{
  "type": "excalidraw",
  "version": 2,
  "source": "https://excalidraw.com",
  "elements": [
    {
      "id": "text-0001",
      "type": "text",
      "x": 80,
      "y": 40,
      "width": 1540,
      "height": 44,
      "angle": 0,
      "strokeColor": "#1e40af",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 19000,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "TurboQuant：论文算法与 vLLM 工程实现",
      "fontSize": 30,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "TurboQuant：论文算法与 vLLM 工程实现",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "text-0002",
      "type": "text",
      "x": 80,
      "y": 90,
      "width": 1540,
      "height": 32,
      "angle": 0,
      "strokeColor": "#64748b",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 19097,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "不要把论文里的 QJL 残差修正和当前 vLLM 的 KV-cache preset 混为一谈",
      "fontSize": 17,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "不要把论文里的 QJL 残差修正和当前 vLLM 的 KV-cache preset 混为一谈",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0003",
      "type": "arrow",
      "x": 90,
      "y": 132,
      "width": 1520,
      "height": 0,
      "angle": 0,
      "strokeColor": "#94a3b8",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "dashed",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 19194,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          1520,
          0
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": null
    },
    {
      "id": "rectangle-0004",
      "type": "rectangle",
      "x": 70,
      "y": 170,
      "width": 750,
      "height": 850,
      "angle": 0,
      "strokeColor": "#64748b",
      "backgroundColor": "#e2e8f0",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 25,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 19291,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "rectangle-0005",
      "type": "rectangle",
      "x": 880,
      "y": 170,
      "width": 750,
      "height": 850,
      "angle": 0,
      "strokeColor": "#64748b",
      "backgroundColor": "#e2e8f0",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 25,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 19388,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0006",
      "type": "text",
      "x": 100,
      "y": 195,
      "width": 690,
      "height": 38,
      "angle": 0,
      "strokeColor": "#1e40af",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 19485,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "论文 TurboQuant",
      "fontSize": 26,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "论文 TurboQuant",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "text-0007",
      "type": "text",
      "x": 910,
      "y": 195,
      "width": 690,
      "height": 38,
      "angle": 0,
      "strokeColor": "#1e40af",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 19582,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "vLLM TurboQuant KV-cache",
      "fontSize": 26,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "vLLM TurboQuant KV-cache",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "rectangle-0008",
      "type": "rectangle",
      "x": 135,
      "y": 260,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "#bfdbfe",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 19679,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0009",
      "type": "text",
      "x": 147,
      "y": 270,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 19776,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "输入向量 x\n来自每个 head 的 K 或 V",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "输入向量 x\n来自每个 head 的 K 或 V",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0010",
      "type": "arrow",
      "x": 445,
      "y": 342,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 19873,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0011",
      "type": "rectangle",
      "x": 135,
      "y": 375,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "#99f6e4",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 19970,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0012",
      "type": "text",
      "x": 147,
      "y": 385,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 20067,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "随机旋转 / Hadamard\n让坐标分布更集中、更接近独立",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "随机旋转 / Hadamard\n让坐标分布更集中、更接近独立",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0013",
      "type": "arrow",
      "x": 445,
      "y": 457,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 20164,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0014",
      "type": "rectangle",
      "x": 135,
      "y": 490,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "#bbf7d0",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 20261,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0015",
      "type": "text",
      "x": 147,
      "y": 500,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 20358,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "逐坐标 MSE 标量量化\n用少量 centroid 逼近旋转后坐标",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "逐坐标 MSE 标量量化\n用少量 centroid 逼近旋转后坐标",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0016",
      "type": "arrow",
      "x": 445,
      "y": 572,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 20455,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0017",
      "type": "rectangle",
      "x": 135,
      "y": 605,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#7c3aed",
      "backgroundColor": "#fed7aa",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 20552,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0018",
      "type": "text",
      "x": 147,
      "y": 615,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 20649,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "残差 r = x - x_hat\nMSE 最优不等于内积无偏",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "残差 r = x - x_hat\nMSE 最优不等于内积无偏",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0019",
      "type": "arrow",
      "x": 445,
      "y": 687,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 20746,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0020",
      "type": "rectangle",
      "x": 135,
      "y": 720,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#7c3aed",
      "backgroundColor": "#ddd6fe",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 20843,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0021",
      "type": "text",
      "x": 147,
      "y": 730,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 20940,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "1-bit QJL 残差修正\n用符号 sketch 修正注意力内积偏差",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "1-bit QJL 残差修正\n用符号 sketch 修正注意力内积偏差",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0022",
      "type": "arrow",
      "x": 445,
      "y": 802,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 21037,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0023",
      "type": "rectangle",
      "x": 135,
      "y": 835,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#7c3aed",
      "backgroundColor": "#fef3c7",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 21134,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0024",
      "type": "text",
      "x": 147,
      "y": 845,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 21231,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "目标\n3.5 bit/channel 质量中性\n2.5 bit/channel 小幅退化",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "目标\n3.5 bit/channel 质量中性\n2.5 bit/channel 小幅退化",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "rectangle-0025",
      "type": "rectangle",
      "x": 135,
      "y": 960,
      "width": 620,
      "height": 90,
      "angle": 0,
      "strokeColor": "#db2777",
      "backgroundColor": "#fbcfe8",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 21328,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0026",
      "type": "text",
      "x": 147,
      "y": 970,
      "width": 596,
      "height": 70,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 21425,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "关键直觉\nAttention score 是 q dot k，K 的方向和内积误差最敏感。\nTurboQuant 不是普通 per-tensor affine int4，而是先改变几何分布再量化。",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "关键直觉\nAttention score 是 q dot k，K 的方向和内积误差最敏感。\nTurboQuant 不是普通 per-tensor affine int4，而是先改变几何分布再量化。",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "rectangle-0027",
      "type": "rectangle",
      "x": 945,
      "y": 260,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "#bfdbfe",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 21522,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0028",
      "type": "text",
      "x": 957,
      "y": 270,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 21619,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "命令入口\n--kv-cache-dtype turboquant_4bit_nc 等 preset",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "命令入口\n--kv-cache-dtype turboquant_4bit_nc 等 preset",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0029",
      "type": "arrow",
      "x": 1255,
      "y": 342,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 21716,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0030",
      "type": "rectangle",
      "x": 945,
      "y": 375,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "#99f6e4",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 21813,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0031",
      "type": "text",
      "x": 957,
      "y": 385,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 21910,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "Keys\nFP8 或 Hadamard + Lloyd-Max MSE 3/4 bit",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "Keys\nFP8 或 Hadamard + Lloyd-Max MSE 3/4 bit",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0032",
      "type": "arrow",
      "x": 1255,
      "y": 457,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 22007,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0033",
      "type": "rectangle",
      "x": 945,
      "y": 490,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "#bbf7d0",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 22104,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0034",
      "type": "text",
      "x": 957,
      "y": 500,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 22201,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "Values\n3/4 bit uniform quantization",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "Values\n3/4 bit uniform quantization",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0035",
      "type": "arrow",
      "x": 1255,
      "y": 572,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 22298,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0036",
      "type": "rectangle",
      "x": 945,
      "y": 605,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#dc2626",
      "backgroundColor": "#fed7aa",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 22395,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0037",
      "type": "text",
      "x": 957,
      "y": 615,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 22492,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "NC\nnorm correction 修正 centroid norm distortion",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "NC\nnorm correction 修正 centroid norm distortion",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0038",
      "type": "arrow",
      "x": 1255,
      "y": 687,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 22589,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0039",
      "type": "rectangle",
      "x": 945,
      "y": 720,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#dc2626",
      "backgroundColor": "#fecaca",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 22686,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0040",
      "type": "text",
      "x": 957,
      "y": 730,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 22783,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "当前实现取舍\nQJL 故意省略，避免 softmax 方差放大",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "当前实现取舍\nQJL 故意省略，避免 softmax 方差放大",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0041",
      "type": "arrow",
      "x": 1255,
      "y": 802,
      "width": 0,
      "height": 33,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 22880,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          0,
          33
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "rectangle-0042",
      "type": "rectangle",
      "x": 945,
      "y": 835,
      "width": 620,
      "height": 82,
      "angle": 0,
      "strokeColor": "#dc2626",
      "backgroundColor": "#fef3c7",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "roundness": {
        "type": 3
      },
      "seed": 22977,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false
    },
    {
      "id": "text-0043",
      "type": "text",
      "x": 957,
      "y": 845,
      "width": 596,
      "height": 62,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 23074,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "工程结论\nFP8 默认更稳；4bit_nc 适合显存压力场景\nk3v4_nc 和 3bit_nc 要谨慎验证",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "工程结论\nFP8 默认更稳；4bit_nc 适合显存压力场景\nk3v4_nc 和 3bit_nc 要谨慎验证",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "arrow-0044",
      "type": "arrow",
      "x": 790,
      "y": 600,
      "width": 120,
      "height": 0,
      "angle": 0,
      "strokeColor": "#64748b",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "dashed",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 23171,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "points": [
        [
          0,
          0
        ],
        [
          120,
          0
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "text-0045",
      "type": "text",
      "x": 805,
      "y": 525,
      "width": 220,
      "height": 90,
      "angle": 0,
      "strokeColor": "#64748b",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 23268,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "同名不等于同实现\n读论文看算法保证\n跑服务看 kernel 和 preset",
      "fontSize": 18,
      "fontFamily": 5,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": null,
      "originalText": "同名不等于同实现\n读论文看算法保证\n跑服务看 kernel 和 preset",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "practice-interview-panel-01-turboquant-pipeline-bg",
      "type": "rectangle",
      "x": 80,
      "y": 1130.0,
      "width": 1550.0,
      "height": 300,
      "angle": 0,
      "strokeColor": "#64748b",
      "backgroundColor": "#e2e8f0",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 70,
      "groupIds": [],
      "seed": 93001,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "roundness": {
        "type": 3
      }
    },
    {
      "id": "practice-interview-panel-01-turboquant-pipeline-title",
      "type": "text",
      "x": 108,
      "y": 1152.0,
      "width": 1494.0,
      "height": 38,
      "angle": 0,
      "strokeColor": "#1e40af",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 91001,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "练习题与面试题补充：KV cache 量化",
      "fontSize": 24,
      "fontFamily": 5,
      "textAlign": "left",
      "verticalAlign": "top",
      "containerId": null,
      "originalText": "练习题与面试题补充：KV cache 量化",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "practice-interview-panel-01-turboquant-pipeline-practice-box",
      "type": "rectangle",
      "x": 108,
      "y": 1208.0,
      "width": 733.0,
      "height": 190,
      "angle": 0,
      "strokeColor": "#2563eb",
      "backgroundColor": "#bfdbfe",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 93002,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "roundness": {
        "type": 3
      }
    },
    {
      "id": "practice-interview-panel-01-turboquant-pipeline-practice-text",
      "type": "text",
      "x": 130,
      "y": 1228.0,
      "width": 710.0,
      "height": 150,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 91001,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "练习题\n1. 计算不同 batch/context/kv_heads 下的 KV cache 显存\n2. 对比 FP8、INT4、2bit 在显存、质量、kernel 复杂度上的取舍\n3. 设计一套长上下文 needle 评测协议验证量化退化",
      "fontSize": 17,
      "fontFamily": 5,
      "textAlign": "left",
      "verticalAlign": "top",
      "containerId": null,
      "originalText": "练习题\n1. 计算不同 batch/context/kv_heads 下的 KV cache 显存\n2. 对比 FP8、INT4、2bit 在显存、质量、kernel 复杂度上的取舍\n3. 设计一套长上下文 needle 评测协议验证量化退化",
      "autoResize": false,
      "lineHeight": 1.25
    },
    {
      "id": "practice-interview-panel-01-turboquant-pipeline-interview-box",
      "type": "rectangle",
      "x": 869.0,
      "y": 1208.0,
      "width": 733.0,
      "height": 190,
      "angle": 0,
      "strokeColor": "#7c3aed",
      "backgroundColor": "#ddd6fe",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 93003,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "roundness": {
        "type": 3
      }
    },
    {
      "id": "practice-interview-panel-01-turboquant-pipeline-interview-text",
      "type": "text",
      "x": 891.0,
      "y": 1228.0,
      "width": 710.0,
      "height": 150,
      "angle": 0,
      "strokeColor": "#243042",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 1,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "seed": 91001,
      "version": 1,
      "isDeleted": false,
      "boundElements": null,
      "updated": 1,
      "link": null,
      "locked": false,
      "text": "面试题\n1. 为什么 K 和 V 的量化敏感性不同\n2. FP8 为什么常是工程默认首选而不是最低 bit\n3. TurboQuant 的旋转和 inner-product bias 目标解决什么问题",
      "fontSize": 17,
      "fontFamily": 5,
      "textAlign": "left",
      "verticalAlign": "top",
      "containerId": null,
      "originalText": "面试题\n1. 为什么 K 和 V 的量化敏感性不同\n2. FP8 为什么常是工程默认首选而不是最低 bit\n3. TurboQuant 的旋转和 inner-product bias 目标解决什么问题",
      "autoResize": false,
      "lineHeight": 1.25
    }
  ],
  "appState": {
    "gridSize": null,
    "viewBackgroundColor": "#ffffff"
  },
  "files": {}
}