{
  "device": {
    "jax": "0.11.2",
    "backend": "gpu",
    "devices": [
      "cuda:0"
    ],
    "python": "3.13.18",
    "device_kind": "NVIDIA H100"
  },
  "config": {
    "output": "/home/tu/tu_tu/tu_iivgl01/tabdiffpfn/experiments/results/h100_branch/kernels.json",
    "families": [
      "ssd",
      "mamba",
      "rla",
      "kernel_mv",
      "mla"
    ],
    "seq_lens": [
      512,
      2048,
      8192
    ],
    "kmv_sizes": [
      1024,
      4096,
      16384,
      65536
    ]
  },
  "results": [
    {
      "family": "ssd",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.0,
      "compile_s": 0.33449030411429703,
      "inner": 11,
      "median_ms": 4.19429992325604,
      "q1_ms": 4.191699543629181,
      "q3_ms": 4.200701813467525
    },
    {
      "family": "ssd",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 6.735324859619141e-06,
      "compile_s": 12.500873720971867,
      "inner": 180,
      "median_ms": 0.13621907241435513,
      "q1_ms": 0.13615446385099655,
      "q3_ms": 0.13635901979998583
    },
    {
      "family": "ssd",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.0,
      "compile_s": 0.4864776241593063,
      "inner": 1,
      "median_ms": 24.96928395703435,
      "q1_ms": 24.943533935584128,
      "q3_ms": 25.03484406042844
    },
    {
      "family": "ssd",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.00016498565673828125,
      "compile_s": 22.551091604167596,
      "inner": 72,
      "median_ms": 0.5989024034028666,
      "q1_ms": 0.5980060846518933,
      "q3_ms": 0.5993871939911818
    },
    {
      "family": "ssd",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.0,
      "compile_s": 0.2749130860902369,
      "inner": 3,
      "median_ms": 16.52792903284232,
      "q1_ms": 16.4923690026626,
      "q3_ms": 16.53314579743892
    },
    {
      "family": "ssd",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 8.64267349243164e-06,
      "compile_s": 11.781082422938198,
      "inner": 79,
      "median_ms": 0.515448963976925,
      "q1_ms": 0.5132067459340715,
      "q3_ms": 0.5177546652104659
    },
    {
      "family": "ssd",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.0,
      "compile_s": 0.656510901870206,
      "inner": 1,
      "median_ms": 97.0373300369829,
      "q1_ms": 96.83405992109329,
      "q3_ms": 97.28381549939513
    },
    {
      "family": "ssd",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.000347137451171875,
      "compile_s": 23.672851047012955,
      "inner": 23,
      "median_ms": 2.091464566309815,
      "q1_ms": 2.0791160641238093,
      "q3_ms": 2.10367850250686
    },
    {
      "family": "ssd",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.0,
      "compile_s": 0.41478908713907003,
      "inner": 1,
      "median_ms": 66.09429488889873,
      "q1_ms": 66.05973944533616,
      "q3_ms": 66.16519403178245
    },
    {
      "family": "ssd",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 1.1622905731201172e-05,
      "compile_s": 9.95056066592224,
      "inner": 24,
      "median_ms": 1.9793018388251462,
      "q1_ms": 1.9688053580466658,
      "q3_ms": 1.986449962714687
    },
    {
      "family": "ssd",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.0,
      "compile_s": 1.3562647451180965,
      "inner": 1,
      "median_ms": 397.82210998237133,
      "q1_ms": 397.1815629629418,
      "q3_ms": 397.9586249915883
    },
    {
      "family": "ssd",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "dk": 128,
      "dv": 128,
      "max_abs_err": 0.0003032684326171875,
      "compile_s": 17.941247476032004,
      "inner": 6,
      "median_ms": 7.608328169832627,
      "q1_ms": 7.598409991866598,
      "q3_ms": 7.6452808182997005
    },
    {
      "family": "mamba",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.0,
      "compile_s": 0.223642559023574,
      "inner": 13,
      "median_ms": 3.515568472301731,
      "q1_ms": 3.4902472263918476,
      "q3_ms": 3.519460764856866
    },
    {
      "family": "mamba",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "dim": 512,
      "state": 16,
      "max_abs_err": 3.0517578125e-05,
      "compile_s": 0.75611452315934,
      "inner": 200,
      "median_ms": 0.14761096914298832,
      "q1_ms": 0.14759177283849567,
      "q3_ms": 0.1476278225891292
    },
    {
      "family": "mamba",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.0,
      "compile_s": 0.43431816017255187,
      "inner": 1,
      "median_ms": 25.036724051460624,
      "q1_ms": 25.025994516909122,
      "q3_ms": 25.05683503113687
    },
    {
      "family": "mamba",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.0234375,
      "compile_s": 0.9215409751050174,
      "inner": 75,
      "median_ms": 0.5339047623177369,
      "q1_ms": 0.5334903718903661,
      "q3_ms": 0.5341630258286993
    },
    {
      "family": "mamba",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.0,
      "compile_s": 0.21006575995124876,
      "inner": 3,
      "median_ms": 13.74606997705996,
      "q1_ms": 13.7360366837432,
      "q3_ms": 13.748718542046845
    },
    {
      "family": "mamba",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.003082275390625,
      "compile_s": 0.6838711660820991,
      "inner": 124,
      "median_ms": 0.2931108859728181,
      "q1_ms": 0.2930108429638729,
      "q3_ms": 0.2932043505772468
    },
    {
      "family": "mamba",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.0,
      "compile_s": 0.5371111340355128,
      "inner": 1,
      "median_ms": 99.06991687603295,
      "q1_ms": 98.97568600717932,
      "q3_ms": 99.11315643694252
    },
    {
      "family": "mamba",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "dim": 512,
      "state": 16,
      "max_abs_err": 7.6875,
      "compile_s": 0.9116635229438543,
      "inner": 43,
      "median_ms": 1.0504329761187003,
      "q1_ms": 1.0492612559045122,
      "q3_ms": 1.0515002095205492
    },
    {
      "family": "mamba",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.0,
      "compile_s": 0.3108331891708076,
      "inner": 1,
      "median_ms": 54.96438010595739,
      "q1_ms": 54.80359448119998,
      "q3_ms": 55.0274410052225
    },
    {
      "family": "mamba",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.01910400390625,
      "compile_s": 0.6556455630343407,
      "inner": 58,
      "median_ms": 0.7490287053173985,
      "q1_ms": 0.7468282670617618,
      "q3_ms": 0.7492879321316964
    },
    {
      "family": "mamba",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "dim": 512,
      "state": 16,
      "max_abs_err": 0.0,
      "compile_s": 1.2456730229314417,
      "inner": 1,
      "median_ms": 401.143808150664,
      "q1_ms": 400.924866553396,
      "q3_ms": 401.3092725072056
    },
    {
      "family": "mamba",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "dim": 512,
      "state": 16,
      "max_abs_err": 270.5,
      "compile_s": 0.9405811370816082,
      "inner": 16,
      "median_ms": 2.9237126873340458,
      "q1_ms": 2.9209692147560418,
      "q3_ms": 2.9252476306282915
    },
    {
      "family": "rla",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0,
      "compile_s": 0.680105066858232,
      "inner": 11,
      "median_ms": 4.084020726044069,
      "q1_ms": 4.076748493720185,
      "q3_ms": 4.0963244074108935
    },
    {
      "family": "rla",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 8.499622344970703e-05,
      "compile_s": 1.1668058109935373,
      "inner": 200,
      "median_ms": 0.10569053003564477,
      "q1_ms": 0.10520235518924892,
      "q3_ms": 0.1057245081756264
    },
    {
      "family": "rla",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0,
      "compile_s": 0.6772336610592902,
      "inner": 3,
      "median_ms": 14.875352693100771,
      "q1_ms": 14.862455699282387,
      "q3_ms": 14.907564308183888
    },
    {
      "family": "rla",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0005169510841369629,
      "compile_s": 10.203703058883548,
      "inner": 101,
      "median_ms": 0.3832185544110466,
      "q1_ms": 0.382098311669018,
      "q3_ms": 0.38337613810960314
    },
    {
      "family": "rla",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0,
      "compile_s": 0.7922758620698005,
      "inner": 3,
      "median_ms": 16.099295346066356,
      "q1_ms": 16.087218342969813,
      "q3_ms": 16.13094195878754
    },
    {
      "family": "rla",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0001558065414428711,
      "compile_s": 1.222227422054857,
      "inner": 94,
      "median_ms": 0.4241937979500979,
      "q1_ms": 0.42162655897319634,
      "q3_ms": 0.4259248727318296
    },
    {
      "family": "rla",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0,
      "compile_s": 0.8596830989699811,
      "inner": 1,
      "median_ms": 59.5350400544703,
      "q1_ms": 59.482570039108396,
      "q3_ms": 59.613165562041104
    },
    {
      "family": "rla",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0045604705810546875,
      "compile_s": 9.730496234958991,
      "inner": 28,
      "median_ms": 1.6088220706608678,
      "q1_ms": 1.60472298323709,
      "q3_ms": 1.610093895578757
    },
    {
      "family": "rla",
      "impl": "scan",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0,
      "compile_s": 1.0008351379074156,
      "inner": 1,
      "median_ms": 64.23196010291576,
      "q1_ms": 64.2138154944405,
      "q3_ms": 64.25755599047989
    },
    {
      "family": "rla",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.00028336048126220703,
      "compile_s": 1.2429741979576647,
      "inner": 27,
      "median_ms": 1.707655593269953,
      "q1_ms": 1.685141262391375,
      "q3_ms": 1.7080220730147428
    },
    {
      "family": "rla",
      "impl": "scan",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.0,
      "compile_s": 1.344883580924943,
      "inner": 1,
      "median_ms": 242.33168014325202,
      "q1_ms": 241.81460856925696,
      "q3_ms": 243.3582270750776
    },
    {
      "family": "rla",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "dim": 64,
      "max_abs_err": 0.03706502914428711,
      "compile_s": 8.825100774876773,
      "inner": 7,
      "median_ms": 6.3939952690686495,
      "q1_ms": 6.372318935713597,
      "q3_ms": 6.395514705218375
    },
    {
      "family": "kernel_mv",
      "impl": "dense",
      "mode": "fwd",
      "batch": 1,
      "n": 1024,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0,
      "compile_s": 0.6456710810307413,
      "inner": 200,
      "median_ms": 0.034144375240430236,
      "q1_ms": 0.034119850606657565,
      "q3_ms": 0.034231600584462285
    },
    {
      "family": "kernel_mv",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 1,
      "n": 1024,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.00020736455917358398,
      "compile_s": 0.1795209120027721,
      "inner": 200,
      "median_ms": 0.023074200144037604,
      "q1_ms": 0.022908052196726203,
      "q3_ms": 0.023639550199732184
    },
    {
      "family": "kernel_mv",
      "impl": "dense",
      "mode": "fwd_bwd",
      "batch": 1,
      "n": 1024,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0,
      "compile_s": 1.2126731709577143,
      "inner": 200,
      "median_ms": 0.060256130527704954,
      "q1_ms": 0.06020753527991474,
      "q3_ms": 0.06030553253367543
    },
    {
      "family": "kernel_mv",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 1,
      "n": 1024,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0093231201171875,
      "compile_s": 0.34079722594469786,
      "inner": 200,
      "median_ms": 0.0983987154904753,
      "q1_ms": 0.09826161549426615,
      "q3_ms": 0.09907091502100229
    },
    {
      "family": "kernel_mv",
      "impl": "dense",
      "mode": "fwd",
      "batch": 1,
      "n": 4096,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0,
      "compile_s": 1.1061887929681689,
      "inner": 200,
      "median_ms": 0.132280089892447,
      "q1_ms": 0.13222556270193309,
      "q3_ms": 0.13230826472863555
    },
    {
      "family": "kernel_mv",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 1,
      "n": 4096,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0011659562587738037,
      "compile_s": 0.17641796614043415,
      "inner": 200,
      "median_ms": 0.06451318971812725,
      "q1_ms": 0.06451056804507971,
      "q3_ms": 0.06453594018239528
    },
    {
      "family": "kernel_mv",
      "impl": "dense",
      "mode": "fwd_bwd",
      "batch": 1,
      "n": 4096,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0,
      "compile_s": 2.397001778939739,
      "inner": 98,
      "median_ms": 0.4069254785890178,
      "q1_ms": 0.4066653307337238,
      "q3_ms": 0.4069499234307785
    },
    {
      "family": "kernel_mv",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 1,
      "n": 4096,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0743865966796875,
      "compile_s": 0.3322007870301604,
      "inner": 106,
      "median_ms": 0.36280824504089804,
      "q1_ms": 0.36031975002326777,
      "q3_ms": 0.36517584494332656
    },
    {
      "family": "kernel_mv",
      "impl": "dense",
      "mode": "fwd",
      "batch": 1,
      "n": 16384,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0,
      "compile_s": 1.1409418790135533,
      "inner": 24,
      "median_ms": 1.989986417659869,
      "q1_ms": 1.9893351854989305,
      "q3_ms": 1.990035181127799
    },
    {
      "family": "kernel_mv",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 1,
      "n": 16384,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0014400482177734375,
      "compile_s": 0.17422437202185392,
      "inner": 113,
      "median_ms": 0.33833304486047905,
      "q1_ms": 0.3359265299926026,
      "q3_ms": 0.33968569921016434
    },
    {
      "family": "kernel_mv",
      "impl": "dense",
      "mode": "fwd_bwd",
      "batch": 1,
      "n": 16384,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0,
      "compile_s": 2.4624588820151985,
      "inner": 8,
      "median_ms": 5.905346624786034,
      "q1_ms": 5.901375363464467,
      "q3_ms": 5.905612240894698
    },
    {
      "family": "kernel_mv",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 1,
      "n": 16384,
      "dim": 8,
      "out": 16,
      "max_abs_err": 1.181640625,
      "compile_s": 0.3459841380827129,
      "inner": 25,
      "median_ms": 1.834685243666172,
      "q1_ms": 1.8296406185254455,
      "q3_ms": 1.8407532386481762
    },
    {
      "family": "kernel_mv",
      "impl": "dense",
      "mode": "fwd",
      "batch": 1,
      "n": 65536,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0,
      "compile_s": 1.8999424509238452,
      "inner": 1,
      "median_ms": 31.88534895889461,
      "q1_ms": 31.85496013611555,
      "q3_ms": 31.8953194655478
    },
    {
      "family": "kernel_mv",
      "impl": "pallas",
      "mode": "fwd",
      "batch": 1,
      "n": 65536,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.03714752197265625,
      "compile_s": 0.1823815491516143,
      "inner": 10,
      "median_ms": 4.879103694111109,
      "q1_ms": 4.823222057893872,
      "q3_ms": 4.940158792305738
    },
    {
      "family": "kernel_mv",
      "impl": "dense",
      "mode": "fwd_bwd",
      "batch": 1,
      "n": 65536,
      "dim": 8,
      "out": 16,
      "max_abs_err": 0.0,
      "compile_s": 4.598857602104545,
      "inner": 1,
      "median_ms": 93.87300489470363,
      "q1_ms": 93.84493040852249,
      "q3_ms": 93.9063458936289
    },
    {
      "family": "kernel_mv",
      "impl": "pallas",
      "mode": "fwd_bwd",
      "batch": 1,
      "n": 65536,
      "dim": 8,
      "out": 16,
      "max_abs_err": 20.0078125,
      "compile_s": 0.3538718849886209,
      "inner": 1,
      "median_ms": 26.622078148648143,
      "q1_ms": 26.368893100880086,
      "q3_ms": 26.984558906406164
    },
    {
      "family": "mla",
      "impl": "expand_xla",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.0,
      "compile_s": 0.8388488269411027,
      "inner": 200,
      "median_ms": 0.04871415439993143,
      "q1_ms": 0.048677207669243217,
      "q3_ms": 0.04878475738223642
    },
    {
      "family": "mla",
      "impl": "pallas_mla",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.00390625,
      "compile_s": 0.7054827050305903,
      "inner": 200,
      "median_ms": 0.03481487510725856,
      "q1_ms": 0.034774227533489466,
      "q3_ms": 0.03484140208456665
    },
    {
      "family": "mla",
      "impl": "expand_cudnn",
      "mode": "fwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.00390625,
      "compile_s": 0.2898731338791549,
      "inner": 200,
      "median_ms": 0.0761189148761332,
      "q1_ms": 0.07557809236459434,
      "q3_ms": 0.07646876736544073
    },
    {
      "family": "mla",
      "impl": "expand_xla",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.0,
      "compile_s": 2.219065578887239,
      "inner": 188,
      "median_ms": 0.1573719943754692,
      "q1_ms": 0.15717593115813872,
      "q3_ms": 0.15754619628151362
    },
    {
      "family": "mla",
      "impl": "pallas_mla",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.0625,
      "compile_s": 1.4966588071547449,
      "inner": 200,
      "median_ms": 0.09727546479552984,
      "q1_ms": 0.0971684348769486,
      "q3_ms": 0.09770571195986122
    },
    {
      "family": "mla",
      "impl": "expand_cudnn",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 512,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.0625,
      "compile_s": 0.3224862858187407,
      "inner": 164,
      "median_ms": 0.11802270082847738,
      "q1_ms": 0.11695934127357493,
      "q3_ms": 0.11849163097284007
    },
    {
      "family": "mla",
      "impl": "expand_xla",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.0,
      "compile_s": 0.7072437291499227,
      "inner": 55,
      "median_ms": 0.8140721443024549,
      "q1_ms": 0.8136972382834011,
      "q3_ms": 0.8149042450399562
    },
    {
      "family": "mla",
      "impl": "pallas_mla",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.001953125,
      "compile_s": 0.7040609319228679,
      "inner": 94,
      "median_ms": 0.4237113604718383,
      "q1_ms": 0.4224366709054943,
      "q3_ms": 0.42601434038357533
    },
    {
      "family": "mla",
      "impl": "expand_cudnn",
      "mode": "fwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.001953125,
      "compile_s": 0.17825972102582455,
      "inner": 200,
      "median_ms": 0.1277074299287051,
      "q1_ms": 0.1230873673921451,
      "q3_ms": 0.1314711378654465
    },
    {
      "family": "mla",
      "impl": "expand_xla",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.0,
      "compile_s": 2.8728893911466002,
      "inner": 21,
      "median_ms": 2.2851602351736453,
      "q1_ms": 2.28412143353905,
      "q3_ms": 2.2855483305950957
    },
    {
      "family": "mla",
      "impl": "pallas_mla",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.25,
      "compile_s": 1.6056027659215033,
      "inner": 44,
      "median_ms": 1.0580693593841386,
      "q1_ms": 1.047862635459751,
      "q3_ms": 1.0684143741276455
    },
    {
      "family": "mla",
      "impl": "expand_cudnn",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 2048,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.25,
      "compile_s": 0.4279358971398324,
      "inner": 83,
      "median_ms": 0.45402556747557166,
      "q1_ms": 0.4461327292788101,
      "q3_ms": 0.45850437908736336
    },
    {
      "family": "mla",
      "impl": "expand_xla",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.0,
      "compile_s": 0.7123958801385015,
      "inner": 4,
      "median_ms": 11.889156012330204,
      "q1_ms": 11.886308493558317,
      "q3_ms": 11.89296625670977
    },
    {
      "family": "mla",
      "impl": "pallas_mla",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.001953125,
      "compile_s": 0.6969728369731456,
      "inner": 7,
      "median_ms": 6.5122385775404315,
      "q1_ms": 6.482949929444917,
      "q3_ms": 6.530958560428449
    },
    {
      "family": "mla",
      "impl": "expand_cudnn",
      "mode": "fwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.001953125,
      "compile_s": 0.18305338104255497,
      "inner": 31,
      "median_ms": 1.5630927687931444,
      "q1_ms": 1.553755497650033,
      "q3_ms": 1.5896252739513592
    },
    {
      "family": "mla",
      "impl": "expand_xla",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 0.0,
      "compile_s": 2.8671856089495122,
      "inner": 1,
      "median_ms": 34.69787607900798,
      "q1_ms": 34.67542107682675,
      "q3_ms": 34.719191025942564
    },
    {
      "family": "mla",
      "impl": "pallas_mla",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 1.0,
      "compile_s": 1.5542019340209663,
      "inner": 3,
      "median_ms": 15.356286972140273,
      "q1_ms": 15.317525171364348,
      "q3_ms": 15.729187871329486
    },
    {
      "family": "mla",
      "impl": "expand_cudnn",
      "mode": "fwd_bwd",
      "batch": 4,
      "seq": 8192,
      "heads": 8,
      "rank": 128,
      "dtype": "bfloat16",
      "max_abs_err": 1.0,
      "compile_s": 0.4178292849101126,
      "inner": 9,
      "median_ms": 5.470045221348603,
      "q1_ms": 5.27584594157007,
      "q3_ms": 5.6026473434435
    }
  ]
}
