{
  "status": "pass",
  "audit_sha256": "97b95dc36314e9eda57a64b162b9b302eb92befc037bb482a9f6e1d0c0d4a503",
  "source_reports_sha256": {
    "reference": "9db930c33983bd52311fc19417706a59cb4526b1ec4887994489672fbde58cc3",
    "candidate": "898a666aa7a9b99e42c06ab658958f35e9b5f663277924b880ee2ecd5937c344"
  },
  "replay_sha256": "804a28f4b442761d61103f20869e43e88744f76befe9124ff21123dabffdc5c1",
  "executed_output_validations": 240,
  "replay_pairs": 48,
  "recorded_unchanged_artifacts": 21,
  "current_artifacts_checked": true,
  "torch_probe_control_ratio_range": [
    0.7616954820704699,
    4.557749997843478
  ],
  "summaries": {
    "gpu_throughput": {
      "positive_medians": 2,
      "all_pairs_positive": 2,
      "all_pairs_negative": 8,
      "faster_torch_all_pairs": 10
    },
    "e2e_throughput": {
      "positive_medians": 2,
      "all_pairs_positive": 1,
      "all_pairs_negative": 9,
      "faster_torch_all_pairs": 12
    },
    "gpu_latency": {
      "positive_medians": 3,
      "all_pairs_positive": 1,
      "all_pairs_negative": 4,
      "faster_torch_all_pairs": 8
    },
    "e2e_latency": {
      "positive_medians": 6,
      "all_pairs_positive": 1,
      "all_pairs_negative": 1,
      "faster_torch_all_pairs": 8
    }
  },
  "records": [
    {
      "name": "softmax_37x1537",
      "gpu_throughput": {
        "reference_us": 4.090820311835937,
        "candidate_us": 4.621094376492188,
        "torch_us": 16.58333394516376,
        "reference_rounds_us": [
          4.0957029341406255,
          3.8925782064375,
          4.100260412093751,
          4.08593768953125
        ],
        "candidate_rounds_us": [
          4.676432581625,
          4.365234417484375,
          4.565756171359375,
          4.956380507796874
        ],
        "torch_rounds_us": [
          16.087241419882048,
          16.718750885047484,
          16.447917005280033,
          20.930989194312133
        ],
        "gain": 0.8837702066605939,
        "gain_min": 0.824379339540955,
        "gain_max": 0.8980462946782742,
        "native_torch_ratio": 0.2693434180534514,
        "ratio_min": 0.2367962862043682,
        "ratio_max": 0.2906920123574106
      },
      "e2e_throughput": {
        "reference_us": 4.4046901453250005,
        "candidate_us": 5.597846156195,
        "torch_us": 30.726977347765185,
        "reference_rounds_us": [
          5.59391189697,
          4.35920250931,
          4.45017778134,
          4.07747933884
        ],
        "candidate_rounds_us": [
          6.47326796195,
          6.34030354927,
          4.68943898471,
          4.85538876312
        ],
        "torch_rounds_us": [
          31.65197152245345,
          35.16386217948718,
          29.154097237569058,
          29.801983173076923
        ],
        "gain": 0.8519700466210031,
        "gain_min": 0.6875384554438096,
        "gain_max": 0.9489787149059589,
        "native_torch_ratio": 0.17161451882049916,
        "ratio_min": 0.1608500838320253,
        "ratio_max": 0.20451389441437978
      },
      "gpu_latency": {
        "reference_us": 6.520829629155,
        "candidate_us": 7.374997949225,
        "torch_us": 70.18749602138996,
        "reference_rounds_us": [
          6.49999128655,
          6.79166987538,
          6.49999128655,
          6.54166797176
        ],
        "candidate_rounds_us": [
          7.54167558625,
          7.95838423073,
          7.16664362699,
          7.208320312200001
        ],
        "torch_rounds_us": [
          55.374985095113516,
          70.25001104921103,
          70.1249809935689,
          89.41662963479757
        ],
        "gain": 0.8844273612309768,
        "gain_min": 0.8533980866562181,
        "gain_max": 0.9075162712578547,
        "native_torch_ratio": 0.10774237236797632,
        "ratio_min": 0.08061498562002156,
        "ratio_max": 0.1361928237686425
      },
      "e2e_latency": {
        "reference_us": 236.70850000000002,
        "candidate_us": 207.83350000000002,
        "torch_us": 330.35400000000004,
        "reference_rounds_us": [
          307.125,
          257.584,
          215.833,
          208.5
        ],
        "candidate_rounds_us": [
          208.125,
          247.542,
          207.542,
          207.166
        ],
        "torch_rounds_us": [
          337.75,
          341.75,
          322.958,
          309.042
        ],
        "gain": 1.0402576969190176,
        "gain_min": 1.0064392805769287,
        "gain_max": 1.4756756756756757,
        "native_torch_ratio": 0.6564887336426377,
        "ratio_min": 0.6162102146558105,
        "ratio_max": 0.72433650329188
      }
    },
    {
      "name": "softmax_256x3072",
      "gpu_throughput": {
        "reference_us": 14.454101346929686,
        "candidate_us": 15.126302059820313,
        "torch_us": 29.387043923634337,
        "reference_rounds_us": [
          14.349609045890626,
          17.2122390721875,
          13.45963482890625,
          14.558593647968749
        ],
        "candidate_rounds_us": [
          18.0338538484375,
          14.567708603859375,
          15.684895515781252,
          14.386718248718749
        ],
        "torch_rounds_us": [
          30.007159693923313,
          30.15689981111791,
          28.76692815334536,
          28.582679988176096
        ],
        "gain": 0.9350369848644009,
        "gain_min": 0.7957039669107618,
        "gain_max": 1.1815337291705243,
        "native_torch_ratio": 0.524288733725197,
        "ratio_min": 0.48306386581848554,
        "ratio_max": 0.6009850326517073
      },
      "e2e_throughput": {
        "reference_us": 17.052239446,
        "candidate_us": 17.82352689755,
        "torch_us": 39.32840541239263,
        "reference_rounds_us": [
          20.1908852061,
          19.8677641357,
          14.1664228972,
          14.2367147563
        ],
        "candidate_rounds_us": [
          20.8618613139,
          22.5713026101,
          14.7851924812,
          14.7282949131
        ],
        "torch_rounds_us": [
          45.70940959409594,
          45.00211558669002,
          33.62697963800905,
          33.654695238095236
        ],
        "gain": 0.9623863947454419,
        "gain_min": 0.8802223105550743,
        "gain_max": 0.9678371887482092,
        "native_torch_ratio": 0.4480421695632879,
        "ratio_min": 0.4376297217640049,
        "ratio_max": 0.5015609225441785
      },
      "gpu_latency": {
        "reference_us": 18.145830836149997,
        "candidate_us": 18.666672986000002,
        "torch_us": 50.79165566712618,
        "reference_rounds_us": [
          17.7499605343,
          18.083374016,
          18.208287656299998,
          18.333375919599998
        ],
        "candidate_rounds_us": [
          18.666672986000002,
          18.7084078789,
          18.666672986000002,
          18.624996300800003
        ],
        "torch_rounds_us": [
          52.91670095175505,
          52.24999040365219,
          49.333320930600166,
          48.666668590158224
        ],
        "gain": 0.9710172010306827,
        "gain_min": 0.9508904209985607,
        "gain_max": 0.9843425267586507,
        "native_torch_ratio": 0.3682171602729421,
        "ratio_min": 0.3527557963792695,
        "ratio_max": 0.3827053883151251
      },
      "e2e_latency": {
        "reference_us": 251.2705,
        "candidate_us": 235.2495,
        "torch_us": 337.77049999999997,
        "reference_rounds_us": [
          244.791,
          240.583,
          261.25,
          257.75
        ],
        "candidate_rounds_us": [
          234.666,
          256.792,
          235.833,
          234.458
        ],
        "torch_rounds_us": [
          487.79200000000003,
          334.541,
          341.0,
          310.66700000000003
        ],
        "gain": 1.07124522418234,
        "gain_min": 0.9368788747313002,
        "gain_max": 1.1077754173504133,
        "native_torch_ratio": 0.7231423493290337,
        "ratio_min": 0.48107800045921206,
        "ratio_max": 0.767595003303033
      }
    },
    {
      "name": "softmax_768x6144",
      "gpu_throughput": {
        "reference_us": 60.71419284125,
        "candidate_us": 66.22460887226563,
        "torch_us": 148.51009200356202,
        "reference_rounds_us": [
          63.23502566375,
          68.59700533828125,
          58.19336001875,
          56.34049466609375
        ],
        "candidate_rounds_us": [
          69.52929561515626,
          74.06901022484375,
          62.71809888859375,
          62.919922129375
        ],
        "torch_rounds_us": [
          165.7018237892771,
          166.31184644211316,
          131.31836021784693,
          123.21744725340977
        ],
        "gain": 0.9177980036155734,
        "gain_min": 0.8954317290833143,
        "gain_max": 0.9278559307436749,
        "native_torch_ratio": 0.46148281480274045,
        "ratio_min": 0.419604890430033,
        "ratio_max": 0.5106413379914737
      },
      "e2e_throughput": {
        "reference_us": 72.40254934275,
        "candidate_us": 77.97195856265,
        "torch_us": 160.53757861695962,
        "reference_rounds_us": [
          86.1267882038,
          82.9557683284,
          61.8493303571,
          56.7273901345
        ],
        "candidate_rounds_us": [
          91.8474789474,
          100.797694534,
          64.0964381779,
          63.293685022
        ],
        "torch_rounds_us": [
          181.47480239520957,
          194.56586451612904,
          139.60035483870968,
          132.04232978723405
        ],
        "gain": 0.9169860173376954,
        "gain_min": 0.8229927153782098,
        "gain_max": 0.9649417676757149,
        "native_torch_ratio": 0.4927304388874359,
        "ratio_min": 0.4591423728969401,
        "ratio_max": 0.5180646398826251
      },
      "gpu_latency": {
        "reference_us": 79.604156781,
        "candidate_us": 85.5416583361,
        "torch_us": 150.208332343027,
        "reference_rounds_us": [
          97.8750176728,
          377.041636966,
          61.3332958892,
          60.333288274699996
        ],
        "candidate_rounds_us": [
          102.9582927,
          375.333358534,
          68.1250239722,
          66.8332795613
        ],
        "torch_rounds_us": [
          161.50000737980008,
          318.9999843016267,
          138.9166573062539,
          137.37502740696073
        ],
        "gain": 0.9266855087495613,
        "gain_min": 0.9003049439546397,
        "gain_max": 1.0045513631899714,
        "native_torch_ratio": 0.5639573721956398,
        "ratio_min": 0.4865023929226426,
        "ratio_max": 1.1765936583216505
      },
      "e2e_latency": {
        "reference_us": 291.29150000000004,
        "candidate_us": 313.312,
        "torch_us": 498.47950000000003,
        "reference_rounds_us": [
          285.375,
          295,
          531.084,
          287.583
        ],
        "candidate_rounds_us": [
          325.458,
          333.708,
          301.166,
          279.5
        ],
        "torch_rounds_us": [
          569.4169999999999,
          566.875,
          430.084,
          425.25
        ],
        "gain": 0.956462937968993,
        "gain_min": 0.8768412514057112,
        "gain_max": 1.7634261503622586,
        "native_torch_ratio": 0.6229702395698231,
        "ratio_min": 0.5715635465748302,
        "ratio_max": 0.7002492536341738
      }
    },
    {
      "name": "softmax_1024x4096",
      "gpu_throughput": {
        "reference_us": 54.40917993835937,
        "candidate_us": 57.43391966460938,
        "torch_us": 137.70572923021973,
        "reference_rounds_us": [
          56.99544271921875,
          61.356770856875,
          51.8229171575,
          51.13997394796875
        ],
        "candidate_rounds_us": [
          60.843750361562506,
          63.680989114843754,
          53.9765624125,
          54.024088967656255
        ],
        "torch_rounds_us": [
          153.44010535045527,
          155.69791503367014,
          121.9713531099842,
          120.76692837581504
        ],
        "gain": 0.9533573201946028,
        "gain_min": 0.9367509790327637,
        "gain_max": 0.9635021646134106,
        "native_torch_ratio": 0.42576911576269494,
        "ratio_min": 0.39653094751594536,
        "ratio_max": 0.44734174905515933
      },
      "e2e_throughput": {
        "reference_us": 63.98709626695,
        "candidate_us": 65.0165471182,
        "torch_us": 148.77312313045854,
        "reference_rounds_us": [
          76.3673069544,
          75.8282755299,
          52.145917004,
          51.7450123239
        ],
        "candidate_rounds_us": [
          75.6175765199,
          76.96875,
          54.4155177165,
          53.9874130019
        ],
        "torch_rounds_us": [
          169.35828571428573,
          168.59791875000002,
          128.94832751091704,
          128.22118502202642
        ],
        "gain": 0.9718235047884571,
        "gain_min": 0.9582912961643696,
        "gain_max": 1.0099147641197241,
        "native_torch_ratio": 0.4342447255467104,
        "ratio_min": 0.4210490878915664,
        "ratio_max": 0.4565225393685353
      },
      "gpu_latency": {
        "reference_us": 66.29165727645001,
        "candidate_us": 57.02080670745,
        "torch_us": 136.85415615327656,
        "reference_rounds_us": [
          54.583302699,
          427.041668445,
          77.41665467620001,
          55.1666598767
        ],
        "candidate_rounds_us": [
          58.2082429901,
          54.7083909623,
          58.9583069086,
          55.8333704248
        ],
        "torch_rounds_us": [
          140.62499394640326,
          142.12500536814332,
          133.08331836014986,
          127.00003571808338
        ],
        "gain": 1.1505667567878015,
        "gain_min": 0.9377246227528887,
        "gain_max": 7.805780081144736,
        "native_torch_ratio": 0.4267790078249433,
        "ratio_min": 0.3849314961894991,
        "ratio_max": 0.44301801033430144
      },
      "e2e_latency": {
        "reference_us": 304.08299999999997,
        "candidate_us": 338.875,
        "torch_us": 429.854,
        "reference_rounds_us": [
          321.25,
          287.416,
          320.75,
          271.125
        ],
        "candidate_rounds_us": [
          324.75,
          353,
          358.25,
          314.584
        ],
        "torch_rounds_us": [
          482.208,
          413.58299999999997,
          446.125,
          388.5
        ],
        "gain": 0.8785884861309059,
        "gain_min": 0.8142096317280453,
        "gain_max": 0.9892224788298691,
        "native_torch_ratio": 0.8063830417296374,
        "ratio_min": 0.6734645630101532,
        "ratio_max": 0.8535167064410288
      }
    },
    {
      "name": "rmsnorm_37x1537",
      "gpu_throughput": {
        "reference_us": 4.076497134515625,
        "candidate_us": 4.569987140710937,
        "torch_us": 5.014973794459365,
        "reference_rounds_us": [
          4.587890543921875,
          4.213540705671875,
          3.939453563359375,
          3.854166607203125
        ],
        "candidate_rounds_us": [
          9.55078121478125,
          4.315755177234375,
          4.55468762084375,
          4.585286660578125
        ],
        "torch_rounds_us": [
          5.724609764001798,
          4.77018238598248,
          5.259765202936251,
          4.735026777780149
        ],
        "gain": 0.8527368689466578,
        "gain_min": 0.4803680914417173,
        "gain_max": 0.9763159708174176,
        "native_torch_ratio": 0.9365559809685358,
        "ratio_min": 0.8659488484963372,
        "ratio_max": 1.6683724495667211
      },
      "e2e_throughput": {
        "reference_us": 4.91881185937,
        "candidate_us": 5.51554660946,
        "torch_us": 8.722819035230218,
        "reference_rounds_us": [
          5.84887393573,
          5.66664126016,
          4.16416669025,
          4.17098245858
        ],
        "candidate_rounds_us": [
          6.32488496241,
          6.3086094161,
          4.48615724001,
          4.72248380282
        ],
        "torch_rounds_us": [
          9.834085983510011,
          9.375360255447031,
          8.070277815013405,
          7.5278315585672795
        ],
        "gain": 0.911489663500157,
        "gain_min": 0.8832179490143143,
        "gain_max": 0.9282257547978229,
        "native_torch_ratio": 0.6352479761326113,
        "ratio_min": 0.5558863452834614,
        "ratio_max": 0.6728924803113282
      },
      "gpu_latency": {
        "reference_us": 6.979156751185,
        "candidate_us": 7.166672730815,
        "torch_us": 8.124974556267262,
        "reference_rounds_us": [
          7.9165911302000005,
          6.9999368861299995,
          6.95837661624,
          6.79166987538
        ],
        "candidate_rounds_us": [
          7.08334846422,
          7.20837851986,
          7.12502514943,
          7.208320312200001
        ],
        "torch_rounds_us": [
          8.375034667551517,
          8.04167939350009,
          8.124974556267262,
          8.124974556267262
        ],
        "gain": 0.9738471187865552,
        "gain_min": 0.942198678919023,
        "gain_max": 1.1176340074456235,
        "native_torch_ratio": 0.8820547905946281,
        "ratio_min": 0.8457694499658533,
        "ratio_max": 0.8963772574278915
      },
      "e2e_latency": {
        "reference_us": 205.35399999999998,
        "candidate_us": 218.45850000000002,
        "torch_us": 224.542,
        "reference_rounds_us": [
          199.833,
          251.625,
          201.083,
          209.625
        ],
        "candidate_rounds_us": [
          230.708,
          206.209,
          236.208,
          203.333
        ],
        "torch_rounds_us": [
          220.125,
          247.709,
          210.916,
          228.959
        ],
        "gain": 0.9485585687731186,
        "gain_min": 0.8512963151121046,
        "gain_max": 1.220242569431984,
        "native_torch_ratio": 0.9680766299635231,
        "ratio_min": 0.8324647065710168,
        "ratio_max": 1.1199150372660207
      }
    },
    {
      "name": "rmsnorm_256x3072",
      "gpu_throughput": {
        "reference_us": 16.64029878160156,
        "candidate_us": 18.009439827414063,
        "torch_us": 16.991862139548175,
        "reference_rounds_us": [
          18.43163954625,
          22.43945345984375,
          14.641927919,
          14.848958016953125
        ],
        "candidate_rounds_us": [
          20.9895824809375,
          22.69466131109375,
          15.029297173890626,
          14.464191735890624
        ],
        "torch_rounds_us": [
          16.892578059923835,
          16.725911336834542,
          17.091146219172515,
          18.54622405517148
        ],
        "gain": 0.9814902212843428,
        "gain_min": 0.8781327386092318,
        "gain_max": 1.0266012984401862,
        "native_torch_ratio": 1.0609471899253544,
        "ratio_min": 0.7798995468221679,
        "ratio_max": 1.3568564877606737
      },
      "e2e_throughput": {
        "reference_us": 18.316391744100002,
        "candidate_us": 18.8084229451,
        "torch_us": 23.11723958019426,
        "reference_rounds_us": [
          21.3798687553,
          20.8371212121,
          15.7956622761,
          15.4653151351
        ],
        "candidate_rounds_us": [
          21.9010779962,
          21.4260257903,
          16.1908200999,
          15.6171527697
        ],
        "torch_rounds_us": [
          26.67684851217313,
          25.909060165975106,
          20.325418994413408,
          20.321328810853952
        ],
        "gain": 0.9758976871917159,
        "gain_min": 0.9725145211732823,
        "gain_max": 0.9902775085293017,
        "native_torch_ratio": 0.8087784268419126,
        "ratio_min": 0.7685104116497847,
        "ratio_max": 0.8269703977312762
      },
      "gpu_latency": {
        "reference_us": 18.562481273000003,
        "candidate_us": 19.229191821100002,
        "torch_us": 21.208339603617787,
        "reference_rounds_us": [
          18.2500225492,
          18.541642930400002,
          18.5833196156,
          18.7083496712
        ],
        "candidate_rounds_us": [
          25.8750515059,
          17.374986782700002,
          19.2500301637,
          19.2083534785
        ],
        "torch_rounds_us": [
          25.124987587332726,
          21.249987185001373,
          21.083338651806116,
          21.1666920222342
        ],
        "gain": 0.9696676002043434,
        "gain_min": 0.705313477155346,
        "gain_max": 1.0671457286437547,
        "native_torch_ratio": 0.9102624291277301,
        "ratio_min": 0.8176469299220841,
        "ratio_max": 1.0298533050398575
      },
      "e2e_latency": {
        "reference_us": 286.29150000000004,
        "candidate_us": 274.8125,
        "torch_us": 270.39549999999997,
        "reference_rounds_us": [
          316.25,
          359.833,
          256.333,
          249.291
        ],
        "candidate_rounds_us": [
          300.417,
          269.458,
          258.833,
          280.167
        ],
        "torch_rounds_us": [
          284.125,
          324.20799999999997,
          256.666,
          255.042
        ],
        "gain": 1.0215223357205754,
        "gain_min": 0.8897943012560366,
        "gain_max": 1.335395497628573,
        "native_torch_ratio": 1.0328919190786614,
        "ratio_min": 0.8311269308592016,
        "ratio_max": 1.0985131860634718
      }
    },
    {
      "name": "rmsnorm_768x6144",
      "gpu_throughput": {
        "reference_us": 62.33235717445312,
        "candidate_us": 69.43457037781249,
        "torch_us": 80.50748738241964,
        "reference_rounds_us": [
          63.85026063078125,
          70.76822839734375,
          60.5533859925,
          60.814453718124994
        ],
        "candidate_rounds_us": [
          72.8704426365625,
          75.47395762234375,
          65.9986981190625,
          65.90625071109375
        ],
        "torch_rounds_us": [
          81.9667975520133,
          82.73893308796687,
          78.50976544432342,
          79.04817721282598
        ],
        "gain": 0.9201177315724314,
        "gain_min": 0.8762161765536551,
        "gain_max": 0.9376509544054056,
        "native_torch_ratio": 0.864833538826131,
        "ratio_min": 0.8337478868570307,
        "ratio_max": 0.9121939914562458
      },
      "e2e_throughput": {
        "reference_us": 77.9055938999,
        "candidate_us": 81.38937100325,
        "torch_us": 100.6529917715063,
        "reference_rounds_us": [
          90.1441456311,
          91.5876099707,
          65.6670421687,
          60.0632780083
        ],
        "candidate_rounds_us": [
          95.8948220065,
          98.5059148265,
          66.88392,
          66.1798329571
        ],
        "torch_rounds_us": [
          120.36084466019418,
          117.98071764705884,
          83.32526589595376,
          81.05970790378007
        ],
        "gain": 0.9348995195650529,
        "gain_min": 0.9075767544961173,
        "gain_max": 0.9818061227377222,
        "native_torch_ratio": 0.8095589504740976,
        "ratio_min": 0.7967277255093441,
        "ratio_max": 0.8349323244598493
      },
      "gpu_latency": {
        "reference_us": 82.83334318544999,
        "candidate_us": 73.18751886485,
        "torch_us": 99.47919170372188,
        "reference_rounds_us": [
          108.416716103,
          61.500002630100006,
          92.9583329707,
          72.7083534002
        ],
        "candidate_rounds_us": [
          228.833348956,
          66.95830961690001,
          77.000062447,
          69.3749752827
        ],
        "torch_rounds_us": [
          116.87504593282938,
          127.58333468809724,
          82.08333747461438,
          80.91668132692575
        ],
        "gain": 0.9832653590602276,
        "gain_min": 0.47378022739092246,
        "gain_max": 1.2072500984617285,
        "native_torch_ratio": 0.8977174306881907,
        "ratio_min": 0.5248201873747295,
        "ratio_max": 1.957931628001374
      },
      "e2e_latency": {
        "reference_us": 307.354,
        "candidate_us": 322.14599999999996,
        "torch_us": 353.2085,
        "reference_rounds_us": [
          312.084,
          297.25,
          309.042,
          305.666
        ],
        "candidate_rounds_us": [
          342.167,
          311.542,
          332.75,
          278
        ],
        "torch_rounds_us": [
          384.625,
          408.58299999999997,
          321.79200000000003,
          321.79200000000003
        ],
        "gain": 0.9414381401475603,
        "gain_min": 0.9120809429313756,
        "gain_max": 1.0995179856115107,
        "native_torch_ratio": 0.8767620259921107,
        "ratio_min": 0.7624937895115558,
        "ratio_max": 1.034053052903739
      }
    },
    {
      "name": "rmsnorm_1024x4096",
      "gpu_throughput": {
        "reference_us": 55.87597661359375,
        "candidate_us": 67.40169283148438,
        "torch_us": 71.93261717475252,
        "reference_rounds_us": [
          58.4316403546875,
          62.8938805675,
          53.3203128725,
          53.05729155218749
        ],
        "candidate_rounds_us": [
          69.4609379934375,
          71.03255211406251,
          63.98958430515625,
          65.34244766953125
        ],
        "torch_rounds_us": [
          72.86328127520392,
          72.39192746055778,
          69.06445378263015,
          71.47330688894726
        ],
        "gain": 0.8372406624955626,
        "gain_min": 0.8119881247871856,
        "gain_max": 0.8854233544432754,
        "native_torch_ratio": 0.9399124553002163,
        "ratio_min": 0.9142216935765695,
        "ratio_max": 0.9812220036932168
      },
      "e2e_throughput": {
        "reference_us": 64.31849031535,
        "candidate_us": 79.5709797049,
        "torch_us": 88.93924334525997,
        "reference_rounds_us": [
          75.1672152231,
          76.4571809045,
          53.4697654076,
          53.4165877863
        ],
        "candidate_rounds_us": [
          93.789975,
          94.5872722063,
          65.0599934354,
          65.3519844098
        ],
        "torch_rounds_us": [
          104.95245953002612,
          111.39916319444445,
          72.64666047745358,
          72.92602716049383
        ],
        "gain": 0.8128458462437311,
        "gain_min": 0.8014418942227035,
        "gain_max": 0.8218532247577264,
        "native_torch_ratio": 0.8946050208768836,
        "ratio_min": 0.8490842255358801,
        "ratio_max": 0.8961407463754324
      },
      "gpu_latency": {
        "reference_us": 54.9374963157,
        "candidate_us": 68.02086136305,
        "torch_us": 108.18752343766391,
        "reference_rounds_us": [
          54.2916823179,
          95.8750024438,
          54.8749812879,
          55.0000113435
        ],
        "candidate_rounds_us": [
          68.2083773427,
          102.499965578,
          67.8333453834,
          65.24997297670001
        ],
        "torch_rounds_us": [
          143.79166532307863,
          344.7500057518482,
          71.8749943189323,
          72.5833815522492
        ],
        "gain": 0.8259400198704381,
        "gain_min": 0.795967950463354,
        "gain_max": 0.9353661916192687,
        "native_torch_ratio": 0.6866606526325949,
        "ratio_min": 0.2973167914949353,
        "ratio_max": 0.9437683581914705
      },
      "e2e_latency": {
        "reference_us": 317.35450000000003,
        "candidate_us": 324.813,
        "torch_us": 368.437,
        "reference_rounds_us": [
          299.375,
          388.916,
          335.334,
          265.334
        ],
        "candidate_rounds_us": [
          366.584,
          373.25,
          280.042,
          283.042
        ],
        "torch_rounds_us": [
          429.417,
          405.791,
          331.083,
          316.459
        ],
        "gain": 0.9897043577745412,
        "gain_min": 0.8166613927503655,
        "gain_max": 1.1974418122995838,
        "native_torch_ratio": 0.8740408708462633,
        "ratio_min": 0.8458362404593409,
        "ratio_max": 0.919808472834538
      }
    },
    {
      "name": "layernorm_37x1537",
      "gpu_throughput": {
        "reference_us": 4.530598744164063,
        "candidate_us": 5.351887921273438,
        "torch_us": 8.37500010675285,
        "reference_rounds_us": [
          4.623047061609375,
          4.6126297092968755,
          4.44856777903125,
          4.207030542609375
        ],
        "candidate_rounds_us": [
          4.9980471885625,
          7.729818207734375,
          5.0006510718906245,
          5.70312477065625
        ],
        "torch_rounds_us": [
          8.945312401920091,
          8.226562385971192,
          8.523437827534508,
          7.567708053102251
        ],
        "gain": 0.8136344279491395,
        "gain_min": 0.5967319780795785,
        "gain_max": 0.9249706709830045,
        "native_torch_ratio": 0.6701536785082504,
        "ratio_min": 0.5587336656335983,
        "ratio_max": 0.9396170411247452
      },
      "e2e_throughput": {
        "reference_us": 5.4343188123,
        "candidate_us": 6.09051892908,
        "torch_us": 13.593591577519433,
        "reference_rounds_us": [
          6.48703510759,
          6.09648250232,
          4.51412140241,
          4.77215512228
        ],
        "candidate_rounds_us": [
          6.7874326215,
          7.06860512951,
          4.93671174909,
          5.39360523666
        ],
        "torch_rounds_us": [
          14.994132799084145,
          14.536413576725614,
          12.434805601317956,
          12.650769578313252
        ],
        "gain": 0.8995893080879535,
        "gain_min": 0.8624732023675244,
        "gain_max": 0.9557421000455378,
        "native_torch_ratio": 0.4395092938384667,
        "ratio_min": 0.3970075534246198,
        "ratio_max": 0.48626885113035095
      },
      "gpu_latency": {
        "reference_us": 8.45832983032,
        "candidate_us": 8.479139069095,
        "torch_us": 10.416668374091387,
        "reference_rounds_us": [
          9.45833744481,
          10.499963536900001,
          7.20826210454,
          7.45832221583
        ],
        "candidate_rounds_us": [
          33.0832554027,
          9.124982170760001,
          7.83329596743,
          7.66658922657
        ],
        "torch_rounds_us": [
          10.25001984089613,
          10.583316907286644,
          10.583316907286644,
          10.25001984089613
        ],
        "gain": 0.9465212628761278,
        "gain_min": 0.2858950042757305,
        "gain_max": 1.150683183858263,
        "native_torch_ratio": 0.8050814273392295,
        "ratio_min": 0.7401550984490272,
        "ratio_max": 3.2276284257228935
      },
      "e2e_latency": {
        "reference_us": 213.8545,
        "candidate_us": 215.667,
        "torch_us": 236.45850000000002,
        "reference_rounds_us": [
          205.125,
          204.708,
          222.584,
          235.667
        ],
        "candidate_rounds_us": [
          213.584,
          217.75,
          213.083,
          238.709
        ],
        "torch_rounds_us": [
          256.0,
          242.209,
          230.708,
          207.334
        ],
        "gain": 0.9738257118643033,
        "gain_min": 0.9401056257175661,
        "gain_max": 1.0445882590352116,
        "native_torch_ratio": 0.9113108472379129,
        "ratio_min": 0.8343125,
        "ratio_max": 1.1513258799810933
      }
    },
    {
      "name": "layernorm_256x3072",
      "gpu_throughput": {
        "reference_us": 18.592773358203125,
        "candidate_us": 17.866536381953125,
        "torch_us": 43.406575514381984,
        "reference_rounds_us": [
          19.79101489265625,
          24.15169274171875,
          16.7154948940625,
          17.39453182375
        ],
        "candidate_rounds_us": [
          19.42122435140625,
          19.96419268828125,
          16.311848412499998,
          15.9511719175
        ],
        "torch_rounds_us": [
          43.263671614113264,
          46.11914027918829,
          43.549479414650705,
          41.91080734017305
        ],
        "gain": 1.0576158680040293,
        "gain_min": 1.019040537020686,
        "gain_max": 1.2097505328074454,
        "native_torch_ratio": 0.4067405390272839,
        "ratio_min": 0.37455897594524273,
        "ratio_max": 0.4489037482678828
      },
      "e2e_throughput": {
        "reference_us": 20.3601597969,
        "candidate_us": 20.1714460696,
        "torch_us": 53.79085678012679,
        "reference_rounds_us": [
          23.4223098495,
          23.1671813685,
          17.5531382253,
          17.3332375479
        ],
        "candidate_rounds_us": [
          23.0996139073,
          23.0559858491,
          16.8110119048,
          17.2869062901
        ],
        "torch_rounds_us": [
          62.82715577889447,
          61.62105910165485,
          45.960654458598725,
          45.827670731707315
        ],
        "gain": 1.0093963014889162,
        "gain_min": 1.0026801358798674,
        "gain_max": 1.0441452498340151,
        "native_torch_ratio": 0.3709134162034168,
        "ratio_min": 0.3657696371565668,
        "ratio_max": 0.37721546860420097
      },
      "gpu_latency": {
        "reference_us": 22.89583790115,
        "candidate_us": 20.18752275035,
        "torch_us": 44.81250653043389,
        "reference_rounds_us": [
          65.0000292808,
          24.8749856837,
          20.9166901186,
          20.541658159300002
        ],
        "candidate_rounds_us": [
          20.0416543521,
          20.208361093,
          20.3333329409,
          20.1666844077
        ],
        "torch_rounds_us": [
          73.45835911110044,
          44.29166438058019,
          44.74999150261283,
          44.87502155825496
        ],
        "gain": 1.129807566771591,
        "gain_min": 1.018593723391478,
        "gain_max": 3.243246697046703,
        "native_torch_ratio": 0.4518864462571164,
        "ratio_min": 0.27283013934177935,
        "ratio_max": 0.4562565298824132
      },
      "e2e_latency": {
        "reference_us": 252.64600000000002,
        "candidate_us": 245.8545,
        "torch_us": 300.97950000000003,
        "reference_rounds_us": [
          250.833,
          254.459,
          245.834,
          272.916
        ],
        "candidate_rounds_us": [
          246.167,
          245.542,
          263.416,
          233.5
        ],
        "torch_rounds_us": [
          326.41700000000003,
          275.54200000000003,
          377.79200000000003,
          274.667
        ],
        "gain": 1.0276350957624958,
        "gain_min": 0.9332538646095909,
        "gain_max": 1.1688051391862955,
        "native_torch_ratio": 0.80213458084193,
        "ratio_min": 0.6972513975944434,
        "ratio_max": 0.8911236762453636
      }
    },
    {
      "name": "layernorm_768x6144",
      "gpu_throughput": {
        "reference_us": 78.633463545,
        "candidate_us": 76.07779934914062,
        "torch_us": 317.6432292093523,
        "reference_rounds_us": [
          81.73828155125,
          114.1803377321875,
          75.47200584671874,
          75.52864553875
        ],
        "candidate_rounds_us": [
          79.63216103234375,
          84.8177087390625,
          72.5234376659375,
          72.0553389328125
        ],
        "torch_rounds_us": [
          349.74674508703174,
          362.537760338455,
          285.53971333167283,
          267.1230477062636
        ],
        "gain": 1.04443004156618,
        "gain_min": 1.0264481145758535,
        "gain_max": 1.346185123715823,
        "native_torch_ratio": 0.24397136736528835,
        "ratio_min": 0.22768520980096013,
        "ratio_max": 0.2697458701206744
      },
      "e2e_throughput": {
        "reference_us": 91.79167315255,
        "candidate_us": 89.7018142158,
        "torch_us": 347.4520728971963,
        "reference_rounds_us": [
          106.551715415,
          110.977257042,
          77.0316308901,
          75.6498502538
        ],
        "candidate_rounds_us": [
          107.041376307,
          109.243542254,
          72.1989296117,
          72.3622521246
        ],
        "torch_rounds_us": [
          386.39480000000003,
          397.4097307692308,
          308.5093457943925,
          274.9961388888889
        ],
        "gain": 1.0306513412539644,
        "gain_min": 0.9954254989155258,
        "gain_max": 1.0669359130999756,
        "native_torch_ratio": 0.2690140517184508,
        "ratio_min": 0.234025097119156,
        "ratio_max": 0.27702592350362887
      },
      "gpu_latency": {
        "reference_us": 105.14582390885,
        "candidate_us": 98.35418313720001,
        "torch_us": 276.2500080280006,
        "reference_rounds_us": [
          112.625013571,
          120.58333959400001,
          93.00000965599999,
          97.66663424670001
        ],
        "candidate_rounds_us": [
          91.0833477974,
          114.916649181,
          105.62501847700001,
          75.66664135079999
        ],
        "torch_rounds_us": [
          276.62498177960515,
          275.87503427639604,
          277.291692327708,
          271.0000262595713
        ],
        "gain": 1.1429081468995468,
        "gain_min": 0.8804733101774639,
        "gain_max": 1.2907489020677594,
        "native_torch_ratio": 0.3550915826035641,
        "ratio_min": 0.27921267165606983,
        "ratio_max": 0.41655327558878097
      },
      "e2e_latency": {
        "reference_us": 332.10400000000004,
        "candidate_us": 324.39549999999997,
        "torch_us": 553.375,
        "reference_rounds_us": [
          330.25,
          344.791,
          333.958,
          307.083
        ],
        "candidate_rounds_us": [
          321.875,
          372.959,
          326.916,
          295.25
        ],
        "torch_rounds_us": [
          513.417,
          581.542,
          622.5,
          525.208
        ],
        "gain": 1.0237800595313402,
        "gain_min": 0.9244742719709137,
        "gain_max": 1.040077900084674,
        "native_torch_ratio": 0.59454263522875,
        "ratio_min": 0.525166265060241,
        "ratio_max": 0.6413277114980517
      }
    },
    {
      "name": "layernorm_1024x4096",
      "gpu_throughput": {
        "reference_us": 64.78971272367187,
        "candidate_us": 79.3085937403125,
        "torch_us": 238.21289050829364,
        "reference_rounds_us": [
          67.608722929375,
          72.20833322203124,
          61.970702517968746,
          61.857422224375
        ],
        "candidate_rounds_us": [
          82.3430991659375,
          86.1686194184375,
          76.05013070140625,
          76.2740883146875
        ],
        "torch_rounds_us": [
          274.417968284979,
          277.6933597488096,
          202.00781273160828,
          200.19726616737898
        ],
        "gain": 0.8179638360142596,
        "gain_min": 0.8109886803125984,
        "gain_max": 0.8379887447353117,
        "native_torch_ratio": 0.343386283409138,
        "ratio_min": 0.30006453178177245,
        "ratio_max": 0.3809946547967194
      },
      "e2e_throughput": {
        "reference_us": 76.5438054915,
        "candidate_us": 94.14215217534999,
        "torch_us": 251.09364599616072,
        "reference_rounds_us": [
          92.3878287037,
          90.7099400428,
          62.158222467,
          62.3776709402
        ],
        "candidate_rounds_us": [
          111.756726316,
          112.089678201,
          76.5275780347,
          76.3707430341
        ],
        "torch_rounds_us": [
          296.2487745098039,
          297.4761904761905,
          205.66753846153844,
          205.93851748251748
        ],
        "gain": 0.8145037085231989,
        "gain_min": 0.8092622041445984,
        "gain_max": 0.8266869632747377,
        "native_torch_ratio": 0.37444789979172954,
        "ratio_min": 0.37084244350056206,
        "ratio_max": 0.3772394552548658
      },
      "gpu_latency": {
        "reference_us": 64.62499732149999,
        "candidate_us": 92.2707840798,
        "torch_us": 200.62498515471816,
        "reference_rounds_us": [
          97.91669435799999,
          64.5833206363,
          64.66667400669999,
          63.8333149254
        ],
        "candidate_rounds_us": [
          106.916588265,
          112.83333879,
          77.2500643507,
          77.62497989459999
        ],
        "torch_rounds_us": [
          200.6666618399322,
          200.541689991951,
          200.58330846950412,
          200.9583404287696
        ],
        "gain": 0.8297189515529468,
        "gain_min": 0.5723779986383225,
        "gain_max": 0.9158232220738921,
        "native_torch_ratio": 0.4595404603245285,
        "ratio_min": 0.3851270823087694,
        "ratio_max": 0.5626428040699603
      },
      "e2e_latency": {
        "reference_us": 350.04150000000004,
        "candidate_us": 342.47900000000004,
        "torch_us": 451.27099999999996,
        "reference_rounds_us": [
          397.708,
          409.792,
          302.375,
          286.833
        ],
        "candidate_rounds_us": [
          365,
          377,
          293.625,
          319.958
        ],
        "torch_rounds_us": [
          534.417,
          445.667,
          456.875,
          432.66700000000003
        ],
        "gain": 1.0583906736090645,
        "gain_min": 0.8964707867907663,
        "gain_max": 1.0896109589041096,
        "native_torch_ratio": 0.7112444983973893,
        "ratio_min": 0.6426812585499316,
        "ratio_max": 0.8459230770956792
      }
    }
  ]
}