{
  "type": "excalidraw",
  "version": 2,
  "source": "https://excalidraw.com",
  "elements": [
    {
      "id": "el1",
      "x": 120,
      "y": 8,
      "width": 514.8,
      "height": 27.5,
      "angle": 0,
      "strokeColor": "#1f2933",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1001,
      "version": 1,
      "versionNonce": 5001,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "Expert Tensor Parallel (ETP)：专家「内部」权重再切",
      "originalText": "Expert Tensor Parallel (ETP)：专家「内部」权重再切",
      "fontSize": 22,
      "fontFamily": 2,
      "textAlign": "left",
      "verticalAlign": "top",
      "containerId": null,
      "lineHeight": 1.25,
      "baseline": 22
    },
    {
      "id": "el2",
      "x": 70,
      "y": 40,
      "width": 483.59999999999997,
      "height": 16.25,
      "angle": 0,
      "strokeColor": "#868e96",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1002,
      "version": 1,
      "versionNonce": 5002,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "MoE 里 Attention 用 TP、Expert 用独立的 ETP（默认=TP，可单独设）；两条路径的 DP 各自推导",
      "originalText": "MoE 里 Attention 用 TP、Expert 用独立的 ETP（默认=TP，可单独设）；两条路径的 DP 各自推导",
      "fontSize": 13,
      "fontFamily": 2,
      "textAlign": "left",
      "verticalAlign": "top",
      "containerId": null,
      "lineHeight": 1.25,
      "baseline": 13
    },
    {
      "id": "el3",
      "x": 40,
      "y": 86,
      "width": 400,
      "height": 76,
      "angle": 0,
      "strokeColor": "#1971c2",
      "backgroundColor": "#a5d8ff",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": {
        "type": 3
      },
      "seed": 1003,
      "version": 1,
      "versionNonce": 5003,
      "isDeleted": false,
      "boundElements": [
        {
          "type": "text",
          "id": "el4"
        }
      ],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "rectangle"
    },
    {
      "id": "el4",
      "x": 45,
      "y": 99.625,
      "width": 390,
      "height": 48.75,
      "angle": 0,
      "strokeColor": "#1971c2",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1004,
      "version": 1,
      "versionNonce": 5004,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "Attention 层\n按 TP 切（如 TP=8）\nDP = world /(TP×CP×PP)",
      "originalText": "Attention 层\n按 TP 切（如 TP=8）\nDP = world /(TP×CP×PP)",
      "fontSize": 13,
      "fontFamily": 2,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": "el3",
      "lineHeight": 1.25,
      "baseline": 13
    },
    {
      "id": "el5",
      "x": 560,
      "y": 86,
      "width": 400,
      "height": 76,
      "angle": 0,
      "strokeColor": "#e8590c",
      "backgroundColor": "#ffd8a8",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": {
        "type": 3
      },
      "seed": 1005,
      "version": 1,
      "versionNonce": 5005,
      "isDeleted": false,
      "boundElements": [
        {
          "type": "text",
          "id": "el6"
        }
      ],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "rectangle"
    },
    {
      "id": "el6",
      "x": 565,
      "y": 101.5,
      "width": 390,
      "height": 45.0,
      "angle": 0,
      "strokeColor": "#e8590c",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1006,
      "version": 1,
      "versionNonce": 5006,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "Expert 层\n按 ETP 切（如 ETP=2，独立于 TP）\nexpert_DP = world /(ETP×EP×PP)",
      "originalText": "Expert 层\n按 ETP 切（如 ETP=2，独立于 TP）\nexpert_DP = world /(ETP×EP×PP)",
      "fontSize": 12,
      "fontFamily": 2,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": "el5",
      "lineHeight": 1.25,
      "baseline": 12
    },
    {
      "id": "el7",
      "x": 40,
      "y": 180,
      "width": 252.0,
      "height": 15.0,
      "angle": 0,
      "strokeColor": "#868e96",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1007,
      "version": 1,
      "versionNonce": 5007,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "为什么独立：专家 FFN 隐藏层往往较小，硬套 TP=8 会切得过碎、",
      "originalText": "为什么独立：专家 FFN 隐藏层往往较小，硬套 TP=8 会切得过碎、",
      "fontSize": 12,
      "fontFamily": 2,
      "textAlign": "left",
      "verticalAlign": "top",
      "containerId": null,
      "lineHeight": 1.25,
      "baseline": 12
    },
    {
      "id": "el8",
      "x": 40,
      "y": 200,
      "width": 273.59999999999997,
      "height": 15.0,
      "angle": 0,
      "strokeColor": "#868e96",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1008,
      "version": 1,
      "versionNonce": 5008,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "通信占比过大 → 让专家用更小的 ETP，Attention 仍用大 TP。",
      "originalText": "通信占比过大 → 让专家用更小的 ETP，Attention 仍用大 TP。",
      "fontSize": 12,
      "fontFamily": 2,
      "textAlign": "left",
      "verticalAlign": "top",
      "containerId": null,
      "lineHeight": 1.25,
      "baseline": 12
    },
    {
      "id": "el9",
      "x": 620,
      "y": 200,
      "width": 160,
      "height": 56,
      "angle": 0,
      "strokeColor": "#2f9e44",
      "backgroundColor": "#b2f2bb",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": {
        "type": 3
      },
      "seed": 1009,
      "version": 1,
      "versionNonce": 5009,
      "isDeleted": false,
      "boundElements": [
        {
          "type": "text",
          "id": "el10"
        }
      ],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "rectangle"
    },
    {
      "id": "el10",
      "x": 625,
      "y": 211.75,
      "width": 150,
      "height": 32.5,
      "angle": 0,
      "strokeColor": "#2f9e44",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1010,
      "version": 1,
      "versionNonce": 5010,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "Expert E0\n权重矩阵",
      "originalText": "Expert E0\n权重矩阵",
      "fontSize": 13,
      "fontFamily": 2,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": "el9",
      "lineHeight": 1.25,
      "baseline": 13
    },
    {
      "id": "el11",
      "x": 660,
      "y": 256,
      "width": 50,
      "height": 34,
      "angle": 0,
      "strokeColor": "#868e96",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": {
        "type": 2
      },
      "seed": 1011,
      "version": 1,
      "versionNonce": 5011,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "arrow",
      "points": [
        [
          0,
          0
        ],
        [
          -50,
          34
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "el12",
      "x": 740,
      "y": 256,
      "width": 60,
      "height": 34,
      "angle": 0,
      "strokeColor": "#868e96",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": {
        "type": 2
      },
      "seed": 1012,
      "version": 1,
      "versionNonce": 5012,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "arrow",
      "points": [
        [
          0,
          0
        ],
        [
          60,
          34
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "el13",
      "x": 540,
      "y": 290,
      "width": 150,
      "height": 58,
      "angle": 0,
      "strokeColor": "#e8590c",
      "backgroundColor": "#ffd8a8",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": {
        "type": 3
      },
      "seed": 1013,
      "version": 1,
      "versionNonce": 5013,
      "isDeleted": false,
      "boundElements": [
        {
          "type": "text",
          "id": "el14"
        }
      ],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "rectangle"
    },
    {
      "id": "el14",
      "x": 545,
      "y": 304.0,
      "width": 140,
      "height": 30.0,
      "angle": 0,
      "strokeColor": "#e8590c",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1014,
      "version": 1,
      "versionNonce": 5014,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "ETP rank0\nW 列片 0",
      "originalText": "ETP rank0\nW 列片 0",
      "fontSize": 12,
      "fontFamily": 2,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": "el13",
      "lineHeight": 1.25,
      "baseline": 12
    },
    {
      "id": "el15",
      "x": 720,
      "y": 290,
      "width": 150,
      "height": 58,
      "angle": 0,
      "strokeColor": "#e8590c",
      "backgroundColor": "#ffd8a8",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": {
        "type": 3
      },
      "seed": 1015,
      "version": 1,
      "versionNonce": 5015,
      "isDeleted": false,
      "boundElements": [
        {
          "type": "text",
          "id": "el16"
        }
      ],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "rectangle"
    },
    {
      "id": "el16",
      "x": 725,
      "y": 304.0,
      "width": 140,
      "height": 30.0,
      "angle": 0,
      "strokeColor": "#e8590c",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1016,
      "version": 1,
      "versionNonce": 5016,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "ETP rank1\nW 列片 1",
      "originalText": "ETP rank1\nW 列片 1",
      "fontSize": 12,
      "fontFamily": 2,
      "textAlign": "center",
      "verticalAlign": "middle",
      "containerId": "el15",
      "lineHeight": 1.25,
      "baseline": 12
    },
    {
      "id": "el17",
      "x": 690,
      "y": 319,
      "width": 30,
      "height": 0,
      "angle": 0,
      "strokeColor": "#2f9e44",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": {
        "type": 2
      },
      "seed": 1017,
      "version": 1,
      "versionNonce": 5017,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "arrow",
      "points": [
        [
          0,
          0
        ],
        [
          30,
          0
        ]
      ],
      "lastCommittedPoint": null,
      "startBinding": null,
      "endBinding": null,
      "startArrowhead": null,
      "endArrowhead": "arrow"
    },
    {
      "id": "el18",
      "x": 540,
      "y": 360,
      "width": 439.2,
      "height": 15.0,
      "angle": 0,
      "strokeColor": "#868e96",
      "backgroundColor": "transparent",
      "fillStyle": "solid",
      "strokeWidth": 2,
      "strokeStyle": "solid",
      "roughness": 1,
      "opacity": 100,
      "groupIds": [],
      "frameId": null,
      "roundness": null,
      "seed": 1018,
      "version": 1,
      "versionNonce": 5018,
      "isDeleted": false,
      "boundElements": [],
      "updated": 1700000000000,
      "link": null,
      "locked": false,
      "type": "text",
      "text": "AllReduce / ReduceScatter 合并（用 expt_tp 组，不是 attention 的 tp 组）",
      "originalText": "AllReduce / ReduceScatter 合并（用 expt_tp 组，不是 attention 的 tp 组）",
      "fontSize": 12,
      "fontFamily": 2,
      "textAlign": "left",
      "verticalAlign": "top",
      "containerId": null,
      "lineHeight": 1.25,
      "baseline": 12
    }
  ],
  "appState": {
    "gridSize": null,
    "viewBackgroundColor": "#ffffff"
  },
  "files": {}
}