vllm

A deep dive into how large language models actually run in production — why prefill is fast and decode is slow, how the KV cache works, sampling strategies like temperature and top-p, speculative decoding, and continuous batching with vLLM.