Inference

A source-grounded look at what's grown around vLLM's V1 engine since the canonical PagedAttention story: fault-tolerant engine cores, tiered KV offload, dual-batch overlap, async scheduling, and adaptive speculative decoding — with the real config flags for each.