{
  "title": "Reschedule unfinished requests",
  "caption": "The scheduler selects work. GPU kernels implement model forward, and K/V stays in memory. Unfinished requests return to scheduling. Model and GPU boxes show the same execution, not two serial runs. The author made this diagram. The playback is an example, and animation time is not measured runtime.",
  "alt": "Feedback flow: a scheduler selects work. GPU kernels implement model forward and update K/V. Logits supply input for sampling before unfinished work returns.",
  "sources": [
    {
      "path": "content/blog/2026-06-08_llm-inference-full-stack-systems/post.md",
      "sha256": "09bb02fb1815396601a7302f5924ed7b6b047a559ec5350a812ea9153627a653",
      "description": "Current article explanation, equations, assumptions and retained citations."
    },
    {
      "path": "research/dynamic-mechanisms-2026-10-21/before/content/blog/2026-06-08_llm-inference-full-stack-systems/illustrations/03-full-stack-feedback-algorithm.svg",
      "sha256": "8758f85f10f0c6c56a2566491a3a93c34bb7aaad82c1385c08b092bb51a64e87",
      "description": "Frozen prior illustration; exact pre-migration bytes retained."
    },
    {
      "path": "content/blog/2026-06-08_llm-inference-full-stack-systems/illustrations/source-notes.json",
      "description": "Existing article illustration provenance, retained unchanged."
    }
  ],
  "limitations": [
    "The playback is an example. Playback duration is not measured runtime."
  ]
}