside that interval reveals overlap quality. In a case study of a DeepSeek-V3.2-style 671B MoE/MLA model on 16 NVIDIA H20 GPUs, TP16 is favored for single-stream latency, while EP16+DP-attention offers roughly an order-of-magnitude larger capacity wall.