{
 "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/attention.json",
  "batch": 4,
  "heads": 8,
  "head_dims": [
   64,
   128
  ],
  "seq_lens": [
   128,
   512,
   1024,
   2048,
   4096,
   8192
  ],
  "ragged_seq_lens": [
   600,
   1500,
   3000
  ],
  "dtypes": [
   "bfloat16",
   "float32"
  ],
  "modes": [
   "fwd",
   "fwd_bwd"
  ],
  "causal": [
   0,
   1
  ],
  "impls": [
   "probjax_mha",
   "probjax_flex",
   "probjax_default",
   "jax_xla",
   "jax_cudnn",
   "torch_flash",
   "torch_efficient",
   "torch_math"
  ],
  "append": false,
  "tag": "h100_nvl",
  "torch_note": "torch rows: torch 2.14.1+cu130, separate process run with XLA_PYTHON_CLIENT_PREALLOCATE=false and the platform allocator (JAX only builds inputs and references)"
 },
 "results": [
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0021501779556274414,
   "rel_err": 0.002294705016538501,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.58245162316598,
   "inner": 200,
   "median_ms": 0.032379410695284605,
   "q1_ms": 0.03235680749639869,
   "q3_ms": 0.03259640536271036,
   "tflops": 4.145156601616163
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0021501779556274414,
   "rel_err": 0.002294705016538501,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.3656611740589142,
   "inner": 200,
   "median_ms": 0.03260710509493947,
   "q1_ms": 0.0323298602597788,
   "q3_ms": 0.032728229998610914,
   "tflops": 4.11621110212664
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0021501779556274414,
   "rel_err": 0.002294705016538501,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.3736522209364921,
   "inner": 200,
   "median_ms": 0.024089179933071136,
   "q1_ms": 0.023997932439669967,
   "q3_ms": 0.024731452576816082,
   "tflops": 5.571701833474932
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0021272897720336914,
   "rel_err": 0.002270278288051486,
   "xla_temp_mb": 3.0,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.21017099916934967,
   "inner": 200,
   "median_ms": 0.04397152573801577,
   "q1_ms": 0.04381847509648651,
   "q3_ms": 0.04421169694978744,
   "tflops": 3.052378232214978
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0021501779556274414,
   "rel_err": 0.002294705016538501,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 1.8839773400686681,
   "inner": 200,
   "median_ms": 0.0483914150390774,
   "q1_ms": 0.0482587848091498,
   "q3_ms": 0.057820758083835244,
   "tflops": 2.773585519076379
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0032709836959838867,
   "rel_err": 0.0035336874425411224,
   "xla_temp_mb": 1.031494140625,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 0.49407712602987885,
   "inner": 200,
   "median_ms": 0.03703694441355765,
   "q1_ms": 0.03648591984529048,
   "q3_ms": 0.043835199903696775,
   "tflops": 12.683607015594951
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0032709836959838867,
   "rel_err": 0.0035336874425411224,
   "xla_temp_mb": 0.53173828125,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 0.5104248500429094,
   "inner": 200,
   "median_ms": 0.041541935643181205,
   "q1_ms": 0.04106380802113563,
   "q3_ms": 0.04165493242908269,
   "tflops": 11.308140574742522
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0032709836959838867,
   "rel_err": 0.0035336874425411224,
   "xla_temp_mb": 0.53173828125,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 0.48620230006054044,
   "inner": 200,
   "median_ms": 0.04087468492798507,
   "q1_ms": 0.04078475700225681,
   "q3_ms": 0.04143867990933359,
   "tflops": 11.49273808049282
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00531381368637085,
   "rel_err": 0.005376721732318401,
   "xla_temp_mb": 5.0,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 0.03203871799632907,
   "inner": 200,
   "median_ms": 0.05848812987096608,
   "q1_ms": 0.0583962828386575,
   "q3_ms": 0.058601482305675745,
   "tflops": 8.031750186514225
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0032709836959838867,
   "rel_err": 0.0035336874425411224,
   "xla_temp_mb": 2.031982421875,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 1.385311946971342,
   "inner": 200,
   "median_ms": 0.09394488530233502,
   "q1_ms": 0.09378393238876015,
   "q3_ms": 0.0957827502861619,
   "tflops": 5.000400463401533
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005635261535644531,
   "rel_err": 0.0019494958687573671,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.5003675778862089,
   "inner": 200,
   "median_ms": 0.03290495602414012,
   "q1_ms": 0.032707182108424604,
   "q3_ms": 0.033099629799835384,
   "tflops": 2.039475875633045
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005635261535644531,
   "rel_err": 0.0019494958687573671,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.4955592919141054,
   "inner": 200,
   "median_ms": 0.03272150410339236,
   "q1_ms": 0.029213064699433744,
   "q3_ms": 0.03321520518511534,
   "tflops": 2.050910122833949
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005635261535644531,
   "rel_err": 0.0019494958687573671,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.49652559007517993,
   "inner": 200,
   "median_ms": 0.03315305453725159,
   "q1_ms": 0.033122305176220834,
   "q3_ms": 0.03340060240589082,
   "tflops": 2.024213603744862
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007589459419250488,
   "rel_err": 0.0026255426928400993,
   "xla_temp_mb": 3.0,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.26251308782957494,
   "inner": 200,
   "median_ms": 0.034733699867501855,
   "q1_ms": 0.03456257516518235,
   "q3_ms": 0.039627409423701465,
   "tflops": 1.932096616715156
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005635261535644531,
   "rel_err": 0.0019494958687573671,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 1.5,
   "xla_out_mb": 0.5,
   "compile_s": 0.35177920293062925,
   "inner": 200,
   "median_ms": 0.06298706983216107,
   "q1_ms": 0.05812208284623921,
   "q3_ms": 0.06320609478279948,
   "tflops": 1.0654387349470629
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009996175765991211,
   "rel_err": 0.005375280044972897,
   "xla_temp_mb": 1.031494140625,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 0.6479109970387071,
   "inner": 200,
   "median_ms": 0.04939231090247631,
   "q1_ms": 0.049186357646249235,
   "q3_ms": 0.04964848281815648,
   "tflops": 4.7554167786108605
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009996175765991211,
   "rel_err": 0.005375280044972897,
   "xla_temp_mb": 0.53173828125,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 0.6658460050821304,
   "inner": 200,
   "median_ms": 0.042945980094373226,
   "q1_ms": 0.04292097466532141,
   "q3_ms": 0.043132202699780464,
   "tflops": 5.469220250273763
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009996175765991211,
   "rel_err": 0.005375280044972897,
   "xla_temp_mb": 0.53173828125,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 0.7093678100500256,
   "inner": 200,
   "median_ms": 0.05475329468026757,
   "q1_ms": 0.05173700279556215,
   "q3_ms": 0.05489754257723689,
   "tflops": 4.289806218449321
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011534929275512695,
   "rel_err": 0.005375280044972897,
   "xla_temp_mb": 5.0,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 0.8127072351053357,
   "inner": 200,
   "median_ms": 0.06344031542539597,
   "q1_ms": 0.06328529270831496,
   "q3_ms": 0.06953857489861548,
   "tflops": 3.702393697525251
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009996175765991211,
   "rel_err": 0.005375280044972897,
   "xla_temp_mb": 2.031982421875,
   "xla_arg_mb": 2.5,
   "xla_out_mb": 1.5000228881835938,
   "compile_s": 1.1307117850519717,
   "inner": 200,
   "median_ms": 0.10058536543510854,
   "q1_ms": 0.09757464285939932,
   "q3_ms": 0.10108178772497922,
   "tflops": 2.3351411309583674
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00117570161819458,
   "rel_err": 0.002601401647552848,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.42613860103301704,
   "inner": 200,
   "median_ms": 0.03812893992289901,
   "q1_ms": 0.037962092901580036,
   "q3_ms": 0.03829291497822851,
   "tflops": 56.32161954521821
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00117570161819458,
   "rel_err": 0.002601401647552848,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.427638057153672,
   "inner": 200,
   "median_ms": 0.022344185272231698,
   "q1_ms": 0.02198373491410166,
   "q3_ms": 0.02245718496851623,
   "tflops": 96.10928399653002
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00117570161819458,
   "rel_err": 0.002601401647552848,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.42889614310115576,
   "inner": 200,
   "median_ms": 0.022598389768972993,
   "q1_ms": 0.022451785043813288,
   "q3_ms": 0.0228771852562204,
   "tflops": 95.02817103139091
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001338571310043335,
   "rel_err": 0.002961773192510009,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.2180318171158433,
   "inner": 200,
   "median_ms": 0.047950310399755836,
   "q1_ms": 0.04793259024154395,
   "q3_ms": 0.04803713760338724,
   "tflops": 44.7856047249057
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00117570161819458,
   "rel_err": 0.002601401647552848,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.30465239798650146,
   "inner": 200,
   "median_ms": 0.06459241500124335,
   "q1_ms": 0.06438524054829031,
   "q3_ms": 0.06464693986345083,
   "tflops": 33.246684582991094
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001962810754776001,
   "rel_err": 0.005092823877930641,
   "xla_temp_mb": 4.125244140625,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.5757848720531911,
   "inner": 200,
   "median_ms": 0.044875419698655605,
   "q1_ms": 0.04485132289119065,
   "q3_ms": 0.04511879466008395,
   "tflops": 167.49019437527787
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001962810754776001,
   "rel_err": 0.005092823877930641,
   "xla_temp_mb": 2.12548828125,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.5716010550968349,
   "inner": 200,
   "median_ms": 0.04710156470537186,
   "q1_ms": 0.04706729028839618,
   "q3_ms": 0.04752036300487816,
   "tflops": 159.5741630881063
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001962810754776001,
   "rel_err": 0.005092823877930641,
   "xla_temp_mb": 2.12548828125,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.5824770429171622,
   "inner": 200,
   "median_ms": 0.05467874580062926,
   "q1_ms": 0.049840257270261645,
   "q3_ms": 0.05551491514779627,
   "tflops": 137.46095778066476
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027485042810440063,
   "rel_err": 0.0069497511722147465,
   "xla_temp_mb": 80.0,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.03245029808022082,
   "inner": 181,
   "median_ms": 0.1510475242500147,
   "q1_ms": 0.15102050560449368,
   "q3_ms": 0.15114412709905956,
   "tflops": 49.76044993335446
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001962810754776001,
   "rel_err": 0.005092823877930641,
   "xla_temp_mb": 8.125732421875,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.1330546799581498,
   "inner": 200,
   "median_ms": 0.10925773996859789,
   "q1_ms": 0.10908590978942811,
   "q3_ms": 0.10932043776847422,
   "tflops": 68.79322938732078
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007361173629760742,
   "rel_err": 0.0025465681683272123,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.5092254630289972,
   "inner": 200,
   "median_ms": 0.02354988013394177,
   "q1_ms": 0.022647185251116753,
   "q3_ms": 0.031813884852454066,
   "tflops": 45.59436472258076
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007361173629760742,
   "rel_err": 0.0025465681683272123,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.5063336009625345,
   "inner": 200,
   "median_ms": 0.02224993542768061,
   "q1_ms": 0.02213678730186075,
   "q3_ms": 0.02237271226476878,
   "tflops": 48.258199557028085
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007361173629760742,
   "rel_err": 0.0025465681683272123,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.5078775370493531,
   "inner": 200,
   "median_ms": 0.022002740297466516,
   "q1_ms": 0.021819387329742312,
   "q3_ms": 0.022213064949028194,
   "tflops": 48.800368021597514
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007361173629760742,
   "rel_err": 0.0025465681683272123,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.711287965066731,
   "inner": 200,
   "median_ms": 0.07699678069911897,
   "q1_ms": 0.07692502986174077,
   "q3_ms": 0.07703394978307188,
   "tflops": 13.945282052711669
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007361173629760742,
   "rel_err": 0.0025465681683272123,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.05738696618936956,
   "inner": 200,
   "median_ms": 0.07047769497148693,
   "q1_ms": 0.07027137442491949,
   "q3_ms": 0.07067804457619786,
   "tflops": 15.235200646593256
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01617288589477539,
   "rel_err": 0.0037439705338329077,
   "xla_temp_mb": 4.125244140625,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.6280911031644791,
   "inner": 200,
   "median_ms": 0.04170603002421558,
   "q1_ms": 0.04164363257586956,
   "q3_ms": 0.04175345762632787,
   "tflops": 90.10918521417535
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01617288589477539,
   "rel_err": 0.0037439705338329077,
   "xla_temp_mb": 2.12548828125,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.6857734378427267,
   "inner": 200,
   "median_ms": 0.05390154430642724,
   "q1_ms": 0.05258607503492385,
   "q3_ms": 0.05449037009384483,
   "tflops": 69.7214974516395
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01617288589477539,
   "rel_err": 0.0037439705338329077,
   "xla_temp_mb": 2.12548828125,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.6475662270095199,
   "inner": 200,
   "median_ms": 0.049328404711559415,
   "q1_ms": 0.04931290808599442,
   "q3_ms": 0.04933960735797882,
   "tflops": 76.18524065343112
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01617288589477539,
   "rel_err": 0.00455421581864357,
   "xla_temp_mb": 80.0,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.049216388957574964,
   "inner": 143,
   "median_ms": 0.2504382998897479,
   "q1_ms": 0.2503502980473783,
   "q3_ms": 0.250633968808851,
   "tflops": 15.00607688861668
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01617288589477539,
   "rel_err": 0.0037439705338329077,
   "xla_temp_mb": 8.125732421875,
   "xla_arg_mb": 10.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.3084611869417131,
   "inner": 200,
   "median_ms": 0.09652367560192943,
   "q1_ms": 0.09640619973652065,
   "q3_ms": 0.09660287236329168,
   "tflops": 38.93445168311513
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010502040386199951,
   "rel_err": 0.00308937462978065,
   "xla_temp_mb": 10.0,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.503125251038,
   "inner": 200,
   "median_ms": 0.04834071500226855,
   "q1_ms": 0.047993610496632755,
   "q3_ms": 0.049567632377147675,
   "tflops": 61.006958624041914
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010502040386199951,
   "rel_err": 0.00308937462978065,
   "xla_temp_mb": 10.0,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.49826927413232625,
   "inner": 200,
   "median_ms": 0.05013680434785783,
   "q1_ms": 0.04949445778038353,
   "q3_ms": 0.05104913027025759,
   "tflops": 58.821459372210775
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010502040386199951,
   "rel_err": 0.00308937462978065,
   "xla_temp_mb": 10.0,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.5027708020061255,
   "inner": 200,
   "median_ms": 0.050257855327799916,
   "q1_ms": 0.05003805737942457,
   "q3_ms": 0.050424630171619356,
   "tflops": 58.67978211097096
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012940764427185059,
   "rel_err": 0.0038067714776843786,
   "xla_temp_mb": 65.91798400878906,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.21607089205645025,
   "inner": 184,
   "median_ms": 0.08133889692228125,
   "q1_ms": 0.08121718786677103,
   "q3_ms": 0.08137841106635396,
   "tflops": 36.257191965829875
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010502040386199951,
   "rel_err": 0.00308937462978065,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.40210924996063113,
   "inner": 200,
   "median_ms": 0.06317637045867741,
   "q1_ms": 0.06310004217084497,
   "q3_ms": 0.07228302012663335,
   "tflops": 46.68074437623746
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002047628164291382,
   "rel_err": 0.004772087559103966,
   "xla_temp_mb": 20.00048828125,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.609733286080882,
   "inner": 200,
   "median_ms": 0.11402752483263612,
   "q1_ms": 0.11399537499528378,
   "q3_ms": 0.11458014778327197,
   "tflops": 90.52130189750234
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002047628164291382,
   "rel_err": 0.004772087559103966,
   "xla_temp_mb": 15.078651428222656,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.649293112102896,
   "inner": 200,
   "median_ms": 0.10831573978066444,
   "q1_ms": 0.10830096493009478,
   "q3_ms": 0.10862863739021122,
   "tflops": 95.2947375967844
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002047628164291382,
   "rel_err": 0.004772087559103966,
   "xla_temp_mb": 15.078651428222656,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.6556209339760244,
   "inner": 198,
   "median_ms": 0.10859125265569398,
   "q1_ms": 0.10856236342923961,
   "q3_ms": 0.10859784364436914,
   "tflops": 95.05296004575345
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0025545358657836914,
   "rel_err": 0.005635968875139952,
   "xla_temp_mb": 109.93653869628906,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.03584496700204909,
   "inner": 122,
   "median_ms": 0.24762199084717232,
   "q1_ms": 0.24629798698712324,
   "q3_ms": 0.24771384062764584,
   "tflops": 41.684181460161575
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002047628164291382,
   "rel_err": 0.004772087559103966,
   "xla_temp_mb": 9.522216796875,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.7258995128795505,
   "inner": 200,
   "median_ms": 0.0958570244256407,
   "q1_ms": 0.09581279999110848,
   "q3_ms": 0.09613394970074296,
   "tflops": 107.68037148917591
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004250764846801758,
   "rel_err": 0.0016897449968382716,
   "xla_temp_mb": 10.0,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.5423730579204857,
   "inner": 200,
   "median_ms": 0.036055295495316386,
   "q1_ms": 0.03599955001845956,
   "q3_ms": 0.036076094838790596,
   "tflops": 40.8971825010156
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004250764846801758,
   "rel_err": 0.0016897449968382716,
   "xla_temp_mb": 10.0,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.5810728068463504,
   "inner": 200,
   "median_ms": 0.036241349298506975,
   "q1_ms": 0.03616817237343639,
   "q3_ms": 0.03630174498539418,
   "tflops": 40.687226842868874
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004250764846801758,
   "rel_err": 0.0016897449968382716,
   "xla_temp_mb": 10.0,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.5769043900072575,
   "inner": 200,
   "median_ms": 0.03850923967547715,
   "q1_ms": 0.03742836939636618,
   "q3_ms": 0.0386758369859308,
   "tflops": 38.29107020617201
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008065938949584961,
   "rel_err": 0.0032063359394669533,
   "xla_temp_mb": 65.91798400878906,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.6546666361391544,
   "inner": 200,
   "median_ms": 0.12268134974874556,
   "q1_ms": 0.1224144222214818,
   "q3_ms": 0.12274559703655541,
   "tflops": 12.019430850898978
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004250764846801758,
   "rel_err": 0.0016897449968382716,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 7.03125,
   "xla_out_mb": 2.34375,
   "compile_s": 0.36299878102727234,
   "inner": 200,
   "median_ms": 0.05076125496998429,
   "q1_ms": 0.050716979312710464,
   "q3_ms": 0.05140920169651508,
   "tflops": 29.048927195986863
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014807701110839844,
   "rel_err": 0.0035223369486629963,
   "xla_temp_mb": 20.00048828125,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.6480019560549408,
   "inner": 200,
   "median_ms": 0.08669144939631224,
   "q1_ms": 0.08664187742397189,
   "q3_ms": 0.08729544992092997,
   "tflops": 59.53251486668005
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014807701110839844,
   "rel_err": 0.0035223369486629963,
   "xla_temp_mb": 15.07861328125,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.6623648749664426,
   "inner": 200,
   "median_ms": 0.08538305526599288,
   "q1_ms": 0.08501575444824994,
   "q3_ms": 0.08573622966650873,
   "tflops": 60.44478010211884
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014807701110839844,
   "rel_err": 0.0035223369486629963,
   "xla_temp_mb": 15.07861328125,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.6582728859502822,
   "inner": 200,
   "median_ms": 0.0857161555904895,
   "q1_ms": 0.08485900762025267,
   "q3_ms": 0.08582645503338426,
   "tflops": 60.2098865079365
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014807701110839844,
   "rel_err": 0.0036034160293638706,
   "xla_temp_mb": 109.93653869628906,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.03911998798139393,
   "inner": 115,
   "median_ms": 0.3133883213867312,
   "q1_ms": 0.31334793033159297,
   "q3_ms": 0.31434879132101073,
   "tflops": 16.468258859050497
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014807701110839844,
   "rel_err": 0.0035223369486629963,
   "xla_temp_mb": 9.522216796875,
   "xla_arg_mb": 11.71875,
   "xla_out_mb": 7.031272888183594,
   "compile_s": 0.7875301770400256,
   "inner": 200,
   "median_ms": 0.09768972522579134,
   "q1_ms": 0.09721432346850634,
   "q3_ms": 0.10652009514160454,
   "tflops": 52.83012095766895
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011154413223266602,
   "rel_err": 0.0039638252928853035,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.4229674700181931,
   "inner": 200,
   "median_ms": 0.034384349128231406,
   "q1_ms": 0.03428167721722275,
   "q3_ms": 0.034651875612325966,
   "tflops": 249.82106131964557
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011154413223266602,
   "rel_err": 0.0039638252928853035,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.4270467881578952,
   "inner": 200,
   "median_ms": 0.034298700047656894,
   "q1_ms": 0.03426480514463037,
   "q3_ms": 0.03472172247711569,
   "tflops": 250.44490257836515
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011154413223266602,
   "rel_err": 0.0039638252928853035,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.4258241110946983,
   "inner": 200,
   "median_ms": 0.034288050374016166,
   "q1_ms": 0.03414000500924885,
   "q3_ms": 0.034786424366757274,
   "tflops": 250.52268934221877
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012263059616088867,
   "rel_err": 0.004357793368399143,
   "xla_temp_mb": 192.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.21097166603431106,
   "inner": 153,
   "median_ms": 0.2303781639175867,
   "q1_ms": 0.2295634286246853,
   "q3_ms": 0.23038362402444573,
   "tflops": 37.28623601268426
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011154413223266602,
   "rel_err": 0.0039638252928853035,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.05436793388798833,
   "inner": 200,
   "median_ms": 0.05153810023330152,
   "q1_ms": 0.05014340975321829,
   "q3_ms": 0.05232307536061853,
   "tflops": 166.67154111454002
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015159845352172852,
   "rel_err": 0.004941033665090799,
   "xla_temp_mb": 8.250244140625,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.5621474818326533,
   "inner": 188,
   "median_ms": 0.15017005337521117,
   "q1_ms": 0.13700141472385285,
   "q3_ms": 0.15314807186517468,
   "tflops": 200.20483709145998
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015159845352172852,
   "rel_err": 0.004941033665090799,
   "xla_temp_mb": 4.25048828125,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.5501581460703164,
   "inner": 200,
   "median_ms": 0.14255574555136263,
   "q1_ms": 0.14131499454379082,
   "q3_ms": 0.1491851225728169,
   "tflops": 210.89834685875712
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015159845352172852,
   "rel_err": 0.004941033665090799,
   "xla_temp_mb": 4.25048828125,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.5571218759287149,
   "inner": 200,
   "median_ms": 0.1466561306733638,
   "q1_ms": 0.14220574230421335,
   "q3_ms": 0.14889519778080285,
   "tflops": 205.00180206554754
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015907883644104004,
   "rel_err": 0.004941033665090799,
   "xla_temp_mb": 320.0,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.0332969450391829,
   "inner": 69,
   "median_ms": 0.5886829559407804,
   "q1_ms": 0.5868160000502848,
   "q3_ms": 0.5889148974413241,
   "tflops": 51.071244323615886
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015159845352172852,
   "rel_err": 0.004941033665090799,
   "xla_temp_mb": 16.250732421875,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.13883083290420473,
   "inner": 190,
   "median_ms": 0.13411509020155982,
   "q1_ms": 0.13080346826954106,
   "q3_ms": 0.13486335029531465,
   "tflops": 224.1714263981484
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006038188934326172,
   "rel_err": 0.002593584591522813,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.4977843570522964,
   "inner": 200,
   "median_ms": 0.033763599349185824,
   "q1_ms": 0.03373642801307142,
   "q3_ms": 0.034146204707212746,
   "tflops": 127.20703298191367
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006038188934326172,
   "rel_err": 0.002593584591522813,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.49566772300750017,
   "inner": 200,
   "median_ms": 0.03413255559280515,
   "q1_ms": 0.03412047983147204,
   "q3_ms": 0.03416717750951648,
   "tflops": 125.83198712801168
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006038188934326172,
   "rel_err": 0.002593584591522813,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.4977842268999666,
   "inner": 200,
   "median_ms": 0.033911250066012144,
   "q1_ms": 0.033852802007459104,
   "q3_ms": 0.033949530334211886,
   "tflops": 126.65316930633206
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008179187774658203,
   "rel_err": 0.0035132081247866154,
   "xla_temp_mb": 192.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.6520703558344394,
   "inner": 126,
   "median_ms": 0.2869752934202552,
   "q1_ms": 0.28627580641547135,
   "q3_ms": 0.28770711524264203,
   "tflops": 14.966331229463444
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006038188934326172,
   "rel_err": 0.002593584591522813,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.05847171205095947,
   "inner": 200,
   "median_ms": 0.0652286154218018,
   "q1_ms": 0.05483569228090346,
   "q3_ms": 0.06536951230373234,
   "tflops": 65.84483310317918
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012666940689086914,
   "rel_err": 0.0043540229089558125,
   "xla_temp_mb": 8.250244140625,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.57534185401164,
   "inner": 200,
   "median_ms": 0.10937329032458365,
   "q1_ms": 0.10769379267003387,
   "q3_ms": 0.11170250421855599,
   "tflops": 137.44110185758208
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012666940689086914,
   "rel_err": 0.0043540229089558125,
   "xla_temp_mb": 4.25048828125,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.58005625102669,
   "inner": 200,
   "median_ms": 0.11379657546058297,
   "q1_ms": 0.10999165999237448,
   "q3_ms": 0.11405807570554316,
   "tflops": 132.09875143568746
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012666940689086914,
   "rel_err": 0.0043540229089558125,
   "xla_temp_mb": 4.25048828125,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.5905174519866705,
   "inner": 200,
   "median_ms": 0.11245882487855852,
   "q1_ms": 0.10979018697980791,
   "q3_ms": 0.112516077933833,
   "tflops": 133.67012817565094
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012666940689086914,
   "rel_err": 0.005755797028541565,
   "xla_temp_mb": 320.0,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.039001098833978176,
   "inner": 53,
   "median_ms": 0.7965000003646567,
   "q1_ms": 0.7957084706264004,
   "q3_ms": 0.7976975362255888,
   "tflops": 18.873051511761222
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012666940689086914,
   "rel_err": 0.0043540229089558125,
   "xla_temp_mb": 16.250732421875,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.1318127829581499,
   "inner": 189,
   "median_ms": 0.12007378233921906,
   "q1_ms": 0.11998685397128894,
   "q3_ms": 0.12041484165897287,
   "tflops": 125.19290425558663
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006500333547592163,
   "rel_err": 0.0023746348451822996,
   "xla_temp_mb": 24.0,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.49730531801469624,
   "inner": 200,
   "median_ms": 0.1280789845623076,
   "q1_ms": 0.12762803758960217,
   "q3_ms": 0.1285910332808271,
   "tflops": 143.91119716469362
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006500333547592163,
   "rel_err": 0.0023746348451822996,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.054749823873862624,
   "inner": 200,
   "median_ms": 0.066941810073331,
   "q1_ms": 0.06308451935183257,
   "q3_ms": 0.06697715725749731,
   "tflops": 275.3436152952658
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006500333547592163,
   "rel_err": 0.0023746348451822996,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.05098122381605208,
   "inner": 200,
   "median_ms": 0.06661640945822,
   "q1_ms": 0.06378039252012968,
   "q3_ms": 0.06879970489535481,
   "tflops": 276.6885839375664
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000764697790145874,
   "rel_err": 0.0027935151010751724,
   "xla_temp_mb": 411.98731994628906,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.03004060499370098,
   "inner": 79,
   "median_ms": 0.500483002183558,
   "q1_ms": 0.4967441458160742,
   "q3_ms": 0.5005446448332712,
   "tflops": 36.828423581985795
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006500333547592163,
   "rel_err": 0.0023746348451822996,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.05626824009232223,
   "inner": 200,
   "median_ms": 0.06640736013650894,
   "q1_ms": 0.06370099203195423,
   "q3_ms": 0.06645096000283957,
   "tflops": 277.55959523327886
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012814700603485107,
   "rel_err": 0.003724043257534504,
   "xla_temp_mb": 48.00048828125,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.5683768140152097,
   "inner": 98,
   "median_ms": 0.42179358971058106,
   "q1_ms": 0.41543401549665293,
   "q3_ms": 0.4281139283973191,
   "tflops": 152.94684787472877
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012814700603485107,
   "rel_err": 0.003724043257534504,
   "xla_temp_mb": 23.804443359375,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.1321250929031521,
   "inner": 133,
   "median_ms": 0.2650814218432608,
   "q1_ms": 0.2586698234851535,
   "q3_ms": 0.26891825114958046,
   "tflops": 243.36673445996948
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012814700603485107,
   "rel_err": 0.003724043257534504,
   "xla_temp_mb": 23.804443359375,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.13054726598784328,
   "inner": 129,
   "median_ms": 0.2667661323103794,
   "q1_ms": 0.25601132660959813,
   "q3_ms": 0.26734857711681104,
   "tflops": 241.82979841286982
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013823211193084717,
   "rel_err": 0.004368611145764589,
   "xla_temp_mb": 686.82861328125,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.03328278404660523,
   "inner": 33,
   "median_ms": 1.3868718154051087,
   "q1_ms": 1.3862089112852558,
   "q3_ms": 1.3886479992475926,
   "tflops": 46.516195140324406
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012814700603485107,
   "rel_err": 0.003724043257534504,
   "xla_temp_mb": 23.804443359375,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.1302407579496503,
   "inner": 133,
   "median_ms": 0.26493367591970846,
   "q1_ms": 0.2624478910636521,
   "q3_ms": 0.2668110786804131,
   "tflops": 243.50245311793125
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007173061370849609,
   "rel_err": 0.0024814915377646685,
   "xla_temp_mb": 24.0,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.5361827171873301,
   "inner": 200,
   "median_ms": 0.09341313038021326,
   "q1_ms": 0.09326638479251415,
   "q3_ms": 0.09386808262206614,
   "tflops": 98.6585072407779
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007173061370849609,
   "rel_err": 0.0024814915377646685,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.05724893603473902,
   "inner": 200,
   "median_ms": 0.06410026922821999,
   "q1_ms": 0.06400751473847777,
   "q3_ms": 0.06459881435148418,
   "tflops": 143.77474714166533
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007173061370849609,
   "rel_err": 0.0024814915377646685,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.052480840124189854,
   "inner": 200,
   "median_ms": 0.06029703072272241,
   "q1_ms": 0.05997252941597253,
   "q3_ms": 0.06055142788682133,
   "tflops": 152.84334716878573
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011619091033935547,
   "rel_err": 0.004019577521830797,
   "xla_temp_mb": 411.98731994628906,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.45758337201550603,
   "inner": 83,
   "median_ms": 0.5046537465763739,
   "q1_ms": 0.5034393440169025,
   "q3_ms": 0.5048790438863707,
   "tflops": 18.262026315116753
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007173061370849609,
   "rel_err": 0.0024814915377646685,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 17.578125,
   "xla_out_mb": 5.859375,
   "compile_s": 0.054428393952548504,
   "inner": 200,
   "median_ms": 0.04865206079557538,
   "q1_ms": 0.04860226239543408,
   "q3_ms": 0.04885073518380523,
   "tflops": 189.42671388008586
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014818668365478516,
   "rel_err": 0.005191786680370569,
   "xla_temp_mb": 48.00048828125,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.6177412141114473,
   "inner": 132,
   "median_ms": 0.2680770368456389,
   "q1_ms": 0.25970740049766994,
   "q3_ms": 0.27272350041929516,
   "tflops": 120.32362181984759
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014818668365478516,
   "rel_err": 0.005191786680370569,
   "xla_temp_mb": 23.804443359375,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.1332472988869995,
   "inner": 146,
   "median_ms": 0.20483091684083826,
   "q1_ms": 0.20193576107873287,
   "q3_ms": 0.20545992477197353,
   "tflops": 157.4762272097048
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014818668365478516,
   "rel_err": 0.005191786680370569,
   "xla_temp_mb": 23.804443359375,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.13333006901666522,
   "inner": 144,
   "median_ms": 0.20380261088980156,
   "q1_ms": 0.2010240803226932,
   "q3_ms": 0.20481858276374018,
   "tflops": 158.27078887346144
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014818668365478516,
   "rel_err": 0.004564434289932251,
   "xla_temp_mb": 686.82861328125,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.040124944876879454,
   "inner": 29,
   "median_ms": 1.5571338349375232,
   "q1_ms": 1.5524936727537164,
   "q3_ms": 1.5574439814121561,
   "tflops": 20.714982409520506
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014818668365478516,
   "rel_err": 0.005191786680370569,
   "xla_temp_mb": 23.804443359375,
   "xla_arg_mb": 29.296875,
   "xla_out_mb": 17.578147888183594,
   "compile_s": 0.1321197219658643,
   "inner": 139,
   "median_ms": 0.20165668411863794,
   "q1_ms": 0.1986099711934225,
   "q3_ms": 0.20707498150442144,
   "tflops": 159.95502524986114
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000682830810546875,
   "rel_err": 0.0027175149880349636,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.42018455802462995,
   "inner": 200,
   "median_ms": 0.1296285295393318,
   "q1_ms": 0.12453486793674529,
   "q3_ms": 0.13064770260825753,
   "tflops": 265.0630882731304
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000682830810546875,
   "rel_err": 0.0027175149880349636,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.05532688181847334,
   "inner": 200,
   "median_ms": 0.10517940041609108,
   "q1_ms": 0.09672415035311133,
   "q3_ms": 0.10835523717105389,
   "tflops": 326.6774504520127
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000682830810546875,
   "rel_err": 0.0027175149880349636,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.05246266000904143,
   "inner": 200,
   "median_ms": 0.10416459990665317,
   "q1_ms": 0.10078108753077686,
   "q3_ms": 0.10712809220422059,
   "tflops": 329.86003305145306
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000711590051651001,
   "rel_err": 0.002831970574334264,
   "xla_temp_mb": 768.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.22422304889187217,
   "inner": 55,
   "median_ms": 0.8099185980179093,
   "q1_ms": 0.8093511452898383,
   "q3_ms": 0.810960953293199,
   "tflops": 42.42369350708529
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000682830810546875,
   "rel_err": 0.0027175149880349636,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.053120998200029135,
   "inner": 200,
   "median_ms": 0.09637472569011152,
   "q1_ms": 0.09627612482290715,
   "q3_ms": 0.10356118029449135,
   "tflops": 356.52229484400453
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015923678874969482,
   "rel_err": 0.00450897915288806,
   "xla_temp_mb": 16.500244140625,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.5320834890007973,
   "inner": 78,
   "median_ms": 0.5679877462964028,
   "q1_ms": 0.5647563390457668,
   "q3_ms": 0.5826235954793026,
   "tflops": 211.72830764775534
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015923678874969482,
   "rel_err": 0.00450897915288806,
   "xla_temp_mb": 32.500732421875,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.13277606084011495,
   "inner": 97,
   "median_ms": 0.40614306350651475,
   "q1_ms": 0.4027564217787736,
   "q3_ms": 0.41790818435514404,
   "tflops": 296.1003033013045
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015923678874969482,
   "rel_err": 0.00450897915288806,
   "xla_temp_mb": 32.500732421875,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.13106274488382041,
   "inner": 96,
   "median_ms": 0.4091828959644772,
   "q1_ms": 0.40263431340766453,
   "q3_ms": 0.4126208538461166,
   "tflops": 293.9005649406229
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015923678874969482,
   "rel_err": 0.004729839973151684,
   "xla_temp_mb": 1280.0,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.03454238106496632,
   "inner": 22,
   "median_ms": 2.1262457276779143,
   "q1_ms": 2.1235434793528505,
   "q3_ms": 2.1268321116539566,
   "tflops": 56.559353757919446
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015923678874969482,
   "rel_err": 0.00450897915288806,
   "xla_temp_mb": 32.500732421875,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.13114561489783227,
   "inner": 99,
   "median_ms": 0.4156270920477732,
   "q1_ms": 0.4049195011964802,
   "q3_ms": 0.4189729436554692,
   "tflops": 289.34370879311473
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008076190948486328,
   "rel_err": 0.002793925581499934,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.5005266789812595,
   "inner": 200,
   "median_ms": 0.09738532011397183,
   "q1_ms": 0.09663472534157336,
   "q3_ms": 0.09914409252814949,
   "tflops": 176.4112821510889
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008076190948486328,
   "rel_err": 0.002793925581499934,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.05727774603292346,
   "inner": 200,
   "median_ms": 0.07088224519975483,
   "q1_ms": 0.06726518564391881,
   "q3_ms": 0.0735415902454406,
   "tflops": 242.37196685270098
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008076190948486328,
   "rel_err": 0.002793925581499934,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.05293866898864508,
   "inner": 200,
   "median_ms": 0.06708091008476913,
   "q1_ms": 0.06705885753035545,
   "q3_ms": 0.07131869497243315,
   "tflops": 256.10668016116745
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008959054946899414,
   "rel_err": 0.0030993486288934946,
   "xla_temp_mb": 768.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.5502286360133439,
   "inner": 48,
   "median_ms": 0.9505414588299269,
   "q1_ms": 0.9456578142514143,
   "q3_ms": 0.9525287290064929,
   "tflops": 18.073771558736254
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008076190948486328,
   "rel_err": 0.002793925581499934,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.05629051895812154,
   "inner": 200,
   "median_ms": 0.07195798913016915,
   "q1_ms": 0.06813655490987003,
   "q3_ms": 0.07495368947274983,
   "tflops": 238.74860028289973
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.016491413116455078,
   "rel_err": 0.008016865700483322,
   "xla_temp_mb": 16.500244140625,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.5763023220933974,
   "inner": 106,
   "median_ms": 0.38035617007891526,
   "q1_ms": 0.3691689486896514,
   "q3_ms": 0.38737554722553436,
   "tflops": 158.08746347278787
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.016491413116455078,
   "rel_err": 0.008016865700483322,
   "xla_temp_mb": 32.500732421875,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.13139121397398412,
   "inner": 119,
   "median_ms": 0.30662382350248446,
   "q1_ms": 0.300894769206017,
   "q3_ms": 0.30723302977300493,
   "tflops": 196.10199056666838
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.016491413116455078,
   "rel_err": 0.008016865700483322,
   "xla_temp_mb": 32.500732421875,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.13364165811799467,
   "inner": 117,
   "median_ms": 0.30671630818874407,
   "q1_ms": 0.30212259739955777,
   "q3_ms": 0.3081777523088659,
   "tflops": 196.04285960235956
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01198720932006836,
   "rel_err": 0.0051156082190573215,
   "xla_temp_mb": 1280.0,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.040311614982783794,
   "inner": 17,
   "median_ms": 2.703152884564856,
   "q1_ms": 2.6948849382974647,
   "q3_ms": 2.7035952011561566,
   "tflops": 22.244225432953804
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.016491413116455078,
   "rel_err": 0.008016865700483322,
   "xla_temp_mb": 32.500732421875,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.1314633151050657,
   "inner": 110,
   "median_ms": 0.30844275500964036,
   "q1_ms": 0.29865378217602323,
   "q3_ms": 0.30926029113205994,
   "tflops": 194.94554878463802
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005548298358917236,
   "rel_err": 0.0030985635239630938,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.49461298482492566,
   "inner": 109,
   "median_ms": 0.36728299231430805,
   "q1_ms": 0.36582005009431084,
   "q3_ms": 0.369445597026668,
   "tflops": 200.73894392829965
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005548298358917236,
   "rel_err": 0.0030985635239630938,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.056779767852276564,
   "inner": 136,
   "median_ms": 0.2517208974014091,
   "q1_ms": 0.236822445431779,
   "q3_ms": 0.2527194592403248,
   "tflops": 292.8958253411474
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005548298358917236,
   "rel_err": 0.0030985635239630938,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05281553999520838,
   "inner": 143,
   "median_ms": 0.24073413358284876,
   "q1_ms": 0.2396636849295322,
   "q3_ms": 0.2472949193113229,
   "tflops": 306.26317465955225
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006076693534851074,
   "rel_err": 0.0033936568070203066,
   "xla_temp_mb": 1647.949234008789,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.01861283602192998,
   "inner": 27,
   "median_ms": 1.715882888270749,
   "q1_ms": 1.712611035650803,
   "q3_ms": 1.7192015614084624,
   "tflops": 42.96796739683231
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005548298358917236,
   "rel_err": 0.0030985635239630938,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.056368298130109906,
   "inner": 140,
   "median_ms": 0.2430879494308361,
   "q1_ms": 0.23922017509383814,
   "q3_ms": 0.25274434592574835,
   "tflops": 303.2976343443847
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008320212364196777,
   "rel_err": 0.003266180632635951,
   "xla_temp_mb": 96.00048828125,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.5794230629689991,
   "inner": 32,
   "median_ms": 1.5003231892478652,
   "q1_ms": 1.4548216095136013,
   "q3_ms": 1.5247103292495012,
   "tflops": 171.99494205602684
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008320212364196777,
   "rel_err": 0.003266180632635951,
   "xla_temp_mb": 47.608154296875,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.13375873700715601,
   "inner": 53,
   "median_ms": 0.8858961474923593,
   "q1_ms": 0.8671596988965318,
   "q3_ms": 0.8994389329773357,
   "tflops": 291.28470727684885
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008320212364196777,
   "rel_err": 0.003266180632635951,
   "xla_temp_mb": 47.608154296875,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.1344597260467708,
   "inner": 53,
   "median_ms": 0.8772912628526958,
   "q1_ms": 0.8748784432855418,
   "q3_ms": 0.8884376499204423,
   "tflops": 294.1417644590498
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011211037635803223,
   "rel_err": 0.004401002544909716,
   "xla_temp_mb": 2746.9482421875,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.03855330985970795,
   "inner": 9,
   "median_ms": 5.248559457767342,
   "q1_ms": 5.244629509333107,
   "q3_ms": 5.251248897467223,
   "tflops": 49.16549046960206
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008320212364196777,
   "rel_err": 0.003266180632635951,
   "xla_temp_mb": 47.608154296875,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.1338913170620799,
   "inner": 52,
   "median_ms": 0.8690788661344693,
   "q1_ms": 0.8684478077786759,
   "q3_ms": 0.8880751534329299,
   "tflops": 296.9212692373458
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006508588790893555,
   "rel_err": 0.002251619938760996,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.5371573241427541,
   "inner": 141,
   "median_ms": 0.2554050144732844,
   "q1_ms": 0.2529778218900798,
   "q3_ms": 0.2605811707635826,
   "tflops": 144.33545901995598
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006508588790893555,
   "rel_err": 0.002251619938760996,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05634375894442201,
   "inner": 172,
   "median_ms": 0.14301424380391836,
   "q1_ms": 0.13791384895570405,
   "q3_ms": 0.14608949463446297,
   "tflops": 257.76453463294814
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006508588790893555,
   "rel_err": 0.002251619938760996,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05419873516075313,
   "inner": 186,
   "median_ms": 0.14471184332124007,
   "q1_ms": 0.14013919656375246,
   "q3_ms": 0.14607148978518703,
   "tflops": 254.74072580339583
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007374167442321777,
   "rel_err": 0.0025510634295642376,
   "xla_temp_mb": 1647.949234008789,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.4470303209964186,
   "inner": 26,
   "median_ms": 1.7746041546790647,
   "q1_ms": 1.7740036128088832,
   "q3_ms": 1.7754066771326158,
   "tflops": 20.77308333962895
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006508588790893555,
   "rel_err": 0.002251619938760996,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05737573583610356,
   "inner": 179,
   "median_ms": 0.14345103864396752,
   "q1_ms": 0.1382520090653327,
   "q3_ms": 0.14591681856828526,
   "tflops": 256.97966601338527
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01353597640991211,
   "rel_err": 0.0039185951463878155,
   "xla_temp_mb": 96.00048828125,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.6216016530524939,
   "inner": 52,
   "median_ms": 0.8971941737959591,
   "q1_ms": 0.8688017394608603,
   "q3_ms": 0.9028062721391996,
   "tflops": 143.8083346597197
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01353597640991211,
   "rel_err": 0.0039185951463878155,
   "xla_temp_mb": 47.608154296875,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.1344268750399351,
   "inner": 73,
   "median_ms": 0.5742794521510193,
   "q1_ms": 0.5581366162976785,
   "q3_ms": 0.5748158205963978,
   "tflops": 224.6711065783881
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01353597640991211,
   "rel_err": 0.0039185951463878155,
   "xla_temp_mb": 47.608154296875,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.1347675141878426,
   "inner": 71,
   "median_ms": 0.5738483512328124,
   "q1_ms": 0.5626939565048251,
   "q3_ms": 0.5740876552242209,
   "tflops": 224.83988970747166
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01353597640991211,
   "rel_err": 0.00468168780207634,
   "xla_temp_mb": 2746.9482421875,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.04391791392117739,
   "inner": 8,
   "median_ms": 5.535922886338085,
   "q1_ms": 5.529904199647717,
   "q3_ms": 5.542023514863104,
   "tflops": 23.30668303173332
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01353597640991211,
   "rel_err": 0.0039185951463878155,
   "xla_temp_mb": 47.608154296875,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.13477759389206767,
   "inner": 72,
   "median_ms": 0.5726807220425042,
   "q1_ms": 0.5555184163515352,
   "q3_ms": 0.5763149513061055,
   "tflops": 225.29831201551056
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00045752525329589844,
   "rel_err": 0.0028358343988656998,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.420315558090806,
   "inner": 91,
   "median_ms": 0.4968916495832113,
   "q1_ms": 0.4957884788042405,
   "q3_ms": 0.49957972541860823,
   "tflops": 276.597430420259
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00045752525329589844,
   "rel_err": 0.0028358343988656998,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.05306143802590668,
   "inner": 102,
   "median_ms": 0.3784900878136065,
   "q1_ms": 0.36657199049003275,
   "q3_ms": 0.39375569059641335,
   "tflops": 363.12431394421094
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00045752525329589844,
   "rel_err": 0.0028358343988656998,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.052137140883132815,
   "inner": 99,
   "median_ms": 0.37978185020913985,
   "q1_ms": 0.3696534890832022,
   "q3_ms": 0.39633148881333946,
   "tflops": 361.88920928241976
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006755292415618896,
   "rel_err": 0.0041870675049722195,
   "xla_temp_mb": 3072.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.2624216100666672,
   "inner": 16,
   "median_ms": 3.0240288615459576,
   "q1_ms": 3.0234004370868206,
   "q3_ms": 3.02438728249399,
   "tflops": 45.44895560346532
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00045752525329589844,
   "rel_err": 0.0028358343988656998,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.05437371414154768,
   "inner": 102,
   "median_ms": 0.39318759752181814,
   "q1_ms": 0.37977523421503456,
   "q3_ms": 0.39761195612121736,
   "tflops": 349.5505818043344
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007363855838775635,
   "rel_err": 0.003625117940828204,
   "xla_temp_mb": 33.000244140625,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.5460895490832627,
   "inner": 23,
   "median_ms": 2.1974519603764233,
   "q1_ms": 2.172768996704532,
   "q3_ms": 2.253900940084587,
   "tflops": 218.9064178993922
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007363855838775635,
   "rel_err": 0.003625117940828204,
   "xla_temp_mb": 65.000732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13070700597018003,
   "inner": 33,
   "median_ms": 1.433954087340019,
   "q1_ms": 1.414836123303482,
   "q3_ms": 1.4551521779856447,
   "tflops": 335.46146379367076
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007363855838775635,
   "rel_err": 0.003625117940828204,
   "xla_temp_mb": 65.000732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.1317147130612284,
   "inner": 33,
   "median_ms": 1.4528113092775599,
   "q1_ms": 1.413215043474779,
   "q3_ms": 1.47344578748964,
   "tflops": 331.10723607403986
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008107870817184448,
   "rel_err": 0.0038248642813414335,
   "xla_temp_mb": 5120.000015258789,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.040669484063982964,
   "inner": 6,
   "median_ms": 8.282971180354556,
   "q1_ms": 8.280047254326444,
   "q3_ms": 8.29048043427368,
   "tflops": 58.07533633497552
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007363855838775635,
   "rel_err": 0.003625117940828204,
   "xla_temp_mb": 65.000732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13484337413683534,
   "inner": 33,
   "median_ms": 1.4123326365017528,
   "q1_ms": 1.4084309840461975,
   "q3_ms": 1.4174863816041388,
   "tflops": 340.59705533923835
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004129528999328613,
   "rel_err": 0.0016415518475696445,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.5003275491762906,
   "inner": 125,
   "median_ms": 0.3316164966672659,
   "q1_ms": 0.3294507758691907,
   "q3_ms": 0.33533396013081074,
   "tflops": 207.22574849752147
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004129528999328613,
   "rel_err": 0.0016415518475696445,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.054862982826307416,
   "inner": 138,
   "median_ms": 0.23780765173875767,
   "q1_ms": 0.22488656922823924,
   "q3_ms": 0.23931667016333213,
   "tflops": 288.9708393886813
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004129528999328613,
   "rel_err": 0.0016415518475696445,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.05494763283059001,
   "inner": 147,
   "median_ms": 0.2320342393097829,
   "q1_ms": 0.22583016983176374,
   "q3_ms": 0.2368578598612831,
   "tflops": 296.16093271585845
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0068624019622802734,
   "rel_err": 0.002727911341935396,
   "xla_temp_mb": 3072.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.5487603510264307,
   "inner": 13,
   "median_ms": 3.736681609342878,
   "q1_ms": 3.72361780430835,
   "q3_ms": 3.7486861871842008,
   "tflops": 18.390508991769522
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004129528999328613,
   "rel_err": 0.0016415518475696445,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.05748815601691604,
   "inner": 157,
   "median_ms": 0.23530524929949811,
   "q1_ms": 0.22553203176635844,
   "q3_ms": 0.2380772834869137,
   "tflops": 292.04395966761194
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015132904052734375,
   "rel_err": 0.003882892895489931,
   "xla_temp_mb": 33.000244140625,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.5757806529290974,
   "inner": 35,
   "median_ms": 1.3794780575803347,
   "q1_ms": 1.3759300711431672,
   "q3_ms": 1.3852450298145413,
   "tflops": 174.35447215295287
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015132904052734375,
   "rel_err": 0.003882892895489931,
   "xla_temp_mb": 65.000732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13595590111799538,
   "inner": 50,
   "median_ms": 0.9253255603834987,
   "q1_ms": 0.9119962994009256,
   "q3_ms": 0.9301450382918119,
   "tflops": 259.9281581244961
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015132904052734375,
   "rel_err": 0.003882892895489931,
   "xla_temp_mb": 65.000732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13564767292700708,
   "inner": 50,
   "median_ms": 0.9105873806402087,
   "q1_ms": 0.9003061288967729,
   "q3_ms": 0.9305737400427461,
   "tflops": 264.13518756091077
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015132904052734375,
   "rel_err": 0.003882892895489931,
   "xla_temp_mb": 5120.000015258789,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.049433848122134805,
   "inner": 4,
   "median_ms": 9.989916521590203,
   "q1_ms": 9.976121509680524,
   "q3_ms": 9.996584005421028,
   "tflops": 24.07609393493852
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015132904052734375,
   "rel_err": 0.003882892895489931,
   "xla_temp_mb": 65.000732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13809308595955372,
   "inner": 50,
   "median_ms": 0.9262433601543307,
   "q1_ms": 0.9076946973800659,
   "q3_ms": 0.9314271295443177,
   "tflops": 259.6705994587911
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00031141191720962524,
   "rel_err": 0.0025766564067453146,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.42445946601219475,
   "inner": 25,
   "median_ms": 1.919437712058425,
   "q1_ms": 1.916142520494759,
   "q3_ms": 1.9296598620712757,
   "tflops": 286.41503208688977
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00031141191720962524,
   "rel_err": 0.0025766564067453146,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05602489993907511,
   "inner": 33,
   "median_ms": 1.5016108765407945,
   "q1_ms": 1.474079270282704,
   "q3_ms": 1.5182970338642146,
   "tflops": 366.11070316329364
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00031141191720962524,
   "rel_err": 0.0025766564067453146,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05599656002596021,
   "inner": 33,
   "median_ms": 1.491209061703447,
   "q1_ms": 1.4875948316220082,
   "q3_ms": 1.4913999651925582,
   "tflops": 368.6644803915016
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003448277711868286,
   "rel_err": 0.0028531427960842848,
   "xla_temp_mb": 12288.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.029638994950801134,
   "inner": 4,
   "median_ms": 11.722026509232819,
   "q1_ms": 11.720011505531147,
   "q3_ms": 11.722750234184787,
   "tflops": 46.89938326406159
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00031141191720962524,
   "rel_err": 0.0025766564067453146,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.057911094976589084,
   "inner": 32,
   "median_ms": 1.4877476278343238,
   "q1_ms": 1.4452755349338986,
   "q3_ms": 1.567193499795394,
   "tflops": 369.5222251426241
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006060153245925903,
   "rel_err": 0.0036204783245921135,
   "xla_temp_mb": 66.000244140625,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.5423598398920149,
   "inner": 6,
   "median_ms": 8.490913004303971,
   "q1_ms": 8.4445207612589,
   "q3_ms": 8.744112836817902,
   "tflops": 226.61230277976787
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006060153245925903,
   "rel_err": 0.0036204783245921135,
   "xla_temp_mb": 130.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.13651272095739841,
   "inner": 9,
   "median_ms": 5.49245910305116,
   "q1_ms": 5.434397724457085,
   "q3_ms": 5.577927274215552,
   "tflops": 350.32492960013167
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006060153245925903,
   "rel_err": 0.0036204783245921135,
   "xla_temp_mb": 130.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.13655729009769857,
   "inner": 9,
   "median_ms": 5.460833664983511,
   "q1_ms": 5.262114896645976,
   "q3_ms": 5.476949762346017,
   "tflops": 352.35377355406223
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006311088800430298,
   "rel_err": 0.004330759402364492,
   "xla_temp_mb": 20480.0,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.06460637110285461,
   "inner": 1,
   "median_ms": 32.366849947720766,
   "q1_ms": 32.34734549187124,
   "q3_ms": 32.38391503691673,
   "tflops": 59.448026351526245
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006060153245925903,
   "rel_err": 0.0036204783245921135,
   "xla_temp_mb": 130.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.1376003078185022,
   "inner": 9,
   "median_ms": 5.3235096598251,
   "q1_ms": 5.272343677158156,
   "q3_ms": 5.573592886018257,
   "tflops": 361.443008759604
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007561683654785156,
   "rel_err": 0.003005886683240533,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.5019453940913081,
   "inner": 43,
   "median_ms": 1.114728768507755,
   "q1_ms": 1.1055585576307982,
   "q3_ms": 1.1352169554853857,
   "tflops": 246.58725486377156
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007561683654785156,
   "rel_err": 0.003005886683240533,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05523768602870405,
   "inner": 56,
   "median_ms": 0.7995604662158128,
   "q1_ms": 0.7923019043768623,
   "q3_ms": 0.8228587061499378,
   "tflops": 343.78626577793625
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007561683654785156,
   "rel_err": 0.003005886683240533,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.0552148160059005,
   "inner": 57,
   "median_ms": 0.8255010348205504,
   "q1_ms": 0.8082373941872726,
   "q3_ms": 0.8505399220443347,
   "tflops": 332.9831161311066
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008063316345214844,
   "rel_err": 0.003205293556675315,
   "xla_temp_mb": 12288.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.4955595429055393,
   "inner": 3,
   "median_ms": 14.991421640540162,
   "q1_ms": 14.970621831404667,
   "q3_ms": 15.000418371831376,
   "tflops": 18.335679799750853
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007561683654785156,
   "rel_err": 0.003005886683240533,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05901136598549783,
   "inner": 56,
   "median_ms": 0.8316203768897269,
   "q1_ms": 0.7954952847545168,
   "q3_ms": 0.8471008685384211,
   "tflops": 330.53291451569237
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01398777961730957,
   "rel_err": 0.0038633455988019705,
   "xla_temp_mb": 66.000244140625,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.6113709609489888,
   "inner": 9,
   "median_ms": 5.241883246021138,
   "q1_ms": 5.23975377695428,
   "q3_ms": 5.260532062190274,
   "tflops": 183.53569302297282
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01398777961730957,
   "rel_err": 0.0038633455988019705,
   "xla_temp_mb": 130.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.1386431148275733,
   "inner": 16,
   "median_ms": 3.027764134458266,
   "q1_ms": 3.0261634310591035,
   "q3_ms": 3.1469928435399197,
   "tflops": 317.75020496308775
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01398777961730957,
   "rel_err": 0.0038633455988019705,
   "xla_temp_mb": 130.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.13939219200983644,
   "inner": 16,
   "median_ms": 3.00752287148498,
   "q1_ms": 3.0017306853551418,
   "q3_ms": 3.1535934103885666,
   "tflops": 319.8887308308221
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01398777961730957,
   "rel_err": 0.004312852397561073,
   "xla_temp_mb": 20480.0,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.0776635438669473,
   "inner": 1,
   "median_ms": 38.71144214645028,
   "q1_ms": 38.58794807456434,
   "q3_ms": 38.741967058740556,
   "tflops": 24.8524110950028
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01398777961730957,
   "rel_err": 0.0038633455988019705,
   "xla_temp_mb": 130.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.14467269787564874,
   "inner": 16,
   "median_ms": 3.0280853679869324,
   "q1_ms": 3.0055788083700463,
   "q3_ms": 3.0974554683780298,
   "tflops": 317.7164965278323
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027510523796081543,
   "rel_err": 0.0036955196410417557,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.27958805300295353,
   "inner": 200,
   "median_ms": 0.037999694468453526,
   "q1_ms": 0.03786759451031685,
   "q3_ms": 0.038041192456148565,
   "tflops": 7.064147745262766
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027510523796081543,
   "rel_err": 0.0036955196410417557,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.28169999714009464,
   "inner": 200,
   "median_ms": 0.025178685318678617,
   "q1_ms": 0.024908980121836066,
   "q3_ms": 0.025400230078957975,
   "tflops": 10.661218113753666
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027510523796081543,
   "rel_err": 0.0036955196410417557,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.2824166249483824,
   "inner": 200,
   "median_ms": 0.025109935086220503,
   "q1_ms": 0.024999277666211128,
   "q3_ms": 0.025545277749188244,
   "tflops": 10.690408202102777
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0032942891120910645,
   "rel_err": 0.004425255581736565,
   "xla_temp_mb": 3.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.17832715390250087,
   "inner": 200,
   "median_ms": 0.03465610556304455,
   "q1_ms": 0.034621680388227105,
   "q3_ms": 0.0347095524193719,
   "tflops": 7.745690164513046
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027510523796081543,
   "rel_err": 0.0036955196410417557,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.3834223640151322,
   "inner": 200,
   "median_ms": 0.05230060080066323,
   "q1_ms": 0.05147501069586724,
   "q3_ms": 0.052455299883149564,
   "tflops": 5.1325501407355905
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.003998756408691406,
   "rel_err": 0.004262624774128199,
   "xla_temp_mb": 2.031494140625,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.6263463089708239,
   "inner": 200,
   "median_ms": 0.0365924509242177,
   "q1_ms": 0.03651264763902873,
   "q3_ms": 0.03693357284646481,
   "tflops": 25.67535303786394
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.003998756408691406,
   "rel_err": 0.004262624774128199,
   "xla_temp_mb": 1.03173828125,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.4525914320256561,
   "inner": 200,
   "median_ms": 0.03957533976063132,
   "q1_ms": 0.03941594040952623,
   "q3_ms": 0.039755639736540616,
   "tflops": 23.740139735568814
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.003998756408691406,
   "rel_err": 0.004262624774128199,
   "xla_temp_mb": 1.03173828125,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.4463101190049201,
   "inner": 200,
   "median_ms": 0.03865198930725455,
   "q1_ms": 0.038511594757437706,
   "q3_ms": 0.038985792780295014,
   "tflops": 24.307263683933126
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004075884819030762,
   "rel_err": 0.004262624774128199,
   "xla_temp_mb": 5.0,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.03179212100803852,
   "inner": 200,
   "median_ms": 0.05856673466041684,
   "q1_ms": 0.058350362232886255,
   "q3_ms": 0.05869013548363,
   "tflops": 16.041940897807827
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.003998756408691406,
   "rel_err": 0.004262624774128199,
   "xla_temp_mb": 4.031982421875,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.9537154780700803,
   "inner": 200,
   "median_ms": 0.1113095402251929,
   "q1_ms": 0.11128184269182384,
   "q3_ms": 0.1114249147940427,
   "tflops": 8.440643040113427
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005361318588256836,
   "rel_err": 0.001854726462624967,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.36887998413294554,
   "inner": 200,
   "median_ms": 0.038452944718301296,
   "q1_ms": 0.038310542004182935,
   "q3_ms": 0.03866141720209271,
   "tflops": 3.4904408227575976
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005361318588256836,
   "rel_err": 0.001854726462624967,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.36772252805531025,
   "inner": 200,
   "median_ms": 0.026051325257867575,
   "q1_ms": 0.02570465498138219,
   "q3_ms": 0.026344250072725117,
   "tflops": 5.1520499119124805
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005361318588256836,
   "rel_err": 0.001854726462624967,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.36704474897123873,
   "inner": 200,
   "median_ms": 0.03338115522637963,
   "q1_ms": 0.033101479639299214,
   "q3_ms": 0.03342098440043628,
   "tflops": 4.0207634244465495
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006750822067260742,
   "rel_err": 0.002335419412702322,
   "xla_temp_mb": 3.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.2825764052104205,
   "inner": 200,
   "median_ms": 0.034562755608931184,
   "q1_ms": 0.03448815492447466,
   "q3_ms": 0.03474157827440649,
   "tflops": 3.8833051831468404
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005361318588256836,
   "rel_err": 0.001854726462624967,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.4309009318239987,
   "inner": 200,
   "median_ms": 0.06660821498371661,
   "q1_ms": 0.06548994511831552,
   "q3_ms": 0.06696883530821651,
   "tflops": 2.0150326507445304
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01621723175048828,
   "rel_err": 0.00388345867395401,
   "xla_temp_mb": 2.031494140625,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.784612599061802,
   "inner": 200,
   "median_ms": 0.04989226465113461,
   "q1_ms": 0.04976418800652027,
   "q3_ms": 0.05032873013988137,
   "tflops": 9.415528665310186
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01621723175048828,
   "rel_err": 0.00388345867395401,
   "xla_temp_mb": 1.03173828125,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.46985696302726865,
   "inner": 200,
   "median_ms": 0.04019833984784782,
   "q1_ms": 0.04003636247944087,
   "q3_ms": 0.04034166282508522,
   "tflops": 11.686105689390818
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01621723175048828,
   "rel_err": 0.00388345867395401,
   "xla_temp_mb": 1.03173828125,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.47549624810926616,
   "inner": 200,
   "median_ms": 0.06039123050868511,
   "q1_ms": 0.04796329187229276,
   "q3_ms": 0.06064690765924752,
   "tflops": 7.778646734685122
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01621723175048828,
   "rel_err": 0.005187384784221649,
   "xla_temp_mb": 5.0,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.7761802719905972,
   "inner": 200,
   "median_ms": 0.07628843537531793,
   "q1_ms": 0.07619636482559144,
   "q3_ms": 0.07634163717739284,
   "tflops": 6.157709824417044
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01621723175048828,
   "rel_err": 0.00388345867395401,
   "xla_temp_mb": 4.031982421875,
   "xla_arg_mb": 5.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.02151871006935835,
   "inner": 200,
   "median_ms": 0.11332123540341854,
   "q1_ms": 0.11305733525659889,
   "q3_ms": 0.11358155985362828,
   "tflops": 4.145401753940186
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011283457279205322,
   "rel_err": 0.0027316263876855373,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.3224048928823322,
   "inner": 200,
   "median_ms": 0.02197809051722288,
   "q1_ms": 0.021757989889010787,
   "q3_ms": 0.02222591545432806,
   "tflops": 195.4204025429005
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011283457279205322,
   "rel_err": 0.0027316263876855373,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.3263449619989842,
   "inner": 200,
   "median_ms": 0.022545489482581615,
   "q1_ms": 0.022341214935295284,
   "q3_ms": 0.022670990438200533,
   "tflops": 190.50228646923998
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011283457279205322,
   "rel_err": 0.0027316263876855373,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.34105254197493196,
   "inner": 200,
   "median_ms": 0.022214240161702037,
   "q1_ms": 0.022196187637746334,
   "q3_ms": 0.022351534571498632,
   "tflops": 193.34297570999715
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012100040912628174,
   "rel_err": 0.002929314039647579,
   "xla_temp_mb": 52.00001525878906,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.018514588940888643,
   "inner": 200,
   "median_ms": 0.05962362978607416,
   "q1_ms": 0.05949671263806522,
   "q3_ms": 0.05990566045511514,
   "tflops": 72.03464987640089
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011283457279205322,
   "rel_err": 0.0027316263876855373,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.05701475217938423,
   "inner": 200,
   "median_ms": 0.0509457103908062,
   "q1_ms": 0.050903654191643,
   "q3_ms": 0.05115668289363384,
   "tflops": 84.30478764655878
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017972588539123535,
   "rel_err": 0.003949459176510572,
   "xla_temp_mb": 8.125244140625,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.6129108360037208,
   "inner": 200,
   "median_ms": 0.07436739513650537,
   "q1_ms": 0.07160822744481266,
   "q3_ms": 0.07553141273092479,
   "tflops": 202.1367765861268
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017972588539123535,
   "rel_err": 0.003949459176510572,
   "xla_temp_mb": 4.12548828125,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.4400263971183449,
   "inner": 200,
   "median_ms": 0.07323745056055486,
   "q1_ms": 0.06977405748330057,
   "q3_ms": 0.07493106706533581,
   "tflops": 205.25544541683334
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017972588539123535,
   "rel_err": 0.003949459176510572,
   "xla_temp_mb": 4.12548828125,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.44478632300160825,
   "inner": 200,
   "median_ms": 0.07253379444591701,
   "q1_ms": 0.06943180749658495,
   "q3_ms": 0.07270892267115414,
   "tflops": 207.24664483406443
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0020554959774017334,
   "rel_err": 0.004599467851221561,
   "xla_temp_mb": 80.00001525878906,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.03494559298269451,
   "inner": 160,
   "median_ms": 0.1749440751154907,
   "q1_ms": 0.1748879840306472,
   "q3_ms": 0.17500988760730252,
   "tflops": 85.92680561531594
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017972588539123535,
   "rel_err": 0.003949459176510572,
   "xla_temp_mb": 16.125732421875,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.1340898780617863,
   "inner": 200,
   "median_ms": 0.09597973432391882,
   "q1_ms": 0.0958810601150617,
   "q3_ms": 0.09623625490348786,
   "tflops": 156.62041202643573
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006621599197387695,
   "rel_err": 0.002405368722975254,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.37553706602193415,
   "inner": 200,
   "median_ms": 0.022253035567700863,
   "q1_ms": 0.02220946247689426,
   "q3_ms": 0.02229266450740397,
   "tflops": 96.50295311247164
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006621599197387695,
   "rel_err": 0.002405368722975254,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.3727867540437728,
   "inner": 200,
   "median_ms": 0.030558659927919507,
   "q1_ms": 0.030349865555763245,
   "q3_ms": 0.030654040165245533,
   "tflops": 70.27414333826793
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006621599197387695,
   "rel_err": 0.002405368722975254,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.37328379205428064,
   "inner": 200,
   "median_ms": 0.03056391491554678,
   "q1_ms": 0.030374042107723653,
   "q3_ms": 0.030696315225213766,
   "tflops": 70.26206079731138
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01019430160522461,
   "rel_err": 0.0037031921092420816,
   "xla_temp_mb": 52.00001525878906,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.7401493121869862,
   "inner": 200,
   "median_ms": 0.0892982049845159,
   "q1_ms": 0.08923475223127753,
   "q3_ms": 0.0894355500349775,
   "tflops": 24.048452579448472
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006621599197387695,
   "rel_err": 0.002405368722975254,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.05873396713286638,
   "inner": 200,
   "median_ms": 0.06720691453665495,
   "q1_ms": 0.06673776195384562,
   "q3_ms": 0.06740931246895343,
   "tflops": 31.953314072003007
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013897895812988281,
   "rel_err": 0.006165427155792713,
   "xla_temp_mb": 8.125244140625,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.763766375835985,
   "inner": 200,
   "median_ms": 0.05852438509464264,
   "q1_ms": 0.058058887952938676,
   "q3_ms": 0.059711160720326006,
   "tflops": 128.4283936660795
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013897895812988281,
   "rel_err": 0.006165427155792713,
   "xla_temp_mb": 4.12548828125,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.45436320709995925,
   "inner": 200,
   "median_ms": 0.06612751516513526,
   "q1_ms": 0.06599019223358482,
   "q3_ms": 0.06626828981097788,
   "tflops": 113.66210796262912
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013897895812988281,
   "rel_err": 0.006165427155792713,
   "xla_temp_mb": 4.12548828125,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.4727944650221616,
   "inner": 200,
   "median_ms": 0.06621401524171233,
   "q1_ms": 0.06617324019316584,
   "q3_ms": 0.06628029281273484,
   "tflops": 113.51362306850531
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013897895812988281,
   "rel_err": 0.006165427155792713,
   "xla_temp_mb": 80.00001525878906,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.03929809993132949,
   "inner": 126,
   "median_ms": 0.26944568306798977,
   "q1_ms": 0.2692013143724392,
   "q3_ms": 0.27062404727090206,
   "tflops": 27.895020185212704
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013897895812988281,
   "rel_err": 0.006165427155792713,
   "xla_temp_mb": 16.125732421875,
   "xla_arg_mb": 20.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 1.0341696948744357,
   "inner": 200,
   "median_ms": 0.09695098036900163,
   "q1_ms": 0.09692902967799455,
   "q3_ms": 0.09706963028293103,
   "tflops": 77.52570153899309
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011000633239746094,
   "rel_err": 0.002210719743743539,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.37245421507395804,
   "inner": 200,
   "median_ms": 0.059816784923896194,
   "q1_ms": 0.05956918466836214,
   "q3_ms": 0.05984425777569413,
   "tflops": 98.6050990119951
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011000633239746094,
   "rel_err": 0.002210719743743539,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.3748027279507369,
   "inner": 200,
   "median_ms": 0.05981593509204686,
   "q1_ms": 0.05969343241304159,
   "q3_ms": 0.05993193248286843,
   "tflops": 98.60649993891396
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011000633239746094,
   "rel_err": 0.002210719743743539,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.37193914502859116,
   "inner": 200,
   "median_ms": 0.06005578557960689,
   "q1_ms": 0.05974023544695228,
   "q3_ms": 0.060088507598266006,
   "tflops": 98.21268580662547
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002395838499069214,
   "rel_err": 0.004814747720956802,
   "xla_temp_mb": 65.91798400878906,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.018599048955366015,
   "inner": 198,
   "median_ms": 0.09621937895626431,
   "q1_ms": 0.09399746436475202,
   "q3_ms": 0.096800237320213,
   "tflops": 61.29991758397229
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011000633239746094,
   "rel_err": 0.002210719743743539,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.4061016011983156,
   "inner": 200,
   "median_ms": 0.06456141942180693,
   "q1_ms": 0.06449947250075638,
   "q3_ms": 0.06483854493126273,
   "tflops": 91.35858617767238
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002751857042312622,
   "rel_err": 0.0050126416608691216,
   "xla_temp_mb": 40.00048828125,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.7448338789399713,
   "inner": 168,
   "median_ms": 0.17960128579510465,
   "q1_ms": 0.17912090143432752,
   "q3_ms": 0.18058015155561624,
   "tflops": 114.9426069452042
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002751857042312622,
   "rel_err": 0.0050126416608691216,
   "xla_temp_mb": 30.078651428222656,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.5032411299180239,
   "inner": 157,
   "median_ms": 0.17533117199352213,
   "q1_ms": 0.1752242537891599,
   "q3_ms": 0.1767288372929024,
   "tflops": 117.74198372872746
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002751857042312622,
   "rel_err": 0.0050126416608691216,
   "xla_temp_mb": 30.078651428222656,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.5026589019689709,
   "inner": 159,
   "median_ms": 0.17737547732280487,
   "q1_ms": 0.17625252487805654,
   "q3_ms": 0.17767912948375228,
   "tflops": 116.3849722159189
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027889013290405273,
   "rel_err": 0.0050126416608691216,
   "xla_temp_mb": 109.93678283691406,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.03658725810237229,
   "inner": 120,
   "median_ms": 0.2917953592259437,
   "q1_ms": 0.2899518158907691,
   "q3_ms": 0.2921739340915034,
   "tflops": 70.74766389281403
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002751857042312622,
   "rel_err": 0.0050126416608691216,
   "xla_temp_mb": 18.897216796875,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.9846655121073127,
   "inner": 170,
   "median_ms": 0.13148004654794931,
   "q1_ms": 0.13095178257893114,
   "q3_ms": 0.1315901323002489,
   "tflops": 157.0112008780847
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006784915924072266,
   "rel_err": 0.0023472141474485397,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.3982698619365692,
   "inner": 200,
   "median_ms": 0.04916146514005959,
   "q1_ms": 0.049103485071100295,
   "q3_ms": 0.04921846557408571,
   "tflops": 59.98844809848613
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006784915924072266,
   "rel_err": 0.0023472141474485397,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.40467746485956013,
   "inner": 200,
   "median_ms": 0.04923961008898914,
   "q1_ms": 0.049221289809793234,
   "q3_ms": 0.04926063760649413,
   "tflops": 59.89324437521238
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006784915924072266,
   "rel_err": 0.0023472141474485397,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.3992117098532617,
   "inner": 200,
   "median_ms": 0.04921286483295262,
   "q1_ms": 0.049173664883710444,
   "q3_ms": 0.04923788772430271,
   "tflops": 59.92579399737135
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007506132125854492,
   "rel_err": 0.0025967159308493137,
   "xla_temp_mb": 65.91798400878906,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.6863689008168876,
   "inner": 200,
   "median_ms": 0.13532397453673184,
   "q1_ms": 0.13507205236237496,
   "q3_ms": 0.13542567321565002,
   "tflops": 21.79303416187722
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006784915924072266,
   "rel_err": 0.0023472141474485397,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.44837676291354,
   "inner": 200,
   "median_ms": 0.05382339935749769,
   "q1_ms": 0.053558400250039995,
   "q3_ms": 0.053998049697838724,
   "tflops": 54.79252583828455
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012672662734985352,
   "rel_err": 0.004155972506850958,
   "xla_temp_mb": 40.00048828125,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.7845551690552384,
   "inner": 200,
   "median_ms": 0.12939748936332762,
   "q1_ms": 0.1293225446715951,
   "q3_ms": 0.12990485993213952,
   "tflops": 79.76909019476943
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012672662734985352,
   "rel_err": 0.004155972506850958,
   "xla_temp_mb": 30.07861328125,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.5035284098703414,
   "inner": 200,
   "median_ms": 0.12481459998525679,
   "q1_ms": 0.12449557718355209,
   "q3_ms": 0.12534635490737855,
   "tflops": 82.69801770962079
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012672662734985352,
   "rel_err": 0.004155972506850958,
   "xla_temp_mb": 30.07861328125,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.5097135729156435,
   "inner": 200,
   "median_ms": 0.12495310511440039,
   "q1_ms": 0.12471385241951793,
   "q3_ms": 0.12510890024714172,
   "tflops": 82.60635052286057
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012672662734985352,
   "rel_err": 0.00421340623870492,
   "xla_temp_mb": 109.93678283691406,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.03925111098214984,
   "inner": 105,
   "median_ms": 0.3520147342767034,
   "q1_ms": 0.3512800671160221,
   "q3_ms": 0.35264078039853347,
   "tflops": 29.322408964524733
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012672662734985352,
   "rel_err": 0.004155972506850958,
   "xla_temp_mb": 18.897216796875,
   "xla_arg_mb": 23.4375,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.02264803694561124,
   "inner": 172,
   "median_ms": 0.11191928358531968,
   "q1_ms": 0.11184742149008914,
   "q3_ms": 0.11208730541321255,
   "tflops": 92.22646597921856
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007505267858505249,
   "rel_err": 0.002722856355831027,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.3217971450649202,
   "inner": 200,
   "median_ms": 0.059329235227778554,
   "q1_ms": 0.05913146014790982,
   "q3_ms": 0.05936431000009179,
   "tflops": 289.5683572869689
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007505267858505249,
   "rel_err": 0.002722856355831027,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.32511990517377853,
   "inner": 200,
   "median_ms": 0.062470374396070845,
   "q1_ms": 0.05964471260085702,
   "q3_ms": 0.06266957789193839,
   "tflops": 275.00826351827584
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007505267858505249,
   "rel_err": 0.002722856355831027,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.3270147400908172,
   "inner": 200,
   "median_ms": 0.06181302480399608,
   "q1_ms": 0.05992871010676026,
   "q3_ms": 0.06397602264769375,
   "tflops": 277.93283435774134
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009716004133224487,
   "rel_err": 0.003524895291775465,
   "xla_temp_mb": 192.00001525878906,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.019132476998493075,
   "inner": 138,
   "median_ms": 0.25218733977796376,
   "q1_ms": 0.2518176333978772,
   "q3_ms": 0.2524019413006803,
   "tflops": 68.12344029294204
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007505267858505249,
   "rel_err": 0.002722856355831027,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.056836241856217384,
   "inner": 200,
   "median_ms": 0.04725011996924877,
   "q1_ms": 0.04720576689578593,
   "q3_ms": 0.04728476982563734,
   "tflops": 363.5941918281047
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015575289726257324,
   "rel_err": 0.0038550400640815496,
   "xla_temp_mb": 16.250244140625,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.6156550680752844,
   "inner": 134,
   "median_ms": 0.2654745597948334,
   "q1_ms": 0.2593301009011469,
   "q3_ms": 0.26804891783299284,
   "tflops": 226.49832130984564
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015575289726257324,
   "rel_err": 0.0038550400640815496,
   "xla_temp_mb": 8.25048828125,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.4295217760372907,
   "inner": 150,
   "median_ms": 0.25617195448527735,
   "q1_ms": 0.24674168011794487,
   "q3_ms": 0.2599996468052268,
   "tflops": 234.72336097375464
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015575289726257324,
   "rel_err": 0.0038550400640815496,
   "xla_temp_mb": 8.25048828125,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.42133169807493687,
   "inner": 148,
   "median_ms": 0.25591476367333454,
   "q1_ms": 0.25554851320496685,
   "q3_ms": 0.26349105708603116,
   "tflops": 234.95925471792265
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0019018948078155518,
   "rel_err": 0.003942885436117649,
   "xla_temp_mb": 320.00001525878906,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.03595892991870642,
   "inner": 64,
   "median_ms": 0.6579028595297132,
   "q1_ms": 0.6558339537150459,
   "q3_ms": 0.6581563757208642,
   "tflops": 91.39577564229198
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015575289726257324,
   "rel_err": 0.0038550400640815496,
   "xla_temp_mb": 32.250732421875,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.1531360240187496,
   "inner": 144,
   "median_ms": 0.2273536746441904,
   "q1_ms": 0.21365454530395153,
   "q3_ms": 0.22846530191600323,
   "tflops": 264.47578750641713
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005032062530517578,
   "rel_err": 0.0017408216372132301,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.3756416959222406,
   "inner": 200,
   "median_ms": 0.05456989980302751,
   "q1_ms": 0.0544668227666989,
   "q3_ms": 0.0545738241635263,
   "tflops": 157.41158812835928
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005032062530517578,
   "rel_err": 0.0017408216372132301,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.3736600410193205,
   "inner": 200,
   "median_ms": 0.05411020014435053,
   "q1_ms": 0.054038624512031674,
   "q3_ms": 0.054383547976613045,
   "tflops": 158.748897048699
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005032062530517578,
   "rel_err": 0.0017408216372132301,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.3742310998495668,
   "inner": 200,
   "median_ms": 0.05444959970191121,
   "q1_ms": 0.0544349750271067,
   "q3_ms": 0.05448004754725844,
   "tflops": 157.75937084985566
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005884408950805664,
   "rel_err": 0.002035687444731593,
   "xla_temp_mb": 192.00001525878906,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.6201782459393144,
   "inner": 115,
   "median_ms": 0.3114912254006966,
   "q1_ms": 0.3109534433030564,
   "q3_ms": 0.3118960014746889,
   "tflops": 27.576810810481305
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005032062530517578,
   "rel_err": 0.0017408216372132301,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.05923623498529196,
   "inner": 200,
   "median_ms": 0.06679031415842474,
   "q1_ms": 0.050929537974298,
   "q3_ms": 0.06684969004709274,
   "tflops": 128.6104834246612
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013173818588256836,
   "rel_err": 0.003089054487645626,
   "xla_temp_mb": 16.250244140625,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.7463977299630642,
   "inner": 179,
   "median_ms": 0.18800322913341017,
   "q1_ms": 0.18350905010362242,
   "q3_ms": 0.18824465304883664,
   "tflops": 159.9162483037222
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013173818588256836,
   "rel_err": 0.003089054487645626,
   "xla_temp_mb": 8.25048828125,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.48212246084585786,
   "inner": 166,
   "median_ms": 0.1943730244255928,
   "q1_ms": 0.19007144041404306,
   "q3_ms": 0.19473169652565595,
   "tflops": 154.67563547383594
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013173818588256836,
   "rel_err": 0.003089054487645626,
   "xla_temp_mb": 8.25048828125,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.47479035099968314,
   "inner": 174,
   "median_ms": 0.19419337381962043,
   "q1_ms": 0.19076335626579394,
   "q3_ms": 0.19442570133498688,
   "tflops": 154.81872774879605
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013173818588256836,
   "rel_err": 0.003089054487645626,
   "xla_temp_mb": 320.00001525878906,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.04005301999859512,
   "inner": 50,
   "median_ms": 0.8498550811782479,
   "q1_ms": 0.8494162815622985,
   "q3_ms": 0.852910962421447,
   "tflops": 35.376350318830696
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013173818588256836,
   "rel_err": 0.003089054487645626,
   "xla_temp_mb": 32.250732421875,
   "xla_arg_mb": 40.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.1360454980749637,
   "inner": 152,
   "median_ms": 0.19544242777077384,
   "q1_ms": 0.19310980018130258,
   "q3_ms": 0.1964791307201315,
   "tflops": 153.82929599739572
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007283985614776611,
   "rel_err": 0.0028605633415281773,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.37466551503166556,
   "inner": 170,
   "median_ms": 0.18842336447799907,
   "q1_ms": 0.1856142853605835,
   "q3_ms": 0.18889348173294873,
   "tflops": 195.64452689891525
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007283985614776611,
   "rel_err": 0.0028605633415281773,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05553427804261446,
   "inner": 200,
   "median_ms": 0.10026662494055927,
   "q1_ms": 0.09557444020174444,
   "q3_ms": 0.10180499753914773,
   "tflops": 367.6597274701723
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007283985614776611,
   "rel_err": 0.0028605633415281773,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05269227595999837,
   "inner": 200,
   "median_ms": 0.0992582249455154,
   "q1_ms": 0.09644718433264643,
   "q3_ms": 0.10063837748020887,
   "tflops": 371.394914831847
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008879303932189941,
   "rel_err": 0.0034870761446654797,
   "xla_temp_mb": 411.98731994628906,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.021270492114126682,
   "inner": 73,
   "median_ms": 0.5621536705030562,
   "q1_ms": 0.560012438704502,
   "q3_ms": 0.562361068385717,
   "tflops": 65.57637516982038
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007283985614776611,
   "rel_err": 0.0028605633415281773,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.057312542805448174,
   "inner": 200,
   "median_ms": 0.0936925399582833,
   "q1_ms": 0.0936227454803884,
   "q3_ms": 0.10051379969809204,
   "tflops": 393.45715268700934
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016910731792449951,
   "rel_err": 0.004783402197062969,
   "xla_temp_mb": 96.00048828125,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.7539875479415059,
   "inner": 70,
   "median_ms": 0.6342162710747549,
   "q1_ms": 0.6105350496779595,
   "q3_ms": 0.6420282430813781,
   "tflops": 203.43848917870474
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016910731792449951,
   "rel_err": 0.004783402197062969,
   "xla_temp_mb": 47.241943359375,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.13748215418308973,
   "inner": 89,
   "median_ms": 0.42832984453004397,
   "q1_ms": 0.4248539550920551,
   "q3_ms": 0.42880798841753365,
   "tflops": 301.22579980753585
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016910731792449951,
   "rel_err": 0.004783402197062969,
   "xla_temp_mb": 47.241943359375,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.13438790291547775,
   "inner": 91,
   "median_ms": 0.4265183956454416,
   "q1_ms": 0.4178842409401313,
   "q3_ms": 0.4293552592171581,
   "tflops": 302.5051236178233
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016910731792449951,
   "rel_err": 0.004783402197062969,
   "xla_temp_mb": 686.8286285400391,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.037204637890681624,
   "inner": 30,
   "median_ms": 1.5478514367714524,
   "q1_ms": 1.5448987794419131,
   "q3_ms": 1.5524460817687213,
   "tflops": 83.35683705480258
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016910731792449951,
   "rel_err": 0.004783402197062969,
   "xla_temp_mb": 47.241943359375,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.13624040805734694,
   "inner": 95,
   "median_ms": 0.42740377891612685,
   "q1_ms": 0.4174838001888833,
   "q3_ms": 0.42818193201367793,
   "tflops": 301.87847268734487
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0072650909423828125,
   "rel_err": 0.0025133287999778986,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.3997378770727664,
   "inner": 184,
   "median_ms": 0.1512858915426161,
   "q1_ms": 0.1508136085011875,
   "q3_ms": 0.15130170314754968,
   "tflops": 121.83555129995611
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0072650909423828125,
   "rel_err": 0.0025133287999778986,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05643430491909385,
   "inner": 200,
   "median_ms": 0.07312340079806745,
   "q1_ms": 0.06688439229037613,
   "q3_ms": 0.0742910971166566,
   "tflops": 252.06705102379667
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0072650909423828125,
   "rel_err": 0.0025133287999778986,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05389420292340219,
   "inner": 200,
   "median_ms": 0.0714100559707731,
   "q1_ms": 0.06693886767607182,
   "q3_ms": 0.07317447452805936,
   "tflops": 258.1149076195081
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009048938751220703,
   "rel_err": 0.003130443627014756,
   "xla_temp_mb": 411.98731994628906,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.467581151984632,
   "inner": 74,
   "median_ms": 0.5684839718899614,
   "q1_ms": 0.5669600135218855,
   "q3_ms": 0.5692276282462518,
   "tflops": 32.423077714436936
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0072650909423828125,
   "rel_err": 0.0025133287999778986,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.05976511584594846,
   "inner": 200,
   "median_ms": 0.07443269481882453,
   "q1_ms": 0.07051678199786693,
   "q3_ms": 0.0747494725510478,
   "tflops": 247.63311398122892
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01283717155456543,
   "rel_err": 0.004822417162358761,
   "xla_temp_mb": 96.00048828125,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.7824872699566185,
   "inner": 97,
   "median_ms": 0.43716200041709485,
   "q1_ms": 0.4351092670496899,
   "q3_ms": 0.4373207105054837,
   "tflops": 147.5700082313863
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01283717155456543,
   "rel_err": 0.004822417162358761,
   "xla_temp_mb": 47.241943359375,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 1.0657460750080645,
   "inner": 108,
   "median_ms": 0.32576150907617474,
   "q1_ms": 0.32298884320038335,
   "q3_ms": 0.33164219887651225,
   "tflops": 198.03444606745967
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01283717155456543,
   "rel_err": 0.004822417162358761,
   "xla_temp_mb": 47.241943359375,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.02214453904889524,
   "inner": 104,
   "median_ms": 0.3341112012724177,
   "q1_ms": 0.328795063139226,
   "q3_ms": 0.33433591418613035,
   "tflops": 193.08541513817764
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01283717155456543,
   "rel_err": 0.004428453277796507,
   "xla_temp_mb": 686.8286285400391,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.04123534681275487,
   "inner": 27,
   "median_ms": 1.7220604802585309,
   "q1_ms": 1.7146330706223294,
   "q3_ms": 1.723667313517244,
   "tflops": 37.46209888651233
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01283717155456543,
   "rel_err": 0.004822417162358761,
   "xla_temp_mb": 47.241943359375,
   "xla_arg_mb": 58.59375,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.021121161989867687,
   "inner": 110,
   "median_ms": 0.3342668991535902,
   "q1_ms": 0.33090427750721574,
   "q3_ms": 0.33438158623704856,
   "tflops": 192.99547805467205
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005683600902557373,
   "rel_err": 0.0023447908461093903,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.32379062403924763,
   "inner": 145,
   "median_ms": 0.23869859407944927,
   "q1_ms": 0.23673066211029375,
   "q3_ms": 0.23993458598852158,
   "tflops": 287.89225592643066
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005683600902557373,
   "rel_err": 0.0023447908461093903,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.054182460997253656,
   "inner": 184,
   "median_ms": 0.15216474370706987,
   "q1_ms": 0.14822250016985217,
   "q3_ms": 0.1626568365558658,
   "tflops": 451.6123450271166
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005683600902557373,
   "rel_err": 0.0023447908461093903,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.05436867079697549,
   "inner": 185,
   "median_ms": 0.16231960491151423,
   "q1_ms": 0.15541308660160852,
   "q3_ms": 0.16355109463974432,
   "tflops": 423.3590685084605
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006561279296875,
   "rel_err": 0.002706880448386073,
   "xla_temp_mb": 768.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.017817370826378465,
   "inner": 49,
   "median_ms": 0.8741931239028975,
   "q1_ms": 0.8731820995976427,
   "q3_ms": 0.8772356644728962,
   "tflops": 78.60903369863743
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005683600902557373,
   "rel_err": 0.0023447908461093903,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.05587040679529309,
   "inner": 200,
   "median_ms": 0.1582311699166894,
   "q1_ms": 0.15545489732176065,
   "q3_ms": 0.16213527997024357,
   "tflops": 434.2979753747736
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010292530059814453,
   "rel_err": 0.0036079285200685263,
   "xla_temp_mb": 32.500244140625,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.605848893057555,
   "inner": 51,
   "median_ms": 0.915514154578833,
   "q1_ms": 0.904058617557965,
   "q3_ms": 0.9411097773551649,
   "tflops": 262.7137629419246
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010292530059814453,
   "rel_err": 0.0036079285200685263,
   "xla_temp_mb": 64.500732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13250372698530555,
   "inner": 62,
   "median_ms": 0.700946322702352,
   "q1_ms": 0.6934764040183397,
   "q3_ms": 0.7017693715170026,
   "tflops": 343.1335050717328
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010292530059814453,
   "rel_err": 0.0036079285200685263,
   "xla_temp_mb": 64.500732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13262801803648472,
   "inner": 64,
   "median_ms": 0.7012852656771429,
   "q1_ms": 0.6916910788277164,
   "q3_ms": 0.702684561474598,
   "tflops": 342.9676628722006
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000937044620513916,
   "rel_err": 0.0036079285200685263,
   "xla_temp_mb": 1280.000015258789,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.03664973005652428,
   "inner": 20,
   "median_ms": 2.2997527034021914,
   "q1_ms": 2.296073955949396,
   "q3_ms": 2.3004309739917517,
   "tflops": 104.58436170992819
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010292530059814453,
   "rel_err": 0.0036079285200685263,
   "xla_temp_mb": 64.500732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.1327230369206518,
   "inner": 64,
   "median_ms": 0.699947624525521,
   "q1_ms": 0.6848161792731844,
   "q3_ms": 0.7006825308053521,
   "tflops": 343.6230942837215
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007552385330200195,
   "rel_err": 0.0026127169840037823,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.37431973591446877,
   "inner": 194,
   "median_ms": 0.16859742769606642,
   "q1_ms": 0.16342743825570671,
   "q3_ms": 0.1691343837667295,
   "tflops": 203.79752430114715
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007552385330200195,
   "rel_err": 0.0026127169840037823,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.05492996913380921,
   "inner": 192,
   "median_ms": 0.11417354107834399,
   "q1_ms": 0.10406211125276363,
   "q3_ms": 0.11724480949245239,
   "tflops": 300.94309104788925
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007552385330200195,
   "rel_err": 0.0026127169840037823,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.053986432030797005,
   "inner": 198,
   "median_ms": 0.1154826104998438,
   "q1_ms": 0.10950477779701803,
   "q3_ms": 0.1163451369108651,
   "tflops": 297.5317081877576
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008072614669799805,
   "rel_err": 0.0027926883194595575,
   "xla_temp_mb": 768.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.5443709208630025,
   "inner": 44,
   "median_ms": 1.0204847309399736,
   "q1_ms": 1.011431886581704,
   "q3_ms": 1.0212043628350576,
   "tflops": 33.670017126420966
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007552385330200195,
   "rel_err": 0.0026127169840037823,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.05896112788468599,
   "inner": 200,
   "median_ms": 0.11307169101200998,
   "q1_ms": 0.10838635265827179,
   "q3_ms": 0.11517153528984636,
   "tflops": 303.8756921425227
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013039588928222656,
   "rel_err": 0.00369092496111989,
   "xla_temp_mb": 32.500244140625,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.765973465051502,
   "inner": 74,
   "median_ms": 0.5975255264422378,
   "q1_ms": 0.566169196374815,
   "q3_ms": 0.5980682295017146,
   "tflops": 201.2618356307583
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013039588928222656,
   "rel_err": 0.00369092496111989,
   "xla_temp_mb": 64.500732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13500462099909782,
   "inner": 78,
   "median_ms": 0.5059511794780309,
   "q1_ms": 0.4987567752742996,
   "q3_ms": 0.5067985501880639,
   "tflops": 237.68910749860567
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013039588928222656,
   "rel_err": 0.00369092496111989,
   "xla_temp_mb": 64.500732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13602404785342515,
   "inner": 82,
   "median_ms": 0.5052490983294641,
   "q1_ms": 0.5034285671327535,
   "q3_ms": 0.5069598967277604,
   "tflops": 238.0193941673918
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013039588928222656,
   "rel_err": 0.00369092496111989,
   "xla_temp_mb": 1280.000015258789,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.04162506596185267,
   "inner": 16,
   "median_ms": 2.9172820650273934,
   "q1_ms": 2.8988289341214113,
   "q3_ms": 2.923328247561585,
   "tflops": 41.22298824980805
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013039588928222656,
   "rel_err": 0.00369092496111989,
   "xla_temp_mb": 64.500732421875,
   "xla_arg_mb": 80.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.13756865402683616,
   "inner": 80,
   "median_ms": 0.504381739301607,
   "q1_ms": 0.5012396315578371,
   "q3_ms": 0.5072774249128997,
   "tflops": 238.42870373244864
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005130171775817871,
   "rel_err": 0.0019032340496778488,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.3661382079590112,
   "inner": 69,
   "median_ms": 0.6224103339448356,
   "q1_ms": 0.6204923543998081,
   "q3_ms": 0.6286884283484973,
   "tflops": 236.9112335674508
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005130171775817871,
   "rel_err": 0.0019032340496778488,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.05497039994224906,
   "inner": 108,
   "median_ms": 0.3607254906523007,
   "q1_ms": 0.35757861095825555,
   "q3_ms": 0.3652338141196028,
   "tflops": 408.77621299607904
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005130171775817871,
   "rel_err": 0.0019032340496778488,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.053051005117595196,
   "inner": 107,
   "median_ms": 0.36748487798270774,
   "q1_ms": 0.35185497210613476,
   "q3_ms": 0.3733720660871156,
   "tflops": 401.25732740202346
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006972551345825195,
   "rel_err": 0.002586735412478447,
   "xla_temp_mb": 1647.949234008789,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.019150747917592525,
   "inner": 25,
   "median_ms": 1.8437785562127829,
   "q1_ms": 1.8248236225917935,
   "q3_ms": 1.8520873179659247,
   "tflops": 79.97489693278692
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005130171775817871,
   "rel_err": 0.0019032340496778488,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.058127901051193476,
   "inner": 113,
   "median_ms": 0.35925786653780833,
   "q1_ms": 0.3489287347590502,
   "q3_ms": 0.36146547398076645,
   "tflops": 410.4461272373606
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009522885084152222,
   "rel_err": 0.003991737496107817,
   "xla_temp_mb": 192.00048828125,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.7494788798503578,
   "inner": 22,
   "median_ms": 2.217314862222834,
   "q1_ms": 2.1549752564169466,
   "q3_ms": 2.2786730921573257,
   "tflops": 232.75720051892827
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009522885084152222,
   "rel_err": 0.003991737496107817,
   "xla_temp_mb": 94.483154296875,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.1348008611239493,
   "inner": 33,
   "median_ms": 1.3871158783634503,
   "q1_ms": 1.3564087125952495,
   "q3_ms": 1.41508218296098,
   "tflops": 372.06408494790026
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009522885084152222,
   "rel_err": 0.003991737496107817,
   "xla_temp_mb": 94.483154296875,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.13496070099063218,
   "inner": 33,
   "median_ms": 1.3785150367766619,
   "q1_ms": 1.370454439893365,
   "q3_ms": 1.4422642257572575,
   "tflops": 374.38547003939175
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009522885084152222,
   "rel_err": 0.003991737496107817,
   "xla_temp_mb": 2746.948257446289,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.03949791193008423,
   "inner": 9,
   "median_ms": 5.522362680898772,
   "q1_ms": 5.510699883517292,
   "q3_ms": 5.5299548872022175,
   "tflops": 93.45565110837751
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009522885084152222,
   "rel_err": 0.003991737496107817,
   "xla_temp_mb": 94.483154296875,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.13363748393021524,
   "inner": 34,
   "median_ms": 1.3839300338398008,
   "q1_ms": 1.3746528660275918,
   "q3_ms": 1.3899216325679684,
   "tflops": 372.92058657622977
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0070171356201171875,
   "rel_err": 0.0024275495670735836,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.39901529904454947,
   "inner": 95,
   "median_ms": 0.427930726130542,
   "q1_ms": 0.42130084169146265,
   "q3_ms": 0.43685643713137035,
   "tflops": 172.289568142646
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0070171356201171875,
   "rel_err": 0.0024275495670735836,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.05727446312084794,
   "inner": 142,
   "median_ms": 0.2222515062206018,
   "q1_ms": 0.21441913091800582,
   "q3_ms": 0.2228316086882227,
   "tflops": 331.73228498536815
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0070171356201171875,
   "rel_err": 0.0024275495670735836,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.054976239101961255,
   "inner": 140,
   "median_ms": 0.2226214639709464,
   "q1_ms": 0.2173129078333399,
   "q3_ms": 0.22268003168781952,
   "tflops": 331.18100422527993
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0070171356201171875,
   "rel_err": 0.0024275495670735836,
   "xla_temp_mb": 1647.949234008789,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.43164451001212,
   "inner": 25,
   "median_ms": 1.8820304423570633,
   "q1_ms": 1.8785502621904016,
   "q3_ms": 1.8848444428294897,
   "tflops": 39.174711705333905
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0070171356201171875,
   "rel_err": 0.0024275495670735836,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.059482666896656156,
   "inner": 147,
   "median_ms": 0.22191013729035045,
   "q1_ms": 0.21516672750542154,
   "q3_ms": 0.22226789820508486,
   "tflops": 332.2425955851364
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011957883834838867,
   "rel_err": 0.004573028534650803,
   "xla_temp_mb": 192.00048828125,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.7846149348188192,
   "inner": 35,
   "median_ms": 1.3715828037155526,
   "q1_ms": 1.322959253697523,
   "q3_ms": 1.3800546393862794,
   "tflops": 188.13884170970957
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011957883834838867,
   "rel_err": 0.004573028534650803,
   "xla_temp_mb": 94.483154296875,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.13818088197149336,
   "inner": 48,
   "median_ms": 0.9411184582859278,
   "q1_ms": 0.9243816263430441,
   "q3_ms": 0.9423480538922983,
   "tflops": 274.1929007215377
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011957883834838867,
   "rel_err": 0.004573028534650803,
   "xla_temp_mb": 94.483154296875,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.13848496112041175,
   "inner": 47,
   "median_ms": 0.9352802150347765,
   "q1_ms": 0.917511318433792,
   "q3_ms": 0.9396689265292693,
   "tflops": 275.9044785208088
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011933326721191406,
   "rel_err": 0.004184063524007797,
   "xla_temp_mb": 2746.948257446289,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.04498838703148067,
   "inner": 8,
   "median_ms": 5.854256625752896,
   "q1_ms": 5.828710942296311,
   "q3_ms": 5.8592465065885335,
   "tflops": 44.07869632240683
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011957883834838867,
   "rel_err": 0.004573028534650803,
   "xla_temp_mb": 94.483154296875,
   "xla_arg_mb": 117.1875,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.13841020199470222,
   "inner": 48,
   "median_ms": 0.9416644995023186,
   "q1_ms": 0.929530479576594,
   "q3_ms": 0.9422885608122064,
   "tflops": 274.0339050016026
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000495031476020813,
   "rel_err": 0.003562052734196186,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.32331868587061763,
   "inner": 52,
   "median_ms": 0.884610098392631,
   "q1_ms": 0.8774993363588762,
   "q3_ms": 0.8934252590943987,
   "tflops": 310.73340383911875
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000495031476020813,
   "rel_err": 0.003562052734196186,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05337549396790564,
   "inner": 75,
   "median_ms": 0.5764762933055559,
   "q1_ms": 0.5733494320884347,
   "q3_ms": 0.5886249182124933,
   "tflops": 476.8243033340202
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000495031476020813,
   "rel_err": 0.003562052734196186,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05411775200627744,
   "inner": 76,
   "median_ms": 0.5865222100462569,
   "q1_ms": 0.5760730342253259,
   "q3_ms": 0.5977890855559197,
   "tflops": 468.6572856675305
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000495031476020813,
   "rel_err": 0.003562052734196186,
   "xla_temp_mb": 3072.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.020255984971299767,
   "inner": 15,
   "median_ms": 3.219590460260709,
   "q1_ms": 3.215200236688058,
   "q3_ms": 3.222222168308993,
   "tflops": 85.37666834860157
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000495031476020813,
   "rel_err": 0.003562052734196186,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05811002105474472,
   "inner": 77,
   "median_ms": 0.5781998428479804,
   "q1_ms": 0.5720283718199118,
   "q3_ms": 0.5861100326473643,
   "tflops": 475.40294302063745
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000783964991569519,
   "rel_err": 0.004214399494230747,
   "xla_temp_mb": 65.000244140625,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.6291449191048741,
   "inner": 14,
   "median_ms": 3.516918214570199,
   "q1_ms": 3.4785101472932314,
   "q3_ms": 3.5399403256763304,
   "tflops": 273.55560055910325
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000783964991569519,
   "rel_err": 0.004214399494230747,
   "xla_temp_mb": 129.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.1351878708228469,
   "inner": 20,
   "median_ms": 2.3853269522078335,
   "q1_ms": 2.3363581218291074,
   "q3_ms": 2.4176076462026685,
   "tflops": 403.32947792063294
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000783964991569519,
   "rel_err": 0.004214399494230747,
   "xla_temp_mb": 129.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.13440350210294127,
   "inner": 20,
   "median_ms": 2.301324694417417,
   "q1_ms": 2.2917957277968526,
   "q3_ms": 2.4183601257391274,
   "tflops": 418.0516885069752
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010238289833068848,
   "rel_err": 0.004597060848027468,
   "xla_temp_mb": 5120.000015258789,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.04208004497922957,
   "inner": 5,
   "median_ms": 8.734190184623003,
   "q1_ms": 8.732028002850711,
   "q3_ms": 8.737523993477225,
   "tflops": 110.15018610400527
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000783964991569519,
   "rel_err": 0.004214399494230747,
   "xla_temp_mb": 129.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.13510232092812657,
   "inner": 20,
   "median_ms": 2.370409050490707,
   "q1_ms": 2.3349906085059047,
   "q3_ms": 2.4243179010227323,
   "tflops": 405.8677864501225
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006810665130615234,
   "rel_err": 0.002356122015044093,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.376660960027948,
   "inner": 83,
   "median_ms": 0.5327075809599405,
   "q1_ms": 0.5278992831303053,
   "q3_ms": 0.5469304938654763,
   "tflops": 258.00074634630624
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006810665130615234,
   "rel_err": 0.002356122015044093,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05828284099698067,
   "inner": 105,
   "median_ms": 0.3637194855227357,
   "q1_ms": 0.3552797904592895,
   "q3_ms": 0.36757628189488534,
   "tflops": 377.87074639257634
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006810665130615234,
   "rel_err": 0.002356122015044093,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05639642593450844,
   "inner": 111,
   "median_ms": 0.3646044060587883,
   "q1_ms": 0.3550488969546046,
   "q3_ms": 0.36748615722801237,
   "tflops": 376.9536275155148
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006975531578063965,
   "rel_err": 0.002413156908005476,
   "xla_temp_mb": 3072.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.5593992099165916,
   "inner": 12,
   "median_ms": 3.9074518329774337,
   "q1_ms": 3.8999089641341316,
   "q3_ms": 3.9157867819691696,
   "tflops": 35.17355026927436
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006810665130615234,
   "rel_err": 0.002356122015044093,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.05873015895485878,
   "inner": 108,
   "median_ms": 0.35995836973328277,
   "q1_ms": 0.3510655186363254,
   "q3_ms": 0.3622200747486204,
   "tflops": 381.81902416614923
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015851497650146484,
   "rel_err": 0.003916752990335226,
   "xla_temp_mb": 65.000244140625,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.7542233171407133,
   "inner": 25,
   "median_ms": 1.9766114000231028,
   "q1_ms": 1.9506106618791819,
   "q3_ms": 2.0357394218444824,
   "tflops": 243.36414185731076
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015851497650146484,
   "rel_err": 0.003916752990335226,
   "xla_temp_mb": 129.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.13595719798468053,
   "inner": 31,
   "median_ms": 1.5126749077030728,
   "q1_ms": 1.5069573381615262,
   "q3_ms": 1.5131395813378115,
   "tflops": 318.0037790687171
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015851497650146484,
   "rel_err": 0.003916752990335226,
   "xla_temp_mb": 129.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.13677743612788618,
   "inner": 32,
   "median_ms": 1.5053289971547201,
   "q1_ms": 1.4700105202791747,
   "q3_ms": 1.5220020795823075,
   "tflops": 319.5556174505541
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015851497650146484,
   "rel_err": 0.004538213834166527,
   "xla_temp_mb": 5152.000015258789,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.04875202709808946,
   "inner": 4,
   "median_ms": 10.54205879336223,
   "q1_ms": 10.525441117351875,
   "q3_ms": 10.55330625968054,
   "tflops": 45.630208157716105
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015851497650146484,
   "rel_err": 0.003916752990335226,
   "xla_temp_mb": 129.000732421875,
   "xla_arg_mb": 160.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.1372115151025355,
   "inner": 31,
   "median_ms": 1.508261355000638,
   "q1_ms": 1.4803365789233676,
   "q3_ms": 1.5142207116549533,
   "tflops": 318.93433824126356
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003512650728225708,
   "rel_err": 0.003266321960836649,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.32673251698724926,
   "inner": 15,
   "median_ms": 3.355496795848012,
   "q1_ms": 3.347341188540061,
   "q3_ms": 3.3813527319580317,
   "tflops": 327.67476611406744
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003512650728225708,
   "rel_err": 0.003266321960836649,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.05560096795670688,
   "inner": 22,
   "median_ms": 2.2701883095909245,
   "q1_ms": 2.236415044700896,
   "q3_ms": 2.3073680032129316,
   "tflops": 484.32617819890277
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003512650728225708,
   "rel_err": 0.003266321960836649,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.05486185988411307,
   "inner": 22,
   "median_ms": 2.2735128560188143,
   "q1_ms": 2.2694910664788703,
   "q3_ms": 2.3095568652603435,
   "tflops": 483.61795046163627
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000379122793674469,
   "rel_err": 0.0035253637470304966,
   "xla_temp_mb": 12288.000015258789,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.029108510119840503,
   "inner": 4,
   "median_ms": 12.416483485139906,
   "q1_ms": 12.413351010764018,
   "q3_ms": 12.43111613439396,
   "tflops": 88.55257844073965
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003512650728225708,
   "rel_err": 0.003266321960836649,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.05757425189949572,
   "inner": 22,
   "median_ms": 2.277453768659722,
   "q1_ms": 2.2418611333705485,
   "q3_ms": 2.3196902475319803,
   "tflops": 482.78109655023235
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006157159805297852,
   "rel_err": 0.005359946750104427,
   "xla_temp_mb": 130.000244140625,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.6228717670310289,
   "inner": 3,
   "median_ms": 13.586166392390927,
   "q1_ms": 13.454311682532232,
   "q3_ms": 13.929998657355707,
   "tflops": 283.2506673384536
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006157159805297852,
   "rel_err": 0.005359946750104427,
   "xla_temp_mb": 258.000732421875,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.1421828221064061,
   "inner": 5,
   "median_ms": 8.8779057841748,
   "q1_ms": 8.85526379570365,
   "q3_ms": 9.086013119667768,
   "tflops": 433.4682965520677
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006157159805297852,
   "rel_err": 0.005359946750104427,
   "xla_temp_mb": 258.000732421875,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.14296656800433993,
   "inner": 5,
   "median_ms": 8.79976600408554,
   "q1_ms": 8.60941840801388,
   "q3_ms": 9.044585190713406,
   "tflops": 437.31738951118945
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006582215428352356,
   "rel_err": 0.005359946750104427,
   "xla_temp_mb": 20480.00001525879,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.06629871902987361,
   "inner": 1,
   "median_ms": 33.73347781598568,
   "q1_ms": 33.67433499079198,
   "q3_ms": 33.79080246668309,
   "tflops": 114.07927514050643
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006157159805297852,
   "rel_err": 0.005359946750104427,
   "xla_temp_mb": 258.000732421875,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.14206700399518013,
   "inner": 5,
   "median_ms": 8.864306379109621,
   "q1_ms": 8.80007571540773,
   "q3_ms": 8.880838518962264,
   "tflops": 434.13331315862564
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004915475845336914,
   "rel_err": 0.0017004889668896794,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.3745421590283513,
   "inner": 26,
   "median_ms": 1.8891427003277035,
   "q1_ms": 1.8793088652623389,
   "q3_ms": 1.9298939240308335,
   "tflops": 291.0080926086926
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004915475845336914,
   "rel_err": 0.0017004889668896794,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.05522048403508961,
   "inner": 38,
   "median_ms": 1.2741321575288709,
   "q1_ms": 1.2361847483062822,
   "q3_ms": 1.2820064868336838,
   "tflops": 431.474718410867
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004915475845336914,
   "rel_err": 0.0017004889668896794,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.054860214004293084,
   "inner": 38,
   "median_ms": 1.2749687364128859,
   "q1_ms": 1.2504540684044754,
   "q3_ms": 1.3021243166325516,
   "tflops": 431.1916035170662
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011793375015258789,
   "rel_err": 0.004079870413988829,
   "xla_temp_mb": 12288.000015258789,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.514892578125,
   "inner": 3,
   "median_ms": 15.848161264633138,
   "q1_ms": 15.724598507707316,
   "q3_ms": 15.879462977560857,
   "tflops": 34.688933606123676
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004915475845336914,
   "rel_err": 0.0017004889668896794,
   "xla_temp_mb": 0.00048828125,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.0586824850179255,
   "inner": 37,
   "median_ms": 1.2912784602392364,
   "q1_ms": 1.2390187444366716,
   "q3_ms": 1.3006019745827526,
   "tflops": 425.7453607536722
  },
  {
   "impl": "probjax_mha",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01760244369506836,
   "rel_err": 0.005058928392827511,
   "xla_temp_mb": 130.000244140625,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.7510640430264175,
   "inner": 7,
   "median_ms": 7.358489125700934,
   "q1_ms": 7.281649990805558,
   "q3_ms": 7.520703632118446,
   "tflops": 261.4864703526643
  },
  {
   "impl": "probjax_flex",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01760244369506836,
   "rel_err": 0.005058928392827511,
   "xla_temp_mb": 258.000732421875,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.13877141289412975,
   "inner": 10,
   "median_ms": 5.011330684646964,
   "q1_ms": 4.986435291357338,
   "q3_ms": 5.088557046838105,
   "tflops": 383.95896612908336
  },
  {
   "impl": "probjax_default",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01760244369506836,
   "rel_err": 0.005058928392827511,
   "xla_temp_mb": 258.000732421875,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.14005058002658188,
   "inner": 10,
   "median_ms": 4.889067099429667,
   "q1_ms": 4.881027084775269,
   "q3_ms": 5.062533658929169,
   "tflops": 393.5608388014271
  },
  {
   "impl": "jax_xla",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01760244369506836,
   "rel_err": 0.00520186685025692,
   "xla_temp_mb": 20480.00001525879,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.0792325499933213,
   "inner": 1,
   "median_ms": 40.3525831643492,
   "q1_ms": 40.21145903971046,
   "q3_ms": 40.41439283173531,
   "tflops": 47.68332527241896
  },
  {
   "impl": "jax_cudnn",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01760244369506836,
   "rel_err": 0.005058928392827511,
   "xla_temp_mb": 258.000732421875,
   "xla_arg_mb": 320.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.1422828531358391,
   "inner": 10,
   "median_ms": 5.043369601480663,
   "q1_ms": 5.0140852108597755,
   "q3_ms": 5.083770107012242,
   "tflops": 381.5197974074908
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004020603373646736,
   "rel_err": 0.00043028933578170836,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.6630640050861984,
   "inner": 200,
   "median_ms": 0.03833124530501664,
   "q1_ms": 0.03803140018135309,
   "q3_ms": 0.038388899993151426,
   "tflops": 3.5015227637917135
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004020603373646736,
   "rel_err": 0.00043028933578170836,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.6649310300126672,
   "inner": 200,
   "median_ms": 0.038431850261986256,
   "q1_ms": 0.038255450199358165,
   "q3_ms": 0.038559322129003704,
   "tflops": 3.492356654312778
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004020603373646736,
   "rel_err": 0.00043028933578170836,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.6666091170627624,
   "inner": 200,
   "median_ms": 0.02383923507295549,
   "q1_ms": 0.023670162190683186,
   "q3_ms": 0.023991313064470887,
   "tflops": 5.630118902274
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00037729740142822266,
   "rel_err": 0.0004037877661176026,
   "xla_temp_mb": 4.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.17435061908327043,
   "inner": 200,
   "median_ms": 0.03456091042608023,
   "q1_ms": 0.03437363542616367,
   "q3_ms": 0.034847757196985185,
   "tflops": 3.8835125100962937
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005728304386138916,
   "rel_err": 0.0007045819074846804,
   "xla_temp_mb": 1.031494140625,
   "xla_arg_mb": 4.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.8486492349766195,
   "inner": 200,
   "median_ms": 0.0343070097733289,
   "q1_ms": 0.03427083254791796,
   "q3_ms": 0.034395832917653024,
   "tflops": 13.692888162033999
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005728304386138916,
   "rel_err": 0.0007045819074846804,
   "xla_temp_mb": 0.03173828125,
   "xla_arg_mb": 4.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.8691200409084558,
   "inner": 200,
   "median_ms": 0.036857599625363946,
   "q1_ms": 0.036656277370639145,
   "q3_ms": 0.03716780222021043,
   "tflops": 12.745323970493411
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005728304386138916,
   "rel_err": 0.0007045819074846804,
   "xla_temp_mb": 0.03173828125,
   "xla_arg_mb": 4.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.8487190550658852,
   "inner": 200,
   "median_ms": 0.037257850635796785,
   "q1_ms": 0.03701087727677077,
   "q3_ms": 0.03728932701051235,
   "tflops": 12.608404402927626
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005763769149780273,
   "rel_err": 0.0007045819074846804,
   "xla_temp_mb": 6.0,
   "xla_arg_mb": 4.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.03122011781670153,
   "inner": 200,
   "median_ms": 0.05862234509550035,
   "q1_ms": 0.05853809474501759,
   "q3_ms": 0.058677069609984756,
   "tflops": 8.013361581402469
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009530782699584961,
   "rel_err": 0.00031446287175640464,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 1.100318348966539,
   "inner": 200,
   "median_ms": 0.023094939533621073,
   "q1_ms": 0.022867265506647527,
   "q3_ms": 0.023132787318900228,
   "tflops": 2.9057821910425217
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009530782699584961,
   "rel_err": 0.00031446287175640464,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.03865455696359277,
   "inner": 200,
   "median_ms": 0.024037935072556138,
   "q1_ms": 0.023910337476991117,
   "q3_ms": 0.024571759859099984,
   "tflops": 2.79178988533909
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009530782699584961,
   "rel_err": 0.00031446287175640464,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.03872656705789268,
   "inner": 200,
   "median_ms": 0.023288684897124767,
   "q1_ms": 0.023195089888758957,
   "q3_ms": 0.023563890135847032,
   "tflops": 2.881608141311805
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013991594314575195,
   "rel_err": 0.0004616449004970491,
   "xla_temp_mb": 4.0,
   "xla_arg_mb": 3.0,
   "xla_out_mb": 1.0,
   "compile_s": 0.282288053072989,
   "inner": 200,
   "median_ms": 0.03435685997828841,
   "q1_ms": 0.03411031037103385,
   "q3_ms": 0.03443255787715316,
   "tflops": 1.9532886312197622
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016770362854003906,
   "rel_err": 0.0007809422677382827,
   "xla_temp_mb": 1.031494140625,
   "xla_arg_mb": 4.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 1.1731115658767521,
   "inner": 200,
   "median_ms": 0.03499935497529805,
   "q1_ms": 0.03494198259431869,
   "q3_ms": 0.035437155165709555,
   "tflops": 6.71101007906503
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016770362854003906,
   "rel_err": 0.0007809422677382827,
   "xla_temp_mb": 0.03173828125,
   "xla_arg_mb": 4.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 1.1963365951087326,
   "inner": 200,
   "median_ms": 0.05995368934236467,
   "q1_ms": 0.03955436986871064,
   "q3_ms": 0.06020159227773547,
   "tflops": 3.9177075935846974
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016770362854003906,
   "rel_err": 0.0007809422677382827,
   "xla_temp_mb": 0.03173828125,
   "xla_arg_mb": 4.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 1.2055549500510097,
   "inner": 200,
   "median_ms": 0.045800129882991314,
   "q1_ms": 0.04555818042717874,
   "q3_ms": 0.046112252748571336,
   "tflops": 5.128392094085027
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016134977340698242,
   "rel_err": 0.0007285389583557844,
   "xla_temp_mb": 6.0,
   "xla_arg_mb": 4.0,
   "xla_out_mb": 3.0000228881835938,
   "compile_s": 0.7476080220658332,
   "inner": 200,
   "median_ms": 0.05851370049640536,
   "q1_ms": 0.058463921886868775,
   "q3_ms": 0.05875304515939206,
   "tflops": 4.014120146348107
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003509223461151123,
   "rel_err": 0.0008548463811166584,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.9365642420016229,
   "inner": 200,
   "median_ms": 0.04150834050960839,
   "q1_ms": 0.04150351509451866,
   "q3_ms": 0.04151873989030719,
   "tflops": 51.73619618695425
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003509223461151123,
   "rel_err": 0.0008548463811166584,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.9402725519612432,
   "inner": 200,
   "median_ms": 0.04162663943134248,
   "q1_ms": 0.04138051706831902,
   "q3_ms": 0.04166546277701855,
   "tflops": 51.589166873342826
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003509223461151123,
   "rel_err": 0.0008548463811166584,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.944615620886907,
   "inner": 200,
   "median_ms": 0.04132283967919648,
   "q1_ms": 0.041316040442325175,
   "q3_ms": 0.04161346470937133,
   "tflops": 51.96844323070873
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004189014434814453,
   "rel_err": 0.0010204433929175138,
   "xla_temp_mb": 68.00001525878906,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.02075071493163705,
   "inner": 200,
   "median_ms": 0.07451220415532589,
   "q1_ms": 0.07386907760519534,
   "q3_ms": 0.0751804024912417,
   "tflops": 28.820562649353665
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00036704540252685547,
   "rel_err": 0.0007074366440065205,
   "xla_temp_mb": 4.125244140625,
   "xla_arg_mb": 16.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 1.0255694061052054,
   "inner": 200,
   "median_ms": 0.13854203512892127,
   "q1_ms": 0.13662021548952907,
   "q3_ms": 0.1391879323637113,
   "tflops": 54.2520741882112
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00036704540252685547,
   "rel_err": 0.0007074366440065205,
   "xla_temp_mb": 0.12548828125,
   "xla_arg_mb": 16.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 1.0183339659124613,
   "inner": 200,
   "median_ms": 0.13984228018671274,
   "q1_ms": 0.13973818044178188,
   "q3_ms": 0.14189262525178492,
   "tflops": 53.7476416857951
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00036704540252685547,
   "rel_err": 0.0007074366440065205,
   "xla_temp_mb": 0.12548828125,
   "xla_arg_mb": 16.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 1.0164034711197019,
   "inner": 200,
   "median_ms": 0.13871543458662927,
   "q1_ms": 0.1386265578912571,
   "q3_ms": 0.13935368217062205,
   "tflops": 54.184257075632466
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000362396240234375,
   "rel_err": 0.0007074366440065205,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 16.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.03555755619890988,
   "inner": 156,
   "median_ms": 0.20126452615771157,
   "q1_ms": 0.19958626571445703,
   "q3_ms": 0.20153129410643417,
   "tflops": 37.344846165837914
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012368559837341309,
   "rel_err": 0.0004327311762608588,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 1.1136323739774525,
   "inner": 200,
   "median_ms": 0.0403639895375818,
   "q1_ms": 0.04028999479487538,
   "q3_ms": 0.04046909511089325,
   "tflops": 26.601479098102242
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012368559837341309,
   "rel_err": 0.0004327311762608588,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.038917966885492206,
   "inner": 200,
   "median_ms": 0.04022514563985169,
   "q1_ms": 0.040170992724597454,
   "q3_ms": 0.04025669244583696,
   "tflops": 26.69329860514481
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012368559837341309,
   "rel_err": 0.0004327311762608588,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.03960388619452715,
   "inner": 200,
   "median_ms": 0.04018069477751851,
   "q1_ms": 0.04014434525743127,
   "q3_ms": 0.040189792634919286,
   "tflops": 26.72282871028823
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017076730728149414,
   "rel_err": 0.0005974530358798802,
   "xla_temp_mb": 68.00001525878906,
   "xla_arg_mb": 12.0,
   "xla_out_mb": 4.0,
   "compile_s": 0.6701728869229555,
   "inner": 200,
   "median_ms": 0.10426916996948421,
   "q1_ms": 0.1040872506564483,
   "q3_ms": 0.10429349960759282,
   "tflops": 10.29778816033776
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0020322799682617188,
   "rel_err": 0.000902797793969512,
   "xla_temp_mb": 4.125244140625,
   "xla_arg_mb": 16.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 1.1966907540336251,
   "inner": 200,
   "median_ms": 0.10613407008349895,
   "q1_ms": 0.10601942020002753,
   "q3_ms": 0.10715468961279839,
   "tflops": 35.408953798185536
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0020322799682617188,
   "rel_err": 0.000902797793969512,
   "xla_temp_mb": 0.12548828125,
   "xla_arg_mb": 16.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 1.212104853009805,
   "inner": 200,
   "median_ms": 0.12367747491225599,
   "q1_ms": 0.12147467990871519,
   "q3_ms": 0.12420694576576351,
   "tflops": 30.38626384203116
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0020322799682617188,
   "rel_err": 0.000902797793969512,
   "xla_temp_mb": 0.12548828125,
   "xla_arg_mb": 16.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 1.2122641720343381,
   "inner": 200,
   "median_ms": 0.12520771473646164,
   "q1_ms": 0.12278757523745298,
   "q3_ms": 0.12553969281725585,
   "tflops": 30.01489478431961
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012679100036621094,
   "rel_err": 0.0006988496752455831,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 16.0,
   "xla_out_mb": 12.000022888183594,
   "compile_s": 0.03832619916647673,
   "inner": 122,
   "median_ms": 0.2966234021530044,
   "q1_ms": 0.29659171428409264,
   "q3_ms": 0.29669200035850285,
   "tflops": 12.669588295199636
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00024455785751342773,
   "rel_err": 0.0005569846834987402,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 1.116810865001753,
   "inner": 200,
   "median_ms": 0.11393139488063753,
   "q1_ms": 0.1139017246896401,
   "q3_ms": 0.11440064990893006,
   "tflops": 25.885051289767002
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00024455785751342773,
   "rel_err": 0.0005569846834987402,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 1.1850355549249798,
   "inner": 200,
   "median_ms": 0.11446584947407246,
   "q1_ms": 0.11387902253773063,
   "q3_ms": 0.11449932237155735,
   "tflops": 25.76419092288309
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00024455785751342773,
   "rel_err": 0.0005569846834987402,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.04143515112809837,
   "inner": 200,
   "median_ms": 0.11355954455211759,
   "q1_ms": 0.11290294991340488,
   "q3_ms": 0.11376350012142211,
   "tflops": 25.96981179901189
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003241300582885742,
   "rel_err": 0.0007382117328234017,
   "xla_temp_mb": 87.89064025878906,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.021855521947145462,
   "inner": 197,
   "median_ms": 0.12191013210845478,
   "q1_ms": 0.12134473066455519,
   "q3_ms": 0.12203066746898109,
   "tflops": 24.19093433002252
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000510871410369873,
   "rel_err": 0.0005970043712295592,
   "xla_temp_mb": 40.00048828125,
   "xla_arg_mb": 18.75,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 1.2177663380280137,
   "inner": 111,
   "median_ms": 0.3213767121705386,
   "q1_ms": 0.3195247384258085,
   "q3_ms": 0.3219791488755528,
   "tflops": 32.1178218866172
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000510871410369873,
   "rel_err": 0.0005970043712295592,
   "xla_temp_mb": 30.078651428222656,
   "xla_arg_mb": 18.75,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 1.2376245758496225,
   "inner": 117,
   "median_ms": 0.3303502383840899,
   "q1_ms": 0.32704687104202235,
   "q3_ms": 0.33055467861425925,
   "tflops": 31.24538384016228
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000510871410369873,
   "rel_err": 0.0005970043712295592,
   "xla_temp_mb": 30.078651428222656,
   "xla_arg_mb": 18.75,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 1.2615070620086044,
   "inner": 115,
   "median_ms": 0.32789034809431306,
   "q1_ms": 0.32695591368752974,
   "q3_ms": 0.33112643185355095,
   "tflops": 31.479792131700822
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005156397819519043,
   "rel_err": 0.0005970043712295592,
   "xla_temp_mb": 131.9091796875,
   "xla_arg_mb": 18.75,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.031787126092240214,
   "inner": 114,
   "median_ms": 0.3374705087034064,
   "q1_ms": 0.33560963633486574,
   "q3_ms": 0.33878134216361666,
   "tflops": 30.586139333056963
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010406970977783203,
   "rel_err": 0.00044006036478094757,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 1.1803848869167268,
   "inner": 200,
   "median_ms": 0.08504388039000332,
   "q1_ms": 0.08501277538016438,
   "q3_ms": 0.08564631978515536,
   "tflops": 17.33881371872738
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010406970977783203,
   "rel_err": 0.00044006036478094757,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 1.1696537958923727,
   "inner": 200,
   "median_ms": 0.08326962939463556,
   "q1_ms": 0.08284187759272754,
   "q3_ms": 0.08328508003614843,
   "tflops": 17.708257028642365
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010406970977783203,
   "rel_err": 0.00044006036478094757,
   "xla_temp_mb": 20.0,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.042617907049134374,
   "inner": 200,
   "median_ms": 0.08258778019808233,
   "q1_ms": 0.0825522851664573,
   "q3_ms": 0.08261695504188538,
   "tflops": 17.85445735995504
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000980377197265625,
   "rel_err": 0.0004145540005993098,
   "xla_temp_mb": 87.89064025878906,
   "xla_arg_mb": 14.0625,
   "xla_out_mb": 4.6875,
   "compile_s": 0.49110779096372426,
   "inner": 200,
   "median_ms": 0.12559406459331512,
   "q1_ms": 0.12509521679021418,
   "q3_ms": 0.12577234243508428,
   "tflops": 11.740682211175805
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0018483996391296387,
   "rel_err": 0.0007540053920820355,
   "xla_temp_mb": 40.00048828125,
   "xla_arg_mb": 18.75,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 1.273083272157237,
   "inner": 155,
   "median_ms": 0.22333844474727108,
   "q1_ms": 0.2218573812335249,
   "q3_ms": 0.2234267288698785,
   "tflops": 23.1082472426103
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0018483996391296387,
   "rel_err": 0.0007540053920820355,
   "xla_temp_mb": 30.07861328125,
   "xla_arg_mb": 18.75,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 1.2725585428997874,
   "inner": 160,
   "median_ms": 0.2141683697118424,
   "q1_ms": 0.21384036881499924,
   "q3_ms": 0.2162483062420506,
   "tflops": 24.097676080477843
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0018483996391296387,
   "rel_err": 0.0007540053920820355,
   "xla_temp_mb": 30.07861328125,
   "xla_arg_mb": 18.75,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 1.2653799219988286,
   "inner": 148,
   "median_ms": 0.213988554802396,
   "q1_ms": 0.21381811883829133,
   "q3_ms": 0.2156423374922392,
   "tflops": 24.117925394495035
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0014612674713134766,
   "rel_err": 0.0005960851558484137,
   "xla_temp_mb": 131.9091796875,
   "xla_arg_mb": 18.75,
   "xla_out_mb": 14.062522888183594,
   "compile_s": 0.03890814702026546,
   "inner": 98,
   "median_ms": 0.38319143789763355,
   "q1_ms": 0.38262168943349806,
   "q3_ms": 0.3840704092147703,
   "tflops": 13.468359387974395
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00022867321968078613,
   "rel_err": 0.0007638643146492541,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.9381521979812533,
   "inner": 186,
   "median_ms": 0.14680767712253395,
   "q1_ms": 0.14617859125537896,
   "q3_ms": 0.14718396524067529,
   "tflops": 58.511480873240416
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00022867321968078613,
   "rel_err": 0.0007638643146492541,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.9400831428356469,
   "inner": 190,
   "median_ms": 0.14688061109106793,
   "q1_ms": 0.14599669528634923,
   "q3_ms": 0.14722703184050165,
   "tflops": 58.4824268376316
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00022867321968078613,
   "rel_err": 0.0007638643146492541,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.9422805269714445,
   "inner": 200,
   "median_ms": 0.1454237150028348,
   "q1_ms": 0.14517389179673046,
   "q3_ms": 0.14859495451673865,
   "tflops": 59.06832040312375
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003216266632080078,
   "rel_err": 0.001074367668479681,
   "xla_temp_mb": 256.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.01739074499346316,
   "inner": 117,
   "median_ms": 0.3144517361831206,
   "q1_ms": 0.31360903850350624,
   "q3_ms": 0.31458023979336536,
   "tflops": 27.317179724514737
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00035509467124938965,
   "rel_err": 0.0007818220183253288,
   "xla_temp_mb": 8.250244140625,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 1.0087425210513175,
   "inner": 83,
   "median_ms": 0.4952661693096161,
   "q1_ms": 0.49490657921162357,
   "q3_ms": 0.49673114953764597,
   "tflops": 60.70426961306332
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00035509467124938965,
   "rel_err": 0.0007818220183253288,
   "xla_temp_mb": 0.25048828125,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 1.0378280910663307,
   "inner": 81,
   "median_ms": 0.5052046056974817,
   "q1_ms": 0.5033449815977136,
   "q3_ms": 0.5087558220336099,
   "tflops": 59.510089046976915
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00035509467124938965,
   "rel_err": 0.0007818220183253288,
   "xla_temp_mb": 0.25048828125,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 1.065715879900381,
   "inner": 79,
   "median_ms": 0.5053983046378516,
   "q1_ms": 0.5019302671492288,
   "q3_ms": 0.5085272080842642,
   "tflops": 59.48728121188144
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003515779972076416,
   "rel_err": 0.0007826067158021033,
   "xla_temp_mb": 384.0,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.03235615510493517,
   "inner": 55,
   "median_ms": 0.7755083285949447,
   "q1_ms": 0.7735909640111707,
   "q3_ms": 0.7760625078596852,
   "tflops": 38.76782487490616
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009534358978271484,
   "rel_err": 0.0003247723507229239,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 1.1300215320661664,
   "inner": 200,
   "median_ms": 0.12583511997945607,
   "q1_ms": 0.1246572972740978,
   "q3_ms": 0.12597524444572628,
   "tflops": 34.13170581234555
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009534358978271484,
   "rel_err": 0.0003247723507229239,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.03835535002872348,
   "inner": 200,
   "median_ms": 0.12500616954639554,
   "q1_ms": 0.12377644947264344,
   "q3_ms": 0.12516892224084586,
   "tflops": 34.35804257969795
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009534358978271484,
   "rel_err": 0.0003247723507229239,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.03922384697943926,
   "inner": 200,
   "median_ms": 0.12457357021048666,
   "q1_ms": 0.12447982269804925,
   "q3_ms": 0.12531497224699706,
   "tflops": 34.47735574041088
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001054137945175171,
   "rel_err": 0.00035907485289499164,
   "xla_temp_mb": 256.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.64617003290914,
   "inner": 112,
   "median_ms": 0.35881476651411504,
   "q1_ms": 0.35799691977445036,
   "q3_ms": 0.3590589684400973,
   "tflops": 11.969873307404825
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013262033462524414,
   "rel_err": 0.000557274033781141,
   "xla_temp_mb": 8.250244140625,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 1.2062800908461213,
   "inner": 115,
   "median_ms": 0.3298445220064858,
   "q1_ms": 0.32748961821198463,
   "q3_ms": 0.3306143046559199,
   "tflops": 45.57415549773603
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013262033462524414,
   "rel_err": 0.000557274033781141,
   "xla_temp_mb": 0.25048828125,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 1.211765965912491,
   "inner": 102,
   "median_ms": 0.35639523583299976,
   "q1_ms": 0.35314867729503735,
   "q3_ms": 0.3568156378106305,
   "tflops": 42.17897442109439
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013262033462524414,
   "rel_err": 0.000557274033781141,
   "xla_temp_mb": 0.25048828125,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 1.207650267984718,
   "inner": 109,
   "median_ms": 0.35676300961421714,
   "q1_ms": 0.354681178978687,
   "q3_ms": 0.35865093611283316,
   "tflops": 42.135493677596095
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013262033462524414,
   "rel_err": 0.0004974621115252376,
   "xla_temp_mb": 384.0,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.03916597692295909,
   "inner": 48,
   "median_ms": 0.9133138519246131,
   "q1_ms": 0.9105391776150402,
   "q3_ms": 0.9169282178239275,
   "tflops": 16.459167354488788
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00014570355415344238,
   "rel_err": 0.0005782486987300217,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 1.1261195910628885,
   "inner": 101,
   "median_ms": 0.39063263361775635,
   "q1_ms": 0.38790734632328006,
   "q3_ms": 0.39196803031655236,
   "tflops": 47.18499790787107
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00014570355415344238,
   "rel_err": 0.0005782486987300217,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 1.1263665209989995,
   "inner": 101,
   "median_ms": 0.39200293186056145,
   "q1_ms": 0.3912447672798344,
   "q3_ms": 0.39505772292613983,
   "tflops": 47.020056489160154
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00014570355415344238,
   "rel_err": 0.0005782486987300217,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.04235825897194445,
   "inner": 93,
   "median_ms": 0.3944143417581756,
   "q1_ms": 0.39128294610168024,
   "q3_ms": 0.3949172421789137,
   "tflops": 46.73258055940846
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00019633769989013672,
   "rel_err": 0.0007791986572556198,
   "xla_temp_mb": 549.3166656494141,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.04736873600631952,
   "inner": 63,
   "median_ms": 0.6591928753233145,
   "q1_ms": 0.6563379595588361,
   "q3_ms": 0.6606104831758237,
   "tflops": 27.96146725790938
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00029903650283813477,
   "rel_err": 0.0005960106500424445,
   "xla_temp_mb": 96.00048828125,
   "xla_arg_mb": 46.875,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 1.2216632801573724,
   "inner": 37,
   "median_ms": 1.2251492149221737,
   "q1_ms": 1.223477311558216,
   "q3_ms": 1.2260358371595677,
   "tflops": 52.65644316157691
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00029903650283813477,
   "rel_err": 0.0005960106500424445,
   "xla_temp_mb": 72.18807983398438,
   "xla_arg_mb": 46.875,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 1.2511975830420852,
   "inner": 36,
   "median_ms": 1.2159162498493161,
   "q1_ms": 1.2143586063757539,
   "q3_ms": 1.2266249622270051,
   "tflops": 53.05628574993938
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00029903650283813477,
   "rel_err": 0.0005960106500424445,
   "xla_temp_mb": 72.18807983398438,
   "xla_arg_mb": 46.875,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 1.2351302551105618,
   "inner": 37,
   "median_ms": 1.217568972827615,
   "q1_ms": 1.2144101878024034,
   "q3_ms": 1.2261479860171676,
   "tflops": 52.98426737187701
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00031238794326782227,
   "rel_err": 0.0006205877871252596,
   "xla_temp_mb": 824.1582183837891,
   "xla_arg_mb": 46.875,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.0368137639015913,
   "inner": 26,
   "median_ms": 1.8087448462700615,
   "q1_ms": 1.799579114259149,
   "q3_ms": 1.810979075694027,
   "tflops": 35.66672222067954
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013152360916137695,
   "rel_err": 0.00044277688721194863,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 1.179918410954997,
   "inner": 129,
   "median_ms": 0.2898091627212689,
   "q1_ms": 0.2879549815976458,
   "q3_ms": 0.2900399727920114,
   "tflops": 31.80023679535528
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013152360916137695,
   "rel_err": 0.00044277688721194863,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 1.1692618278320879,
   "inner": 127,
   "median_ms": 0.28249130796963773,
   "q1_ms": 0.2813811494609503,
   "q3_ms": 0.2837302795735987,
   "tflops": 32.624012633303884
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013152360916137695,
   "rel_err": 0.00044277688721194863,
   "xla_temp_mb": 48.0,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.04308899608440697,
   "inner": 125,
   "median_ms": 0.281408304348588,
   "q1_ms": 0.280608544126153,
   "q3_ms": 0.2830575406551361,
   "tflops": 32.74956658202913
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013723373413085938,
   "rel_err": 0.000462000141851604,
   "xla_temp_mb": 549.3166656494141,
   "xla_arg_mb": 35.15625,
   "xla_out_mb": 11.71875,
   "compile_s": 0.4189130300655961,
   "inner": 68,
   "median_ms": 0.6554871044286034,
   "q1_ms": 0.6531650536036229,
   "q3_ms": 0.6575462860329186,
   "tflops": 14.059773163705037
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0014880597591400146,
   "rel_err": 0.0005150618962943554,
   "xla_temp_mb": 96.00048828125,
   "xla_arg_mb": 46.875,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 1.2861243919469416,
   "inner": 56,
   "median_ms": 0.7579790918888258,
   "q1_ms": 0.7573556878404426,
   "q3_ms": 0.7630494281849158,
   "tflops": 42.55526352266593
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0014880597591400146,
   "rel_err": 0.0005150618962943554,
   "xla_temp_mb": 72.18798828125,
   "xla_arg_mb": 46.875,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 1.2637640300672501,
   "inner": 56,
   "median_ms": 0.7757079646190894,
   "q1_ms": 0.7749184982718101,
   "q3_ms": 0.7794154454521569,
   "tflops": 41.582659288330596
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0014880597591400146,
   "rel_err": 0.0005150618962943554,
   "xla_temp_mb": 72.18798828125,
   "xla_arg_mb": 46.875,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 1.2716774391010404,
   "inner": 56,
   "median_ms": 0.7772324315737933,
   "q1_ms": 0.7749827782390639,
   "q3_ms": 0.7797036793947751,
   "tflops": 41.50109888580672
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001489013433456421,
   "rel_err": 0.0005425030831247568,
   "xla_temp_mb": 823.9751129150391,
   "xla_arg_mb": 46.875,
   "xla_out_mb": 35.156272888183594,
   "compile_s": 0.038827777840197086,
   "inner": 25,
   "median_ms": 1.8417587596923113,
   "q1_ms": 1.8377189617604017,
   "q3_ms": 1.8455433612689376,
   "tflops": 17.513694358857705
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00013641268014907837,
   "rel_err": 0.0006847892655059695,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.9380263648927212,
   "inner": 73,
   "median_ms": 0.5519816983644276,
   "q1_ms": 0.5503730060634752,
   "q3_ms": 0.5541203977310494,
   "tflops": 62.24796668043716
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00013641268014907837,
   "rel_err": 0.0006847892655059695,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.9424483641050756,
   "inner": 72,
   "median_ms": 0.5482632938461999,
   "q1_ms": 0.5469830761689487,
   "q3_ms": 0.5541414656262431,
   "tflops": 62.67014179803669
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00013641268014907837,
   "rel_err": 0.0006847892655059695,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.9488540180027485,
   "inner": 74,
   "median_ms": 0.5503526095905014,
   "q1_ms": 0.547708959218014,
   "q3_ms": 0.5519480815761395,
   "tflops": 62.43222575716668
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00015877187252044678,
   "rel_err": 0.0007970320293679833,
   "xla_temp_mb": 1024.000015258789,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.018277692142874002,
   "inner": 40,
   "median_ms": 1.1213263031095266,
   "q1_ms": 1.1183290771441534,
   "q3_ms": 1.1217887979000807,
   "tflops": 30.642051535505523
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00025177001953125,
   "rel_err": 0.0006236736080609262,
   "xla_temp_mb": 16.500244140625,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 1.030411692801863,
   "inner": 24,
   "median_ms": 1.9225811993237585,
   "q1_ms": 1.9118624768452719,
   "q3_ms": 1.9610046195642403,
   "tflops": 62.55084795913924
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00025177001953125,
   "rel_err": 0.0006236736080609262,
   "xla_temp_mb": 0.50048828125,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 1.0493711731396616,
   "inner": 24,
   "median_ms": 1.9255037089654554,
   "q1_ms": 1.9144007916717478,
   "q3_ms": 1.951219436402122,
   "tflops": 62.45590892816998
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00025177001953125,
   "rel_err": 0.0006236736080609262,
   "xla_temp_mb": 0.50048828125,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 1.0464642709121108,
   "inner": 24,
   "median_ms": 1.9263020464374374,
   "q1_ms": 1.907082502536165,
   "q3_ms": 1.9435515811589237,
   "tflops": 62.4300246736543
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0002517402172088623,
   "rel_err": 0.000623599742539227,
   "xla_temp_mb": 1536.000503540039,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.0344984601251781,
   "inner": 16,
   "median_ms": 2.8391563100740314,
   "q1_ms": 2.8372200613375753,
   "q3_ms": 2.8397659698384814,
   "tflops": 42.35733124706481
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008454322814941406,
   "rel_err": 0.00032341157202608883,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 1.115656099980697,
   "inner": 104,
   "median_ms": 0.3621403929383422,
   "q1_ms": 0.36135539882637274,
   "q3_ms": 0.36836413290494907,
   "tflops": 47.43980378605551
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008454322814941406,
   "rel_err": 0.00032341157202608883,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.03909953706897795,
   "inner": 100,
   "median_ms": 0.3632518486119807,
   "q1_ms": 0.3605189046356827,
   "q3_ms": 0.36339359590783715,
   "tflops": 47.294650391032796
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008454322814941406,
   "rel_err": 0.00032341157202608883,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.03977192612364888,
   "inner": 103,
   "median_ms": 0.36118642734454093,
   "q1_ms": 0.3593928436070536,
   "q3_ms": 0.36272783402028014,
   "tflops": 47.56510179606465
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013530254364013672,
   "rel_err": 0.0005175861879251897,
   "xla_temp_mb": 1024.000015258789,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.515314576216042,
   "inner": 36,
   "median_ms": 1.2737049708246355,
   "q1_ms": 1.2683070866235842,
   "q3_ms": 1.2764022071173207,
   "tflops": 13.488107197129981
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001447916030883789,
   "rel_err": 0.0005423835827969015,
   "xla_temp_mb": 16.500244140625,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 1.200056267902255,
   "inner": 41,
   "median_ms": 1.1044327033365646,
   "q1_ms": 1.1042527082126314,
   "q3_ms": 1.1098507433993423,
   "tflops": 54.443826194520184
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001447916030883789,
   "rel_err": 0.0005423835827969015,
   "xla_temp_mb": 0.50048828125,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 1.2020838619209826,
   "inner": 39,
   "median_ms": 1.1578631014204943,
   "q1_ms": 1.150603895672621,
   "q3_ms": 1.1611573326473055,
   "tflops": 51.93147797026405
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001447916030883789,
   "rel_err": 0.0005423835827969015,
   "xla_temp_mb": 0.50048828125,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 1.2244977939408273,
   "inner": 39,
   "median_ms": 1.1573579783240955,
   "q1_ms": 1.1535173592468104,
   "q3_ms": 1.1633325768157077,
   "tflops": 51.95414320387732
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0014529973268508911,
   "rel_err": 0.000590299372561276,
   "xla_temp_mb": 1536.000503540039,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.040480653988197446,
   "inner": 14,
   "median_ms": 3.3214355020650794,
   "q1_ms": 3.2831159935865015,
   "q3_ms": 3.3234833889374773,
   "tflops": 18.103480289355275
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011336803436279297,
   "rel_err": 0.0006243938696570694,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 1.1201440969016403,
   "inner": 32,
   "median_ms": 1.4123147193458863,
   "q1_ms": 1.4118669205345213,
   "q3_ms": 1.4273757769842632,
   "tflops": 52.20366182556473
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011336803436279297,
   "rel_err": 0.0006243938696570694,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 1.1823693050537258,
   "inner": 30,
   "median_ms": 1.4621497985596457,
   "q1_ms": 1.4550148509442806,
   "q3_ms": 1.4708403148688376,
   "tflops": 50.42438201108325
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011336803436279297,
   "rel_err": 0.0006243938696570694,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.04339800588786602,
   "inner": 30,
   "median_ms": 1.4546838278571765,
   "q1_ms": 1.4537343328508237,
   "q3_ms": 1.4637591627736888,
   "tflops": 50.68317842551746
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011709332466125488,
   "rel_err": 0.0006449115462601185,
   "xla_temp_mb": 2197.265884399414,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.019015519879758358,
   "inner": 19,
   "median_ms": 2.357126421932327,
   "q1_ms": 2.3557491061326705,
   "q3_ms": 2.3578208944711245,
   "tflops": 31.27876354614837
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00015710480511188507,
   "rel_err": 0.0007145414128899574,
   "xla_temp_mb": 192.00048828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.2253672908991575,
   "inner": 10,
   "median_ms": 4.571020998992026,
   "q1_ms": 4.570599994622171,
   "q3_ms": 4.586304456461221,
   "tflops": 56.45303315318464
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00015710480511188507,
   "rel_err": 0.0007145414128899574,
   "xla_temp_mb": 144.37548828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.2988526080735028,
   "inner": 10,
   "median_ms": 4.570539016276598,
   "q1_ms": 4.545201547443867,
   "q3_ms": 4.58229900104925,
   "tflops": 56.458986364855384
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00015710480511188507,
   "rel_err": 0.0007145414128899574,
   "xla_temp_mb": 144.37548828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.2937759410124272,
   "inner": 10,
   "median_ms": 4.55215300898999,
   "q1_ms": 4.53742160461843,
   "q3_ms": 4.579158953856677,
   "tflops": 56.687022490321446
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00015683472156524658,
   "rel_err": 0.0007214891957119107,
   "xla_temp_mb": 3296.264907836914,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.03869562805630267,
   "inner": 7,
   "median_ms": 6.525501409279449,
   "q1_ms": 6.522800706859146,
   "q3_ms": 6.544686338332083,
   "tflops": 39.544547432485174
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009112358093261719,
   "rel_err": 0.00033203110797330737,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 1.1822618381120265,
   "inner": 50,
   "median_ms": 0.8646249398589134,
   "q1_ms": 0.8562058582901955,
   "q3_ms": 0.8673237380571663,
   "tflops": 42.63582774516699
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009112358093261719,
   "rel_err": 0.00033203110797330737,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 1.1845574968028814,
   "inner": 52,
   "median_ms": 0.8406349224969745,
   "q1_ms": 0.8399765580319442,
   "q3_ms": 0.8500901816520267,
   "tflops": 43.8525678786949
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009112358093261719,
   "rel_err": 0.00033203110797330737,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.04346047714352608,
   "inner": 50,
   "median_ms": 0.8463439997285604,
   "q1_ms": 0.8388822129927576,
   "q3_ms": 0.8469141880050302,
   "tflops": 43.556757077291294
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001251220703125,
   "rel_err": 0.00045591293019242585,
   "xla_temp_mb": 2197.265884399414,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.4325488011818379,
   "inner": 20,
   "median_ms": 2.3637103498913348,
   "q1_ms": 2.3627485788892955,
   "q3_ms": 2.36422031885013,
   "tflops": 15.595819513881946
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001356363296508789,
   "rel_err": 0.0007203135173767805,
   "xla_temp_mb": 192.00048828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.291944975964725,
   "inner": 18,
   "median_ms": 2.547875551196436,
   "q1_ms": 2.5475806065110698,
   "q3_ms": 2.550076137089895,
   "tflops": 50.639835976059615
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001356363296508789,
   "rel_err": 0.0007203135173767805,
   "xla_temp_mb": 144.37548828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.2975841220468283,
   "inner": 19,
   "median_ms": 2.511991371743773,
   "q1_ms": 2.5021073984374342,
   "q3_ms": 2.527755789311701,
   "tflops": 51.36323374806585
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001356363296508789,
   "rel_err": 0.0007203135173767805,
   "xla_temp_mb": 144.37548828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.290293080965057,
   "inner": 19,
   "median_ms": 2.509715060066236,
   "q1_ms": 2.504787156603446,
   "q3_ms": 2.5222099717020203,
   "tflops": 51.40982020349148
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013564825057983398,
   "rel_err": 0.0005977512337267399,
   "xla_temp_mb": 3296.264907836914,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.0678949519060552,
   "inner": 7,
   "median_ms": 6.674471144963588,
   "q1_ms": 6.647113344765135,
   "q3_ms": 6.681726131189082,
   "tflops": 19.330969779884168
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011709332466125488,
   "rel_err": 0.0005204850458540022,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.9394409470260143,
   "inner": 21,
   "median_ms": 2.111457853711077,
   "q1_ms": 2.1062871175152917,
   "q3_ms": 2.1180604096679456,
   "tflops": 65.0919710428691
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011709332466125488,
   "rel_err": 0.0005204850458540022,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.9413465620018542,
   "inner": 21,
   "median_ms": 2.107610659939902,
   "q1_ms": 2.1052983134896275,
   "q3_ms": 2.120858523994684,
   "tflops": 65.21078873074168
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011709332466125488,
   "rel_err": 0.0005204850458540022,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.9424188800621778,
   "inner": 22,
   "median_ms": 2.116953086277301,
   "q1_ms": 2.1093313169496306,
   "q3_ms": 2.121399458370764,
   "tflops": 64.92300389787513
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000124320387840271,
   "rel_err": 0.0005526096792891622,
   "xla_temp_mb": 4096.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.020992256002500653,
   "inner": 11,
   "median_ms": 4.211733625693755,
   "q1_ms": 4.210951325314289,
   "q3_ms": 4.212942175482484,
   "tflops": 32.63239456397509
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00012206658720970154,
   "rel_err": 0.0006644873647019267,
   "xla_temp_mb": 33.000244140625,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 1.0275019768159837,
   "inner": 6,
   "median_ms": 7.579628494568169,
   "q1_ms": 7.539544487372041,
   "q3_ms": 7.599346844169001,
   "tflops": 63.4643686688243
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00012206658720970154,
   "rel_err": 0.0006644873647019267,
   "xla_temp_mb": 1.00048828125,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 1.059117894852534,
   "inner": 6,
   "median_ms": 7.560941856354475,
   "q1_ms": 7.482142905549457,
   "q3_ms": 7.620285168134918,
   "tflops": 63.62121892892491
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00012206658720970154,
   "rel_err": 0.0006644873647019267,
   "xla_temp_mb": 1.00048828125,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 1.052788021042943,
   "inner": 6,
   "median_ms": 7.512786968921621,
   "q1_ms": 7.450952233436206,
   "q3_ms": 7.66289490275085,
   "tflops": 64.0290133530896
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00012206658720970154,
   "rel_err": 0.0006644873647019267,
   "xla_temp_mb": 6144.0,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.04243549797683954,
   "inner": 4,
   "median_ms": 10.686997033189982,
   "q1_ms": 10.685394488973543,
   "q3_ms": 10.68781825597398,
   "tflops": 45.01136621055228
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009582042694091797,
   "rel_err": 0.000337931647663936,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 1.11674726405181,
   "inner": 37,
   "median_ms": 1.1921036497664612,
   "q1_ms": 1.1850801497899197,
   "q3_ms": 1.2008398362259203,
   "tflops": 57.64555519099574
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009582042694091797,
   "rel_err": 0.000337931647663936,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.04137500189244747,
   "inner": 37,
   "median_ms": 1.189507166477474,
   "q1_ms": 1.18709312487595,
   "q3_ms": 1.1918806609370418,
   "tflops": 57.771385219562156
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009582042694091797,
   "rel_err": 0.000337931647663936,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.04065299406647682,
   "inner": 37,
   "median_ms": 1.188954193698796,
   "q1_ms": 1.1887809614071974,
   "q3_ms": 1.196365526600464,
   "tflops": 57.79825421382808
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001003861427307129,
   "rel_err": 0.0003540336329024285,
   "xla_temp_mb": 4096.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.5023693279363215,
   "inner": 10,
   "median_ms": 4.8450424103066325,
   "q1_ms": 4.831437394022942,
   "q3_ms": 4.854979296214879,
   "tflops": 14.183462375853773
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001342177391052246,
   "rel_err": 0.0006075144046917558,
   "xla_temp_mb": 33.000244140625,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 1.2254456309601665,
   "inner": 12,
   "median_ms": 4.083973491409172,
   "q1_ms": 4.037983201366539,
   "q3_ms": 4.136401703969264,
   "tflops": 58.893175747085806
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001342177391052246,
   "rel_err": 0.0006075144046917558,
   "xla_temp_mb": 1.00048828125,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 1.2117465469054878,
   "inner": 11,
   "median_ms": 4.125566539269957,
   "q1_ms": 4.120645600117065,
   "q3_ms": 4.175525878302076,
   "tflops": 58.29942779654237
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001342177391052246,
   "rel_err": 0.0006075144046917558,
   "xla_temp_mb": 1.00048828125,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 1.213885940844193,
   "inner": 11,
   "median_ms": 4.169669997116381,
   "q1_ms": 4.131215592761608,
   "q3_ms": 4.174255459061401,
   "tflops": 57.682782748355415
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013422966003417969,
   "rel_err": 0.0005387319251894951,
   "xla_temp_mb": 6144.0,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.049199221888557076,
   "inner": 4,
   "median_ms": 12.313162966165692,
   "q1_ms": 12.279666785616428,
   "q3_ms": 12.345293973339722,
   "tflops": 19.533418767939622
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.511062383651733e-05,
   "rel_err": 0.0005790345603600144,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.9478051150217652,
   "inner": 5,
   "median_ms": 8.190814638510346,
   "q1_ms": 8.188112708739936,
   "q3_ms": 8.22895357850939,
   "tflops": 67.11857588660844
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.511062383651733e-05,
   "rel_err": 0.0005790345603600144,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.9477268350310624,
   "inner": 5,
   "median_ms": 8.175568608567119,
   "q1_ms": 8.174620592035353,
   "q3_ms": 8.19998059887439,
   "tflops": 67.24374049187416
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.511062383651733e-05,
   "rel_err": 0.0005790345603600144,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.9472633260302246,
   "inner": 5,
   "median_ms": 8.184004621580243,
   "q1_ms": 8.177696610800922,
   "q3_ms": 8.19216959644109,
   "tflops": 67.17442612854342
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 7.735192775726318e-05,
   "rel_err": 0.0006878975545987487,
   "xla_temp_mb": 16384.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.03300302312709391,
   "inner": 3,
   "median_ms": 16.30902065274616,
   "q1_ms": 16.307754131654896,
   "q3_ms": 16.31238750026872,
   "tflops": 33.70869567176804
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00010075420141220093,
   "rel_err": 0.0006305324495770037,
   "xla_temp_mb": 66.000244140625,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.0383139690384269,
   "inner": 1,
   "median_ms": 30.137700960040092,
   "q1_ms": 29.917001957073808,
   "q3_ms": 30.269065871834755,
   "tflops": 63.84512711036736
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00010075420141220093,
   "rel_err": 0.0006305324495770037,
   "xla_temp_mb": 2.00048828125,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.0529542409349233,
   "inner": 1,
   "median_ms": 30.289700953289866,
   "q1_ms": 29.793627094477415,
   "q3_ms": 30.61029000673443,
   "tflops": 63.52473903836981
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00010075420141220093,
   "rel_err": 0.0006305324495770037,
   "xla_temp_mb": 2.00048828125,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.0476899140048772,
   "inner": 1,
   "median_ms": 30.82504984922707,
   "q1_ms": 29.80337687768042,
   "q3_ms": 31.207983614876866,
   "tflops": 62.42148376140412
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00010014325380325317,
   "rel_err": 0.0006270954618230462,
   "xla_temp_mb": 24576.0,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.07256955094635487,
   "inner": 1,
   "median_ms": 41.41590092331171,
   "q1_ms": 41.33753199130297,
   "q3_ms": 41.429082048125565,
   "tflops": 46.45909676505332
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007331967353820801,
   "rel_err": 0.0002779076457954943,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 1.134107128949836,
   "inner": 11,
   "median_ms": 4.263250644064762,
   "q1_ms": 4.2498061877929345,
   "q3_ms": 4.281018310311166,
   "tflops": 64.47613098391979
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007331967353820801,
   "rel_err": 0.0002779076457954943,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.04359711613506079,
   "inner": 11,
   "median_ms": 4.280082542787899,
   "q1_ms": 4.2690088104626,
   "q3_ms": 4.291875594803555,
   "tflops": 64.22257145652007
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007331967353820801,
   "rel_err": 0.0002779076457954943,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.04397983499802649,
   "inner": 11,
   "median_ms": 4.266842543570833,
   "q1_ms": 4.254722036421299,
   "q3_ms": 4.288981552235782,
   "tflops": 64.42185389713498
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008221864700317383,
   "rel_err": 0.00031163793755695224,
   "xla_temp_mb": 16384.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.49155392684042454,
   "inner": 2,
   "median_ms": 18.603841541334987,
   "q1_ms": 18.376746971625835,
   "q3_ms": 18.794523319229484,
   "tflops": 14.775330478560674
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016331672668457031,
   "rel_err": 0.0008911209879443049,
   "xla_temp_mb": 66.000244140625,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.2756553289946169,
   "inner": 3,
   "median_ms": 15.362080031385025,
   "q1_ms": 15.35468501970172,
   "q3_ms": 15.621735791986186,
   "tflops": 62.62645893905428
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016331672668457031,
   "rel_err": 0.0008911209879443049,
   "xla_temp_mb": 2.00048828125,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.2515066920313984,
   "inner": 3,
   "median_ms": 16.05522135893504,
   "q1_ms": 15.899320016615093,
   "q3_ms": 16.192234470508993,
   "tflops": 59.922728737002934
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016331672668457031,
   "rel_err": 0.0008911209879443049,
   "xla_temp_mb": 2.00048828125,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.2347067662049085,
   "inner": 3,
   "median_ms": 16.037428053095937,
   "q1_ms": 15.96838499729832,
   "q3_ms": 16.08534968302896,
   "tflops": 59.9892121803332
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0014424324035644531,
   "rel_err": 0.0007870484841987491,
   "xla_temp_mb": 24576.0,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.065112448995933,
   "inner": 1,
   "median_ms": 47.10189695470035,
   "q1_ms": 46.83802754152566,
   "q3_ms": 47.49793093651533,
   "tflops": 20.425348797082655
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005143880844116211,
   "rel_err": 0.0006051270756870508,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.5320823909714818,
   "inner": 200,
   "median_ms": 0.035517209907993674,
   "q1_ms": 0.03511785762384534,
   "q3_ms": 0.03573048219550401,
   "tflops": 7.557898176556504
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005143880844116211,
   "rel_err": 0.0006051270756870508,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.5310147739946842,
   "inner": 200,
   "median_ms": 0.03687035525217652,
   "q1_ms": 0.03659368027001619,
   "q3_ms": 0.03694082726724446,
   "tflops": 7.280522635706196
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005143880844116211,
   "rel_err": 0.0006051270756870508,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.5353186919819564,
   "inner": 200,
   "median_ms": 0.029342324705794454,
   "q1_ms": 0.02929292561020702,
   "q3_ms": 0.0295221246778965,
   "tflops": 9.148404521165633
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004310011863708496,
   "rel_err": 0.0005070305778644979,
   "xla_temp_mb": 4.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.18917072610929608,
   "inner": 200,
   "median_ms": 0.0343177595641464,
   "q1_ms": 0.03423910995479673,
   "q3_ms": 0.03449320967774838,
   "tflops": 7.822056550581142
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000563204288482666,
   "rel_err": 0.0008504130528308451,
   "xla_temp_mb": 2.031494140625,
   "xla_arg_mb": 8.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 1.448246518848464,
   "inner": 200,
   "median_ms": 0.06538397981785238,
   "q1_ms": 0.06531252758577466,
   "q3_ms": 0.06539192982017994,
   "tflops": 14.369331732594736
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000563204288482666,
   "rel_err": 0.0008504130528308451,
   "xla_temp_mb": 0.03173828125,
   "xla_arg_mb": 8.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.8985309940762818,
   "inner": 200,
   "median_ms": 0.06501063471660018,
   "q1_ms": 0.0647705077426508,
   "q3_ms": 0.0650205573765561,
   "tflops": 14.451852379163077
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000563204288482666,
   "rel_err": 0.0008504130528308451,
   "xla_temp_mb": 0.03173828125,
   "xla_arg_mb": 8.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.8907003139611334,
   "inner": 200,
   "median_ms": 0.06503448006696999,
   "q1_ms": 0.0647990801371634,
   "q3_ms": 0.06504832999780774,
   "tflops": 14.44655350565599
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005519390106201172,
   "rel_err": 0.0008334030280821025,
   "xla_temp_mb": 4.0,
   "xla_arg_mb": 8.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.03295764420181513,
   "inner": 200,
   "median_ms": 0.0582879944704473,
   "q1_ms": 0.058110199752263725,
   "q3_ms": 0.058367797173559666,
   "tflops": 16.118655385824773
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009992122650146484,
   "rel_err": 0.0003183129010722041,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.615049023879692,
   "inner": 200,
   "median_ms": 0.022647290024906397,
   "q1_ms": 0.02234796469565481,
   "q3_ms": 0.022838115110062063,
   "tflops": 5.9264365781686825
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009992122650146484,
   "rel_err": 0.0003183129010722041,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.610185147030279,
   "inner": 200,
   "median_ms": 0.02365258987993002,
   "q1_ms": 0.023381815408356488,
   "q3_ms": 0.02401703502982855,
   "tflops": 5.674546790915613
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009992122650146484,
   "rel_err": 0.0003183129010722041,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.6128234299831092,
   "inner": 200,
   "median_ms": 0.023090639151632786,
   "q1_ms": 0.02242614224087447,
   "q3_ms": 0.023356362362392247,
   "tflops": 5.812646723142317
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011777877807617188,
   "rel_err": 0.0003752006159629673,
   "xla_temp_mb": 4.0,
   "xla_arg_mb": 6.0,
   "xla_out_mb": 2.0,
   "compile_s": 0.2589289741590619,
   "inner": 200,
   "median_ms": 0.03439466585405171,
   "q1_ms": 0.03431855991948396,
   "q3_ms": 0.03463678527623415,
   "tflops": 3.9022832368696814
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001806020736694336,
   "rel_err": 0.0005865312996320426,
   "xla_temp_mb": 2.031494140625,
   "xla_arg_mb": 8.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 1.682232677936554,
   "inner": 200,
   "median_ms": 0.06050228956155479,
   "q1_ms": 0.06013761798385531,
   "q3_ms": 0.06066359463147819,
   "tflops": 7.764368115723388
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001806020736694336,
   "rel_err": 0.0005865312996320426,
   "xla_temp_mb": 0.03173828125,
   "xla_arg_mb": 8.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.9933245091233402,
   "inner": 200,
   "median_ms": 0.06734602502547204,
   "q1_ms": 0.0667049502953887,
   "q3_ms": 0.06750129454303533,
   "tflops": 6.975349292290429
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001806020736694336,
   "rel_err": 0.0005865312996320426,
   "xla_temp_mb": 0.03173828125,
   "xla_arg_mb": 8.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 1.0087986579164863,
   "inner": 200,
   "median_ms": 0.06707837455905974,
   "q1_ms": 0.06695002492051572,
   "q3_ms": 0.06760684540495276,
   "tflops": 7.003181742073877
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001806020736694336,
   "rel_err": 0.0006100231548771262,
   "xla_temp_mb": 4.0,
   "xla_arg_mb": 8.0,
   "xla_out_mb": 6.000022888183594,
   "compile_s": 0.7180497571825981,
   "inner": 200,
   "median_ms": 0.07078916067257524,
   "q1_ms": 0.0705362152075395,
   "q3_ms": 0.0709807372186333,
   "tflops": 6.636073143638681
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00029471516609191895,
   "rel_err": 0.00043253504554741085,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.5411123381927609,
   "inner": 200,
   "median_ms": 0.04924532026052475,
   "q1_ms": 0.049232947058044374,
   "q3_ms": 0.04927299451082945,
   "tflops": 87.21574503482036
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00029471516609191895,
   "rel_err": 0.00043253504554741085,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.5440514711663127,
   "inner": 200,
   "median_ms": 0.04921552026644349,
   "q1_ms": 0.04918652237392962,
   "q3_ms": 0.049231472657993436,
   "tflops": 87.26855416234272
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00029471516609191895,
   "rel_err": 0.00043253504554741085,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.5429858539719135,
   "inner": 200,
   "median_ms": 0.04922157037071884,
   "q1_ms": 0.049202697118744254,
   "q3_ms": 0.04922504769638181,
   "tflops": 87.2578274860367
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003890395164489746,
   "rel_err": 0.0005709690158255398,
   "xla_temp_mb": 72.00001525878906,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.018136641941964626,
   "inner": 200,
   "median_ms": 0.10086274007335305,
   "q1_ms": 0.09908323991112411,
   "q3_ms": 0.10194240952841938,
   "tflops": 42.58229840748386
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005009770393371582,
   "rel_err": 0.0006858734996058047,
   "xla_temp_mb": 8.125244140625,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 1.5380317179951817,
   "inner": 93,
   "median_ms": 0.42852006542185944,
   "q1_ms": 0.42767366719862787,
   "q3_ms": 0.4350721976289185,
   "tflops": 35.07977046816062
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005009770393371582,
   "rel_err": 0.0006858734996058047,
   "xla_temp_mb": 0.12548828125,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.9145427639596164,
   "inner": 96,
   "median_ms": 0.4053942490524302,
   "q1_ms": 0.404495667074419,
   "q3_ms": 0.410931953714074,
   "tflops": 37.08090475169972
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005009770393371582,
   "rel_err": 0.0006858734996058047,
   "xla_temp_mb": 0.12548828125,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.911350553156808,
   "inner": 96,
   "median_ms": 0.4055619695767139,
   "q1_ms": 0.404849880093631,
   "q3_ms": 0.411247463489417,
   "tflops": 37.06556990954881
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005003213882446289,
   "rel_err": 0.0006710723391734064,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.031869968166574836,
   "inner": 126,
   "median_ms": 0.269158261029848,
   "q1_ms": 0.2657198217800922,
   "q3_ms": 0.26998215536069536,
   "tflops": 55.84961605296223
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009677410125732422,
   "rel_err": 0.0002774132299236953,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.62613774696365,
   "inner": 200,
   "median_ms": 0.05144226481206715,
   "q1_ms": 0.0514163903426379,
   "q3_ms": 0.0514547893544659,
   "tflops": 41.7455113192499
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009677410125732422,
   "rel_err": 0.0002774132299236953,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.6236200227867812,
   "inner": 200,
   "median_ms": 0.05121021531522274,
   "q1_ms": 0.05120106739923358,
   "q3_ms": 0.051241592736914754,
   "tflops": 41.934673283079114
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009677410125732422,
   "rel_err": 0.0002774132299236953,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.6268553847912699,
   "inner": 200,
   "median_ms": 0.05125731462612748,
   "q1_ms": 0.05122066708281636,
   "q3_ms": 0.05127866461407393,
   "tflops": 41.8961403589676
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00148773193359375,
   "rel_err": 0.0004264741437509656,
   "xla_temp_mb": 72.00001525878906,
   "xla_arg_mb": 24.0,
   "xla_out_mb": 8.0,
   "compile_s": 0.6891566319391131,
   "inner": 200,
   "median_ms": 0.1269210153259337,
   "q1_ms": 0.12685141758993268,
   "q3_ms": 0.12705079221632332,
   "tflops": 16.919842962847824
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001967310905456543,
   "rel_err": 0.000911962240934372,
   "xla_temp_mb": 8.125244140625,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 1.6707123520318419,
   "inner": 134,
   "median_ms": 0.2621230443439154,
   "q1_ms": 0.260744470609491,
   "q3_ms": 0.2636758388200802,
   "tflops": 28.674292208122182
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001967310905456543,
   "rel_err": 0.000911962240934372,
   "xla_temp_mb": 0.12548828125,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.9868635660968721,
   "inner": 124,
   "median_ms": 0.2928266598422441,
   "q1_ms": 0.2919381931273928,
   "q3_ms": 0.29503338267996665,
   "tflops": 25.66772018657466
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001967310905456543,
   "rel_err": 0.000911962240934372,
   "xla_temp_mb": 0.12548828125,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.9967790090013295,
   "inner": 123,
   "median_ms": 0.2941831386065095,
   "q1_ms": 0.291796430483521,
   "q3_ms": 0.2950760980734859,
   "tflops": 25.549366301558948
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0019021034240722656,
   "rel_err": 0.0004917786573059857,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 32.0,
   "xla_out_mb": 24.000022888183594,
   "compile_s": 0.03864142904058099,
   "inner": 101,
   "median_ms": 0.35732540550946007,
   "q1_ms": 0.35683604966996624,
   "q3_ms": 0.3575419403133121,
   "tflops": 21.0345882271755
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0002888590097427368,
   "rel_err": 0.0006460706354118884,
   "xla_temp_mb": 40.0,
   "xla_arg_mb": 28.125,
   "xla_out_mb": 9.375,
   "compile_s": 0.6413314370438457,
   "inner": 184,
   "median_ms": 0.15725567936897278,
   "q1_ms": 0.15722138050477952,
   "q3_ms": 0.15954086088838623,
   "tflops": 37.50732580004833
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0002888590097427368,
   "rel_err": 0.0006460706354118884,
   "xla_temp_mb": 40.0,
   "xla_arg_mb": 28.125,
   "xla_out_mb": 9.375,
   "compile_s": 0.6343865250237286,
   "inner": 193,
   "median_ms": 0.1600620148635899,
   "q1_ms": 0.15979261352280105,
   "q3_ms": 0.16100414053879564,
   "tflops": 36.84971731129759
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0002888590097427368,
   "rel_err": 0.0006460706354118884,
   "xla_temp_mb": 40.0,
   "xla_arg_mb": 28.125,
   "xla_out_mb": 9.375,
   "compile_s": 0.6366270000580698,
   "inner": 186,
   "median_ms": 0.16216591866787083,
   "q1_ms": 0.15952410013665275,
   "q3_ms": 0.1623284732622485,
   "tflops": 36.37163744670717
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003879666328430176,
   "rel_err": 0.0008677377481944859,
   "xla_temp_mb": 87.89088439941406,
   "xla_arg_mb": 28.125,
   "xla_out_mb": 9.375,
   "compile_s": 0.018378312001004815,
   "inner": 174,
   "median_ms": 0.15299528805089407,
   "q1_ms": 0.15198038262047475,
   "q3_ms": 0.1531849172362394,
   "tflops": 38.551775516367165
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003733038902282715,
   "rel_err": 0.0008604301838204265,
   "xla_temp_mb": 80.00048828125,
   "xla_arg_mb": 37.5,
   "xla_out_mb": 28.125022888183594,
   "compile_s": 1.7133067410904914,
   "inner": 51,
   "median_ms": 0.8294954667707868,
   "q1_ms": 0.8254291183844793,
   "q3_ms": 0.8354529227111853,
   "tflops": 24.887224616628895
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003733038902282715,
   "rel_err": 0.0008604301838204265,
   "xla_temp_mb": 60.078651428222656,
   "xla_arg_mb": 37.5,
   "xla_out_mb": 28.125022888183594,
   "compile_s": 0.9940975462086499,
   "inner": 54,
   "median_ms": 0.7959729258867877,
   "q1_ms": 0.7933816372382421,
   "q3_ms": 0.8026481847520227,
   "tflops": 25.935354493371552
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003733038902282715,
   "rel_err": 0.0008604301838204265,
   "xla_temp_mb": 60.078651428222656,
   "xla_arg_mb": 37.5,
   "xla_out_mb": 28.125022888183594,
   "compile_s": 0.9658215600065887,
   "inner": 54,
   "median_ms": 0.7983847624725766,
   "q1_ms": 0.7944429353638379,
   "q3_ms": 0.8032862420400811,
   "tflops": 25.85700650907536
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003719925880432129,
   "rel_err": 0.0008574077510274947,
   "xla_temp_mb": 131.90968322753906,
   "xla_arg_mb": 37.5,
   "xla_out_mb": 28.125022888183594,
   "compile_s": 0.03258794592693448,
   "inner": 94,
   "median_ms": 0.4258941081253455,
   "q1_ms": 0.42153150462446376,
   "q3_ms": 0.4307157397171126,
   "tflops": 48.47176705699878
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012070536613464355,
   "rel_err": 0.0003694213228300214,
   "xla_temp_mb": 40.0,
   "xla_arg_mb": 28.125,
   "xla_out_mb": 9.375,
   "compile_s": 0.6441521099768579,
   "inner": 200,
   "median_ms": 0.1263868692331016,
   "q1_ms": 0.1263076445320621,
   "q3_ms": 0.1272801449522376,
   "tflops": 23.334069574591574
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012070536613464355,
   "rel_err": 0.0003694213228300214,
   "xla_temp_mb": 40.0,
   "xla_arg_mb": 28.125,
   "xla_out_mb": 9.375,
   "compile_s": 0.6517216891516,
   "inner": 200,
   "median_ms": 0.12652197037823498,
   "q1_ms": 0.12646852468606085,
   "q3_ms": 0.12727859255392104,
   "tflops": 23.30915327340906
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012070536613464355,
   "rel_err": 0.0003694213228300214,
   "xla_temp_mb": 40.0,
   "xla_arg_mb": 28.125,
   "xla_out_mb": 9.375,
   "compile_s": 0.6466877029743046,
   "inner": 200,
   "median_ms": 0.12713732081465423,
   "q1_ms": 0.1264283445198089,
   "q3_ms": 0.1271568424999714,
   "tflops": 23.196335907529015
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001227259635925293,
   "rel_err": 0.00037560539203695953,
   "xla_temp_mb": 87.89088439941406,
   "xla_arg_mb": 28.125,
   "xla_out_mb": 9.375,
   "compile_s": 0.4684437080286443,
   "inner": 185,
   "median_ms": 0.15611467580940272,
   "q1_ms": 0.15544245913121346,
   "q3_ms": 0.15619291382766254,
   "tflops": 18.890728784528378
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0019232630729675293,
   "rel_err": 0.0007195310317911208,
   "xla_temp_mb": 80.00048828125,
   "xla_arg_mb": 37.5,
   "xla_out_mb": 28.125022888183594,
   "compile_s": 1.7126034130342305,
   "inner": 81,
   "median_ms": 0.510692987361072,
   "q1_ms": 0.5097145367022833,
   "q3_ms": 0.5153301976805116,
   "tflops": 20.211595333111866
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0019232630729675293,
   "rel_err": 0.0007195310317911208,
   "xla_temp_mb": 60.07861328125,
   "xla_arg_mb": 37.5,
   "xla_out_mb": 28.125022888183594,
   "compile_s": 0.9975453468505293,
   "inner": 81,
   "median_ms": 0.4871376174973485,
   "q1_ms": 0.4855932978292306,
   "q3_ms": 0.48998600722830976,
   "tflops": 21.18891998739182
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0019232630729675293,
   "rel_err": 0.0007195310317911208,
   "xla_temp_mb": 60.07861328125,
   "xla_arg_mb": 37.5,
   "xla_out_mb": 28.125022888183594,
   "compile_s": 1.0215109349228442,
   "inner": 78,
   "median_ms": 0.48565898890583176,
   "q1_ms": 0.48510469824601055,
   "q3_ms": 0.493190000550105,
   "tflops": 21.253431390727123
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017398595809936523,
   "rel_err": 0.0005386921693570912,
   "xla_temp_mb": 131.90968322753906,
   "xla_arg_mb": 37.5,
   "xla_out_mb": 28.125022888183594,
   "compile_s": 0.037959781009703875,
   "inner": 84,
   "median_ms": 0.4663012860830696,
   "q1_ms": 0.4611332634729999,
   "q3_ms": 0.466713124686586,
   "tflops": 22.135731356660237
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00017982721328735352,
   "rel_err": 0.00048050330951809883,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.5399880819022655,
   "inner": 160,
   "median_ms": 0.19744692399399355,
   "q1_ms": 0.1952195532794576,
   "q3_ms": 0.19903594729839824,
   "tflops": 87.01006243339917
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00017982721328735352,
   "rel_err": 0.00048050330951809883,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.5412535180803388,
   "inner": 163,
   "median_ms": 0.19586880897040382,
   "q1_ms": 0.19538197509281657,
   "q3_ms": 0.19850251850517797,
   "tflops": 87.71110251962533
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00017982721328735352,
   "rel_err": 0.00048050330951809883,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.5430864740628749,
   "inner": 168,
   "median_ms": 0.19528448908767176,
   "q1_ms": 0.1951789253369151,
   "q3_ms": 0.19948227938065038,
   "tflops": 87.97354702496214
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00016279518604278564,
   "rel_err": 0.00043499324237927794,
   "xla_temp_mb": 256.00025939941406,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.01998933800496161,
   "inner": 103,
   "median_ms": 0.37146048412377974,
   "q1_ms": 0.3711183067158009,
   "q3_ms": 0.3717715537754222,
   "tflops": 46.24952025388317
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004494190216064453,
   "rel_err": 0.000607359572313726,
   "xla_temp_mb": 16.250244140625,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 1.5061539320740849,
   "inner": 28,
   "median_ms": 1.5852622870754982,
   "q1_ms": 1.5807351779325733,
   "q3_ms": 1.5978672683039414,
   "tflops": 37.93034290554363
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004494190216064453,
   "rel_err": 0.000607359572313726,
   "xla_temp_mb": 0.25048828125,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.9464004219044,
   "inner": 30,
   "median_ms": 1.5140277333557606,
   "q1_ms": 1.5100858833951254,
   "q3_ms": 1.5222546954949698,
   "tflops": 39.71495423715002
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004494190216064453,
   "rel_err": 0.000607359572313726,
   "xla_temp_mb": 0.25048828125,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.9360280279070139,
   "inner": 29,
   "median_ms": 1.510879859842103,
   "q1_ms": 1.5104512386987436,
   "q3_ms": 1.527976533719178,
   "tflops": 39.79769916999485
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004520416259765625,
   "rel_err": 0.000610903836786747,
   "xla_temp_mb": 384.00001525878906,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.03676208411343396,
   "inner": 49,
   "median_ms": 0.9114404674619436,
   "q1_ms": 0.9093189574017817,
   "q3_ms": 0.9139945507239626,
   "tflops": 65.97199081081031
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011320114135742188,
   "rel_err": 0.0003317581140436232,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.6254824879579246,
   "inner": 166,
   "median_ms": 0.18116362611707076,
   "q1_ms": 0.17971594899286886,
   "q3_ms": 0.1817814211910927,
   "tflops": 47.415338145467736
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011320114135742188,
   "rel_err": 0.0003317581140436232,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.6296691379975528,
   "inner": 171,
   "median_ms": 0.18102280136568155,
   "q1_ms": 0.17830976884136773,
   "q3_ms": 0.1810721580581794,
   "tflops": 47.452224400436705
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011320114135742188,
   "rel_err": 0.0003317581140436232,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.6236564428545535,
   "inner": 171,
   "median_ms": 0.17983871321493422,
   "q1_ms": 0.17955228046388827,
   "q3_ms": 0.1818958127990975,
   "tflops": 47.76465777829349
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013012886047363281,
   "rel_err": 0.0003813680959865451,
   "xla_temp_mb": 256.00025939941406,
   "xla_arg_mb": 48.0,
   "xla_out_mb": 16.0,
   "compile_s": 0.604048253968358,
   "inner": 94,
   "median_ms": 0.4167075323416514,
   "q1_ms": 0.4151914667516471,
   "q3_ms": 0.41713039985204,
   "tflops": 20.613821266271852
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017702579498291016,
   "rel_err": 0.0008775067399255931,
   "xla_temp_mb": 16.250244140625,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 1.673870462924242,
   "inner": 48,
   "median_ms": 0.8939241233747452,
   "q1_ms": 0.8903127818484791,
   "q3_ms": 0.900009634885161,
   "tflops": 33.632352328181256
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017702579498291016,
   "rel_err": 0.0008775067399255931,
   "xla_temp_mb": 0.25048828125,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.9561426560394466,
   "inner": 46,
   "median_ms": 0.9547311954362238,
   "q1_ms": 0.9522679459263126,
   "q3_ms": 0.9623649884420244,
   "tflops": 31.490299275560155
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017702579498291016,
   "rel_err": 0.0008775067399255931,
   "xla_temp_mb": 0.25048828125,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.9723222241736948,
   "inner": 45,
   "median_ms": 0.9502319567319419,
   "q1_ms": 0.9499709768634703,
   "q3_ms": 0.960621054077314,
   "tflops": 31.639402210171298
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001769423484802246,
   "rel_err": 0.0006764822173863649,
   "xla_temp_mb": 384.00001525878906,
   "xla_arg_mb": 64.0,
   "xla_out_mb": 48.000022888183594,
   "compile_s": 0.038582568988204,
   "inner": 43,
   "median_ms": 1.0515612084418535,
   "q1_ms": 1.0467261065144178,
   "q3_ms": 1.052384348694495,
   "tflops": 28.59060493164097
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00018177181482315063,
   "rel_err": 0.0008210560190491378,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.642299324972555,
   "inner": 79,
   "median_ms": 0.5070223425856873,
   "q1_ms": 0.50341052480655,
   "q3_ms": 0.5106848732367912,
   "tflops": 72.70685511017682
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00018177181482315063,
   "rel_err": 0.0008210560190491378,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.6399649509694427,
   "inner": 77,
   "median_ms": 0.5165661818214825,
   "q1_ms": 0.5143926170497358,
   "q3_ms": 0.5216098121403681,
   "tflops": 71.3635566889271
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00018177181482315063,
   "rel_err": 0.0008210560190491378,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.6336204069666564,
   "inner": 77,
   "median_ms": 0.5167752879296804,
   "q1_ms": 0.5149114483894852,
   "q3_ms": 0.5248041487867374,
   "tflops": 71.33468039403661
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0001644045114517212,
   "rel_err": 0.0007426086231134832,
   "xla_temp_mb": 549.3166656494141,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.02184364292770624,
   "inner": 58,
   "median_ms": 0.743663757397183,
   "q1_ms": 0.7397966892940218,
   "q3_ms": 0.7454085761906002,
   "tflops": 49.570790069188924
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00042620301246643066,
   "rel_err": 0.0011019905796274543,
   "xla_temp_mb": 192.00048828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.7275641430169344,
   "inner": 13,
   "median_ms": 3.4482172296310845,
   "q1_ms": 3.43609988116301,
   "q3_ms": 3.4729247697843952,
   "tflops": 37.417596226617064
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00042620301246643066,
   "rel_err": 0.0011019905796274543,
   "xla_temp_mb": 144.18807983398438,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.9749578549526632,
   "inner": 14,
   "median_ms": 3.356789138966373,
   "q1_ms": 3.3556662882412116,
   "q3_ms": 3.389050493881639,
   "tflops": 38.43673065497621
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00042620301246643066,
   "rel_err": 0.0011019905796274543,
   "xla_temp_mb": 144.18807983398438,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.9701613748911768,
   "inner": 14,
   "median_ms": 3.371576279667871,
   "q1_ms": 3.3571630649800808,
   "q3_ms": 3.3893851380396103,
   "tflops": 38.268153913074144
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00042435526847839355,
   "rel_err": 0.001103577553294599,
   "xla_temp_mb": 824.1582183837891,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.03380066202953458,
   "inner": 22,
   "median_ms": 2.139868553389203,
   "q1_ms": 2.134396959180859,
   "q3_ms": 2.143557405691933,
   "tflops": 60.295292341974466
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008329153060913086,
   "rel_err": 0.0003060622839257121,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.6444049701094627,
   "inner": 93,
   "median_ms": 0.40858367868568946,
   "q1_ms": 0.4071648012325969,
   "q3_ms": 0.41196823635849583,
   "tflops": 45.11193413131697
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008329153060913086,
   "rel_err": 0.0003060622839257121,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.6506771328859031,
   "inner": 96,
   "median_ms": 0.4123775749273288,
   "q1_ms": 0.41075987489117927,
   "q3_ms": 0.41582131719527143,
   "tflops": 44.69690187020518
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008329153060913086,
   "rel_err": 0.0003060622839257121,
   "xla_temp_mb": 96.0,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.6472583319991827,
   "inner": 96,
   "median_ms": 0.4130635109807675,
   "q1_ms": 0.41050575479554635,
   "q3_ms": 0.4143357497620551,
   "tflops": 44.622677893371716
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012900829315185547,
   "rel_err": 0.00047405267832800746,
   "xla_temp_mb": 549.3166656494141,
   "xla_arg_mb": 70.3125,
   "xla_out_mb": 23.4375,
   "compile_s": 0.4298571851104498,
   "inner": 58,
   "median_ms": 0.7446408271789551,
   "q1_ms": 0.7423563551254058,
   "q3_ms": 0.7457641888698883,
   "tflops": 24.752873233971023
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002327442169189453,
   "rel_err": 0.0008719622856006026,
   "xla_temp_mb": 192.00048828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 1.709251469001174,
   "inner": 22,
   "median_ms": 2.033117449503731,
   "q1_ms": 2.031155629083514,
   "q3_ms": 2.0458517511459915,
   "tflops": 31.730582026014726
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002327442169189453,
   "rel_err": 0.0008719622856006026,
   "xla_temp_mb": 144.18798828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.9881407681386918,
   "inner": 22,
   "median_ms": 2.0411269612271674,
   "q1_ms": 2.0367674296721816,
   "q3_ms": 2.055592633868483,
   "tflops": 31.606069208558228
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002327442169189453,
   "rel_err": 0.0008719622856006026,
   "xla_temp_mb": 144.18798828125,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.996055677998811,
   "inner": 22,
   "median_ms": 2.050735596144064,
   "q1_ms": 2.0334120014343746,
   "q3_ms": 2.0566094529137695,
   "tflops": 31.45798030779783
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017006397247314453,
   "rel_err": 0.0006371345371007919,
   "xla_temp_mb": 823.9751129150391,
   "xla_arg_mb": 93.75,
   "xla_out_mb": 70.3125228881836,
   "compile_s": 0.9908906309865415,
   "inner": 22,
   "median_ms": 2.2131874522363595,
   "q1_ms": 2.1878422756510028,
   "q3_ms": 2.217357205650346,
   "tflops": 29.148909160321036
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00010614842176437378,
   "rel_err": 0.0005807357956655324,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.541711196070537,
   "inner": 57,
   "median_ms": 0.7209067371788255,
   "q1_ms": 0.7204833060601039,
   "q3_ms": 0.7370017701759934,
   "tflops": 95.32367113799589
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00010614842176437378,
   "rel_err": 0.0005807357956655324,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.541806774912402,
   "inner": 57,
   "median_ms": 0.7215965449352536,
   "q1_ms": 0.7212451503922541,
   "q3_ms": 0.7292763513856029,
   "tflops": 95.23254679963297
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00010614842176437378,
   "rel_err": 0.0005807357956655324,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.5457619251683354,
   "inner": 57,
   "median_ms": 0.7258854896287646,
   "q1_ms": 0.7233572470252974,
   "q3_ms": 0.7305661759649713,
   "tflops": 94.66985870063995
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011555105447769165,
   "rel_err": 0.0006321773980744183,
   "xla_temp_mb": 1024.000015258789,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.01813064212910831,
   "inner": 36,
   "median_ms": 1.2610192449452977,
   "q1_ms": 1.259196181005488,
   "q3_ms": 1.2615836781656575,
   "tflops": 54.49518475745468
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00021569430828094482,
   "rel_err": 0.0008874887716956437,
   "xla_temp_mb": 32.500244140625,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 1.5151424719952047,
   "inner": 7,
   "median_ms": 6.096101299460445,
   "q1_ms": 6.035498659392553,
   "q3_ms": 6.123837715546999,
   "tflops": 39.45442451838651
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00021569430828094482,
   "rel_err": 0.0008874887716956437,
   "xla_temp_mb": 0.50048828125,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.9489481698255986,
   "inner": 8,
   "median_ms": 5.799400008982047,
   "q1_ms": 5.78291185956914,
   "q3_ms": 5.841296064318158,
   "tflops": 41.47293999439391
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00021569430828094482,
   "rel_err": 0.0008874887716956437,
   "xla_temp_mb": 0.50048828125,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.947917551966384,
   "inner": 8,
   "median_ms": 5.797973746666685,
   "q1_ms": 5.779893748695031,
   "q3_ms": 5.839464807650074,
   "tflops": 41.483142057046464
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00021566450595855713,
   "rel_err": 0.0008873661863617599,
   "xla_temp_mb": 1568.000259399414,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.03408810100518167,
   "inner": 15,
   "median_ms": 3.178007062524557,
   "q1_ms": 3.166382131166756,
   "q3_ms": 3.180517434763412,
   "tflops": 75.68207491173297
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008929818868637085,
   "rel_err": 0.00025550625286996365,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.6289706500247121,
   "inner": 82,
   "median_ms": 0.5110776954807523,
   "q1_ms": 0.5075794746872129,
   "q3_ms": 0.5111616579020715,
   "tflops": 67.2299704562122
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008929818868637085,
   "rel_err": 0.00025550625286996365,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.623072185087949,
   "inner": 78,
   "median_ms": 0.5108693178790884,
   "q1_ms": 0.505269269267909,
   "q3_ms": 0.5125378330166523,
   "tflops": 67.25739277247455
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008929818868637085,
   "rel_err": 0.00025550625286996365,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.6236765128560364,
   "inner": 79,
   "median_ms": 0.5113124682367602,
   "q1_ms": 0.5076902659161936,
   "q3_ms": 0.512248037760205,
   "tflops": 67.19910133716888
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011209994554519653,
   "rel_err": 0.0003207482513971627,
   "xla_temp_mb": 1024.000015258789,
   "xla_arg_mb": 96.0,
   "xla_out_mb": 32.0,
   "compile_s": 0.5312968520447612,
   "inner": 33,
   "median_ms": 1.4244108460843563,
   "q1_ms": 1.4201893478236864,
   "q3_ms": 1.4259752422345406,
   "tflops": 24.12207016146601
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002022385597229004,
   "rel_err": 0.0009868371998891234,
   "xla_temp_mb": 32.500244140625,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 1.6596710770390928,
   "inner": 15,
   "median_ms": 3.132779197767377,
   "q1_ms": 3.1300122306371727,
   "q3_ms": 3.16173043102026,
   "tflops": 38.3873476859475
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002022385597229004,
   "rel_err": 0.0009868371998891234,
   "xla_temp_mb": 0.50048828125,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.9611204979009926,
   "inner": 14,
   "median_ms": 3.2947429322770665,
   "q1_ms": 3.28376900870353,
   "q3_ms": 3.3202707168779204,
   "tflops": 36.50029357673935
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002022385597229004,
   "rel_err": 0.0009868371998891234,
   "xla_temp_mb": 0.50048828125,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.9959395180922002,
   "inner": 14,
   "median_ms": 3.2808608541797315,
   "q1_ms": 3.2772111423712755,
   "q3_ms": 3.3177871456635852,
   "tflops": 36.65473472756184
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013439655303955078,
   "rel_err": 0.0006167421233840287,
   "xla_temp_mb": 1568.000259399414,
   "xla_arg_mb": 128.0,
   "xla_out_mb": 96.0000228881836,
   "compile_s": 0.04078252403996885,
   "inner": 13,
   "median_ms": 3.6776505518131533,
   "q1_ms": 3.6767116085124703,
   "q3_ms": 3.6825804272666574,
   "tflops": 32.69997586603489
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.392201900482178e-05,
   "rel_err": 0.0005303416983224452,
   "xla_temp_mb": 192.0,
   "xla_arg_mb": 140.625,
   "xla_out_mb": 46.875,
   "compile_s": 0.6469288419466466,
   "inner": 25,
   "median_ms": 1.7681758385151625,
   "q1_ms": 1.7626650212332606,
   "q3_ms": 1.7878029588609934,
   "tflops": 83.39442084211893
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.392201900482178e-05,
   "rel_err": 0.0005303416983224452,
   "xla_temp_mb": 192.0,
   "xla_arg_mb": 140.625,
   "xla_out_mb": 46.875,
   "compile_s": 0.631896052043885,
   "inner": 25,
   "median_ms": 1.812836080789566,
   "q1_ms": 1.7977691581472754,
   "q3_ms": 1.823994885198772,
   "tflops": 81.33995211292174
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.392201900482178e-05,
   "rel_err": 0.0005303416983224452,
   "xla_temp_mb": 192.0,
   "xla_arg_mb": 140.625,
   "xla_out_mb": 46.875,
   "compile_s": 0.629877737024799,
   "inner": 25,
   "median_ms": 1.8066437169909477,
   "q1_ms": 1.7990473238751292,
   "q3_ms": 1.8264116765931249,
   "tflops": 81.61874896152469
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00012783706188201904,
   "rel_err": 0.0007218469399958849,
   "xla_temp_mb": 2197.265884399414,
   "xla_arg_mb": 140.625,
   "xla_out_mb": 46.875,
   "compile_s": 0.018467431887984276,
   "inner": 18,
   "median_ms": 2.527601784095168,
   "q1_ms": 2.5272539997887282,
   "q3_ms": 2.529592887084517,
   "tflops": 58.33830349695941
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00014616549015045166,
   "rel_err": 0.0006567650125361979,
   "xla_temp_mb": 384.00048828125,
   "xla_arg_mb": 187.5,
   "xla_out_mb": 140.6250228881836,
   "compile_s": 1.6993258642032743,
   "inner": 3,
   "median_ms": 13.03408636401097,
   "q1_ms": 13.031986192800105,
   "q3_ms": 13.149639319938919,
   "tflops": 39.59587082566961
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00014616549015045166,
   "rel_err": 0.0006567650125361979,
   "xla_temp_mb": 288.37548828125,
   "xla_arg_mb": 187.5,
   "xla_out_mb": 140.6250228881836,
   "compile_s": 1.0072060588281602,
   "inner": 3,
   "median_ms": 12.642177287489176,
   "q1_ms": 12.636112320857743,
   "q3_ms": 12.73723381261031,
   "tflops": 40.82334777180619
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00014616549015045166,
   "rel_err": 0.0006567650125361979,
   "xla_temp_mb": 288.37548828125,
   "xla_arg_mb": 187.5,
   "xla_out_mb": 140.6250228881836,
   "compile_s": 0.9832481809426099,
   "inner": 3,
   "median_ms": 12.639897332216302,
   "q1_ms": 12.630115534799794,
   "q3_ms": 12.75451535669466,
   "tflops": 40.8307113922979
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00014568865299224854,
   "rel_err": 0.0006590303382836282,
   "xla_temp_mb": 3296.265151977539,
   "xla_arg_mb": 187.5,
   "xla_out_mb": 140.6250228881836,
   "compile_s": 0.039638967951759696,
   "inner": 6,
   "median_ms": 7.500875663633148,
   "q1_ms": 7.489998824894428,
   "q3_ms": 7.50241382047534,
   "tflops": 68.80476668906975
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000994861125946045,
   "rel_err": 0.0002707694948185235,
   "xla_temp_mb": 192.0,
   "xla_arg_mb": 140.625,
   "xla_out_mb": 46.875,
   "compile_s": 0.6427697730250657,
   "inner": 37,
   "median_ms": 1.1662199174532213,
   "q1_ms": 1.1638903650581032,
   "q3_ms": 1.1751127580331788,
   "tflops": 63.219637134140555
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000994861125946045,
   "rel_err": 0.0002707694948185235,
   "xla_temp_mb": 192.0,
   "xla_arg_mb": 140.625,
   "xla_out_mb": 46.875,
   "compile_s": 0.6478092798497528,
   "inner": 38,
   "median_ms": 1.181195104053538,
   "q1_ms": 1.1730397236533463,
   "q3_ms": 1.1863187629435408,
   "tflops": 62.418138838355915
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000994861125946045,
   "rel_err": 0.0002707694948185235,
   "xla_temp_mb": 192.0,
   "xla_arg_mb": 140.625,
   "xla_out_mb": 46.875,
   "compile_s": 0.6420365059748292,
   "inner": 38,
   "median_ms": 1.176551946340815,
   "q1_ms": 1.173105118428602,
   "q3_ms": 1.1846445666282976,
   "tflops": 62.66446647706535
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010943412780761719,
   "rel_err": 0.00029784481739625335,
   "xla_temp_mb": 2197.265884399414,
   "xla_arg_mb": 140.625,
   "xla_out_mb": 46.875,
   "compile_s": 0.4276209711097181,
   "inner": 18,
   "median_ms": 2.532670610687799,
   "q1_ms": 2.5321348285716443,
   "q3_ms": 2.533513166579521,
   "tflops": 29.11077330343311
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016607046127319336,
   "rel_err": 0.0007745111361145973,
   "xla_temp_mb": 384.00048828125,
   "xla_arg_mb": 187.5,
   "xla_out_mb": 140.6250228881836,
   "compile_s": 1.733971324050799,
   "inner": 6,
   "median_ms": 7.146249835689862,
   "q1_ms": 7.12258490966633,
   "q3_ms": 7.191819682096441,
   "tflops": 36.10956878547045
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016607046127319336,
   "rel_err": 0.0007745111361145973,
   "xla_temp_mb": 288.37548828125,
   "xla_arg_mb": 187.5,
   "xla_out_mb": 140.6250228881836,
   "compile_s": 1.0315270361024886,
   "inner": 7,
   "median_ms": 6.779676702405725,
   "q1_ms": 6.7435989289411475,
   "q3_ms": 6.822551667158093,
   "tflops": 38.06199193959106
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016607046127319336,
   "rel_err": 0.0007745111361145973,
   "xla_temp_mb": 288.37548828125,
   "xla_arg_mb": 187.5,
   "xla_out_mb": 140.6250228881836,
   "compile_s": 1.0158285561483353,
   "inner": 7,
   "median_ms": 6.797708138557417,
   "q1_ms": 6.772985931352846,
   "q3_ms": 6.849016495315092,
   "tflops": 37.96102962060415
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015103816986083984,
   "rel_err": 0.0006280419765971601,
   "xla_temp_mb": 3296.265151977539,
   "xla_arg_mb": 187.5,
   "xla_out_mb": 140.6250228881836,
   "compile_s": 0.94312704494223,
   "inner": 6,
   "median_ms": 7.564132179444035,
   "q1_ms": 7.562299588850388,
   "q3_ms": 7.573660307874283,
   "tflops": 34.1146867715057
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011526048183441162,
   "rel_err": 0.0007247517351061106,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.5449543870054185,
   "inner": 16,
   "median_ms": 2.8027808730257675,
   "q1_ms": 2.7924168171011843,
   "q3_ms": 2.8370907821226865,
   "tflops": 98.07327771837298
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011526048183441162,
   "rel_err": 0.0007247517351061106,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.5439134398475289,
   "inner": 16,
   "median_ms": 2.7876633685082197,
   "q1_ms": 2.782008123176638,
   "q3_ms": 2.805431460728869,
   "tflops": 98.60512931699397
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011526048183441162,
   "rel_err": 0.0007247517351061106,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.550090943928808,
   "inner": 16,
   "median_ms": 2.7914265083381906,
   "q1_ms": 2.78130062361015,
   "q3_ms": 2.821318026690278,
   "tflops": 98.47219911501163
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.158385753631592e-05,
   "rel_err": 0.0005129949422553182,
   "xla_temp_mb": 4096.000259399414,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.021007294999435544,
   "inner": 10,
   "median_ms": 4.499579407274723,
   "q1_ms": 4.494767799042165,
   "q3_ms": 4.503576306160539,
   "tflops": 61.08968907173622
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0001957714557647705,
   "rel_err": 0.0008156931144185364,
   "xla_temp_mb": 65.000244140625,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.528247238136828,
   "inner": 2,
   "median_ms": 23.631118470802903,
   "q1_ms": 23.596198996528983,
   "q3_ms": 23.749095969833434,
   "tflops": 40.71210914086337
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0001957714557647705,
   "rel_err": 0.0008156931144185364,
   "xla_temp_mb": 1.00048828125,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.9375571189448237,
   "inner": 2,
   "median_ms": 22.748056100681424,
   "q1_ms": 22.65639853430912,
   "q3_ms": 22.858297976199538,
   "tflops": 42.29252249273206
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0001957714557647705,
   "rel_err": 0.0008156931144185364,
   "xla_temp_mb": 1.00048828125,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.9298848900943995,
   "inner": 2,
   "median_ms": 22.732580895535648,
   "q1_ms": 22.676675987895578,
   "q3_ms": 22.855000745039433,
   "tflops": 42.321313128723425
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00019532442092895508,
   "rel_err": 0.000810497731436044,
   "xla_temp_mb": 6144.000503540039,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.042555700056254864,
   "inner": 4,
   "median_ms": 11.638532218057662,
   "q1_ms": 11.63372749579139,
   "q3_ms": 11.646867264062166,
   "tflops": 82.66271521861705
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009086132049560547,
   "rel_err": 0.0003274537157267332,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.6313751770649105,
   "inner": 27,
   "median_ms": 1.6528972234852888,
   "q1_ms": 1.6449639030421774,
   "q3_ms": 1.6581538711088124,
   "tflops": 83.15033234927765
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009086132049560547,
   "rel_err": 0.0003274537157267332,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.6293572941794991,
   "inner": 27,
   "median_ms": 1.6473375985191927,
   "q1_ms": 1.641638352777119,
   "q3_ms": 1.6646620006886899,
   "tflops": 83.43095768320056
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009086132049560547,
   "rel_err": 0.0003274537157267332,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.6386108801234514,
   "inner": 27,
   "median_ms": 1.6543005196446623,
   "q1_ms": 1.6471492385284767,
   "q3_ms": 1.6559194040450231,
   "tflops": 83.07979828327768
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013452768325805664,
   "rel_err": 0.0004848222597502172,
   "xla_temp_mb": 4096.000259399414,
   "xla_arg_mb": 192.0,
   "xla_out_mb": 64.0,
   "compile_s": 0.49407247197814286,
   "inner": 9,
   "median_ms": 5.168681110565861,
   "q1_ms": 5.14389277021918,
   "q3_ms": 5.172295508802765,
   "tflops": 26.590720249900144
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0018148422241210938,
   "rel_err": 0.0005610198713839054,
   "xla_temp_mb": 65.000244140625,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 1.7054514531046152,
   "inner": 4,
   "median_ms": 11.776766972616315,
   "q1_ms": 11.749882076401263,
   "q3_ms": 11.890386784216389,
   "tflops": 40.84621341922786
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0018148422241210938,
   "rel_err": 0.0005610198713839054,
   "xla_temp_mb": 1.00048828125,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.9878709260374308,
   "inner": 3,
   "median_ms": 12.245591652269164,
   "q1_ms": 12.196756821746629,
   "q3_ms": 12.329029850661755,
   "tflops": 39.282408789359046
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0018148422241210938,
   "rel_err": 0.0005610198713839054,
   "xla_temp_mb": 1.00048828125,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.9795288189779967,
   "inner": 3,
   "median_ms": 12.228675031413635,
   "q1_ms": 12.21265197576334,
   "q3_ms": 12.404587857114773,
   "tflops": 39.33675037698603
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0018155574798583984,
   "rel_err": 0.0004941796069033444,
   "xla_temp_mb": 6144.000503540039,
   "xla_arg_mb": 256.0,
   "xla_out_mb": 192.0000228881836,
   "compile_s": 0.05770303099416196,
   "inner": 3,
   "median_ms": 13.349995715543628,
   "q1_ms": 13.291968808819849,
   "q3_ms": 13.370943837799132,
   "tflops": 36.032695994944866
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.854666233062744e-05,
   "rel_err": 0.0005574174574576318,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 384.0,
   "xla_out_mb": 128.0,
   "compile_s": 0.5531918709166348,
   "inner": 4,
   "median_ms": 10.807484504766762,
   "q1_ms": 10.797157156048343,
   "q3_ms": 10.927760624326766,
   "tflops": 101.7361280778194
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.854666233062744e-05,
   "rel_err": 0.0005574174574576318,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 384.0,
   "xla_out_mb": 128.0,
   "compile_s": 0.5532487500458956,
   "inner": 4,
   "median_ms": 10.863259492907673,
   "q1_ms": 10.820389521541074,
   "q3_ms": 10.890014236792922,
   "tflops": 101.21378657058143
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.854666233062744e-05,
   "rel_err": 0.0005574174574576318,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 384.0,
   "xla_out_mb": 128.0,
   "compile_s": 0.557009199867025,
   "inner": 4,
   "median_ms": 10.916371771600097,
   "q1_ms": 10.819679504493251,
   "q3_ms": 10.963445500237867,
   "tflops": 100.72134320640089
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.587058305740356e-05,
   "rel_err": 0.0006271477905102074,
   "xla_temp_mb": 16384.00001525879,
   "xla_arg_mb": 384.0,
   "xla_out_mb": 128.0,
   "compile_s": 0.03300835518166423,
   "inner": 2,
   "median_ms": 17.207315540872514,
   "q1_ms": 17.18812552280724,
   "q3_ms": 17.227022966835648,
   "tflops": 63.897917438913204
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0001177862286567688,
   "rel_err": 0.0006473481771536171,
   "xla_temp_mb": 130.000244140625,
   "xla_arg_mb": 512.0,
   "xla_out_mb": 384.0000228881836,
   "compile_s": 1.6244349819608033,
   "inner": 1,
   "median_ms": 93.11620891094208,
   "q1_ms": 92.93037944007665,
   "q3_ms": 93.24985952116549,
   "tflops": 41.32782833648834
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0001177862286567688,
   "rel_err": 0.0006473481771536171,
   "xla_temp_mb": 2.00048828125,
   "xla_arg_mb": 512.0,
   "xla_out_mb": 384.0000228881836,
   "compile_s": 1.0339836480561644,
   "inner": 1,
   "median_ms": 89.34917999431491,
   "q1_ms": 89.0511239413172,
   "q3_ms": 89.69058294314891,
   "tflops": 43.070240795280476
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0001177862286567688,
   "rel_err": 0.0006473481771536171,
   "xla_temp_mb": 2.00048828125,
   "xla_arg_mb": 512.0,
   "xla_out_mb": 384.0000228881836,
   "compile_s": 1.0083862540777773,
   "inner": 1,
   "median_ms": 89.36970890499651,
   "q1_ms": 89.26827902905643,
   "q3_ms": 89.53973837196827,
   "tflops": 43.060347229136475
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00011768937110900879,
   "rel_err": 0.0006468158680945635,
   "xla_temp_mb": 24576.00050354004,
   "xla_arg_mb": 512.0,
   "xla_out_mb": 384.0000228881836,
   "compile_s": 0.07774913893081248,
   "inner": 1,
   "median_ms": 44.77674514055252,
   "q1_ms": 44.6724594803527,
   "q3_ms": 44.78795395698398,
   "tflops": 85.94395785438088
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012829303741455078,
   "rel_err": 0.00042788597056642175,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 384.0,
   "xla_out_mb": 128.0,
   "compile_s": 0.6421466800384223,
   "inner": 8,
   "median_ms": 5.8045136102009565,
   "q1_ms": 5.7808206911431625,
   "q3_ms": 5.833356757648289,
   "tflops": 94.71177962643576
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012829303741455078,
   "rel_err": 0.00042788597056642175,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 384.0,
   "xla_out_mb": 128.0,
   "compile_s": 0.632449665106833,
   "inner": 8,
   "median_ms": 5.815433745738119,
   "q1_ms": 5.811393741169013,
   "q3_ms": 5.8191212301608175,
   "tflops": 94.53393124646848
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012829303741455078,
   "rel_err": 0.00042788597056642175,
   "xla_temp_mb": 0.0,
   "xla_arg_mb": 384.0,
   "xla_out_mb": 128.0,
   "compile_s": 0.6316442978568375,
   "inner": 8,
   "median_ms": 5.819944868562743,
   "q1_ms": 5.80175629875157,
   "q3_ms": 5.840738682309166,
   "tflops": 94.46065663913484
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011343955993652344,
   "rel_err": 0.00037834630347788334,
   "xla_temp_mb": 16384.00001525879,
   "xla_arg_mb": 384.0,
   "xla_out_mb": 128.0,
   "compile_s": 0.48432000912725925,
   "inner": 2,
   "median_ms": 19.775293534621596,
   "q1_ms": 19.615144236013293,
   "q3_ms": 19.778891233727336,
   "tflops": 27.800134188932013
  },
  {
   "impl": "probjax_mha",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001294255256652832,
   "rel_err": 0.0005421916139312088,
   "xla_temp_mb": 130.000244140625,
   "xla_arg_mb": 512.0,
   "xla_out_mb": 384.0000228881836,
   "compile_s": 1.7186019390355796,
   "inner": 1,
   "median_ms": 45.56924197822809,
   "q1_ms": 45.36722297780216,
   "q3_ms": 45.58686702512205,
   "tflops": 42.22465121380143
  },
  {
   "impl": "probjax_flex",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001294255256652832,
   "rel_err": 0.0005421916139312088,
   "xla_temp_mb": 2.00048828125,
   "xla_arg_mb": 512.0,
   "xla_out_mb": 384.0000228881836,
   "compile_s": 1.0417713881470263,
   "inner": 1,
   "median_ms": 46.954998979344964,
   "q1_ms": 46.74274940043688,
   "q3_ms": 47.01076354831457,
   "tflops": 40.97849835870324
  },
  {
   "impl": "probjax_default",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001294255256652832,
   "rel_err": 0.0005421916139312088,
   "xla_temp_mb": 2.00048828125,
   "xla_arg_mb": 512.0,
   "xla_out_mb": 384.0000228881836,
   "compile_s": 1.019721585093066,
   "inner": 1,
   "median_ms": 46.7894789762795,
   "q1_ms": 46.780164004303515,
   "q3_ms": 46.87690897844732,
   "tflops": 41.12346174197557
  },
  {
   "impl": "jax_xla",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001294255256652832,
   "rel_err": 0.0004761589225381613,
   "xla_temp_mb": 24576.00050354004,
   "xla_arg_mb": 512.0,
   "xla_out_mb": 384.0000228881836,
   "compile_s": 1.0114301559515297,
   "inner": 1,
   "median_ms": 50.909418147057295,
   "q1_ms": 50.790804089047015,
   "q3_ms": 51.053018076345325,
   "tflops": 37.79546925973305
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0021501779556274414,
   "rel_err": 0.002294705016538501,
   "peak_mb": 3.5166015625,
   "compile_s": 0.0001264710444957018,
   "inner": 200,
   "median_ms": 0.018255264731124043,
   "q1_ms": 0.018203314975835383,
   "q3_ms": 0.018353061750531197,
   "tflops": 7.352275082111928
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0021501779556274414,
   "rel_err": 0.002294705016538501,
   "peak_mb": 3.5009765625,
   "compile_s": 9.628990665078163e-05,
   "inner": 200,
   "median_ms": 0.023398185148835182,
   "q1_ms": 0.02336325531359762,
   "q3_ms": 0.023423630045726895,
   "tflops": 5.736245232108597
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001876235008239746,
   "rel_err": 0.002002348657697439,
   "peak_mb": 43.50439453125,
   "compile_s": 0.0003122908528894186,
   "inner": 200,
   "median_ms": 0.15614709467627108,
   "q1_ms": 0.15610851522069424,
   "q3_ms": 0.1562904694583267,
   "tflops": 0.8595595600306511
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0032709836959838867,
   "rel_err": 0.0035336874425411224,
   "peak_mb": 40.533203125,
   "compile_s": 0.0005487820599228144,
   "inner": 189,
   "median_ms": 0.1811436981100727,
   "q1_ms": 0.18080607950983066,
   "q3_ms": 0.1817318227723556,
   "tflops": 2.593311569219191
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0038875937461853027,
   "rel_err": 0.003933617379516363,
   "peak_mb": 40.0341796875,
   "compile_s": 0.0003298309165984392,
   "inner": 194,
   "median_ms": 0.18432455679843412,
   "q1_ms": 0.1835250205686796,
   "q3_ms": 0.18495695903109982,
   "tflops": 2.5485592161965838
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027950406074523926,
   "rel_err": 0.0032178577966988087,
   "peak_mb": 79.5009765625,
   "compile_s": 0.0006930120289325714,
   "inner": 80,
   "median_ms": 0.5457642750116065,
   "q1_ms": 0.5450822689454071,
   "q3_ms": 0.5461065884446725,
   "tflops": 0.8607416599227017
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005635261535644531,
   "rel_err": 0.0019494958687573671,
   "peak_mb": 67.5166015625,
   "compile_s": 7.71600753068924e-05,
   "inner": 200,
   "median_ms": 0.018451964715495706,
   "q1_ms": 0.018431664793752134,
   "q3_ms": 0.018632264691405,
   "tflops": 3.6369495083438403
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005635261535644531,
   "rel_err": 0.0019494958687573671,
   "peak_mb": 67.5009765625,
   "compile_s": 7.827999070286751e-05,
   "inner": 200,
   "median_ms": 0.0234483799431473,
   "q1_ms": 0.023424680111929774,
   "q3_ms": 0.02356097975280136,
   "tflops": 2.861982966955988
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0038852691650390625,
   "rel_err": 0.0013440931215882301,
   "peak_mb": 75.56689453125,
   "compile_s": 0.0003415709361433983,
   "inner": 194,
   "median_ms": 0.19711774255427503,
   "q1_ms": 0.19700604625353496,
   "q3_ms": 0.19817973712708842,
   "tflops": 0.34045065213509146
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009996175765991211,
   "rel_err": 0.005375280044972897,
   "peak_mb": 72.533203125,
   "compile_s": 0.0004657318349927664,
   "inner": 186,
   "median_ms": 0.18098197316610684,
   "q1_ms": 0.18081213134811613,
   "q3_ms": 0.1813932666204549,
   "tflops": 1.29781447229788
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012074887752532959,
   "rel_err": 0.005375280044972897,
   "peak_mb": 72.0341796875,
   "compile_s": 0.00036331103183329105,
   "inner": 200,
   "median_ms": 0.18479963997378945,
   "q1_ms": 0.1835015625692904,
   "q3_ms": 0.21260496228933334,
   "tflops": 1.2710036882826918
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009835958480834961,
   "rel_err": 0.003037187270820141,
   "peak_mb": 79.5009765625,
   "compile_s": 0.0008276330772787333,
   "inner": 66,
   "median_ms": 0.6673186805776574,
   "q1_ms": 0.6663665602294107,
   "q3_ms": 0.668714371523961,
   "tflops": 0.3519772948610965
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001060187816619873,
   "rel_err": 0.002345811342820525,
   "peak_mb": 78.0634765625,
   "compile_s": 8.422997780144215e-05,
   "inner": 200,
   "median_ms": 0.018421615241095424,
   "q1_ms": 0.018401240231469274,
   "q3_ms": 0.018478889833204448,
   "tflops": 116.57412338139258
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010578334331512451,
   "rel_err": 0.0023406019899994135,
   "peak_mb": 78.0009765625,
   "compile_s": 8.393009193241596e-05,
   "inner": 200,
   "median_ms": 0.03055760520510376,
   "q1_ms": 0.030533130047842864,
   "q3_ms": 0.030560606974177063,
   "tflops": 70.27656891258368
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009667873382568359,
   "rel_err": 0.002139149932190776,
   "peak_mb": 164.01611328125,
   "compile_s": 0.00030711106956005096,
   "inner": 180,
   "median_ms": 0.2173220337782469,
   "q1_ms": 0.21626538816942936,
   "q3_ms": 0.21738631120469004,
   "tflops": 9.881573491030686
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001962810754776001,
   "rel_err": 0.005092823877930641,
   "peak_mb": 98.126953125,
   "compile_s": 0.0003705208655446768,
   "inner": 199,
   "median_ms": 0.18196324607235703,
   "q1_ms": 0.18186977626896234,
   "q3_ms": 0.18228206551872456,
   "tflops": 41.306104008560126
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0030225813388824463,
   "rel_err": 0.007842566817998886,
   "peak_mb": 96.130859375,
   "compile_s": 0.0003563119098544121,
   "inner": 139,
   "median_ms": 0.18386229564496082,
   "q1_ms": 0.18307592103378378,
   "q3_ms": 0.18416617899838328,
   "tflops": 40.87946765613007
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016031265258789062,
   "rel_err": 0.0033890039194375277,
   "peak_mb": 222.0009765625,
   "compile_s": 0.000748553080484271,
   "inner": 75,
   "median_ms": 0.544713626926144,
   "q1_ms": 0.5442396939421694,
   "q3_ms": 0.5449965596199036,
   "tflops": 13.798429847283952
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007361173629760742,
   "rel_err": 0.0025465681683272123,
   "peak_mb": 78.0634765625,
   "compile_s": 8.883094415068626e-05,
   "inner": 200,
   "median_ms": 0.018537763971835375,
   "q1_ms": 0.01849333755671978,
   "q3_ms": 0.018609915277920663,
   "tflops": 57.92186294050068
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007361173629760742,
   "rel_err": 0.0025465681683272123,
   "peak_mb": 78.0009765625,
   "compile_s": 8.430005982518196e-05,
   "inner": 200,
   "median_ms": 0.03019075491465628,
   "q1_ms": 0.030186105286702514,
   "q3_ms": 0.03020017989911139,
   "tflops": 35.5652525760045
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007361173629760742,
   "rel_err": 0.0025465681683272123,
   "peak_mb": 165.01611328125,
   "compile_s": 0.0003682610113173723,
   "inner": 146,
   "median_ms": 0.2580620619557696,
   "q1_ms": 0.2573787749821499,
   "q3_ms": 0.258170486880747,
   "tflops": 4.160789136777622
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01617288589477539,
   "rel_err": 0.0037439705338329077,
   "peak_mb": 98.126953125,
   "compile_s": 0.0003773909993469715,
   "inner": 181,
   "median_ms": 0.2162085977274904,
   "q1_ms": 0.21542553038681078,
   "q3_ms": 0.21721111373842092,
   "tflops": 17.381808232884012
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01617288589477539,
   "rel_err": 0.005760819651186466,
   "peak_mb": 96.130859375,
   "compile_s": 0.00036759208887815475,
   "inner": 174,
   "median_ms": 0.21624644206644134,
   "q1_ms": 0.21615701435861745,
   "q3_ms": 0.21628299179411997,
   "tflops": 17.37876631905616
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01617288589477539,
   "rel_err": 0.0037439705338329077,
   "peak_mb": 222.0009765625,
   "compile_s": 0.000866543035954237,
   "inner": 61,
   "median_ms": 0.6588881980383493,
   "q1_ms": 0.6582545818852596,
   "q3_ms": 0.6589945901918117,
   "tflops": 5.703693578347061
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009658932685852051,
   "rel_err": 0.002841358305886388,
   "peak_mb": 80.48046875,
   "compile_s": 8.392008021473885e-05,
   "inner": 200,
   "median_ms": 0.02268158015795052,
   "q1_ms": 0.022675730288028717,
   "q3_ms": 0.02268967975396663,
   "tflops": 130.02268710834292
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010502040386199951,
   "rel_err": 0.00308937462978065,
   "peak_mb": 80.4072265625,
   "compile_s": 9.455997496843338e-05,
   "inner": 200,
   "median_ms": 0.042531900107860565,
   "q1_ms": 0.042515869718044996,
   "q3_ms": 0.04253454739227891,
   "tflops": 69.3390135996994
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009658932685852051,
   "rel_err": 0.002841358305886388,
   "peak_mb": 195.81787109375,
   "compile_s": 0.00040515209548175335,
   "inner": 132,
   "median_ms": 0.32185089322879457,
   "q1_ms": 0.3207694848463165,
   "q3_ms": 0.3222987350344545,
   "tflops": 9.16300082443318
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002047628164291382,
   "rel_err": 0.004772087559103966,
   "peak_mb": 104.3095703125,
   "compile_s": 0.0006928930524736643,
   "inner": 124,
   "median_ms": 0.2180284993242352,
   "q1_ms": 0.2175279754033733,
   "q3_ms": 0.2184930565436521,
   "tflops": 47.34206781219934
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00208398699760437,
   "rel_err": 0.004597815684974194,
   "peak_mb": 102.8994140625,
   "compile_s": 0.00035999203100800514,
   "inner": 150,
   "median_ms": 0.21633181255310774,
   "q1_ms": 0.215789246528099,
   "q3_ms": 0.21641322023545703,
   "tflops": 47.71337085462662
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015577971935272217,
   "rel_err": 0.003416633466258645,
   "peak_mb": 275.1572265625,
   "compile_s": 0.0010801439639180899,
   "inner": 50,
   "median_ms": 0.7431053835898638,
   "q1_ms": 0.7407562807202339,
   "q3_ms": 0.7436846802011132,
   "tflops": 13.890250599633516
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004250764846801758,
   "rel_err": 0.0016897449968382716,
   "peak_mb": 80.48046875,
   "compile_s": 7.864017970860004e-05,
   "inner": 200,
   "median_ms": 0.023153580259531736,
   "q1_ms": 0.0230793550144881,
   "q3_ms": 0.02332390460651368,
   "tflops": 63.686046972928146
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004250764846801758,
   "rel_err": 0.0016897449968382716,
   "peak_mb": 80.4072265625,
   "compile_s": 9.079999290406704e-05,
   "inner": 200,
   "median_ms": 0.03908613580279052,
   "q1_ms": 0.03905613499227911,
   "q3_ms": 0.03910611500032246,
   "tflops": 37.72590893712049
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.003871440887451172,
   "rel_err": 0.0015389578184112906,
   "peak_mb": 198.16162109375,
   "compile_s": 0.0005051109474152327,
   "inner": 105,
   "median_ms": 0.386899725223581,
   "q1_ms": 0.38647429485406193,
   "q3_ms": 0.387002781610049,
   "tflops": 3.8112200755580368
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014807701110839844,
   "rel_err": 0.0035223369486629963,
   "peak_mb": 104.3095703125,
   "compile_s": 0.0007111618760973215,
   "inner": 161,
   "median_ms": 0.21815467018472112,
   "q1_ms": 0.21805588202261778,
   "q3_ms": 0.21887787325620095,
   "tflops": 23.65734364352589
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014807701110839844,
   "rel_err": 0.004309217445552349,
   "peak_mb": 102.8994140625,
   "compile_s": 0.0003500410821288824,
   "inner": 159,
   "median_ms": 0.21671716990710804,
   "q1_ms": 0.21660339333257583,
   "q3_ms": 0.21727031401399546,
   "tflops": 23.814264472963323
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014807701110839844,
   "rel_err": 0.0035223369486629963,
   "peak_mb": 275.1572265625,
   "compile_s": 0.0010221328120678663,
   "inner": 54,
   "median_ms": 0.8007507400449227,
   "q1_ms": 0.7986307415800791,
   "q3_ms": 0.8012144524535094,
   "tflops": 6.445151708146366
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00095328688621521,
   "rel_err": 0.003387594362720847,
   "peak_mb": 92.1259765625,
   "compile_s": 8.84511973708868e-05,
   "inner": 200,
   "median_ms": 0.03380881971679628,
   "q1_ms": 0.03373166488017887,
   "q3_ms": 0.03384831768926233,
   "tflops": 254.07377908943997
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011154413223266602,
   "rel_err": 0.0039638252928853035,
   "peak_mb": 92.0009765625,
   "compile_s": 0.0001327809877693653,
   "inner": 200,
   "median_ms": 0.08246143581345677,
   "q1_ms": 0.07953952706884593,
   "q3_ms": 0.08526449440978467,
   "tflops": 104.16911259503222
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00095328688621521,
   "rel_err": 0.003387594362720847,
   "peak_mb": 408.03173828125,
   "compile_s": 0.0008773228619247675,
   "inner": 58,
   "median_ms": 0.8041388274909094,
   "q1_ms": 0.8020607136382625,
   "q3_ms": 0.8044292656544211,
   "tflops": 10.682153750494168
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0015018582344055176,
   "rel_err": 0.004941033665090799,
   "peak_mb": 132.251953125,
   "compile_s": 0.0006766521837562323,
   "inner": 150,
   "median_ms": 0.21570982101062935,
   "q1_ms": 0.21545337978750467,
   "q3_ms": 0.21684178694461784,
   "tflops": 139.37599563683528
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002331942319869995,
   "rel_err": 0.005864798091351986,
   "peak_mb": 128.259765625,
   "compile_s": 0.0009197229519486427,
   "inner": 92,
   "median_ms": 0.4366847823131019,
   "q1_ms": 0.43509509819357295,
   "q3_ms": 0.4375154891496767,
   "tflops": 68.84776454253364
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010400712490081787,
   "rel_err": 0.002775006927549839,
   "peak_mb": 636.0009765625,
   "compile_s": 0.0018440170679241419,
   "inner": 27,
   "median_ms": 1.7563134811266705,
   "q1_ms": 1.7490336731835097,
   "q3_ms": 1.7574468489598345,
   "tflops": 17.118112111007385
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006038188934326172,
   "rel_err": 0.002593584591522813,
   "peak_mb": 92.1259765625,
   "compile_s": 8.431007154285908e-05,
   "inner": 200,
   "median_ms": 0.03343511954881251,
   "q1_ms": 0.033340115332975984,
   "q3_ms": 0.033462466672062874,
   "tflops": 128.45676504101334
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006038188934326172,
   "rel_err": 0.002593584591522813,
   "peak_mb": 92.0009765625,
   "compile_s": 0.00012141000479459763,
   "inner": 200,
   "median_ms": 0.0681469903793186,
   "q1_ms": 0.06809973565395921,
   "q3_ms": 0.06859663757495582,
   "tflops": 63.02504735856165
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006038188934326172,
   "rel_err": 0.002593584591522813,
   "peak_mb": 412.03173828125,
   "compile_s": 0.001048052916303277,
   "inner": 48,
   "median_ms": 0.9600320821239924,
   "q1_ms": 0.9582769780536182,
   "q3_ms": 0.960729164944496,
   "tflops": 4.473774758128641
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012666940689086914,
   "rel_err": 0.0043540229089558125,
   "peak_mb": 132.251953125,
   "compile_s": 0.0005305020604282618,
   "inner": 151,
   "median_ms": 0.240490310794569,
   "q1_ms": 0.24036056938772368,
   "q3_ms": 0.2412149098251533,
   "tflops": 62.50723983986584
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012666940689086914,
   "rel_err": 0.004499658010900021,
   "peak_mb": 128.259765625,
   "compile_s": 0.0006344220601022243,
   "inner": 112,
   "median_ms": 0.32479425135534257,
   "q1_ms": 0.3246415321232884,
   "q3_ms": 0.3249288482558248,
   "tflops": 46.282794332938344
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009896278381347656,
   "rel_err": 0.0034243951085954905,
   "peak_mb": 636.0009765625,
   "compile_s": 0.0020024869590997696,
   "inner": 25,
   "median_ms": 1.8908610008656979,
   "q1_ms": 1.8890599580481648,
   "q3_ms": 1.8966828007251024,
   "tflops": 7.9500214606561155
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006342977285385132,
   "rel_err": 0.002317151054739952,
   "peak_mb": 105.48095703125,
   "compile_s": 0.00014872010797262192,
   "inner": 200,
   "median_ms": 0.09493192424997687,
   "q1_ms": 0.09198124462272972,
   "q3_ms": 0.09625698497984558,
   "tflops": 194.1601852656483
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006347447633743286,
   "rel_err": 0.0023187841288745403,
   "peak_mb": 105.2978515625,
   "compile_s": 0.00023049116134643555,
   "inner": 200,
   "median_ms": 0.18806560547091067,
   "q1_ms": 0.18543261976446956,
   "q3_ms": 0.19136351475026459,
   "tflops": 98.00835168050438
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0004876554012298584,
   "rel_err": 0.0017814523307606578,
   "peak_mb": 765.46484375,
   "compile_s": 0.0016973060555756092,
   "inner": 29,
   "median_ms": 1.687340036933792,
   "q1_ms": 1.6830689474490694,
   "q3_ms": 1.6907436195118675,
   "tflops": 10.923702156380015
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0012814700603485107,
   "rel_err": 0.003724043257534504,
   "peak_mb": 165.38818359375,
   "compile_s": 0.0007788820657879114,
   "inner": 104,
   "median_ms": 0.40101842247307873,
   "q1_ms": 0.39778407349681055,
   "q3_ms": 0.40385707800921333,
   "tflops": 160.87041488556758
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013823211193084717,
   "rel_err": 0.004368611145764589,
   "peak_mb": 159.53662109375,
   "compile_s": 0.0009363829158246517,
   "inner": 62,
   "median_ms": 0.8111778203578245,
   "q1_ms": 0.8059314268100406,
   "q3_ms": 0.8223768860672511,
   "tflops": 79.52880167697712
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010690689086914062,
   "rel_err": 0.003709746291860938,
   "peak_mb": 1251.6494140625,
   "compile_s": 0.0036995229311287403,
   "inner": 13,
   "median_ms": 3.760279227907841,
   "q1_ms": 3.749705340641622,
   "q3_ms": 3.7661149936656537,
   "tflops": 17.156172744089922
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007173061370849609,
   "rel_err": 0.0024814915377646685,
   "peak_mb": 105.76220703125,
   "compile_s": 0.00012219115160405636,
   "inner": 200,
   "median_ms": 0.06593963014893234,
   "q1_ms": 0.06558940222021192,
   "q3_ms": 0.06658498256001621,
   "tflops": 139.7642052159618
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007173061370849609,
   "rel_err": 0.0024814915377646685,
   "peak_mb": 105.5791015625,
   "compile_s": 0.00018487009219825268,
   "inner": 200,
   "median_ms": 0.12499098433181642,
   "q1_ms": 0.1248902577208355,
   "q3_ms": 0.1256324880523607,
   "tflops": 73.73331804103626
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007173061370849609,
   "rel_err": 0.0024814915377646685,
   "peak_mb": 774.32958984375,
   "compile_s": 0.002054997021332383,
   "inner": 24,
   "median_ms": 2.002118630722786,
   "q1_ms": 1.997721546407168,
   "q3_ms": 2.0072898963311063,
   "tflops": 4.603123840205676
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014818668365478516,
   "rel_err": 0.005191786680370569,
   "peak_mb": 165.38818359375,
   "compile_s": 0.0005801618099212646,
   "inner": 120,
   "median_ms": 0.3123844255848477,
   "q1_ms": 0.3120120003586635,
   "q3_ms": 0.3126282128505409,
   "tflops": 103.25738851932248
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.014818668365478516,
   "rel_err": 0.005191786680370569,
   "peak_mb": 159.53662109375,
   "compile_s": 0.0008567830082029104,
   "inner": 73,
   "median_ms": 0.6057344383419785,
   "q1_ms": 0.6051339457857692,
   "q3_ms": 0.6064642268577463,
   "tflops": 53.25105848082767
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01280665397644043,
   "rel_err": 0.003944694995880127,
   "peak_mb": 1251.6494140625,
   "compile_s": 0.004035813966766,
   "inner": 12,
   "median_ms": 4.020092232773702,
   "q1_ms": 4.0073239263923215,
   "q3_ms": 4.028484002143765,
   "tflops": 8.023696505526358
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000682830810546875,
   "rel_err": 0.0027175149880349636,
   "peak_mb": 120.2509765625,
   "compile_s": 0.0001811198890209198,
   "inner": 200,
   "median_ms": 0.14028718462213874,
   "q1_ms": 0.13607482251245528,
   "q3_ms": 0.14224757032934576,
   "tflops": 244.92428485572222
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000682830810546875,
   "rel_err": 0.0027175149880349636,
   "peak_mb": 120.0009765625,
   "compile_s": 0.00033985101617872715,
   "inner": 160,
   "median_ms": 0.3143643436487764,
   "q1_ms": 0.31329655976151116,
   "q3_ms": 0.320857331098523,
   "tflops": 109.29909534011409
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000682830810546875,
   "rel_err": 0.0027175149880349636,
   "peak_mb": 1328.06298828125,
   "compile_s": 0.002957721007987857,
   "inner": 16,
   "median_ms": 2.978512813569978,
   "q1_ms": 2.9714746851823293,
   "q3_ms": 2.9853375599486753,
   "tflops": 11.53587058999998
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011211037635803223,
   "rel_err": 0.0036814571358263493,
   "peak_mb": 200.501953125,
   "compile_s": 0.00090932403691113,
   "inner": 80,
   "median_ms": 0.6202055374160409,
   "q1_ms": 0.6155351991765201,
   "q3_ms": 0.6257402375922538,
   "tflops": 193.90198415356753
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010370314121246338,
   "rel_err": 0.004226936958730221,
   "peak_mb": 192.517578125,
   "compile_s": 0.0013717750553041697,
   "inner": 36,
   "median_ms": 1.3997256949854393,
   "q1_ms": 1.394726374807457,
   "q3_ms": 1.4348651263086747,
   "tflops": 85.91617966208086
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008320212364196777,
   "rel_err": 0.0034037057776004076,
   "peak_mb": 2232.0009765625,
   "compile_s": 0.0064754721242934465,
   "inner": 7,
   "median_ms": 6.581751563187156,
   "q1_ms": 6.557532702572644,
   "q3_ms": 6.607694434933364,
   "tflops": 18.271592771843483
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008076190948486328,
   "rel_err": 0.002793925581499934,
   "peak_mb": 120.2509765625,
   "compile_s": 0.00015183980576694012,
   "inner": 200,
   "median_ms": 0.10159745579585433,
   "q1_ms": 0.09867861750535667,
   "q3_ms": 0.10196662973612547,
   "tflops": 169.09743506294598
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.008076190948486328,
   "rel_err": 0.002793925581499934,
   "peak_mb": 120.0009765625,
   "compile_s": 0.0002556710969656706,
   "inner": 200,
   "median_ms": 0.20848752465099096,
   "q1_ms": 0.2061626425711438,
   "q3_ms": 0.21116628486197442,
   "tflops": 82.4023845683773
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007548809051513672,
   "rel_err": 0.0026114799547940493,
   "peak_mb": 1344.06298828125,
   "compile_s": 0.003618963062763214,
   "inner": 13,
   "median_ms": 3.6034301424828863,
   "q1_ms": 3.5990101482289343,
   "q3_ms": 3.6044944650851765,
   "tflops": 4.7676431912628905
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.016491413116455078,
   "rel_err": 0.008016865700483322,
   "peak_mb": 200.501953125,
   "compile_s": 0.0008074629586189985,
   "inner": 94,
   "median_ms": 0.46349001027881465,
   "q1_ms": 0.4588506912733329,
   "q3_ms": 0.46387820707713356,
   "tflops": 129.73212110403153
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.016491413116455078,
   "rel_err": 0.008016865700483322,
   "peak_mb": 192.517578125,
   "compile_s": 0.0010533039458096027,
   "inner": 52,
   "median_ms": 0.9340872907509598,
   "q1_ms": 0.9269087951487074,
   "q3_ms": 0.9369308558794168,
   "tflops": 64.37250858606463
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011342048645019531,
   "rel_err": 0.004965658765286207,
   "peak_mb": 2232.0009765625,
   "compile_s": 0.007114334963262081,
   "inner": 7,
   "median_ms": 7.109567589525666,
   "q1_ms": 7.097413197958044,
   "q3_ms": 7.12786049449018,
   "tflops": 8.457552641118038
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000499039888381958,
   "rel_err": 0.0027869928162544966,
   "peak_mb": 147.5234375,
   "compile_s": 0.00032907119020819664,
   "inner": 166,
   "median_ms": 0.30521021774375295,
   "q1_ms": 0.3034518673982336,
   "q3_ms": 0.3111600568686354,
   "tflops": 241.56465188167527
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005548298358917236,
   "rel_err": 0.0030985635239630938,
   "peak_mb": 147.1572265625,
   "compile_s": 0.0006928718648850918,
   "inner": 75,
   "median_ms": 0.7032124387721221,
   "q1_ms": 0.6771798199042678,
   "q3_ms": 0.7049405202269554,
   "tflops": 104.84456180658054
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000477522611618042,
   "rel_err": 0.002666825195774436,
   "peak_mb": 2703.45361328125,
   "compile_s": 0.006385911954566836,
   "inner": 7,
   "median_ms": 6.525874137878418,
   "q1_ms": 6.5126112229855995,
   "q3_ms": 6.540420564955899,
   "tflops": 11.297796807336098
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008320212364196777,
   "rel_err": 0.0034725868608802557,
   "peak_mb": 268.1806640625,
   "compile_s": 0.0013875749427825212,
   "inner": 39,
   "median_ms": 1.3016904346071756,
   "q1_ms": 1.2985364330980258,
   "q3_ms": 1.3266652553843765,
   "tflops": 198.2406823768923
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011704117059707642,
   "rel_err": 0.005179732106626034,
   "peak_mb": 255.8916015625,
   "compile_s": 0.0027531099040061235,
   "inner": 18,
   "median_ms": 2.9140317734951773,
   "q1_ms": 2.903073696264376,
   "q3_ms": 2.9655763905288444,
   "tflops": 88.55359860763957
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008320212364196777,
   "rel_err": 0.0033195260912179947,
   "peak_mb": 4637.6884765625,
   "compile_s": 0.01394927897490561,
   "inner": 3,
   "median_ms": 14.382906723767519,
   "q1_ms": 14.264579319084685,
   "q3_ms": 14.402950143752,
   "tflops": 17.941296912784665
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006508588790893555,
   "rel_err": 0.002251619938760996,
   "peak_mb": 147.8046875,
   "compile_s": 0.00027146097272634506,
   "inner": 200,
   "median_ms": 0.21507400088012218,
   "q1_ms": 0.20862687495537102,
   "q3_ms": 0.21520309732295573,
   "tflops": 171.40147042016125
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006508588790893555,
   "rel_err": 0.002251619938760996,
   "peak_mb": 147.4384765625,
   "compile_s": 0.00045557180419564247,
   "inner": 118,
   "median_ms": 0.41468279688792714,
   "q1_ms": 0.4077290028917714,
   "q3_ms": 0.4197537457033739,
   "tflops": 88.89686352231999
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006508588790893555,
   "rel_err": 0.002251619938760996,
   "peak_mb": 2738.0673828125,
   "compile_s": 0.008091457886621356,
   "inner": 6,
   "median_ms": 8.131348333942393,
   "q1_ms": 8.115015671743702,
   "q3_ms": 8.137856668327004,
   "tflops": 4.533565466150299
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01353597640991211,
   "rel_err": 0.0039185951463878155,
   "peak_mb": 268.4619140625,
   "compile_s": 0.001048634061589837,
   "inner": 57,
   "median_ms": 0.8831806154104701,
   "q1_ms": 0.877350587152729,
   "q3_ms": 0.8889814295495551,
   "tflops": 146.09016292781104
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01353597640991211,
   "rel_err": 0.005356759298592806,
   "peak_mb": 256.1728515625,
   "compile_s": 0.0018611859995871782,
   "inner": 27,
   "median_ms": 1.8845780373171523,
   "q1_ms": 1.8776320783352412,
   "q3_ms": 1.8862048896995407,
   "tflops": 68.46307101385729
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013211488723754883,
   "rel_err": 0.003183129010722041,
   "peak_mb": 4637.4072265625,
   "compile_s": 0.01549034402705729,
   "inner": 3,
   "median_ms": 15.754011382038396,
   "q1_ms": 15.748154682417711,
   "q3_ms": 15.806739839414757,
   "tflops": 8.189914103217165
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00045752525329589844,
   "rel_err": 0.0028358343988656998,
   "peak_mb": 176.5009765625,
   "compile_s": 0.0005635421257466078,
   "inner": 96,
   "median_ms": 0.5526864066875229,
   "q1_ms": 0.5474628912149153,
   "q3_ms": 0.5639459692853658,
   "tflops": 248.67438715514683
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00045752525329589844,
   "rel_err": 0.0028358343988656998,
   "peak_mb": 176.0009765625,
   "compile_s": 0.0011932139750570059,
   "inner": 43,
   "median_ms": 1.2451889762265045,
   "q1_ms": 1.2258464649189698,
   "q3_ms": 1.2498109758463363,
   "tflops": 110.37597994844386
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00045752525329589844,
   "rel_err": 0.0028358343988656998,
   "peak_mb": 4896.12548828125,
   "compile_s": 0.011528261005878448,
   "inner": 4,
   "median_ms": 11.722980765625834,
   "q1_ms": 11.67626425740309,
   "q3_ms": 11.76120963646099,
   "tflops": 11.723891407806365
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007363855838775635,
   "rel_err": 0.003625117940828204,
   "peak_mb": 337.001953125,
   "compile_s": 0.002091257832944393,
   "inner": 24,
   "median_ms": 2.1785405018211654,
   "q1_ms": 2.1346642897697166,
   "q3_ms": 2.2031691041775048,
   "tflops": 220.80669914094983
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007723420858383179,
   "rel_err": 0.004020660184323788,
   "peak_mb": 321.033203125,
   "compile_s": 0.004902817076072097,
   "inner": 10,
   "median_ms": 5.214192206040025,
   "q1_ms": 5.182590952608734,
   "q3_ms": 5.251033743843436,
   "tflops": 92.25519853195598
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007155686616897583,
   "rel_err": 0.0036366344429552555,
   "peak_mb": 8496.0009765625,
   "compile_s": 0.02514912700280547,
   "inner": 1,
   "median_ms": 25.71429987438023,
   "q1_ms": 25.609944481402636,
   "q3_ms": 25.797009468078613,
   "tflops": 18.706958365655055
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004129528999328613,
   "rel_err": 0.0016415518475696445,
   "peak_mb": 176.5009765625,
   "compile_s": 0.0003668610006570816,
   "inner": 150,
   "median_ms": 0.34344305905203026,
   "q1_ms": 0.33072891645133495,
   "q3_ms": 0.34487553328896564,
   "tflops": 200.08986912031105
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004129528999328613,
   "rel_err": 0.0016415518475696445,
   "peak_mb": 176.0009765625,
   "compile_s": 0.0007257531397044659,
   "inner": 71,
   "median_ms": 0.7224498467493645,
   "q1_ms": 0.7054351981807971,
   "q3_ms": 0.7253191335467806,
   "tflops": 95.12006548994461
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0038406848907470703,
   "rel_err": 0.001526731881313026,
   "peak_mb": 4960.12548828125,
   "compile_s": 0.014684651046991348,
   "inner": 3,
   "median_ms": 14.720757647107044,
   "q1_ms": 14.698856005755564,
   "q3_ms": 14.731356369641919,
   "tflops": 4.668202437902706
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015132904052734375,
   "rel_err": 0.003882892895489931,
   "peak_mb": 337.001953125,
   "compile_s": 0.001445394940674305,
   "inner": 36,
   "median_ms": 1.4000376363077924,
   "q1_ms": 1.3880827367150537,
   "q3_ms": 1.4044393319636583,
   "tflops": 171.79407348669523
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015132904052734375,
   "rel_err": 0.004317702725529671,
   "peak_mb": 321.033203125,
   "compile_s": 0.0030076501425355673,
   "inner": 16,
   "median_ms": 3.102978938841261,
   "q1_ms": 3.0908635671949014,
   "q3_ms": 3.1239258678397164,
   "tflops": 77.51202096969959
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015132904052734375,
   "rel_err": 0.0036865491420030594,
   "peak_mb": 8496.0009765625,
   "compile_s": 0.02796170790679753,
   "inner": 1,
   "median_ms": 28.33575801923871,
   "q1_ms": 28.31681352108717,
   "q3_ms": 28.395294095389545,
   "tflops": 8.488150146281562
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00033595412969589233,
   "rel_err": 0.0027797212824225426,
   "peak_mb": 289.0009765625,
   "compile_s": 0.002046847017481923,
   "inner": 25,
   "median_ms": 2.135543841868639,
   "q1_ms": 2.12305077817291,
   "q3_ms": 2.163923545740545,
   "tflops": 257.4312936637975
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0002998635172843933,
   "rel_err": 0.002481103641912341,
   "peak_mb": 288.0009765625,
   "compile_s": 0.004426114959642291,
   "inner": 11,
   "median_ms": 4.819575908847831,
   "q1_ms": 4.795282084325498,
   "q3_ms": 4.868545548313043,
   "tflops": 114.06725908782806
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006060153245925903,
   "rel_err": 0.0036204783245921135,
   "peak_mb": 610.001953125,
   "compile_s": 0.00764953694306314,
   "inner": 6,
   "median_ms": 8.174963334264854,
   "q1_ms": 8.144224896871796,
   "q3_ms": 8.20597360143438,
   "tflops": 235.37051726495991
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006060153245925903,
   "rel_err": 0.003806143067777157,
   "peak_mb": 578.064453125,
   "compile_s": 0.018140593077987432,
   "inner": 2,
   "median_ms": 19.50739254243672,
   "q1_ms": 19.4796952418983,
   "q3_ms": 20.111822173930705,
   "tflops": 98.63672679073746
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007561683654785156,
   "rel_err": 0.003005886683240533,
   "peak_mb": 289.0009765625,
   "compile_s": 0.001156164100393653,
   "inner": 45,
   "median_ms": 1.2069239798519344,
   "q1_ms": 1.198076389523016,
   "q3_ms": 1.2186125671077104,
   "tflops": 227.75080413740892
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007561683654785156,
   "rel_err": 0.003005886683240533,
   "peak_mb": 288.0009765625,
   "compile_s": 0.0024661088827997446,
   "inner": 20,
   "median_ms": 2.5547098484821618,
   "q1_ms": 2.5486140977591276,
   "q3_ms": 2.633699675789103,
   "tflops": 107.5965269039512
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01398777961730957,
   "rel_err": 0.0038633455988019705,
   "peak_mb": 610.001953125,
   "compile_s": 0.004469125997275114,
   "inner": 11,
   "median_ms": 4.638597094030542,
   "q1_ms": 4.614819634959779,
   "q3_ms": 4.642388868060979,
   "tflops": 207.40595805186464
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01398777961730957,
   "rel_err": 0.0048573496751487255,
   "peak_mb": 578.064453125,
   "compile_s": 0.01052279700525105,
   "inner": 4,
   "median_ms": 11.108876031357795,
   "q1_ms": 11.068948486354202,
   "q3_ms": 11.385190999135375,
   "tflops": 86.60396169588091
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027510523796081543,
   "rel_err": 0.0036955196410417557,
   "peak_mb": 71.0166015625,
   "compile_s": 7.912004366517067e-05,
   "inner": 200,
   "median_ms": 0.018523664912208915,
   "q1_ms": 0.018477565026842058,
   "q3_ms": 0.01859471492934972,
   "tflops": 14.491487363446886
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027510523796081543,
   "rel_err": 0.0036955196410417557,
   "peak_mb": 71.0009765625,
   "compile_s": 9.134015999734402e-05,
   "inner": 200,
   "median_ms": 0.023920584935694933,
   "q1_ms": 0.02383428276516497,
   "q3_ms": 0.02403033256996423,
   "tflops": 11.221943640660452
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0019525885581970215,
   "rel_err": 0.002622934291139245,
   "peak_mb": 84.0,
   "compile_s": 0.0002440200187265873,
   "inner": 200,
   "median_ms": 0.15167142963036895,
   "q1_ms": 0.1516053796513006,
   "q3_ms": 0.15184697753284127,
   "tflops": 1.7698485248948401
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005645871162414551,
   "rel_err": 0.0045579178258776665,
   "peak_mb": 81.033203125,
   "compile_s": 0.00048444210551679134,
   "inner": 189,
   "median_ms": 0.2297949250393286,
   "q1_ms": 0.1872738517781414,
   "q3_ms": 0.24278383275799492,
   "tflops": 4.0885328335219056
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007427692413330078,
   "rel_err": 0.006798913236707449,
   "peak_mb": 80.03369140625,
   "compile_s": 0.0006694919429719448,
   "inner": 153,
   "median_ms": 0.24193999345347383,
   "q1_ms": 0.24138479737041435,
   "q3_ms": 0.24212058177035228,
   "tflops": 3.8832938803921837
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004670619964599609,
   "rel_err": 0.004275236278772354,
   "peak_mb": 88.0,
   "compile_s": 0.0012094841804355383,
   "inner": 72,
   "median_ms": 0.6213806369083209,
   "q1_ms": 0.6210193209375979,
   "q3_ms": 0.6215141798343717,
   "tflops": 1.5119944848532805
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005361318588256836,
   "rel_err": 0.001854726462624967,
   "peak_mb": 71.0166015625,
   "compile_s": 7.502082735300064e-05,
   "inner": 200,
   "median_ms": 0.018900964641943574,
   "q1_ms": 0.018846640596166253,
   "q3_ms": 0.019007193041034043,
   "tflops": 7.101104654846784
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005361318588256836,
   "rel_err": 0.001854726462624967,
   "peak_mb": 71.0009765625,
   "compile_s": 8.236989378929138e-05,
   "inner": 200,
   "median_ms": 0.023924284614622593,
   "q1_ms": 0.02387577958870679,
   "q3_ms": 0.023959309910424054,
   "tflops": 5.610104133185479
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0039037466049194336,
   "rel_err": 0.0013504852540791035,
   "peak_mb": 84.0625,
   "compile_s": 0.0006702521350234747,
   "inner": 173,
   "median_ms": 0.1970823057469605,
   "q1_ms": 0.19701395043834097,
   "q3_ms": 0.19727430018364867,
   "tflops": 0.6810237351917624
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01621723175048828,
   "rel_err": 0.00388345867395401,
   "peak_mb": 81.033203125,
   "compile_s": 0.00047885114327073097,
   "inner": 187,
   "median_ms": 0.18786541734109907,
   "q1_ms": 0.18480896045837333,
   "q3_ms": 0.2259147966051484,
   "tflops": 2.5005243362437133
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01621723175048828,
   "rel_err": 0.005187384784221649,
   "peak_mb": 80.03369140625,
   "compile_s": 0.0006441220175474882,
   "inner": 158,
   "median_ms": 0.24159818363104818,
   "q1_ms": 0.24142998780560077,
   "q3_ms": 0.24168910105017166,
   "tflops": 1.9443939558642862
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01621723175048828,
   "rel_err": 0.00388345867395401,
   "peak_mb": 88.0625,
   "compile_s": 0.0013636040966957808,
   "inner": 65,
   "median_ms": 0.6669469392643524,
   "q1_ms": 0.6664104759693146,
   "q3_ms": 0.667005478261182,
   "tflops": 0.7043469582724993
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010970830917358398,
   "rel_err": 0.002655942225828767,
   "peak_mb": 92.0634765625,
   "compile_s": 8.169002830982208e-05,
   "inner": 200,
   "median_ms": 0.01927806413732469,
   "q1_ms": 0.019253442296758294,
   "q3_ms": 0.019304364686831832,
   "tflops": 222.79038317361017
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011283457279205322,
   "rel_err": 0.0027316263876855373,
   "peak_mb": 92.0009765625,
   "compile_s": 9.083002805709839e-05,
   "inner": 200,
   "median_ms": 0.0337223696988076,
   "q1_ms": 0.03371659491676837,
   "q3_ms": 0.033745214459486306,
   "tflops": 127.36255886999149
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009626150131225586,
   "rel_err": 0.002330406801775098,
   "peak_mb": 192.01611328125,
   "compile_s": 0.0003526520449668169,
   "inner": 154,
   "median_ms": 0.2825550127218102,
   "q1_ms": 0.2810980228090344,
   "q3_ms": 0.2825985807780321,
   "tflops": 15.200463989745648
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0023974180221557617,
   "rel_err": 0.005364567507058382,
   "peak_mb": 132.126953125,
   "compile_s": 0.00037115090526640415,
   "inner": 160,
   "median_ms": 0.18366351287113503,
   "q1_ms": 0.18284010657225735,
   "q3_ms": 0.1843249840021599,
   "tflops": 81.8474246789958
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0022614002227783203,
   "rel_err": 0.004969405476003885,
   "peak_mb": 128.12890625,
   "compile_s": 0.0003955711144953966,
   "inner": 134,
   "median_ms": 0.19014386613088757,
   "q1_ms": 0.18825927669349224,
   "q3_ms": 0.1915368953469529,
   "tflops": 79.0579567034379
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0017199218273162842,
   "rel_err": 0.003803657367825508,
   "peak_mb": 252.0009765625,
   "compile_s": 0.0008568721823394299,
   "inner": 61,
   "median_ms": 0.684137803456578,
   "q1_ms": 0.6827522874580787,
   "q3_ms": 0.6858145237946119,
   "tflops": 21.972745052311815
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006621599197387695,
   "rel_err": 0.002405368722975254,
   "peak_mb": 92.0634765625,
   "compile_s": 7.808092050254345e-05,
   "inner": 200,
   "median_ms": 0.02087372587993741,
   "q1_ms": 0.02086939464788884,
   "q3_ms": 0.020888375001959503,
   "tflops": 102.87974750420739
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006621599197387695,
   "rel_err": 0.002405368722975254,
   "peak_mb": 92.0009765625,
   "compile_s": 9.877001866698265e-05,
   "inner": 200,
   "median_ms": 0.03311381442472339,
   "q1_ms": 0.033080289722420275,
   "q3_ms": 0.033125043264590204,
   "tflops": 64.85159397392313
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006621599197387695,
   "rel_err": 0.002405368722975254,
   "peak_mb": 193.01611328125,
   "compile_s": 0.00041562202386558056,
   "inner": 128,
   "median_ms": 0.3212770552636357,
   "q1_ms": 0.3197275191269,
   "q3_ms": 0.32142420695890905,
   "tflops": 6.684211065859662
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013897895812988281,
   "rel_err": 0.004398775286972523,
   "peak_mb": 132.126953125,
   "compile_s": 0.0005422618705779314,
   "inner": 154,
   "median_ms": 0.18406992887037915,
   "q1_ms": 0.18392863015148353,
   "q3_ms": 0.18446966884778693,
   "tflops": 40.83335509567591
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013897895812988281,
   "rel_err": 0.006165427155792713,
   "peak_mb": 128.12890625,
   "compile_s": 0.0005542309954762459,
   "inner": 132,
   "median_ms": 0.18724035577509893,
   "q1_ms": 0.18676648544373387,
   "q3_ms": 0.18767792481054185,
   "tflops": 40.141948763587955
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013897895812988281,
   "rel_err": 0.004398775286972523,
   "peak_mb": 252.0009765625,
   "compile_s": 0.0009070828091353178,
   "inner": 58,
   "median_ms": 0.71267438008739,
   "q1_ms": 0.7102383292633399,
   "q3_ms": 0.7139444732556826,
   "tflops": 10.54646129846613
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001120448112487793,
   "rel_err": 0.0022516855970025063,
   "peak_mb": 96.88671875,
   "compile_s": 9.516021236777306e-05,
   "inner": 200,
   "median_ms": 0.035866874968633056,
   "q1_ms": 0.035859100171364844,
   "q3_ms": 0.03587730054277927,
   "tflops": 164.44811557065498
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011000633239746094,
   "rel_err": 0.002210719743743539,
   "peak_mb": 96.8134765625,
   "compile_s": 0.0001154008787125349,
   "inner": 200,
   "median_ms": 0.059124710969626904,
   "q1_ms": 0.059099579812027514,
   "q3_ms": 0.05967728269752115,
   "tflops": 99.7593037373155
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0009724795818328857,
   "rel_err": 0.00195432361215353,
   "peak_mb": 228.63037109375,
   "compile_s": 0.0004891620483249426,
   "inner": 109,
   "median_ms": 0.4272888256493238,
   "q1_ms": 0.4245613844757643,
   "q3_ms": 0.42742102859763925,
   "tflops": 13.803871400186555
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002751857042312622,
   "rel_err": 0.0050126416608691216,
   "peak_mb": 145.0908203125,
   "compile_s": 0.0005493620410561562,
   "inner": 153,
   "median_ms": 0.21736010501246436,
   "q1_ms": 0.21731216396552092,
   "q3_ms": 0.21747941854318567,
   "tflops": 94.97529456390441
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0027889013290405273,
   "rel_err": 0.0050126416608691216,
   "peak_mb": 139.77685546875,
   "compile_s": 0.0006802030839025974,
   "inner": 98,
   "median_ms": 0.35554328600742985,
   "q1_ms": 0.35417149061983333,
   "q3_ms": 0.3577356383546578,
   "tflops": 58.06280363727245
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0021083950996398926,
   "rel_err": 0.003840544493868947,
   "peak_mb": 310.3134765625,
   "compile_s": 0.0011064040008932352,
   "inner": 46,
   "median_ms": 1.0451577583570844,
   "q1_ms": 1.0413810226094464,
   "q3_ms": 1.0461848451877418,
   "tflops": 19.751888970762355
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006784915924072266,
   "rel_err": 0.0023472141474485397,
   "peak_mb": 96.88671875,
   "compile_s": 8.958997204899788e-05,
   "inner": 200,
   "median_ms": 0.03623557509854436,
   "q1_ms": 0.036159674637019634,
   "q3_ms": 0.036262349458411336,
   "tflops": 81.38742084207932
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006784915924072266,
   "rel_err": 0.0023472141474485397,
   "peak_mb": 96.8134765625,
   "compile_s": 9.890994988381863e-05,
   "inner": 200,
   "median_ms": 0.044454854214563966,
   "q1_ms": 0.04444130521733314,
   "q3_ms": 0.044462779769673944,
   "tflops": 66.33966193581244
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006784915924072266,
   "rel_err": 0.0023472141474485397,
   "peak_mb": 230.00390625,
   "compile_s": 0.0005701419431716204,
   "inner": 92,
   "median_ms": 0.48739473973198433,
   "q1_ms": 0.48626039642840624,
   "q3_ms": 0.48781615180854243,
   "tflops": 6.050783399143177
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012672662734985352,
   "rel_err": 0.00421340623870492,
   "peak_mb": 145.0908203125,
   "compile_s": 0.0005179820582270622,
   "inner": 141,
   "median_ms": 0.19288131247228343,
   "q1_ms": 0.19183034797245307,
   "q3_ms": 0.1937003900356432,
   "tflops": 53.51436003673624
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012672662734985352,
   "rel_err": 0.004915619734674692,
   "peak_mb": 139.77685546875,
   "compile_s": 0.0005522121209651232,
   "inner": 138,
   "median_ms": 0.26010713823463605,
   "q1_ms": 0.2596173907859602,
   "q3_ms": 0.2609087353356291,
   "tflops": 39.683339988496805
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012472391128540039,
   "rel_err": 0.0035177457612007856,
   "peak_mb": 310.3134765625,
   "compile_s": 0.0011875440832227468,
   "inner": 42,
   "median_ms": 1.0889175935604034,
   "q1_ms": 1.0851404540950345,
   "q3_ms": 1.0931201201553145,
   "tflops": 9.479064403992872
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007398128509521484,
   "rel_err": 0.002683986909687519,
   "peak_mb": 120.1259765625,
   "compile_s": 0.00011593103408813477,
   "inner": 200,
   "median_ms": 0.06363057065755129,
   "q1_ms": 0.059806156787090003,
   "q3_ms": 0.064861502032727,
   "tflops": 269.99395112231633
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007505267858505249,
   "rel_err": 0.002722856355831027,
   "peak_mb": 120.0009765625,
   "compile_s": 0.00017019989900290966,
   "inner": 200,
   "median_ms": 0.126865494530648,
   "q1_ms": 0.1184797374298796,
   "q3_ms": 0.13074428017716855,
   "tflops": 135.41798144214627
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007398128509521484,
   "rel_err": 0.002683986909687519,
   "peak_mb": 464.03173828125,
   "compile_s": 0.0010388740338385105,
   "inner": 49,
   "median_ms": 1.024195615544307,
   "q1_ms": 1.0190375187263196,
   "q3_ms": 1.0251881630749118,
   "tflops": 16.7740116470522
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0013456344604492188,
   "rel_err": 0.0038550400640815496,
   "peak_mb": 200.251953125,
   "compile_s": 0.0007377930451184511,
   "inner": 123,
   "median_ms": 0.3560585541484075,
   "q1_ms": 0.353202285298063,
   "q3_ms": 0.35824681200632236,
   "tflops": 168.87543198565487
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0019114017486572266,
   "rel_err": 0.0041186693124473095,
   "peak_mb": 192.255859375,
   "compile_s": 0.0008016328793019056,
   "inner": 80,
   "median_ms": 0.5978059489279985,
   "q1_ms": 0.5967223376501352,
   "q3_ms": 0.5987292126519606,
   "tflops": 100.58371324645712
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00145760178565979,
   "rel_err": 0.003629300044849515,
   "peak_mb": 696.0009765625,
   "compile_s": 0.002345428103581071,
   "inner": 21,
   "median_ms": 2.3910445161163807,
   "q1_ms": 2.3856002205450624,
   "q3_ms": 2.3930354786681987,
   "tflops": 25.147813743620524
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005032062530517578,
   "rel_err": 0.0017408216372132301,
   "peak_mb": 120.1259765625,
   "compile_s": 0.00011528981849551201,
   "inner": 200,
   "median_ms": 0.055277489591389894,
   "q1_ms": 0.055271139717660844,
   "q3_ms": 0.05552287038881332,
   "tflops": 155.3966118124507
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005032062530517578,
   "rel_err": 0.0017408216372132301,
   "peak_mb": 120.0009765625,
   "compile_s": 0.00015218998305499554,
   "inner": 200,
   "median_ms": 0.093564975541085,
   "q1_ms": 0.09290412184782326,
   "q3_ms": 0.09367537451907992,
   "tflops": 91.80715906058354
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.005032062530517578,
   "rel_err": 0.0017408216372132301,
   "peak_mb": 468.03173828125,
   "compile_s": 0.0012181440833956003,
   "inner": 41,
   "median_ms": 1.1762992149536928,
   "q1_ms": 1.1709535823816932,
   "q3_ms": 1.1768861838457423,
   "tflops": 7.302508139766257
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01667952537536621,
   "rel_err": 0.004910856951028109,
   "peak_mb": 200.251953125,
   "compile_s": 0.0006899519357830286,
   "inner": 118,
   "median_ms": 0.3158418733973877,
   "q1_ms": 0.3157652079885415,
   "q3_ms": 0.31594800372150234,
   "tflops": 95.18931340105414
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013173818588256836,
   "rel_err": 0.004045982379466295,
   "peak_mb": 192.255859375,
   "compile_s": 0.0007752231322228909,
   "inner": 85,
   "median_ms": 0.49347324735101533,
   "q1_ms": 0.4934097708695952,
   "q3_ms": 0.49399848315207395,
   "tflops": 60.9248246655496
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011829137802124023,
   "rel_err": 0.0024739813525229692,
   "peak_mb": 696.0009765625,
   "compile_s": 0.0025311079807579517,
   "inner": 20,
   "median_ms": 2.5039906962774694,
   "q1_ms": 2.467106096446514,
   "q3_ms": 2.5135692791081965,
   "tflops": 12.006742324041166
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007283985614776611,
   "rel_err": 0.0028605633415281773,
   "peak_mb": 146.77783203125,
   "compile_s": 0.00020290003158152103,
   "inner": 200,
   "median_ms": 0.1522071799263358,
   "q1_ms": 0.14731288712937385,
   "q3_ms": 0.15404161531478167,
   "tflops": 242.19619611795704
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007283985614776611,
   "rel_err": 0.0028605633415281773,
   "peak_mb": 146.5947265625,
   "compile_s": 0.00031171110458672047,
   "inner": 181,
   "median_ms": 0.2709151984022632,
   "q1_ms": 0.27016743127441867,
   "q3_ms": 0.2797224915210796,
   "tflops": 136.07210011622604
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007283985614776611,
   "rel_err": 0.0028605633415281773,
   "peak_mb": 849.46484375,
   "compile_s": 0.0021618669852614403,
   "inner": 23,
   "median_ms": 2.20878682423221,
   "q1_ms": 2.1963198466793354,
   "q3_ms": 2.212448151903632,
   "tflops": 16.689704771674464
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0016910731792449951,
   "rel_err": 0.004783402197062969,
   "peak_mb": 267.81005859375,
   "compile_s": 0.0009645731188356876,
   "inner": 71,
   "median_ms": 0.7026197756050337,
   "q1_ms": 0.6973212322508786,
   "q3_ms": 0.7117471270735415,
   "tflops": 183.6327477246082
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.002030491828918457,
   "rel_err": 0.006870935205370188,
   "peak_mb": 255.24951171875,
   "compile_s": 0.0014179949648678303,
   "inner": 39,
   "median_ms": 1.2904024587418788,
   "q1_ms": 1.279879973317759,
   "q3_ms": 1.3121926398852315,
   "tflops": 99.98741022688091
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0011043697595596313,
   "rel_err": 0.002777058631181717,
   "peak_mb": 1340.6650390625,
   "compile_s": 0.00515755801461637,
   "inner": 9,
   "median_ms": 5.385788778464,
   "q1_ms": 5.378464223920471,
   "q3_ms": 5.393456564181381,
   "tflops": 23.956379521589223
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0072650909423828125,
   "rel_err": 0.0025133287999778986,
   "peak_mb": 147.62158203125,
   "compile_s": 0.00020018103532493114,
   "inner": 200,
   "median_ms": 0.12301922542974354,
   "q1_ms": 0.11934801819734275,
   "q3_ms": 0.1238459552405402,
   "tflops": 149.83023942486568
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0072650909423828125,
   "rel_err": 0.0025133287999778986,
   "peak_mb": 147.4384765625,
   "compile_s": 0.00025853095576167107,
   "inner": 200,
   "median_ms": 0.1841490902006626,
   "q1_ms": 0.18212106020655483,
   "q3_ms": 0.18429106450639665,
   "tflops": 100.09281055863546
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0072650909423828125,
   "rel_err": 0.0025133287999778986,
   "peak_mb": 858.04833984375,
   "compile_s": 0.002511268947273493,
   "inner": 20,
   "median_ms": 2.5269907899200916,
   "q1_ms": 2.518431277712807,
   "q3_ms": 2.5281910493504256,
   "tflops": 7.294051119427648
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01283717155456543,
   "rel_err": 0.0036486154422163963,
   "peak_mb": 267.81005859375,
   "compile_s": 0.0008529829792678356,
   "inner": 80,
   "median_ms": 0.567060973844491,
   "q1_ms": 0.5664798503858037,
   "q3_ms": 0.5682517323293723,
   "tflops": 113.76554369917119
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015437126159667969,
   "rel_err": 0.007170556113123894,
   "peak_mb": 255.24951171875,
   "compile_s": 0.001232924172654748,
   "inner": 50,
   "median_ms": 0.9429510962218046,
   "q1_ms": 0.9394264686852694,
   "q3_ms": 0.9446282917633653,
   "tflops": 68.41500079748063
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011974096298217773,
   "rel_err": 0.0034811771474778652,
   "peak_mb": 1341.2275390625,
   "compile_s": 0.005446049151942134,
   "inner": 9,
   "median_ms": 5.600168343840374,
   "q1_ms": 5.585964445749092,
   "q3_ms": 5.61415232045369,
   "tflops": 11.519653703081403
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006326138973236084,
   "rel_err": 0.0026098722591996193,
   "peak_mb": 176.2509765625,
   "compile_s": 0.00027498090639710426,
   "inner": 200,
   "median_ms": 0.24601425509899852,
   "q1_ms": 0.24410134996287525,
   "q3_ms": 0.2480948151787743,
   "tflops": 279.3312798412702
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005683600902557373,
   "rel_err": 0.0023447908461093903,
   "peak_mb": 176.0009765625,
   "compile_s": 0.0005055719520896673,
   "inner": 108,
   "median_ms": 0.4744059988297522,
   "q1_ms": 0.4730408135974021,
   "q3_ms": 0.4829067318631267,
   "tflops": 144.85372635572645
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000486910343170166,
   "rel_err": 0.0020087668672204018,
   "peak_mb": 1440.06298828125,
   "compile_s": 0.0038699228316545486,
   "inner": 12,
   "median_ms": 3.937010420486331,
   "q1_ms": 3.9223390825403235,
   "q3_ms": 3.96118675901865,
   "tflops": 17.454735801159302
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.001190096139907837,
   "rel_err": 0.004034040495753288,
   "peak_mb": 336.501953125,
   "compile_s": 0.0012560940813273191,
   "inner": 44,
   "median_ms": 1.1574594748460434,
   "q1_ms": 1.1529599353459412,
   "q3_ms": 1.1782964998432859,
   "tflops": 207.7983495776316
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010787546634674072,
   "rel_err": 0.0036132358945906162,
   "peak_mb": 320.509765625,
   "compile_s": 0.00203651818446815,
   "inner": 24,
   "median_ms": 2.1350886672735214,
   "q1_ms": 2.1167102725788327,
   "q3_ms": 2.1411139362802105,
   "tflops": 112.65020149403836
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0007630288600921631,
   "rel_err": 0.0032062882091850042,
   "peak_mb": 2352.0009765625,
   "compile_s": 0.00884337117895484,
   "inner": 5,
   "median_ms": 9.20452419668436,
   "q1_ms": 9.15345789398998,
   "q3_ms": 9.213525103405118,
   "tflops": 26.130429279836008
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007552385330200195,
   "rel_err": 0.0026127169840037823,
   "peak_mb": 176.2509765625,
   "compile_s": 0.0002413308247923851,
   "inner": 200,
   "median_ms": 0.1889779104385525,
   "q1_ms": 0.1865995500702411,
   "q3_ms": 0.1899885101011023,
   "tflops": 181.81880775516515
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007552385330200195,
   "rel_err": 0.0026127169840037823,
   "peak_mb": 176.0009765625,
   "compile_s": 0.0003567810636013746,
   "inner": 163,
   "median_ms": 0.3048628411524318,
   "q1_ms": 0.3035750462919291,
   "q3_ms": 0.30497397564671525,
   "tflops": 112.70556371552047
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.007552385330200195,
   "rel_err": 0.0026127169840037823,
   "peak_mb": 1456.06298828125,
   "compile_s": 0.004507825011387467,
   "inner": 11,
   "median_ms": 4.526233914392916,
   "q1_ms": 4.524630822495303,
   "q3_ms": 4.540353092703629,
   "tflops": 7.591242303836726
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013039588928222656,
   "rel_err": 0.00369092496111989,
   "peak_mb": 336.501953125,
   "compile_s": 0.0012056438717991114,
   "inner": 57,
   "median_ms": 0.851451194632733,
   "q1_ms": 0.8365007203170343,
   "q3_ms": 0.8645573847420644,
   "tflops": 141.2401380678934
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013039588928222656,
   "rel_err": 0.004484264645725489,
   "peak_mb": 320.509765625,
   "compile_s": 0.0015355050563812256,
   "inner": 34,
   "median_ms": 1.4523027040173904,
   "q1_ms": 1.4470198794322857,
   "q3_ms": 1.4554144855698241,
   "tflops": 82.80579796163484
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.013872146606445312,
   "rel_err": 0.003926584962755442,
   "peak_mb": 2352.0009765625,
   "compile_s": 0.009425553027540445,
   "inner": 5,
   "median_ms": 9.652693616226315,
   "q1_ms": 9.645280591212213,
   "q3_ms": 9.66697761323303,
   "tflops": 12.45860368818117
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005313754081726074,
   "rel_err": 0.001971340971067548,
   "peak_mb": 228.9921875,
   "compile_s": 0.0005581420846283436,
   "inner": 100,
   "median_ms": 0.5207774112932384,
   "q1_ms": 0.5142651440110058,
   "q3_ms": 0.5224056204315275,
   "tflops": 283.1459214673402
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0005130171775817871,
   "rel_err": 0.0019032340496778488,
   "peak_mb": 228.0634765625,
   "compile_s": 0.001031004125252366,
   "inner": 50,
   "median_ms": 1.038115220144391,
   "q1_ms": 1.0348128015175462,
   "q3_ms": 1.0464538400992751,
   "tflops": 142.04203650871278
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00048758089542388916,
   "rel_err": 0.0018088683718815446,
   "peak_mb": 2864.14111328125,
   "compile_s": 0.007804816821590066,
   "inner": 6,
   "median_ms": 8.05233969974021,
   "q1_ms": 8.042508057163406,
   "q3_ms": 8.061971418404331,
   "tflops": 18.312193163529514
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.00100022554397583,
   "rel_err": 0.00427584582939744,
   "peak_mb": 464.3056640625,
   "compile_s": 0.0022862779442220926,
   "inner": 22,
   "median_ms": 2.3310122249478646,
   "q1_ms": 2.3076423623768445,
   "q3_ms": 2.3847399014894934,
   "tflops": 221.40424424910213
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0010921657085418701,
   "rel_err": 0.004668879322707653,
   "peak_mb": 440.8720703125,
   "compile_s": 0.004519656067714095,
   "inner": 11,
   "median_ms": 4.785899453881112,
   "q1_ms": 4.745763356120071,
   "q3_ms": 4.827367279424586,
   "tflops": 107.83678281863472
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000692903995513916,
   "rel_err": 0.003122720867395401,
   "peak_mb": 4810.0947265625,
   "compile_s": 0.01800558203831315,
   "inner": 2,
   "median_ms": 18.80805555265397,
   "q1_ms": 18.672622682061046,
   "q3_ms": 18.896743305958807,
   "tflops": 27.440157147301424
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0070171356201171875,
   "rel_err": 0.0024275495670735836,
   "peak_mb": 228.4296875,
   "compile_s": 0.0004134711343795061,
   "inner": 135,
   "median_ms": 0.37025691837900215,
   "q1_ms": 0.3671048735095947,
   "q3_ms": 0.3744570400427889,
   "tflops": 199.12659653406016
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0070171356201171875,
   "rel_err": 0.0024275495670735836,
   "peak_mb": 228.0634765625,
   "compile_s": 0.000635222066193819,
   "inner": 84,
   "median_ms": 0.6059271067247859,
   "q1_ms": 0.6002206604794732,
   "q3_ms": 0.6067078932011056,
   "tflops": 121.67800248865166
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0070171356201171875,
   "rel_err": 0.0024275495670735836,
   "peak_mb": 2898.4736328125,
   "compile_s": 0.009527093032374978,
   "inner": 5,
   "median_ms": 9.625205583870411,
   "q1_ms": 9.59375542588532,
   "q3_ms": 9.639732609502971,
   "tflops": 7.659888337715181
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.011933326721191406,
   "rel_err": 0.004095529671758413,
   "peak_mb": 465.4306640625,
   "compile_s": 0.001643936149775982,
   "inner": 31,
   "median_ms": 1.5861745473117599,
   "q1_ms": 1.5746535462958198,
   "q3_ms": 1.6003084864707724,
   "tflops": 162.68575260984886
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.012387275695800781,
   "rel_err": 0.004737239796668291,
   "peak_mb": 442.5595703125,
   "compile_s": 0.0029563000425696373,
   "inner": 17,
   "median_ms": 2.968290998764774,
   "q1_ms": 2.9459117350223307,
   "q3_ms": 3.0067610581788946,
   "tflops": 86.93487266153628
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.009859561920166016,
   "rel_err": 0.002931643510237336,
   "peak_mb": 4810.6572265625,
   "compile_s": 0.019537438172847033,
   "inner": 2,
   "median_ms": 19.87271453253925,
   "q1_ms": 19.85700154909864,
   "q3_ms": 19.874256569892168,
   "tflops": 12.985040346525208
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000495031476020813,
   "rel_err": 0.003562052734196186,
   "peak_mb": 288.5009765625,
   "compile_s": 0.00097263278439641,
   "inner": 56,
   "median_ms": 0.952361712864201,
   "q1_ms": 0.9361327247461304,
   "q3_ms": 0.9531819835371736,
   "tflops": 288.627633000189
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000495031476020813,
   "rel_err": 0.003562052734196186,
   "peak_mb": 288.0009765625,
   "compile_s": 0.001778926933184266,
   "inner": 29,
   "median_ms": 1.847762309400172,
   "q1_ms": 1.834807259126984,
   "q3_ms": 1.901773691723316,
   "tflops": 148.76259005046595
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.000481531023979187,
   "rel_err": 0.003464908804744482,
   "peak_mb": 5120.12548828125,
   "compile_s": 0.01399792917072773,
   "inner": 3,
   "median_ms": 14.445083681493998,
   "q1_ms": 14.373588336942095,
   "q3_ms": 14.503622311167419,
   "tflops": 19.02916680892294
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008475780487060547,
   "rel_err": 0.004556367173790932,
   "peak_mb": 609.001953125,
   "compile_s": 0.003921613795682788,
   "inner": 13,
   "median_ms": 4.01835861759117,
   "q1_ms": 3.98119345593911,
   "q3_ms": 4.146808310626792,
   "tflops": 239.41931665639146
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0008551478385925293,
   "rel_err": 0.004597060848027468,
   "peak_mb": 577.017578125,
   "compile_s": 0.007407565834000707,
   "inner": 6,
   "median_ms": 7.850967347621918,
   "q1_ms": 7.819253912506004,
   "q3_ms": 7.992347003892064,
   "tflops": 122.5419278549687
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006635189056396484,
   "rel_err": 0.0030714820604771376,
   "peak_mb": 8736.0009765625,
   "compile_s": 0.03236389299854636,
   "inner": 1,
   "median_ms": 33.789418172091246,
   "q1_ms": 33.70981244370341,
   "q3_ms": 34.01364339515567,
   "tflops": 28.47260255870978
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006810665130615234,
   "rel_err": 0.002356122015044093,
   "peak_mb": 288.5009765625,
   "compile_s": 0.0006564720533788204,
   "inner": 89,
   "median_ms": 0.6089940316598402,
   "q1_ms": 0.5998846177493087,
   "q3_ms": 0.610441623485825,
   "tflops": 225.68193828994353
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006810665130615234,
   "rel_err": 0.002356122015044093,
   "peak_mb": 288.0009765625,
   "compile_s": 0.0010580038651823997,
   "inner": 50,
   "median_ms": 1.0706921387463808,
   "q1_ms": 1.0654928116127849,
   "q3_ms": 1.0733142402023077,
   "tflops": 128.3645863253655
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.006810665130615234,
   "rel_err": 0.002356122015044093,
   "peak_mb": 5184.12548828125,
   "compile_s": 0.01716218004003167,
   "inner": 2,
   "median_ms": 17.286985530517995,
   "q1_ms": 17.214060004334897,
   "q3_ms": 17.306322522927076,
   "tflops": 7.9504291381148455
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015851497650146484,
   "rel_err": 0.003916752990335226,
   "peak_mb": 609.001953125,
   "compile_s": 0.0025770689826458693,
   "inner": 20,
   "median_ms": 2.597698546014726,
   "q1_ms": 2.5397711026016623,
   "q3_ms": 2.6076948503032327,
   "tflops": 185.17789059472838
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015851497650146484,
   "rel_err": 0.003916752990335226,
   "peak_mb": 577.017578125,
   "compile_s": 0.004574715858325362,
   "inner": 11,
   "median_ms": 4.7099381812255485,
   "q1_ms": 4.66930540278554,
   "q3_ms": 4.7753006854856554,
   "tflops": 102.13219763042241
  },
  {
   "impl": "torch_math",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.015851497650146484,
   "rel_err": 0.003916752990335226,
   "peak_mb": 8736.0009765625,
   "compile_s": 0.035069752018898726,
   "inner": 1,
   "median_ms": 36.10978601500392,
   "q1_ms": 35.895834560506046,
   "q3_ms": 36.275601014494896,
   "tflops": 13.32149509144489
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003153011202812195,
   "rel_err": 0.002931902650743723,
   "peak_mb": 513.0009765625,
   "compile_s": 0.003563513047993183,
   "inner": 14,
   "median_ms": 3.753031005284616,
   "q1_ms": 3.7450112868100405,
   "q3_ms": 3.880444248872144,
   "tflops": 292.9663054282753
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0003512650728225708,
   "rel_err": 0.003266321960836649,
   "peak_mb": 512.0009765625,
   "compile_s": 0.0067433028016239405,
   "inner": 7,
   "median_ms": 7.277723867446184,
   "q1_ms": 7.228206643568618,
   "q3_ms": 7.480728290309863,
   "tflops": 151.07905270962527
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006157159805297852,
   "rel_err": 0.005359946750104427,
   "peak_mb": 1154.001953125,
   "compile_s": 0.013908217893913388,
   "inner": 3,
   "median_ms": 14.997265379255017,
   "q1_ms": 14.906550330730777,
   "q3_ms": 15.17653469151507,
   "tflops": 256.59949330090217
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.0006157159805297852,
   "rel_err": 0.005359946750104427,
   "peak_mb": 1090.033203125,
   "compile_s": 0.028337948955595493,
   "inner": 1,
   "median_ms": 30.677677132189274,
   "q1_ms": 30.391750391572714,
   "q3_ms": 31.038912944495678,
   "tflops": 125.44270156550056
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004915475845336914,
   "rel_err": 0.0017004889668896794,
   "peak_mb": 513.0009765625,
   "compile_s": 0.0020226668566465378,
   "inner": 25,
   "median_ms": 2.100210525095463,
   "q1_ms": 2.0495289750397205,
   "q3_ms": 2.114935778081417,
   "tflops": 261.76224112723725
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.004915475845336914,
   "rel_err": 0.0017004889668896794,
   "peak_mb": 512.0009765625,
   "compile_s": 0.003631601808592677,
   "inner": 14,
   "median_ms": 3.7831060056175505,
   "q1_ms": 3.7777417455799878,
   "q3_ms": 3.823466890025884,
   "tflops": 145.31863846047804
  },
  {
   "impl": "torch_flash",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01760244369506836,
   "rel_err": 0.005058928392827511,
   "peak_mb": 1154.001953125,
   "compile_s": 0.00797761813737452,
   "inner": 6,
   "median_ms": 8.399719333586594,
   "q1_ms": 8.341048242679488,
   "q3_ms": 8.435636817011982,
   "tflops": 229.07257637933597
  },
  {
   "impl": "torch_efficient",
   "dtype": "bfloat16",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 0.01760244369506836,
   "rel_err": 0.005591186229139566,
   "peak_mb": 1090.033203125,
   "compile_s": 0.015571523923426867,
   "inner": 3,
   "median_ms": 16.697434708476067,
   "q1_ms": 16.619109509823225,
   "q3_ms": 17.007144206824403,
   "tflops": 115.23598577877667
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.642673492431641e-07,
   "rel_err": 9.249482673112652e-07,
   "peak_mb": 70.0009765625,
   "compile_s": 0.0002504209987819195,
   "inner": 200,
   "median_ms": 0.02362438477575779,
   "q1_ms": 0.02359755744691938,
   "q3_ms": 0.02364963002037257,
   "tflops": 5.681321620604816
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.470348358154297e-07,
   "rel_err": 4.784215548170323e-07,
   "peak_mb": 74.50439453125,
   "compile_s": 0.00034217117354273796,
   "inner": 200,
   "median_ms": 0.12004217016510665,
   "q1_ms": 0.11972441745456308,
   "q3_ms": 0.12016419204883277,
   "tflops": 1.1180881503174775
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.3113021850585938e-06,
   "rel_err": 1.412345000062487e-06,
   "peak_mb": 78.048828125,
   "compile_s": 0.0006458619609475136,
   "inner": 168,
   "median_ms": 0.21242061934788667,
   "q1_ms": 0.2115813923800098,
   "q3_ms": 0.21318598217996104,
   "tflops": 2.2114710400625404
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.172325134277344e-07,
   "rel_err": 5.33427794380259e-07,
   "peak_mb": 83.0009765625,
   "compile_s": 0.0009195129387080669,
   "inner": 83,
   "median_ms": 0.5134731076702654,
   "q1_ms": 0.512921494005405,
   "q3_ms": 0.5140231083517512,
   "tflops": 0.9148717644267844
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0728836059570312e-06,
   "rel_err": 3.539919930517499e-07,
   "peak_mb": 70.0009765625,
   "compile_s": 7.76299275457859e-05,
   "inner": 200,
   "median_ms": 0.023842635564506054,
   "q1_ms": 0.023810809943825006,
   "q3_ms": 0.023938632803037763,
   "tflops": 2.8146579608801017
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.960464477539062e-07,
   "rel_err": 1.966622136251317e-07,
   "peak_mb": 74.56689453125,
   "compile_s": 0.0002820009831339121,
   "inner": 200,
   "median_ms": 0.16564138000831008,
   "q1_ms": 0.1652208500308916,
   "q3_ms": 0.16578647715505213,
   "tflops": 0.4051455258138588
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.9073486328125e-06,
   "rel_err": 1.1423012438172009e-06,
   "peak_mb": 78.048828125,
   "compile_s": 0.0006659519858658314,
   "inner": 139,
   "median_ms": 0.21468175628738437,
   "q1_ms": 0.21453711569389639,
   "q3_ms": 0.2147643098259251,
   "tflops": 1.094089353757549
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.6689300537109375e-06,
   "rel_err": 6.106315595388878e-07,
   "peak_mb": 83.0009765625,
   "compile_s": 0.0009549139067530632,
   "inner": 74,
   "median_ms": 0.5657015701856565,
   "q1_ms": 0.564699128861665,
   "q3_ms": 0.5662910265822869,
   "tflops": 0.41520306178912475
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.3213367462158203e-07,
   "rel_err": 1.052677021107229e-06,
   "peak_mb": 88.0009765625,
   "compile_s": 0.00027654110454022884,
   "inner": 200,
   "median_ms": 0.08671590592712164,
   "q1_ms": 0.08331324032042176,
   "q3_ms": 0.08792492968495935,
   "tflops": 24.764587592555426
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.5762786865234375e-07,
   "rel_err": 8.711809869055287e-07,
   "peak_mb": 160.01611328125,
   "compile_s": 0.00029198103584349155,
   "inner": 200,
   "median_ms": 0.1847137906588614,
   "q1_ms": 0.18444031768012792,
   "q3_ms": 0.18473402014933527,
   "tflops": 11.626006051524758
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1324882507324219e-06,
   "rel_err": 1.943797542480752e-06,
   "peak_mb": 120.189453125,
   "compile_s": 0.00046231108717620373,
   "inner": 126,
   "median_ms": 0.3104060001316524,
   "q1_ms": 0.30845571919122616,
   "q3_ms": 0.31622142128143754,
   "tflops": 24.21407049094463
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 7.748603820800781e-07,
   "rel_err": 1.32996672164154e-06,
   "peak_mb": 236.0009765625,
   "compile_s": 0.0009548128582537174,
   "inner": 72,
   "median_ms": 0.5102496943436563,
   "q1_ms": 0.5093980273361214,
   "q3_ms": 0.5104296320092141,
   "tflops": 14.730420912193232
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.5367431640625e-07,
   "rel_err": 3.336561462674581e-07,
   "peak_mb": 88.0009765625,
   "compile_s": 0.00012456090189516544,
   "inner": 200,
   "median_ms": 0.07598516414873302,
   "q1_ms": 0.0759210146497935,
   "q3_ms": 0.07602011493872851,
   "tflops": 14.13094037539042
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.8743019104003906e-07,
   "rel_err": 1.35547807644798e-07,
   "peak_mb": 161.01611328125,
   "compile_s": 0.0003584008663892746,
   "inner": 161,
   "median_ms": 0.2254874341374969,
   "q1_ms": 0.2246417137013274,
   "q3_ms": 0.22552808043865127,
   "tflops": 4.761869893580224
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.337860107421875e-06,
   "rel_err": 8.472996455566317e-07,
   "peak_mb": 120.189453125,
   "compile_s": 0.0004957718774676323,
   "inner": 101,
   "median_ms": 0.283485336118553,
   "q1_ms": 0.28139479356900887,
   "q3_ms": 0.2843596036592037,
   "tflops": 13.25675759972421
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.337860107421875e-06,
   "rel_err": 1.1120807812403655e-06,
   "peak_mb": 236.0009765625,
   "compile_s": 0.000997783849015832,
   "inner": 67,
   "median_ms": 0.5612946567195121,
   "q1_ms": 0.5607912768679324,
   "q3_ms": 0.5623374096076212,
   "tflops": 6.695407374736477
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.556510925292969e-07,
   "rel_err": 1.4932564909031498e-06,
   "peak_mb": 92.1259765625,
   "compile_s": 0.00018202094361186028,
   "inner": 200,
   "median_ms": 0.14346285024657845,
   "q1_ms": 0.13988448772579432,
   "q3_ms": 0.14353267499245703,
   "tflops": 20.556680666326965
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.2782554626464844e-07,
   "rel_err": 7.466282454515749e-07,
   "peak_mb": 191.13037109375,
   "compile_s": 0.00045193196274340153,
   "inner": 146,
   "median_ms": 0.27608109497758626,
   "q1_ms": 0.27508373700771227,
   "q3_ms": 0.27764720159697615,
   "tflops": 10.682078757473146
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1920928955078125e-06,
   "rel_err": 1.6196685237446218e-06,
   "peak_mb": 130.091796875,
   "compile_s": 0.0008710429538041353,
   "inner": 78,
   "median_ms": 0.573317383010036,
   "q1_ms": 0.5722364174942367,
   "q3_ms": 0.5739781533917174,
   "tflops": 18.003849710273503
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 7.152557373046875e-07,
   "rel_err": 1.127220912167104e-06,
   "peak_mb": 291.5634765625,
   "compile_s": 0.0008931539487093687,
   "inner": 62,
   "median_ms": 0.6753675181478743,
   "q1_ms": 0.6706425732362175,
   "q3_ms": 0.6761293569129081,
   "tflops": 15.283411953697152
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.834766387939453e-07,
   "rel_err": 4.1586460497455846e-07,
   "peak_mb": 92.1259765625,
   "compile_s": 0.00016660080291330814,
   "inner": 200,
   "median_ms": 0.11586465407162905,
   "q1_ms": 0.11577532975934446,
   "q3_ms": 0.11592113005463034,
   "tflops": 12.726573188475646
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.172325134277344e-07,
   "rel_err": 1.76427406017865e-07,
   "peak_mb": 192.50390625,
   "compile_s": 0.0004569909069687128,
   "inner": 118,
   "median_ms": 0.33564022860600284,
   "q1_ms": 0.3351817160087117,
   "q3_ms": 0.33592379659856275,
   "tflops": 4.393275520411286
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.291534423828125e-06,
   "rel_err": 1.113206849367998e-06,
   "peak_mb": 130.091796875,
   "compile_s": 0.0006817721296101809,
   "inner": 106,
   "median_ms": 0.3887775841713795,
   "q1_ms": 0.3886285758542143,
   "q3_ms": 0.38894442050665057,
   "tflops": 13.274839420075631
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 2.86102294921875e-06,
   "rel_err": 9.239368523594749e-07,
   "peak_mb": 291.5634765625,
   "compile_s": 0.0009545630309730768,
   "inner": 57,
   "median_ms": 0.7257674363229359,
   "q1_ms": 0.7248962984273308,
   "q3_ms": 0.7274864801091322,
   "tflops": 7.111038249591002
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.917383193969727e-07,
   "rel_err": 1.6426118918388966e-06,
   "peak_mb": 112.0009765625,
   "compile_s": 0.00036197109147906303,
   "inner": 148,
   "median_ms": 0.3467041816612756,
   "q1_ms": 0.3381184093037469,
   "q3_ms": 0.3498712195888967,
   "tflops": 24.775976311679532
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.0547380447387695e-07,
   "rel_err": 1.0204104228250799e-06,
   "peak_mb": 400.03173828125,
   "compile_s": 0.0007442520000040531,
   "inner": 69,
   "median_ms": 0.6803780274492675,
   "q1_ms": 0.6787150430128626,
   "q3_ms": 0.6806785274512959,
   "tflops": 12.625238096244239
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.662441253662109e-07,
   "rel_err": 1.656587301113177e-06,
   "peak_mb": 176.376953125,
   "compile_s": 0.0011464538984000683,
   "inner": 46,
   "median_ms": 1.1192656314486393,
   "q1_ms": 1.1108146433758996,
   "q3_ms": 1.120340957245587,
   "tflops": 26.861158090852722
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.21540641784668e-07,
   "rel_err": 1.420848320776713e-06,
   "peak_mb": 664.0009765625,
   "compile_s": 0.0016887059900909662,
   "inner": 30,
   "median_ms": 1.6110586312909922,
   "q1_ms": 1.605118267859022,
   "q3_ms": 1.6139626347770295,
   "tflops": 18.661500263282253
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.344650268554688e-07,
   "rel_err": 2.8424688025552314e-07,
   "peak_mb": 112.0009765625,
   "compile_s": 0.0002734509762376547,
   "inner": 200,
   "median_ms": 0.2352216700091958,
   "q1_ms": 0.23139487951993942,
   "q3_ms": 0.23611372045706958,
   "tflops": 18.25923307079697
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.172325134277344e-07,
   "rel_err": 1.4212344012776157e-07,
   "peak_mb": 404.03173828125,
   "compile_s": 0.0009740130044519901,
   "inner": 54,
   "median_ms": 0.8359917965545146,
   "q1_ms": 0.8335306839500036,
   "q3_ms": 0.8402919996081403,
   "tflops": 5.137571102612999
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 2.6226043701171875e-06,
   "rel_err": 1.0977464626193978e-06,
   "peak_mb": 176.376953125,
   "compile_s": 0.0010022141505032778,
   "inner": 59,
   "median_ms": 0.7922188655110235,
   "q1_ms": 0.7856414748071614,
   "q3_ms": 0.7994617631469491,
   "tflops": 18.97504110344975
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.7220458984375e-06,
   "rel_err": 1.7685914599496755e-06,
   "peak_mb": 664.0009765625,
   "compile_s": 0.00188536592759192,
   "inner": 27,
   "median_ms": 1.7427072204925396,
   "q1_ms": 1.7358903014273555,
   "q3_ms": 1.743542185674111,
   "tflops": 8.62588124914718
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.129243850708008e-07,
   "rel_err": 1.2418921642165515e-06,
   "peak_mb": 134.8759765625,
   "compile_s": 0.0007993530016392469,
   "inner": 75,
   "median_ms": 0.6883786680797735,
   "q1_ms": 0.684875319711864,
   "q3_ms": 0.7048857941602666,
   "tflops": 26.775960462888698
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.470348358154297e-07,
   "rel_err": 1.774131760612363e-06,
   "peak_mb": 752.90234375,
   "compile_s": 0.0014752349816262722,
   "inner": 34,
   "median_ms": 1.4560915570815698,
   "q1_ms": 1.4518962364972514,
   "q3_ms": 1.4570797656607979,
   "tflops": 12.658544657001569
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.258487701416016e-07,
   "rel_err": 1.2473811921154265e-06,
   "peak_mb": 231.25537109375,
   "compile_s": 0.002198738045990467,
   "inner": 23,
   "median_ms": 2.3383616024385327,
   "q1_ms": 2.3004675850920053,
   "q3_ms": 2.401890764620317,
   "tflops": 27.588547439679314
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.662441253662109e-07,
   "rel_err": 1.207150944537716e-06,
   "peak_mb": 1292.3837890625,
   "compile_s": 0.003490183036774397,
   "inner": 14,
   "median_ms": 3.4938871394842863,
   "q1_ms": 3.479736782277801,
   "q3_ms": 3.498352922698749,
   "tflops": 18.4642483928437
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.3709068298339844e-06,
   "rel_err": 4.6151853894116357e-07,
   "peak_mb": 135.1572265625,
   "compile_s": 0.0004839219618588686,
   "inner": 110,
   "median_ms": 0.4726121900603175,
   "q1_ms": 0.4720527773977003,
   "q3_ms": 0.4744641144167293,
   "tflops": 19.50013180748427
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.76837158203125e-07,
   "rel_err": 1.6052818807565927e-07,
   "peak_mb": 762.04833984375,
   "compile_s": 0.0018467758782207966,
   "inner": 27,
   "median_ms": 1.7767914105206728,
   "q1_ms": 1.773673208878824,
   "q3_ms": 1.7777728477354955,
   "tflops": 5.186877843640258
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.814697265625e-06,
   "rel_err": 1.5930278323139646e-06,
   "peak_mb": 231.81787109375,
   "compile_s": 0.0015991448890417814,
   "inner": 33,
   "median_ms": 1.5325150311444744,
   "q1_ms": 1.5294197276516845,
   "q3_ms": 1.5384794572706926,
   "tflops": 21.04775440663142
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.5299530029296875e-06,
   "rel_err": 1.4902518614690052e-06,
   "peak_mb": 1292.9462890625,
   "compile_s": 0.003793272888287902,
   "inner": 13,
   "median_ms": 3.7606962275906253,
   "q1_ms": 3.7441972750597277,
   "q3_ms": 3.7633573117021184,
   "tflops": 8.577135202612611
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.650784492492676e-07,
   "rel_err": 1.8326874169360963e-06,
   "peak_mb": 160.0009765625,
   "compile_s": 0.0013113049790263176,
   "inner": 43,
   "median_ms": 1.223551228555829,
   "q1_ms": 1.2163784311607828,
   "q3_ms": 1.28655819774627,
   "tflops": 28.081977743224677
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.129243850708008e-07,
   "rel_err": 1.5708749288023682e-06,
   "peak_mb": 1312.06298828125,
   "compile_s": 0.002593049081042409,
   "inner": 19,
   "median_ms": 2.5731552588312248,
   "q1_ms": 2.560485999933199,
   "q3_ms": 2.5740616011286255,
   "tflops": 13.353153973152338
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.046627044677734e-07,
   "rel_err": 2.781386911010486e-06,
   "peak_mb": 288.751953125,
   "compile_s": 0.003842714009806514,
   "inner": 13,
   "median_ms": 4.083483462007,
   "q1_ms": 4.011702070084329,
   "q3_ms": 4.246142889874486,
   "tflops": 29.45012154619909
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.513429641723633e-07,
   "rel_err": 1.648229385864397e-06,
   "peak_mb": 2288.0009765625,
   "compile_s": 0.006075580837205052,
   "inner": 8,
   "median_ms": 6.172191497171298,
   "q1_ms": 6.1595890001626685,
   "q3_ms": 6.179542760946788,
   "tflops": 19.484017037241067
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0728836059570312e-06,
   "rel_err": 4.1042079601538717e-07,
   "peak_mb": 160.0009765625,
   "compile_s": 0.0007774629630148411,
   "inner": 66,
   "median_ms": 0.8227604399011893,
   "q1_ms": 0.7746961906157208,
   "q3_ms": 0.8282217595754473,
   "tflops": 20.880767172110563
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.5762786865234375e-07,
   "rel_err": 1.3680693200512906e-07,
   "peak_mb": 1328.06298828125,
   "compile_s": 0.0032765408977866173,
   "inner": 15,
   "median_ms": 3.1879911199212074,
   "q1_ms": 3.182321429873506,
   "q3_ms": 3.199441839630405,
   "tflops": 5.388932571563942
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.291534423828125e-06,
   "rel_err": 1.452910396437801e-06,
   "peak_mb": 288.751953125,
   "compile_s": 0.002470399020239711,
   "inner": 20,
   "median_ms": 2.562994894105941,
   "q1_ms": 2.5127337779849768,
   "q3_ms": 2.580012724502012,
   "tflops": 23.46065623551514
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.9141387939453125e-06,
   "rel_err": 1.8645683894646936e-06,
   "peak_mb": 2288.0009765625,
   "compile_s": 0.0066794229205697775,
   "inner": 7,
   "median_ms": 6.671917425202472,
   "q1_ms": 6.658732508575278,
   "q3_ms": 6.684344710915216,
   "tflops": 9.012333083869851
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.203749656677246e-07,
   "rel_err": 1.7645199932303512e-06,
   "peak_mb": 204.6259765625,
   "compile_s": 0.002492329105734825,
   "inner": 20,
   "median_ms": 2.6443487033247948,
   "q1_ms": 2.6249506801832467,
   "q3_ms": 2.7629886229988188,
   "tflops": 27.881345568116732
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.501772880554199e-07,
   "rel_err": 1.928661276906496e-06,
   "peak_mb": 2676.64111328125,
   "compile_s": 0.0055732091423124075,
   "inner": 8,
   "median_ms": 5.728391115553677,
   "q1_ms": 5.706727446522564,
   "q3_ms": 5.739556305343285,
   "tflops": 12.870629555969805
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.940696716308594e-07,
   "rel_err": 3.1199194836517563e-06,
   "peak_mb": 397.1640625,
   "compile_s": 0.00815481785684824,
   "inner": 6,
   "median_ms": 8.799173985607922,
   "q1_ms": 8.68706857242311,
   "q3_ms": 8.971464645583183,
   "tflops": 29.326389093120294
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.5367431640625e-07,
   "rel_err": 3.3279141007369617e-06,
   "peak_mb": 4718.0322265625,
   "compile_s": 0.013125655939802527,
   "inner": 3,
   "median_ms": 13.503773603588343,
   "q1_ms": 13.488893668788176,
   "q3_ms": 13.516920308272045,
   "tflops": 19.109325109792216
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0728836059570312e-06,
   "rel_err": 3.909314330030611e-07,
   "peak_mb": 205.7509765625,
   "compile_s": 0.0014793460723012686,
   "inner": 33,
   "median_ms": 1.4972803676783135,
   "q1_ms": 1.4769408964749537,
   "q3_ms": 1.5098863161366545,
   "tflops": 24.620639391112437
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.76837158203125e-07,
   "rel_err": 1.7374730987285147e-07,
   "peak_mb": 2712.0986328125,
   "compile_s": 0.00727512501180172,
   "inner": 6,
   "median_ms": 7.3390121494109435,
   "q1_ms": 7.317342174549898,
   "q3_ms": 7.358233172756931,
   "tflops": 5.02301934504344
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.337860107421875e-06,
   "rel_err": 1.4926331459719222e-06,
   "peak_mb": 397.7265625,
   "compile_s": 0.004741176031529903,
   "inner": 10,
   "median_ms": 5.075963702984154,
   "q1_ms": 4.940533253829926,
   "q3_ms": 5.135991401039064,
   "tflops": 25.41862147756236
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0013580322265625e-05,
   "rel_err": 3.607196958910208e-06,
   "peak_mb": 4717.4697265625,
   "compile_s": 0.01472575101070106,
   "inner": 3,
   "median_ms": 14.942205355813106,
   "q1_ms": 14.864451834000647,
   "q3_ms": 14.952602136569718,
   "tflops": 8.634869949087173
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.0547380447387695e-07,
   "rel_err": 1.3578446669271216e-06,
   "peak_mb": 256.0009765625,
   "compile_s": 0.004493615124374628,
   "inner": 11,
   "median_ms": 4.80164035054093,
   "q1_ms": 4.771486136384985,
   "q3_ms": 4.918959415094419,
   "tflops": 28.623333577350657
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.3213367462158203e-07,
   "rel_err": 1.9208534922654508e-06,
   "peak_mb": 4864.12548828125,
   "compile_s": 0.009986025048419833,
   "inner": 5,
   "median_ms": 10.261655785143375,
   "q1_ms": 10.239298571832478,
   "q3_ms": 10.274386801756918,
   "tflops": 13.393448031162908
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 7.450580596923828e-07,
   "rel_err": 3.5952757571067195e-06,
   "peak_mb": 513.501953125,
   "compile_s": 0.014772261027246714,
   "inner": 3,
   "median_ms": 16.048466010640066,
   "q1_ms": 15.948287366578976,
   "q3_ms": 16.386293650915228,
   "tflops": 29.973976131617498
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.407499313354492e-07,
   "rel_err": 2.838011823769193e-06,
   "peak_mb": 8608.0009765625,
   "compile_s": 0.02360473293811083,
   "inner": 2,
   "median_ms": 24.2412545485422,
   "q1_ms": 24.225266999565065,
   "q3_ms": 24.28581955609843,
   "tflops": 19.843706363824644
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.2516975402832031e-06,
   "rel_err": 4.4143843069832656e-07,
   "peak_mb": 256.0009765625,
   "compile_s": 0.002543109003454447,
   "inner": 19,
   "median_ms": 2.632124472017351,
   "q1_ms": 2.6117693694112334,
   "q3_ms": 2.6394502308807875,
   "tflops": 26.107988990099326
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.364418029785156e-07,
   "rel_err": 1.8918790090083348e-07,
   "peak_mb": 4928.12548828125,
   "compile_s": 0.013117486145347357,
   "inner": 3,
   "median_ms": 13.247382982323566,
   "q1_ms": 13.181196215252083,
   "q3_ms": 13.269819513273736,
   "tflops": 5.187400170108672
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.198883056640625e-06,
   "rel_err": 1.7577269773028092e-06,
   "peak_mb": 513.501953125,
   "compile_s": 0.008394569158554077,
   "inner": 6,
   "median_ms": 9.02271484180043,
   "q1_ms": 8.908431006905932,
   "q3_ms": 9.114861759978037,
   "tflops": 26.656962210722597
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 7.152557373046875e-06,
   "rel_err": 2.6365905796410516e-06,
   "peak_mb": 8608.0009765625,
   "compile_s": 0.026418392080813646,
   "inner": 1,
   "median_ms": 26.856482960283756,
   "q1_ms": 26.774498401209712,
   "q3_ms": 26.880873483605683,
   "tflops": 8.955683770346479
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 2.8312206268310547e-07,
   "rel_err": 2.5178296709782444e-06,
   "peak_mb": 448.0009765625,
   "compile_s": 0.017671552021056414,
   "inner": 2,
   "median_ms": 19.266702001914382,
   "q1_ms": 18.917513254564255,
   "q3_ms": 19.646203552838415,
   "tflops": 28.533986451514902
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.140900611877441e-07,
   "rel_err": 3.7055394841445377e-06,
   "peak_mb": 963.001953125,
   "compile_s": 0.06370388204231858,
   "inner": 1,
   "median_ms": 65.28051709756255,
   "q1_ms": 64.56094444729388,
   "q3_ms": 65.68088906351477,
   "tflops": 29.475032278502646
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0728836059570312e-06,
   "rel_err": 4.066610586050956e-07,
   "peak_mb": 448.0009765625,
   "compile_s": 0.00926714320667088,
   "inner": 5,
   "median_ms": 9.78560820221901,
   "q1_ms": 9.700239822268486,
   "q3_ms": 9.871178423054516,
   "tflops": 28.090017632390797
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 64,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0013580322265625e-05,
   "rel_err": 3.3173118936247192e-06,
   "peak_mb": 963.001953125,
   "compile_s": 0.030912618152797222,
   "inner": 1,
   "median_ms": 33.85834791697562,
   "q1_ms": 32.77042938861996,
   "q3_ms": 34.478735411539674,
   "tflops": 28.41463726059841
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.2516975402832031e-06,
   "rel_err": 1.4724993206982617e-06,
   "peak_mb": 76.0009765625,
   "compile_s": 8.078105747699738e-05,
   "inner": 200,
   "median_ms": 0.023915680358186364,
   "q1_ms": 0.02386606065556407,
   "q3_ms": 0.02404293220024556,
   "tflops": 11.224245013297907
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.556510925292969e-07,
   "rel_err": 7.71309146330168e-07,
   "peak_mb": 80.50439453125,
   "compile_s": 0.00023110094480216503,
   "inner": 200,
   "median_ms": 0.11793151032179594,
   "q1_ms": 0.11782603512983769,
   "q3_ms": 0.11809981253463775,
   "tflops": 2.276197898827283
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0728836059570312e-06,
   "rel_err": 1.3375477010413306e-06,
   "peak_mb": 92.048828125,
   "compile_s": 0.0003423320595175028,
   "inner": 200,
   "median_ms": 0.18094983068294823,
   "q1_ms": 0.18037427798844874,
   "q3_ms": 0.18110623059328645,
   "tflops": 5.192180022794218
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 7.748603820800781e-07,
   "rel_err": 1.1700042250595288e-06,
   "peak_mb": 94.0009765625,
   "compile_s": 0.0006389529444277287,
   "inner": 90,
   "median_ms": 0.4545270206613673,
   "q1_ms": 0.4544341380500959,
   "q3_ms": 0.4549371337311135,
   "tflops": 2.0670368389384848
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.7881393432617188e-06,
   "rel_err": 5.69636540603824e-07,
   "peak_mb": 76.0009765625,
   "compile_s": 8.159107528626919e-05,
   "inner": 200,
   "median_ms": 0.023726829094812274,
   "q1_ms": 0.02366200729738921,
   "q3_ms": 0.02378920733463019,
   "tflops": 5.656791620307405
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.344650268554688e-07,
   "rel_err": 2.6583037993077596e-07,
   "peak_mb": 80.56689453125,
   "compile_s": 0.0002543199807405472,
   "inner": 200,
   "median_ms": 0.16387411975301802,
   "q1_ms": 0.16366726718842983,
   "q3_ms": 0.16408714465796947,
   "tflops": 0.8190294367547818
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 2.0265579223632812e-06,
   "rel_err": 1.1511345974213327e-06,
   "peak_mb": 92.048828125,
   "compile_s": 0.0003652020823210478,
   "inner": 164,
   "median_ms": 0.24139395029079624,
   "q1_ms": 0.24123425615338107,
   "q3_ms": 0.24203748780120982,
   "tflops": 1.946039026388603
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 128,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 2.1457672119140625e-06,
   "rel_err": 5.560196996157174e-07,
   "peak_mb": 94.0009765625,
   "compile_s": 0.0007665730081498623,
   "inner": 77,
   "median_ms": 0.5684124952932069,
   "q1_ms": 0.5682615923102606,
   "q3_ms": 0.5697562803768299,
   "tflops": 0.8264456743824401
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.043081283569336e-06,
   "rel_err": 1.5308652336898376e-06,
   "peak_mb": 112.0009765625,
   "compile_s": 0.00015983101911842823,
   "inner": 200,
   "median_ms": 0.11063883546739817,
   "q1_ms": 0.1091509050456807,
   "q3_ms": 0.11332932277582586,
   "tflops": 38.81970808763251
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.556510925292969e-07,
   "rel_err": 9.62258127401583e-07,
   "peak_mb": 184.01611328125,
   "compile_s": 0.00033552106469869614,
   "inner": 174,
   "median_ms": 0.24457935688096558,
   "q1_ms": 0.24175966337815613,
   "q3_ms": 0.24518093974704203,
   "tflops": 17.56062879047605
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1920928955078125e-06,
   "rel_err": 1.6162218798854155e-06,
   "peak_mb": 176.189453125,
   "compile_s": 0.0007928130216896534,
   "inner": 88,
   "median_ms": 0.5086307487958533,
   "q1_ms": 0.5073242090557787,
   "q3_ms": 0.5089380230691114,
   "tflops": 29.554614170669172
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0132789611816406e-06,
   "rel_err": 1.4112234794083633e-06,
   "peak_mb": 280.0009765625,
   "compile_s": 0.0007927222177386284,
   "inner": 70,
   "median_ms": 0.6212083024105856,
   "q1_ms": 0.6174437951163522,
   "q3_ms": 0.6218113786807018,
   "tflops": 24.19862303460393
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.3709068298339844e-06,
   "rel_err": 3.929849867745361e-07,
   "peak_mb": 112.0009765625,
   "compile_s": 0.00014639017172157764,
   "inner": 200,
   "median_ms": 0.09296397445723414,
   "q1_ms": 0.0927384226815775,
   "q3_ms": 0.09322032448835671,
   "tflops": 23.100170367478196
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1920928955078125e-06,
   "rel_err": 3.417260927562893e-07,
   "peak_mb": 185.01611328125,
   "compile_s": 0.000417881878092885,
   "inner": 138,
   "median_ms": 0.28233526045105595,
   "q1_ms": 0.28083140901087417,
   "q3_ms": 0.28238982633701054,
   "tflops": 7.606147544480282
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 2.86102294921875e-06,
   "rel_err": 1.1809877378254896e-06,
   "peak_mb": 176.189453125,
   "compile_s": 0.0007221931591629982,
   "inner": 93,
   "median_ms": 0.44471853893370394,
   "q1_ms": 0.4436302572108245,
   "q3_ms": 0.4452176827697023,
   "tflops": 16.901010661758065
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 512,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.7220458984375e-06,
   "rel_err": 1.1296403954474954e-06,
   "peak_mb": 280.0009765625,
   "compile_s": 0.0009199630003422499,
   "inner": 62,
   "median_ms": 0.647451450687743,
   "q1_ms": 0.6440295837819576,
   "q3_ms": 0.6481156451627612,
   "tflops": 11.608890149239866
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0132789611816406e-06,
   "rel_err": 2.266329829581082e-06,
   "peak_mb": 120.2509765625,
   "compile_s": 0.0002457709051668644,
   "inner": 200,
   "median_ms": 0.19300852087326348,
   "q1_ms": 0.1910482649691403,
   "q3_ms": 0.19716463750228286,
   "tflops": 30.559479826660098
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.9371509552001953e-06,
   "rel_err": 4.332689513830701e-06,
   "peak_mb": 219.25537109375,
   "compile_s": 0.0004335809499025345,
   "inner": 122,
   "median_ms": 0.37018818109006174,
   "q1_ms": 0.3681385777050965,
   "q3_ms": 0.37046592514656607,
   "tflops": 15.933085661006121
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.344650268554688e-07,
   "rel_err": 2.030101541095064e-06,
   "peak_mb": 196.6494140625,
   "compile_s": 0.0010569228325039148,
   "inner": 56,
   "median_ms": 0.8356834462444697,
   "q1_ms": 0.8294014551211149,
   "q3_ms": 0.8410132868448272,
   "tflops": 24.702942355472818
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.6093254089355469e-06,
   "rel_err": 3.709342990987352e-06,
   "peak_mb": 343.1259765625,
   "compile_s": 0.0010863838251680136,
   "inner": 50,
   "median_ms": 0.9533789195120335,
   "q1_ms": 0.952206018846482,
   "q3_ms": 0.9548228303901851,
   "tflops": 21.65334221000618
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.430511474609375e-06,
   "rel_err": 4.378110531888524e-07,
   "peak_mb": 120.2509765625,
   "compile_s": 0.00018307100981473923,
   "inner": 200,
   "median_ms": 0.12817055103369057,
   "q1_ms": 0.12766657047905028,
   "q3_ms": 0.1289932255167514,
   "tflops": 23.0093416640208
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1920928955078125e-06,
   "rel_err": 3.648425490609952e-07,
   "peak_mb": 220.57421875,
   "compile_s": 0.0005673819687217474,
   "inner": 100,
   "median_ms": 0.4257543897256255,
   "q1_ms": 0.42375633027404547,
   "q3_ms": 0.426843436434865,
   "tflops": 6.926810553616465
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.814697265625e-06,
   "rel_err": 1.427154188604618e-06,
   "peak_mb": 196.029296875,
   "compile_s": 0.001164924120530486,
   "inner": 78,
   "median_ms": 0.588014100988706,
   "q1_ms": 0.5856958913425796,
   "q3_ms": 0.5887311537009783,
   "tflops": 17.5538647502575
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 600,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.0531158447265625e-06,
   "rel_err": 1.3379570873439661e-06,
   "peak_mb": 343.1259765625,
   "compile_s": 0.0011359942145645618,
   "inner": 45,
   "median_ms": 0.9994523754964272,
   "q1_ms": 0.993645020450155,
   "q3_ms": 1.0001786001440554,
   "tflops": 10.327575633479395
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1324882507324219e-06,
   "rel_err": 3.026040076292702e-06,
   "peak_mb": 160.0009765625,
   "compile_s": 0.0004651418421417475,
   "inner": 113,
   "median_ms": 0.43520497248829465,
   "q1_ms": 0.4344467388927541,
   "q3_ms": 0.43684479665110065,
   "tflops": 39.47535131727401
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.5367431640625e-07,
   "rel_err": 2.5482443106739083e-06,
   "peak_mb": 448.03173828125,
   "compile_s": 0.0009045230690389872,
   "inner": 57,
   "median_ms": 0.8726218069919891,
   "q1_ms": 0.8643151142454722,
   "q3_ms": 0.8755334032078583,
   "tflops": 19.687645949647592
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1324882507324219e-06,
   "rel_err": 2.0152297111053485e-06,
   "peak_mb": 288.376953125,
   "compile_s": 0.0019387570209801197,
   "inner": 27,
   "median_ms": 1.8505145864630188,
   "q1_ms": 1.8414349459042705,
   "q3_ms": 1.8537079622210175,
   "tflops": 32.49341701160465
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1324882507324219e-06,
   "rel_err": 1.5304817679862026e-06,
   "peak_mb": 752.0009765625,
   "compile_s": 0.0021943680476397276,
   "inner": 23,
   "median_ms": 2.183836743073619,
   "q1_ms": 2.1772180258741844,
   "q3_ms": 2.1960930848170235,
   "tflops": 27.53389983693162
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 2.384185791015625e-06,
   "rel_err": 6.987323217799712e-07,
   "peak_mb": 160.0009765625,
   "compile_s": 0.0003359310794621706,
   "inner": 161,
   "median_ms": 0.29376835377181165,
   "q1_ms": 0.2872070560798697,
   "q3_ms": 0.2966114060999536,
   "tflops": 29.24050355223879
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.2516975402832031e-06,
   "rel_err": 3.668344845664251e-07,
   "peak_mb": 452.03173828125,
   "compile_s": 0.0011320130433887243,
   "inner": 46,
   "median_ms": 1.0269992154739473,
   "q1_ms": 1.0213150956627466,
   "q3_ms": 1.0286602166319347,
   "tflops": 8.364110179028572
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.337860107421875e-06,
   "rel_err": 1.2475003359213588e-06,
   "peak_mb": 288.376953125,
   "compile_s": 0.0015868251211941242,
   "inner": 38,
   "median_ms": 1.2713744194786016,
   "q1_ms": 1.2636369057489854,
   "q3_ms": 1.2757228687405586,
   "tflops": 23.64745633652889
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1024,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.0067901611328125e-06,
   "rel_err": 1.368881839880487e-06,
   "peak_mb": 752.0009765625,
   "compile_s": 0.0023932689800858498,
   "inner": 21,
   "median_ms": 2.3014999127813747,
   "q1_ms": 2.2897825055267838,
   "q3_ms": 2.3107516213453243,
   "tflops": 13.063120665369294
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.8743019104003906e-07,
   "rel_err": 1.7500067315268097e-06,
   "peak_mb": 204.6259765625,
   "compile_s": 0.0009566931985318661,
   "inner": 54,
   "median_ms": 0.9459149635707339,
   "q1_ms": 0.9428555171729789,
   "q3_ms": 0.9465463506920193,
   "tflops": 38.97179072085096
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.917383193969727e-07,
   "rel_err": 2.2211622763279593e-06,
   "peak_mb": 822.65234375,
   "compile_s": 0.001905607059597969,
   "inner": 26,
   "median_ms": 1.9495956545982223,
   "q1_ms": 1.9356859624027631,
   "q3_ms": 1.9510737351643352,
   "tflops": 18.90853619469983
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.0728836059570312e-06,
   "rel_err": 2.3929278540890664e-06,
   "peak_mb": 396.44287109375,
   "compile_s": 0.0038050829898566008,
   "inner": 13,
   "median_ms": 3.951994612669716,
   "q1_ms": 3.9439075715983143,
   "q3_ms": 3.97830311423884,
   "tflops": 32.647817784559074
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.258487701416016e-07,
   "rel_err": 1.3958745057607302e-06,
   "peak_mb": 1422.1337890625,
   "compile_s": 0.0048929378390312195,
   "inner": 10,
   "median_ms": 5.077809700742364,
   "q1_ms": 4.999782459344715,
   "q3_ms": 5.105415813159198,
   "tflops": 25.409380737749387
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1324882507324219e-06,
   "rel_err": 4.1614308088355756e-07,
   "peak_mb": 205.7509765625,
   "compile_s": 0.0006167721003293991,
   "inner": 86,
   "median_ms": 0.5923821753280801,
   "q1_ms": 0.5725708324462175,
   "q3_ms": 0.5936216059422423,
   "tflops": 31.11504830440209
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1920928955078125e-06,
   "rel_err": 4.3804536176139663e-07,
   "peak_mb": 832.36083984375,
   "compile_s": 0.002259668195620179,
   "inner": 22,
   "median_ms": 2.261708776297217,
   "q1_ms": 2.2437905479984526,
   "q3_ms": 2.2648894968865947,
   "tflops": 8.149590342120069
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.0531158447265625e-06,
   "rel_err": 1.161187242360029e-06,
   "peak_mb": 396.44287109375,
   "compile_s": 0.002469097962602973,
   "inner": 20,
   "median_ms": 2.430452441330999,
   "q1_ms": 2.397373819258064,
   "q3_ms": 2.469054877292365,
   "tflops": 26.543206072640128
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 1500,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.344650268554688e-06,
   "rel_err": 3.1262734410120174e-06,
   "peak_mb": 1421.5712890625,
   "compile_s": 0.005177047802135348,
   "inner": 9,
   "median_ms": 5.312195202956597,
   "q1_ms": 5.234366489781275,
   "q3_ms": 5.322719109244645,
   "tflops": 12.144132046219742
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.917383193969727e-07,
   "rel_err": 2.690290102691506e-06,
   "peak_mb": 256.0009765625,
   "compile_s": 0.0016171259339898825,
   "inner": 31,
   "median_ms": 1.6876803861270027,
   "q1_ms": 1.6739008414949623,
   "q3_ms": 1.6979959177514239,
   "tflops": 40.71830027823092
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 3.948807716369629e-07,
   "rel_err": 2.160384610760957e-06,
   "peak_mb": 1408.06298828125,
   "compile_s": 0.0033906218595802784,
   "inner": 14,
   "median_ms": 3.5070329405633465,
   "q1_ms": 3.465040287535105,
   "q3_ms": 3.5205080057494342,
   "tflops": 19.594762267890577
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.364418029785156e-07,
   "rel_err": 2.2752760742150713e-06,
   "peak_mb": 512.751953125,
   "compile_s": 0.0066629331558942795,
   "inner": 7,
   "median_ms": 6.991048576310277,
   "q1_ms": 6.965533647287105,
   "q3_ms": 7.0492681282173315,
   "tflops": 34.40373299522119
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 7.450580596923828e-07,
   "rel_err": 2.6057530249090632e-06,
   "peak_mb": 2464.0009765625,
   "compile_s": 0.00838470901362598,
   "inner": 6,
   "median_ms": 8.705052003885308,
   "q1_ms": 8.695133593088638,
   "q3_ms": 8.719807005642602,
   "tflops": 27.62972219679446
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.4901161193847656e-06,
   "rel_err": 4.2636250441319135e-07,
   "peak_mb": 256.0009765625,
   "compile_s": 0.0009689729195088148,
   "inner": 53,
   "median_ms": 0.9800873582585241,
   "q1_ms": 0.9694677153777964,
   "q3_ms": 0.9857473020740556,
   "tflops": 35.057832425318054
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1920928955078125e-06,
   "rel_err": 3.410899864775274e-07,
   "peak_mb": 1424.06298828125,
   "compile_s": 0.004049513954669237,
   "inner": 12,
   "median_ms": 4.110767661283414,
   "q1_ms": 4.0819592056019856,
   "q3_ms": 4.11451599211432,
   "tflops": 8.358472479875601
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.76837158203125e-06,
   "rel_err": 2.3267602955456823e-06,
   "peak_mb": 512.751953125,
   "compile_s": 0.004017294151708484,
   "inner": 12,
   "median_ms": 4.129275252732138,
   "q1_ms": 4.109684746557226,
   "q3_ms": 4.187713745826235,
   "tflops": 29.123533048185752
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 2048,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.106231689453125e-06,
   "rel_err": 2.5594363251002505e-06,
   "peak_mb": 2464.0009765625,
   "compile_s": 0.008907550945878029,
   "inner": 5,
   "median_ms": 9.144032001495361,
   "q1_ms": 9.13836790714413,
   "q3_ms": 9.159589908085763,
   "tflops": 13.151647355163842
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 2.980232238769531e-07,
   "rel_err": 1.6828230400278699e-06,
   "peak_mb": 345.2509765625,
   "compile_s": 0.003641453105956316,
   "inner": 14,
   "median_ms": 3.6766542720475366,
   "q1_ms": 3.6451612499409487,
   "q3_ms": 3.7093125616333316,
   "tflops": 40.10602822274106
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.023313522338867e-07,
   "rel_err": 2.271811126774992e-06,
   "peak_mb": 2817.26611328125,
   "compile_s": 0.006907572969794273,
   "inner": 7,
   "median_ms": 7.205962269966092,
   "q1_ms": 7.193533005192876,
   "q3_ms": 7.214637356810272,
   "tflops": 20.463054686614928
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 6.705522537231445e-07,
   "rel_err": 2.276585064464598e-06,
   "peak_mb": 722.1337890625,
   "compile_s": 0.01450857101008296,
   "inner": 3,
   "median_ms": 15.2967000224938,
   "q1_ms": 15.27787318142752,
   "q3_ms": 15.377470292150974,
   "tflops": 33.73904170449056
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 7.748603820800781e-07,
   "rel_err": 3.1148110792855732e-06,
   "peak_mb": 4974.1572265625,
   "compile_s": 0.017096309922635555,
   "inner": 2,
   "median_ms": 17.938082572072744,
   "q1_ms": 17.864822526462376,
   "q3_ms": 17.944912484381348,
   "tflops": 28.770968018816806
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1920928955078125e-06,
   "rel_err": 3.2444970088363334e-07,
   "peak_mb": 345.2509765625,
   "compile_s": 0.0018942670430988073,
   "inner": 26,
   "median_ms": 1.9723553067216506,
   "q1_ms": 1.951237003175685,
   "q3_ms": 1.9853480625897646,
   "tflops": 37.3806888387402
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.430511474609375e-06,
   "rel_err": 3.893396467447019e-07,
   "peak_mb": 2851.5986328125,
   "compile_s": 0.008582910057157278,
   "inner": 5,
   "median_ms": 8.817460807040334,
   "q1_ms": 8.738191495649517,
   "q3_ms": 8.838045783340931,
   "tflops": 8.361590894867557
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 5.245208740234375e-06,
   "rel_err": 1.8902719602920115e-06,
   "peak_mb": 722.1337890625,
   "compile_s": 0.008051138138398528,
   "inner": 6,
   "median_ms": 8.352162355246643,
   "q1_ms": 8.33516069299852,
   "q3_ms": 8.367331756744534,
   "tflops": 30.895951135085394
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 3000,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.106231689453125e-06,
   "rel_err": 3.6137550978310173e-06,
   "peak_mb": 4974.1572265625,
   "compile_s": 0.01854202407412231,
   "inner": 2,
   "median_ms": 19.152006483636796,
   "q1_ms": 19.080254016444087,
   "q3_ms": 19.22895445022732,
   "tflops": 13.473679649203993
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.3213367462158203e-07,
   "rel_err": 2.717233428484178e-06,
   "peak_mb": 448.0009765625,
   "compile_s": 0.006133242044597864,
   "inner": 8,
   "median_ms": 6.603000481845811,
   "q1_ms": 6.555150946951471,
   "q3_ms": 6.658499492914416,
   "tflops": 41.6292422967021
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.313225746154785e-07,
   "rel_err": 5.856106781720882e-06,
   "peak_mb": 5056.12548828125,
   "compile_s": 0.012404982931911945,
   "inner": 4,
   "median_ms": 12.972047785297036,
   "q1_ms": 12.899474997539073,
   "q3_ms": 13.077626877930015,
   "tflops": 21.190016525806826
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 8.642673492431641e-07,
   "rel_err": 3.2987438771669986e-06,
   "peak_mb": 961.501953125,
   "compile_s": 0.02545953798107803,
   "inner": 1,
   "median_ms": 28.199238004162908,
   "q1_ms": 27.85786206368357,
   "q3_ms": 28.277418576180935,
   "tflops": 34.116974159442684
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.6689300537109375e-06,
   "rel_err": 6.3699881138745695e-06,
   "peak_mb": 8960.0009765625,
   "compile_s": 0.03075951710343361,
   "inner": 1,
   "median_ms": 32.2564619127661,
   "q1_ms": 32.12913160677999,
   "q3_ms": 32.28197793941945,
   "tflops": 29.825734666927055
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.9073486328125e-06,
   "rel_err": 6.873864322187728e-07,
   "peak_mb": 448.0009765625,
   "compile_s": 0.0033186920918524265,
   "inner": 15,
   "median_ms": 3.5456183987359204,
   "q1_ms": 3.5354570019990206,
   "q3_ms": 3.5929498340313635,
   "tflops": 38.7630415955083
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.1920928955078125e-06,
   "rel_err": 4.2961653434758773e-07,
   "peak_mb": 5120.12548828125,
   "compile_s": 0.015518823871389031,
   "inner": 3,
   "median_ms": 15.868368713806074,
   "q1_ms": 15.804218361154199,
   "q3_ms": 15.904895340402922,
   "tflops": 8.661189814200812
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.049041748046875e-05,
   "rel_err": 2.534422037570039e-06,
   "peak_mb": 961.501953125,
   "compile_s": 0.013820677995681763,
   "inner": 3,
   "median_ms": 14.587153991063436,
   "q1_ms": 14.490672193157177,
   "q3_ms": 14.610684321572384,
   "tflops": 32.976709332519455
  },
  {
   "impl": "torch_math",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 4096,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.059906005859375e-06,
   "rel_err": 3.395472504053032e-06,
   "peak_mb": 8960.0009765625,
   "compile_s": 0.03335608611814678,
   "inner": 1,
   "median_ms": 34.38678989186883,
   "q1_ms": 34.31901440490037,
   "q3_ms": 34.55825534183532,
   "tflops": 13.988986429516842
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 4.3213367462158203e-07,
   "rel_err": 4.1143052840197925e-06,
   "peak_mb": 832.0009765625,
   "compile_s": 0.024457076098769903,
   "inner": 1,
   "median_ms": 26.495682075619698,
   "q1_ms": 26.008145534433424,
   "q3_ms": 26.889564003795385,
   "tflops": 41.49776649032666
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": false,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 9.5367431640625e-07,
   "rel_err": 5.241354301688261e-06,
   "peak_mb": 1859.001953125,
   "compile_s": 0.10603432892821729,
   "inner": 1,
   "median_ms": 109.30961091071367,
   "q1_ms": 108.35058242082596,
   "q3_ms": 119.50461042579263,
   "tflops": 35.20541940597852
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.430511474609375e-06,
   "rel_err": 4.771075623466459e-07,
   "peak_mb": 832.0009765625,
   "compile_s": 0.012361462926492095,
   "inner": 3,
   "median_ms": 13.414806686341763,
   "q1_ms": 13.363354839384556,
   "q3_ms": 13.429690307627123,
   "tflops": 40.98126993121205
  },
  {
   "impl": "torch_efficient",
   "dtype": "float32",
   "head_dim": 128,
   "seq": 8192,
   "causal": true,
   "mode": "fwd_bwd",
   "batch": 4,
   "heads": 8,
   "tag": "h100_nvl",
   "max_abs_err": 1.4066696166992188e-05,
   "rel_err": 4.82915083921398e-06,
   "peak_mb": 1859.001953125,
   "compile_s": 0.05496867117471993,
   "inner": 1,
   "median_ms": 54.8763619735837,
   "q1_ms": 54.627465549856424,
   "q3_ms": 54.950681631453335,
   "tflops": 35.06328188326774
  }
 ]
}