KV-Centric LLM Serving: vLLM, SGLang, and Disaggregated Attention July 5, 2026 AI ENGINEERINGLARGE LANGUAGE MODELS vllmsglangpaged attentionradix attentionkv cachedisaggregated servingprefill decode The more I look at LLM serving, the more it feels like the main object is not the request, the model, or even the GPU.