{
  "title": "Commit the accepted prefix, then add the correction",
  "caption": "The diagram shows only one rejection branch. It does not show acceptance probabilities, correction sampling or the all-accepted branch.",
  "alt": "Example: after verification accepts d1 and d2 of four draft tokens but rejects d3 and discards d4, the context contains d1, d2 and correction c3.",
  "sources": [
    {
      "file": "content/blog/2026-06-11_life-of-a-request-vllm/post.md",
      "description": "Mechanism and scope from the current article; concrete requests/tokens are illustrative."
    },
    {
      "file": "/Users/jiwenbo/Documents/GitHub/Blog_Illustration_Skill/examples/animated-flow-2026-10-16/02-speculative-decoding/diagram.svg",
      "sha256": "1852283430a294e2358780f0700fd53ca3eea4a8f670d401c777d7871fa2902d",
      "description": "Skill source illustration; imported, not an observed runtime trace."
    }
  ],
  "limitations": [
    "Playback time helps show the example. It is not measured latency or throughput.",
    "This is a constructed execution trace. It is not a captured vLLM serving run.",
    "Playback timing is an example, not measured runtime. The original SVG from the Skill has no changes.",
    "This trace does not show acceptance/correction probabilities or the all-accepted branch. The actual speculative-sampling rule is necessary for distributional correctness."
  ]
}