VLLM on C.CUI's Log

VLLM on C.CUI's Loghttps://cuicaihao.github.io/tags/vllm/Recent content in VLLM on C.CUI's LogHugoen-AUSat, 08 Aug 2026 07:00:00 +1000Optimizing LLM Inference: From PagedAttention to AI Infrahttps://cuicaihao.github.io/posts/2026-08-08-optimizing-llm-inference-the-role-of-pagedattention-in-efficient-gpu-memory-management/Sat, 08 Aug 2026 07:00:00 +1000https://cuicaihao.github.io/posts/2026-08-08-optimizing-llm-inference-the-role-of-pagedattention-in-efficient-gpu-memory-management/Why do expensive GPUs remain underutilized during LLM inference? This post explores how vLLM’s PagedAttention leverages operating system virtual memory concepts, alongside continuous batching, RadixAttention, and Prefill-Decode disaggregation, to reshape modern large language model serving.