{
 "date": "2026-10-21",
 "purpose": "Test whether the role path's combine changes are AFD-specific by applying them to Masked DeepEP's generic instantiation (scratch audit build: AFDEP_AUDIT_GENERIC passes upstream expected top-k ceil(k/R), AFDEP_COMBINE_MAX_UNROLL and AFDEP_AUDIT_SPLIT to the generic combine; not shipped)",
 "scope": "4 x H100, 2 Attention + 2 FFN, E128 top-k 8, capacity 2048, BF16, graphs, 24 SMs, random routing; three fresh-process repeats",
 "arms": {
  "masked": "upstream Masked",
  "masked_u2": "Masked + unroll 2",
  "masked_split": "Masked + split combine (unroll 4)",
  "masked_u2_split": "Masked + unroll 2 + split combine",
  "afd": "role path at 0d33a60~ (E=ceil(k/F)=4, unroll 2, split)",
  "afd_e2": "role path with upstream E=2 (unroll 2, split)",
  "afd_nosplit": "role path without split"
 },
 "screens": [
  {
   "hidden": 2048,
   "native_source_sha256": "a572b48348508ceed0c13f6408c2ccb4fe8af9dac83b161b414ad4ebbaf2d8a8",
   "processes": [
    {
     "order": [
      "masked",
      "masked_u2",
      "masked_split",
      "masked_u2_split",
      "afd",
      "afd_e2",
      "afd_nosplit"
     ],
     "median_us": {
      "masked": {
       "1": 31.369119882583618,
       "8": 35.53928017616272,
       "32": 35.60719966888428,
       "128": 43.9302396774292,
       "512": 78.72639894485474,
       "2048": 200.90999603271484
      },
      "masked_u2": {
       "1": 31.082559823989868,
       "8": 34.046560525894165,
       "32": 35.25295972824097,
       "128": 44.077041149139404,
       "512": 76.60536050796509,
       "2048": 196.8052864074707
      },
      "masked_split": {
       "1": 29.58392024040222,
       "8": 32.01583981513977,
       "32": 33.04207921028137,
       "128": 38.870240449905396,
       "512": 79.85680103302002,
       "2048": 202.27696418762207
      },
      "masked_u2_split": {
       "1": 28.224400281906128,
       "8": 29.006799459457397,
       "32": 30.224720239639282,
       "128": 38.10760021209717,
       "512": 66.71648025512695,
       "2048": 183.4384822845459
      },
      "afd": {
       "1": 19.55743968486786,
       "8": 20.874799489974976,
       "32": 22.038480043411255,
       "128": 30.673359632492065,
       "512": 67.77400016784668,
       "2048": 185.41728019714355
      },
      "afd_e2": {
       "1": 19.522719979286194,
       "8": 21.02735996246338,
       "32": 22.02280044555664,
       "128": 30.440080165863037,
       "512": 70.3217601776123,
       "2048": 190.0320816040039
      },
      "afd_nosplit": {
       "1": 22.237999439239502,
       "8": 25.81112027168274,
       "32": 26.466879844665527,
       "128": 34.72439885139465,
       "512": 66.90976142883301,
       "2048": 185.4305648803711
      }
     }
    },
    {
     "order": [
      "masked_u2",
      "masked_split",
      "masked_u2_split",
      "afd",
      "afd_e2",
      "afd_nosplit",
      "masked"
     ],
     "median_us": {
      "masked_u2": {
       "1": 31.17616057395935,
       "8": 34.88391995429993,
       "32": 35.40152072906494,
       "128": 43.576719760894775,
       "512": 76.51576042175293,
       "2048": 196.72719955444336
      },
      "masked_split": {
       "1": 29.2739999294281,
       "8": 31.79800033569336,
       "32": 32.395039796829224,
       "128": 38.68551969528198,
       "512": 79.01983976364136,
       "2048": 201.69272422790527
      },
      "masked_u2_split": {
       "1": 27.944399118423462,
       "8": 29.104080200195312,
       "32": 30.22960066795349,
       "128": 38.02824020385742,
       "512": 66.71560049057007,
       "2048": 183.18463325500488
      },
      "afd": {
       "1": 19.563039541244507,
       "8": 21.248639822006226,
       "32": 22.239279747009277,
       "128": 30.993599891662598,
       "512": 67.47143983840942,
       "2048": 185.5685520172119
      },
      "afd_e2": {
       "1": 19.613119959831238,
       "8": 21.516720056533813,
       "32": 22.10279941558838,
       "128": 30.98047971725464,
       "512": 70.49175977706909,
       "2048": 190.17224311828613
      },
      "afd_nosplit": {
       "1": 22.4072003364563,
       "8": 25.84015965461731,
       "32": 26.96816086769104,
       "128": 35.19775986671448,
       "512": 67.19559907913208,
       "2048": 185.28928756713867
      },
      "masked": {
       "1": 31.635520458221436,
       "8": 35.501999855041504,
       "32": 35.72167992591858,
       "128": 43.95648002624512,
       "512": 78.77135992050171,
       "2048": 201.1023235321045
      }
     }
    },
    {
     "order": [
      "masked_split",
      "masked_u2_split",
      "afd",
      "afd_e2",
      "afd_nosplit",
      "masked",
      "masked_u2"
     ],
     "median_us": {
      "masked_split": {
       "1": 29.419760704040527,
       "8": 31.404320001602173,
       "32": 32.428959608078,
       "128": 38.853600025177,
       "512": 79.197678565979,
       "2048": 201.40416145324707
      },
      "masked_u2_split": {
       "1": 27.907999753952026,
       "8": 29.53160047531128,
       "32": 29.9452805519104,
       "128": 38.065359592437744,
       "512": 66.81895971298218,
       "2048": 183.41888427734375
      },
      "afd": {
       "1": 19.721839427947998,
       "8": 21.087040901184082,
       "32": 22.466319799423218,
       "128": 31.324080228805542,
       "512": 68.03783893585205,
       "2048": 185.69600105285645
      },
      "afd_e2": {
       "1": 19.379120469093323,
       "8": 21.017600297927856,
       "32": 22.036880254745483,
       "128": 30.81704020500183,
       "512": 70.51271915435791,
       "2048": 190.23239135742188
      },
      "afd_nosplit": {
       "1": 22.270320653915405,
       "8": 25.56615948677063,
       "32": 26.8338406085968,
       "128": 34.55431938171387,
       "512": 67.16423988342285,
       "2048": 185.79872131347656
      },
      "masked": {
       "1": 31.22904062271118,
       "8": 35.57431936264038,
       "32": 35.985599756240845,
       "128": 44.4596791267395,
       "512": 78.86216163635254,
       "2048": 201.173677444458
      },
      "masked_u2": {
       "1": 31.326000690460205,
       "8": 35.11647939682007,
       "32": 35.256160497665405,
       "128": 43.750479221343994,
       "512": 76.61695957183838,
       "2048": 196.31335258483887
      }
     }
    }
   ],
   "table": [
    {
     "tokens": 1,
     "masked": {
      "median_us": 31.37,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 31.18,
      "geomean_change_pct": -0.69,
      "repeats_faster": 2
     },
     "masked_split": {
      "median_us": 29.42,
      "geomean_change_pct": -6.32,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 27.94,
      "geomean_change_pct": -10.78,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 19.56,
      "geomean_change_pct": -37.56,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 19.52,
      "geomean_change_pct": -37.9,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 22.27,
      "geomean_change_pct": -28.99,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 8,
     "masked": {
      "median_us": 35.54,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 34.88,
      "geomean_change_pct": -2.42,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 31.8,
      "geomean_change_pct": -10.69,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 29.1,
      "geomean_change_pct": -17.8,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 21.09,
      "geomean_change_pct": -40.71,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 21.03,
      "geomean_change_pct": -40.39,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 25.81,
      "geomean_change_pct": -27.57,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 32,
     "masked": {
      "median_us": 35.72,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 35.26,
      "geomean_change_pct": -1.31,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 32.43,
      "geomean_change_pct": -8.81,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 30.22,
      "geomean_change_pct": -15.76,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 22.24,
      "geomean_change_pct": -37.81,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 22.04,
      "geomean_change_pct": -38.35,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 26.83,
      "geomean_change_pct": -25.2,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 128,
     "masked": {
      "median_us": 43.96,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 43.75,
      "geomean_change_pct": -0.71,
      "repeats_faster": 2
     },
     "masked_split": {
      "median_us": 38.85,
      "geomean_change_pct": -12.04,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 38.07,
      "geomean_change_pct": -13.71,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 30.99,
      "geomean_change_pct": -29.74,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 30.82,
      "geomean_change_pct": -30.31,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 34.72,
      "geomean_change_pct": -21.06,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 512,
     "masked": {
      "median_us": 78.77,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 76.61,
      "geomean_change_pct": -2.8,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 79.2,
      "geomean_change_pct": 0.72,
      "repeats_faster": 0
     },
     "masked_u2_split": {
      "median_us": 66.72,
      "geomean_change_pct": -15.28,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 67.77,
      "geomean_change_pct": -13.99,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 70.49,
      "geomean_change_pct": -10.59,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 67.16,
      "geomean_change_pct": -14.85,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 2048,
     "masked": {
      "median_us": 201.1,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 196.73,
      "geomean_change_pct": -2.21,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 201.69,
      "geomean_change_pct": 0.36,
      "repeats_faster": 0
     },
     "masked_u2_split": {
      "median_us": 183.42,
      "geomean_change_pct": -8.81,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 185.57,
      "geomean_change_pct": -7.71,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 190.17,
      "geomean_change_pct": -5.43,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 185.43,
      "geomean_change_pct": -7.74,
      "repeats_faster": 3
     }
    }
   ]
  },
  {
   "hidden": 4096,
   "native_source_sha256": "a572b48348508ceed0c13f6408c2ccb4fe8af9dac83b161b414ad4ebbaf2d8a8",
   "processes": [
    {
     "order": [
      "masked",
      "masked_u2",
      "masked_split",
      "masked_u2_split",
      "afd",
      "afd_e2",
      "afd_nosplit"
     ],
     "median_us": {
      "masked": {
       "1": 38.136080503463745,
       "8": 45.95655918121338,
       "32": 54.86511945724487,
       "128": 68.85159969329834,
       "512": 136.05096340179443,
       "2048": 392.89920806884766
      },
      "masked_u2": {
       "1": 35.079439878463745,
       "8": 41.077919006347656,
       "32": 45.29720067977905,
       "128": 55.17735958099365,
       "512": 111.27863883972168,
       "2048": 341.3617515563965
      },
      "masked_split": {
       "1": 31.041680574417114,
       "8": 34.07984018325806,
       "32": 37.56935954093933,
       "128": 56.06775999069214,
       "512": 136.27103805541992,
       "2048": 387.10880279541016
      },
      "masked_u2_split": {
       "1": 28.75048041343689,
       "8": 30.44592022895813,
       "32": 32.45944023132324,
       "128": 46.65247917175293,
       "512": 111.59272193908691,
       "2048": 341.76416397094727
      },
      "afd": {
       "1": 20.561760663986206,
       "8": 22.533040046691895,
       "32": 24.517040252685547,
       "128": 44.70815896987915,
       "512": 104.09303665161133,
       "2048": 332.8269577026367
      },
      "afd_e2": {
       "1": 20.0547194480896,
       "8": 21.807680130004883,
       "32": 23.550400733947754,
       "128": 39.825520515441895,
       "512": 104.169602394104,
       "2048": 333.77817153930664
      },
      "afd_nosplit": {
       "1": 25.894640684127808,
       "8": 32.10856080055237,
       "32": 37.29103922843933,
       "128": 49.655280113220215,
       "512": 103.67487907409668,
       "2048": 332.572078704834
      }
     }
    },
    {
     "order": [
      "masked_u2",
      "masked_split",
      "masked_u2_split",
      "afd",
      "afd_e2",
      "afd_nosplit",
      "masked"
     ],
     "median_us": {
      "masked_u2": {
       "1": 34.74840044975281,
       "8": 41.431119441986084,
       "32": 46.52640104293823,
       "128": 55.08847951889038,
       "512": 111.09255790710449,
       "2048": 340.71184158325195
      },
      "masked_split": {
       "1": 31.11224055290222,
       "8": 34.660959243774414,
       "32": 37.39287972450256,
       "128": 55.94496011734009,
       "512": 136.78607940673828,
       "2048": 387.2947883605957
      },
      "masked_u2_split": {
       "1": 28.72391939163208,
       "8": 30.08504033088684,
       "32": 32.48840093612671,
       "128": 46.531360149383545,
       "512": 111.4686393737793,
       "2048": 341.15840911865234
      },
      "afd": {
       "1": 20.48672080039978,
       "8": 22.60215997695923,
       "32": 24.59127902984619,
       "128": 44.64832067489624,
       "512": 104.07527923583984,
       "2048": 332.17321395874023
      },
      "afd_e2": {
       "1": 19.9726402759552,
       "8": 21.558159589767456,
       "32": 23.467119932174683,
       "128": 40.02840042114258,
       "512": 104.41463947296143,
       "2048": 333.74624252319336
      },
      "afd_nosplit": {
       "1": 25.87056040763855,
       "8": 33.53584051132202,
       "32": 36.52935981750488,
       "128": 48.74760150909424,
       "512": 103.69215965270996,
       "2048": 332.5165557861328
      },
      "masked": {
       "1": 38.248640298843384,
       "8": 47.90143966674805,
       "32": 55.47247886657715,
       "128": 69.1372799873352,
       "512": 136.14327907562256,
       "2048": 391.80383682250977
      }
     }
    },
    {
     "order": [
      "masked_split",
      "masked_u2_split",
      "afd",
      "afd_e2",
      "afd_nosplit",
      "masked",
      "masked_u2"
     ],
     "median_us": {
      "masked_split": {
       "1": 31.233999729156494,
       "8": 34.655680656433105,
       "32": 37.44279980659485,
       "128": 56.10719919204712,
       "512": 137.02271938323975,
       "2048": 386.8976020812988
      },
      "masked_u2_split": {
       "1": 28.767279386520386,
       "8": 30.091919898986816,
       "32": 32.32416033744812,
       "128": 46.49480104446411,
       "512": 111.52047634124756,
       "2048": 341.3852882385254
      },
      "afd": {
       "1": 20.29431939125061,
       "8": 22.157520055770874,
       "32": 24.37432050704956,
       "128": 44.27511930465698,
       "512": 103.80144119262695,
       "2048": 332.57238388061523
      },
      "afd_e2": {
       "1": 20.081839561462402,
       "8": 21.86792016029358,
       "32": 23.577040433883667,
       "128": 40.0927209854126,
       "512": 104.31439876556396,
       "2048": 334.59800720214844
      },
      "afd_nosplit": {
       "1": 25.78511953353882,
       "8": 31.871360540390015,
       "32": 37.05855965614319,
       "128": 49.2464804649353,
       "512": 103.09631824493408,
       "2048": 332.33903884887695
      },
      "masked": {
       "1": 38.28688025474548,
       "8": 47.7675199508667,
       "32": 55.091919898986816,
       "128": 68.69624137878418,
       "512": 136.08335971832275,
       "2048": 391.84167861938477
      },
      "masked_u2": {
       "1": 34.84431982040405,
       "8": 41.00816011428833,
       "32": 46.80503845214844,
       "128": 55.23239850997925,
       "512": 110.86599826812744,
       "2048": 340.9695243835449
      }
     }
    }
   ],
   "table": [
    {
     "tokens": 1,
     "masked": {
      "median_us": 38.25,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 34.84,
      "geomean_change_pct": -8.72,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 31.11,
      "geomean_change_pct": -18.56,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 28.75,
      "geomean_change_pct": -24.79,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 20.49,
      "geomean_change_pct": -46.51,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 20.05,
      "geomean_change_pct": -47.58,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 25.87,
      "geomean_change_pct": -32.37,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 8,
     "masked": {
      "median_us": 47.77,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 41.08,
      "geomean_change_pct": -12.77,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 34.66,
      "geomean_change_pct": -26.98,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 30.09,
      "geomean_change_pct": -36.0,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 22.53,
      "geomean_change_pct": -52.48,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 21.81,
      "geomean_change_pct": -53.93,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 32.11,
      "geomean_change_pct": -31.15,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 32,
     "masked": {
      "median_us": 55.09,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 46.53,
      "geomean_change_pct": -16.21,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 37.44,
      "geomean_change_pct": -32.05,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 32.46,
      "geomean_change_pct": -41.2,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 24.52,
      "geomean_change_pct": -55.58,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 23.55,
      "geomean_change_pct": -57.33,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 37.06,
      "geomean_change_pct": -32.98,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 128,
     "masked": {
      "median_us": 68.85,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 55.18,
      "geomean_change_pct": -19.93,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 56.07,
      "geomean_change_pct": -18.66,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 46.53,
      "geomean_change_pct": -32.42,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 44.65,
      "geomean_change_pct": -35.35,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 40.03,
      "geomean_change_pct": -41.97,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 49.25,
      "geomean_change_pct": -28.56,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 512,
     "masked": {
      "median_us": 136.08,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 111.09,
      "geomean_change_pct": -18.38,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 136.79,
      "geomean_change_pct": 0.44,
      "repeats_faster": 0
     },
     "masked_u2_split": {
      "median_us": 111.52,
      "geomean_change_pct": -18.05,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 104.08,
      "geomean_change_pct": -23.59,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 104.31,
      "geomean_change_pct": -23.36,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 103.67,
      "geomean_change_pct": -23.96,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 2048,
     "masked": {
      "median_us": 391.84,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "masked_u2": {
      "median_us": 340.97,
      "geomean_change_pct": -13.05,
      "repeats_faster": 3
     },
     "masked_split": {
      "median_us": 387.11,
      "geomean_change_pct": -1.3,
      "repeats_faster": 3
     },
     "masked_u2_split": {
      "median_us": 341.39,
      "geomean_change_pct": -12.94,
      "repeats_faster": 3
     },
     "afd": {
      "median_us": 332.57,
      "geomean_change_pct": -15.21,
      "repeats_faster": 3
     },
     "afd_e2": {
      "median_us": 333.78,
      "geomean_change_pct": -14.82,
      "repeats_faster": 3
     },
     "afd_nosplit": {
      "median_us": 332.52,
      "geomean_change_pct": -15.22,
      "repeats_faster": 3
     }
    }
   ]
  }
 ],
 "conclusion": "Unroll 2 and the split combine speed up Masked by up to 20% and 32% on their own: they are generic DeepEP changes, not AFD properties. The AFD expected top-k (4 vs upstream 2) is mixed (better at H2048 512-2048 tokens, worse at H4096 8-128 tokens). All three are removed from the role path."
}
