{
  "replay": {
    "runs": {
      "old-forward": {
        "measurements": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.6438799928125,
                "gpu_single_us": 38.4583836421,
                "e2e_batch_us": 33.3962633745,
                "e2e_single_us": 253.375
              },
              "torch": {
                "gpu_batch_us": 15.537109902652446,
                "gpu_single_us": 20.20830288529396,
                "e2e_batch_us": 29.034377532228362,
                "e2e_single_us": 285.58299999999997
              },
              "system": {
                "gpu_batch_us": 15.85937479828125,
                "gpu_single_us": 17.2916334122,
                "e2e_batch_us": 16.1152030217,
                "e2e_single_us": 264.083
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.296875931718752,
                "gpu_single_us": 37.916644942,
                "e2e_batch_us": 32.9410699301,
                "e2e_single_us": 320.292
              },
              "torch": {
                "gpu_batch_us": 15.362630620074924,
                "gpu_single_us": 24.333305191248655,
                "e2e_batch_us": 29.368130824372756,
                "e2e_single_us": 259.75
              },
              "system": {
                "gpu_batch_us": 15.21549438621875,
                "gpu_single_us": 17.0833081938,
                "e2e_batch_us": 16.0693724066,
                "e2e_single_us": 257.75
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2109.0156224125,
                "gpu_single_us": 2226.41666187,
                "e2e_batch_us": 2272.859375,
                "e2e_single_us": 2365.334
              },
              "torch": {
                "gpu_batch_us": 384.4328231786435,
                "gpu_single_us": 361.49995867162943,
                "e2e_batch_us": 403.5765306122449,
                "e2e_single_us": 663.833
              },
              "system": {
                "gpu_batch_us": 480.26875010749995,
                "gpu_single_us": 466.83329856,
                "e2e_batch_us": 495.940625,
                "e2e_single_us": 714.375
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2065.9404856714286,
                "gpu_single_us": 2240.75000733,
                "e2e_batch_us": 2360.45828571,
                "e2e_single_us": 2650.75
              },
              "torch": {
                "gpu_batch_us": 388.55706604764515,
                "gpu_single_us": 374.8750314116478,
                "e2e_batch_us": 407.5009130434783,
                "e2e_single_us": 638.958
              },
              "system": {
                "gpu_batch_us": 484.42414472307695,
                "gpu_single_us": 469.45834765200004,
                "e2e_batch_us": 501.884615385,
                "e2e_single_us": 707.542
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15312.2500051,
                "gpu_single_us": 15771.7499533,
                "e2e_batch_us": 16903.042,
                "e2e_single_us": 16850.834
              },
              "torch": {
                "gpu_batch_us": 2689.351194671222,
                "gpu_single_us": 2485.4583316482604,
                "e2e_batch_us": 2759.708285714286,
                "e2e_single_us": 3574.292
              },
              "system": {
                "gpu_batch_us": 2869.1041613166667,
                "gpu_single_us": 2608.9583407200003,
                "e2e_batch_us": 2982.125,
                "e2e_single_us": 3289.25
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15769.604156950001,
                "gpu_single_us": 15373.875037799999,
                "e2e_batch_us": 15974.1875,
                "e2e_single_us": 15895.958
              },
              "torch": {
                "gpu_batch_us": 2665.6607182563416,
                "gpu_single_us": 2455.500012729317,
                "e2e_batch_us": 2810.3095714285714,
                "e2e_single_us": 3682.9170000000004
              },
              "system": {
                "gpu_batch_us": 2769.0535727285715,
                "gpu_single_us": 2597.12501429,
                "e2e_batch_us": 2854.07142857,
                "e2e_single_us": 3034.583
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 141850.833315,
                "gpu_single_us": 140299.291699,
                "e2e_batch_us": 137221.625,
                "e2e_single_us": 140039.833
              },
              "torch": {
                "gpu_batch_us": 20446.499984245747,
                "gpu_single_us": 20665.000018198043,
                "e2e_batch_us": 20606.625,
                "e2e_single_us": 21046.167
              },
              "system": {
                "gpu_batch_us": 21981.2500291,
                "gpu_single_us": 22244.8749701,
                "e2e_batch_us": 22236.584,
                "e2e_single_us": 22275.958
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 139695.833321,
                "gpu_single_us": 139235.04168599998,
                "e2e_batch_us": 141215.25,
                "e2e_single_us": 141584.833
              },
              "torch": {
                "gpu_batch_us": 20450.458338018507,
                "gpu_single_us": 20675.37500988692,
                "e2e_batch_us": 20900.959,
                "e2e_single_us": 21502.666999999998
              },
              "system": {
                "gpu_batch_us": 22180.7499765,
                "gpu_single_us": 21256.625011999997,
                "e2e_batch_us": 22479.917,
                "e2e_single_us": 23011.625
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 280.19921774062504,
                "gpu_single_us": 284.50001263999997,
                "e2e_batch_us": 284.054924242,
                "e2e_single_us": 492.833
              },
              "torch": {
                "gpu_batch_us": 290.5716155510163,
                "gpu_single_us": 289.4999925047159,
                "e2e_batch_us": 301.849609375,
                "e2e_single_us": 783.2090000000001
              },
              "system": {
                "gpu_batch_us": 280.65820333749997,
                "gpu_single_us": 271.6666786,
                "e2e_batch_us": 288.256835821,
                "e2e_single_us": 486.583
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 279.2402337953125,
                "gpu_single_us": 271.1249399,
                "e2e_batch_us": 292.477057971,
                "e2e_single_us": 497.125
              },
              "torch": {
                "gpu_batch_us": 288.542319140106,
                "gpu_single_us": 289.9999963119626,
                "e2e_batch_us": 319.1115384615385,
                "e2e_single_us": 1262.833
              },
              "system": {
                "gpu_batch_us": 279.3938801921875,
                "gpu_single_us": 269.125041086,
                "e2e_batch_us": 285.891030769,
                "e2e_single_us": 478.541
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 677.3333339607143,
                "gpu_single_us": 722.9582988660001,
                "e2e_batch_us": 692.491071429,
                "e2e_single_us": 956.667
              },
              "torch": {
                "gpu_batch_us": 288.28389733493077,
                "gpu_single_us": 276.583363302052,
                "e2e_batch_us": 303.02330508474574,
                "e2e_single_us": 695.8330000000001
              },
              "system": {
                "gpu_batch_us": 280.78255218125,
                "gpu_single_us": 269.750016741,
                "e2e_batch_us": 284.823227273,
                "e2e_single_us": 520.541
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 279.0690114125,
                "gpu_single_us": 270.541699138,
                "e2e_batch_us": 285.620656716,
                "e2e_single_us": 499
              },
              "torch": {
                "gpu_batch_us": 286.82682204816956,
                "gpu_single_us": 275.49994410946965,
                "e2e_batch_us": 303.18229687499996,
                "e2e_single_us": 573.25
              },
              "system": {
                "gpu_batch_us": 279.30012947812503,
                "gpu_single_us": 269.250012934,
                "e2e_batch_us": 288.423606061,
                "e2e_single_us": 501.416
              }
            }
          }
        ],
        "rejections": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          }
        ],
        "outputs": 36,
        "elements": 170166552
      },
      "new-forward": {
        "measurements": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "b98cc1fee0af0db51a07709382313ed8f0dd1b46f1a8f71a1cb5c5217be5a1ae",
            "paths": {
              "native": {
                "gpu_batch_us": 298.0774734171875,
                "gpu_single_us": 301.958352793,
                "e2e_batch_us": 303.296153846,
                "e2e_single_us": 546.5
              },
              "torch": {
                "gpu_batch_us": 14.721354091307148,
                "gpu_single_us": 23.79168290644884,
                "e2e_batch_us": 29.03822463768116,
                "e2e_single_us": 302.458
              },
              "system": {
                "gpu_batch_us": 14.59179748053125,
                "gpu_single_us": 17.5000168383,
                "e2e_batch_us": 16.6694868687,
                "e2e_single_us": 248.042
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "a3a4bcd315f1e4284cefb6015c9990704da35c56661864e5ea2ff65cff2d5121",
            "paths": {
              "native": {
                "gpu_batch_us": 295.09244814062504,
                "gpu_single_us": 297.875027172,
                "e2e_batch_us": 301.285892308,
                "e2e_single_us": 519.75
              },
              "torch": {
                "gpu_batch_us": 18.373697457718663,
                "gpu_single_us": 25.12504579499364,
                "e2e_batch_us": 29.498255813953488,
                "e2e_single_us": 289.625
              },
              "system": {
                "gpu_batch_us": 41.9459638578125,
                "gpu_single_us": 29.5416684821,
                "e2e_batch_us": 15.6981261044,
                "e2e_single_us": 227.875
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "a3a4bcd315f1e4284cefb6015c9990704da35c56661864e5ea2ff65cff2d5121",
            "paths": {
              "native": {
                "gpu_batch_us": 293.9023443109375,
                "gpu_single_us": 297.91670385699996,
                "e2e_batch_us": 301.022727273,
                "e2e_single_us": 517
              },
              "torch": {
                "gpu_batch_us": 14.694010133098345,
                "gpu_single_us": 20.16668440774083,
                "e2e_batch_us": 29.199029126213592,
                "e2e_single_us": 313.916
              },
              "system": {
                "gpu_batch_us": 14.81250001234375,
                "gpu_single_us": 16.9999548234,
                "e2e_batch_us": 16.1064167508,
                "e2e_single_us": 239
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "a3a4bcd315f1e4284cefb6015c9990704da35c56661864e5ea2ff65cff2d5121",
            "paths": {
              "native": {
                "gpu_batch_us": 295.137369903125,
                "gpu_single_us": 298.49994462,
                "e2e_batch_us": 301.5611875,
                "e2e_single_us": 531
              },
              "torch": {
                "gpu_batch_us": 12.98372353630839,
                "gpu_single_us": 20.33333294093609,
                "e2e_batch_us": 31.27985996705107,
                "e2e_single_us": 309.458
              },
              "system": {
                "gpu_batch_us": 13.74218754790625,
                "gpu_single_us": 17.2083382495,
                "e2e_batch_us": 16.3465167785,
                "e2e_single_us": 272.125
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "acb1daf6a7178124176490e0436dd3a3eb439c11b777a267a95acb6adc1b9198",
            "paths": {
              "native": {
                "gpu_batch_us": 3772.4666530199997,
                "gpu_single_us": 3712.66662842,
                "e2e_batch_us": 3871.9832,
                "e2e_single_us": 3986.833
              },
              "torch": {
                "gpu_batch_us": 384.0755210452092,
                "gpu_single_us": 370.9582961164415,
                "e2e_batch_us": 404.0946041666666,
                "e2e_single_us": 636.0
              },
              "system": {
                "gpu_batch_us": 487.34294996923074,
                "gpu_single_us": 468.08330807799996,
                "e2e_batch_us": 499.44125641,
                "e2e_single_us": 670.792
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "acb1daf6a7178124176490e0436dd3a3eb439c11b777a267a95acb6adc1b9198",
            "paths": {
              "native": {
                "gpu_batch_us": 3779.1499984399998,
                "gpu_single_us": 3710.5000228599997,
                "e2e_batch_us": 3848.5334,
                "e2e_single_us": 4110.708
              },
              "torch": {
                "gpu_batch_us": 386.45920358248986,
                "gpu_single_us": 382.25000025704503,
                "e2e_batch_us": 413.0842083333334,
                "e2e_single_us": 650.208
              },
              "system": {
                "gpu_batch_us": 486.492709025,
                "gpu_single_us": 468.000012916,
                "e2e_batch_us": 499.242725,
                "e2e_single_us": 760.833
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "12a7ebdf4ccc7b30c32eff03f1a8177fc1cf1dd15160901272710b403c52471f",
            "paths": {
              "native": {
                "gpu_batch_us": 4642.1562438,
                "gpu_single_us": 4367.41660815,
                "e2e_batch_us": 4759.7085,
                "e2e_single_us": 4848.208
              },
              "torch": {
                "gpu_batch_us": 387.5772575459753,
                "gpu_single_us": 379.333330783993,
                "e2e_batch_us": 406.42795833333327,
                "e2e_single_us": 659.5830000000001
              },
              "system": {
                "gpu_batch_us": 480.8365379282052,
                "gpu_single_us": 468.541693408,
                "e2e_batch_us": 490.604692308,
                "e2e_single_us": 682.542
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "acb1daf6a7178124176490e0436dd3a3eb439c11b777a267a95acb6adc1b9198",
            "paths": {
              "native": {
                "gpu_batch_us": 3762.3749958500002,
                "gpu_single_us": 3738.3333547,
                "e2e_batch_us": 3885.8125,
                "e2e_single_us": 4038.084
              },
              "torch": {
                "gpu_batch_us": 382.9496514905865,
                "gpu_single_us": 369.208340998739,
                "e2e_batch_us": 404.28820833333333,
                "e2e_single_us": 651.2090000000001
              },
              "system": {
                "gpu_batch_us": 482.8643167205128,
                "gpu_single_us": 470.875005703,
                "e2e_batch_us": 499.682692308,
                "e2e_single_us": 706
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "8bebfd4aad6def6763a60e05b848663c49d8becfd6b288e18b890406b7f5c0c4",
            "paths": {
              "native": {
                "gpu_batch_us": 10586.10415205,
                "gpu_single_us": 10693.125019300001,
                "e2e_batch_us": 11239.0415,
                "e2e_single_us": 11180.959
              },
              "torch": {
                "gpu_batch_us": 2574.5277816895396,
                "gpu_single_us": 2446.5416790917516,
                "e2e_batch_us": 2770.1806666666666,
                "e2e_single_us": 2944.209
              },
              "system": {
                "gpu_batch_us": 2799.54167005,
                "gpu_single_us": 2616.2500144,
                "e2e_batch_us": 2898.90983333,
                "e2e_single_us": 3174.958
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "e52a83a1f1f76a0899f66da0318074300ab08897414961667365224692b38199",
            "paths": {
              "native": {
                "gpu_batch_us": 20967.2916681,
                "gpu_single_us": 21727.1666625,
                "e2e_batch_us": 21842.417,
                "e2e_single_us": 21840.208
              },
              "torch": {
                "gpu_batch_us": 2655.229162580023,
                "gpu_single_us": 2495.416672900319,
                "e2e_batch_us": 2761.465166666667,
                "e2e_single_us": 2904.541
              },
              "system": {
                "gpu_batch_us": 2796.410716,
                "gpu_single_us": 2626.8333313099997,
                "e2e_batch_us": 2864.97614286,
                "e2e_single_us": 2992.25
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "7fad2e25424286e493540a6f1dd7ba3f48d5aa2fbd0fcacfb9676a1c5f704e10",
            "paths": {
              "native": {
                "gpu_batch_us": 21876.041661,
                "gpu_single_us": 22245.5416806,
                "e2e_batch_us": 22702.583,
                "e2e_single_us": 22695.417
              },
              "torch": {
                "gpu_batch_us": 2598.315485686596,
                "gpu_single_us": 2442.45829526335,
                "e2e_batch_us": 2803.244142857143,
                "e2e_single_us": 2894.875
              },
              "system": {
                "gpu_batch_us": 2789.451381733333,
                "gpu_single_us": 2605.6249626000003,
                "e2e_batch_us": 2912.95833333,
                "e2e_single_us": 3167.834
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "8bebfd4aad6def6763a60e05b848663c49d8becfd6b288e18b890406b7f5c0c4",
            "paths": {
              "native": {
                "gpu_batch_us": 10573.68749755,
                "gpu_single_us": 10843.9583564,
                "e2e_batch_us": 11267,
                "e2e_single_us": 11445.833
              },
              "torch": {
                "gpu_batch_us": 2633.8571499634004,
                "gpu_single_us": 2452.8750218451023,
                "e2e_batch_us": 2744.5832857142855,
                "e2e_single_us": 3448.291
              },
              "system": {
                "gpu_batch_us": 2774.2291664833333,
                "gpu_single_us": 2640.16666915,
                "e2e_batch_us": 2868.35416667,
                "e2e_single_us": 3014
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "2476bca9ed7bc9b4795408a6b0043fe82839646e8fa5bab13eb77ed40e292d7d",
            "paths": {
              "native": {
                "gpu_batch_us": 83534.2082428,
                "gpu_single_us": 82770.9999867,
                "e2e_batch_us": 84604.125,
                "e2e_single_us": 84026.333
              },
              "torch": {
                "gpu_batch_us": 20340.87502397597,
                "gpu_single_us": 21259.874978568405,
                "e2e_batch_us": 21434.417,
                "e2e_single_us": 21808.334
              },
              "system": {
                "gpu_batch_us": 21750.166663,
                "gpu_single_us": 21990.375011199998,
                "e2e_batch_us": 23182.25,
                "e2e_single_us": 22302.625
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "cf4bb653701938496bfc5baf33c9862ac66b2c1a5899007948abea51eea616d4",
            "paths": {
              "native": {
                "gpu_batch_us": 51947.458356200004,
                "gpu_single_us": 52449.0000098,
                "e2e_batch_us": 53474,
                "e2e_single_us": 53111.375
              },
              "torch": {
                "gpu_batch_us": 20338.708360213786,
                "gpu_single_us": 20546.58334236592,
                "e2e_batch_us": 20608.875,
                "e2e_single_us": 20838.041
              },
              "system": {
                "gpu_batch_us": 21697.5833173,
                "gpu_single_us": 22195.5000306,
                "e2e_batch_us": 22853.417,
                "e2e_single_us": 23068.458
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "c810e89b927b3df81973d20a3212484c0ed43116532762fd3ce4f53b1086f346",
            "paths": {
              "native": {
                "gpu_batch_us": 53998.791670900006,
                "gpu_single_us": 53099.7500173,
                "e2e_batch_us": 53017.125,
                "e2e_single_us": 53747.875
              },
              "torch": {
                "gpu_batch_us": 20215.083321090788,
                "gpu_single_us": 20343.750016763806,
                "e2e_batch_us": 21680.459,
                "e2e_single_us": 20401.833
              },
              "system": {
                "gpu_batch_us": 21936.8333346,
                "gpu_single_us": 22444.9166562,
                "e2e_batch_us": 22557.416,
                "e2e_single_us": 22720.25
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "cf4bb653701938496bfc5baf33c9862ac66b2c1a5899007948abea51eea616d4",
            "paths": {
              "native": {
                "gpu_batch_us": 52429.3750059,
                "gpu_single_us": 53069.333313,
                "e2e_batch_us": 53076.208,
                "e2e_single_us": 53793.791
              },
              "torch": {
                "gpu_batch_us": 20634.41666541621,
                "gpu_single_us": 20720.375003293157,
                "e2e_batch_us": 21483.666,
                "e2e_single_us": 21081.0
              },
              "system": {
                "gpu_batch_us": 21778.7500005,
                "gpu_single_us": 21760.458359499997,
                "e2e_batch_us": 23073.416,
                "e2e_single_us": 22506.458
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 278.7363291640625,
                "gpu_single_us": 271.00002626,
                "e2e_batch_us": 284.757855072,
                "e2e_single_us": 491.209
              },
              "torch": {
                "gpu_batch_us": 286.4576817955822,
                "gpu_single_us": 277.25001564249396,
                "e2e_batch_us": 301.29948437499996,
                "e2e_single_us": 569.75
              },
              "system": {
                "gpu_batch_us": 280.039063,
                "gpu_single_us": 271.999975666,
                "e2e_batch_us": 291.486914286,
                "e2e_single_us": 476.125
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 279.71549479843753,
                "gpu_single_us": 269.49995663,
                "e2e_batch_us": 289.542857143,
                "e2e_single_us": 505.333
              },
              "torch": {
                "gpu_batch_us": 287.0579410227947,
                "gpu_single_us": 294.8333276435733,
                "e2e_batch_us": 302.537109375,
                "e2e_single_us": 539.9580000000001
              },
              "system": {
                "gpu_batch_us": 281.339843240625,
                "gpu_single_us": 270.791642834,
                "e2e_batch_us": 289.413552239,
                "e2e_single_us": 527.625
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 675.3994244275863,
                "gpu_single_us": 669.499975629,
                "e2e_batch_us": 689.709758621,
                "e2e_single_us": 982.917
              },
              "torch": {
                "gpu_batch_us": 287.20638056256576,
                "gpu_single_us": 279.8750065267086,
                "e2e_batch_us": 301.9397384615385,
                "e2e_single_us": 540.875
              },
              "system": {
                "gpu_batch_us": 280.3144534484375,
                "gpu_single_us": 270.000018645,
                "e2e_batch_us": 290.019283582,
                "e2e_single_us": 550.5
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 281.4121098703125,
                "gpu_single_us": 269.166601356,
                "e2e_batch_us": 287.75065625,
                "e2e_single_us": 484.625
              },
              "torch": {
                "gpu_batch_us": 285.9960941350437,
                "gpu_single_us": 276.00000612437725,
                "e2e_batch_us": 301.2416769230769,
                "e2e_single_us": 601.25
              },
              "system": {
                "gpu_batch_us": 283.34221264426225,
                "gpu_single_us": 269.791635219,
                "e2e_batch_us": 288.120901639,
                "e2e_single_us": 502.083
              }
            }
          }
        ],
        "rejections": [],
        "outputs": 60,
        "elements": 327750192
      },
      "new-reverse": {
        "measurements": [
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 280.432943328125,
                "gpu_single_us": 270.16666717799995,
                "e2e_batch_us": 282.164855072,
                "e2e_single_us": 527.709
              },
              "torch": {
                "gpu_batch_us": 287.1438773581758,
                "gpu_single_us": 277.9166679829359,
                "e2e_batch_us": 300.5506461538462,
                "e2e_single_us": 532.791
              },
              "system": {
                "gpu_batch_us": 280.033203125,
                "gpu_single_us": 272.791658062,
                "e2e_batch_us": 287.578358209,
                "e2e_single_us": 490.667
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 279.54036522656247,
                "gpu_single_us": 269.625044893,
                "e2e_batch_us": 283.185985507,
                "e2e_single_us": 506.291
              },
              "torch": {
                "gpu_batch_us": 286.7923194571631,
                "gpu_single_us": 285.25001835078,
                "e2e_batch_us": 302.40234375,
                "e2e_single_us": 534.875
              },
              "system": {
                "gpu_batch_us": 282.8925780703125,
                "gpu_single_us": 270.666670986,
                "e2e_batch_us": 287.028985507,
                "e2e_single_us": 490.958
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 675.4598226892857,
                "gpu_single_us": 668.916676659,
                "e2e_batch_us": 693.068464286,
                "e2e_single_us": 908.333
              },
              "torch": {
                "gpu_batch_us": 287.8899722418282,
                "gpu_single_us": 286.5000278688967,
                "e2e_batch_us": 301.9889375,
                "e2e_single_us": 543.0
              },
              "system": {
                "gpu_batch_us": 280.5644535328125,
                "gpu_single_us": 271.291646641,
                "e2e_batch_us": 288.508212121,
                "e2e_single_us": 534.125
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 278.8203128140625,
                "gpu_single_us": 269.41666146700004,
                "e2e_batch_us": 283.556768116,
                "e2e_single_us": 512.625
              },
              "torch": {
                "gpu_batch_us": 287.73763096978655,
                "gpu_single_us": 278.45834847539663,
                "e2e_batch_us": 300.7108923076923,
                "e2e_single_us": 687.5
              },
              "system": {
                "gpu_batch_us": 279.83919244375,
                "gpu_single_us": 269.458338153,
                "e2e_batch_us": 288.134191176,
                "e2e_single_us": 505.666
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "cf4bb653701938496bfc5baf33c9862ac66b2c1a5899007948abea51eea616d4",
            "paths": {
              "native": {
                "gpu_batch_us": 52085.7083029,
                "gpu_single_us": 52768.5416746,
                "e2e_batch_us": 53388.541,
                "e2e_single_us": 53388.25
              },
              "torch": {
                "gpu_batch_us": 19617.749960161746,
                "gpu_single_us": 21159.54167675227,
                "e2e_batch_us": 20496.791,
                "e2e_single_us": 21300.583
              },
              "system": {
                "gpu_batch_us": 22063.000011300002,
                "gpu_single_us": 21731.0416629,
                "e2e_batch_us": 22611,
                "e2e_single_us": 22561.625
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "2476bca9ed7bc9b4795408a6b0043fe82839646e8fa5bab13eb77ed40e292d7d",
            "paths": {
              "native": {
                "gpu_batch_us": 82694.7083697,
                "gpu_single_us": 82872.2916194,
                "e2e_batch_us": 83593.125,
                "e2e_single_us": 82843.25
              },
              "torch": {
                "gpu_batch_us": 20444.875000976026,
                "gpu_single_us": 20516.499993391335,
                "e2e_batch_us": 20629.083000000002,
                "e2e_single_us": 21296.084
              },
              "system": {
                "gpu_batch_us": 21327.2083201,
                "gpu_single_us": 21834.541636100003,
                "e2e_batch_us": 22905.333,
                "e2e_single_us": 22420.167
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "c810e89b927b3df81973d20a3212484c0ed43116532762fd3ce4f53b1086f346",
            "paths": {
              "native": {
                "gpu_batch_us": 53066.3750251,
                "gpu_single_us": 53557.3333618,
                "e2e_batch_us": 53543.375,
                "e2e_single_us": 53785.917
              },
              "torch": {
                "gpu_batch_us": 20246.29170773551,
                "gpu_single_us": 20740.500011015683,
                "e2e_batch_us": 21056.084,
                "e2e_single_us": 20465.416999999998
              },
              "system": {
                "gpu_batch_us": 21868.9999892,
                "gpu_single_us": 21881.875,
                "e2e_batch_us": 23244.708,
                "e2e_single_us": 22089.042
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "cf4bb653701938496bfc5baf33c9862ac66b2c1a5899007948abea51eea616d4",
            "paths": {
              "native": {
                "gpu_batch_us": 51361.5000527,
                "gpu_single_us": 53130.8749923,
                "e2e_batch_us": 53494.5,
                "e2e_single_us": 53262.125
              },
              "torch": {
                "gpu_batch_us": 20354.458305519074,
                "gpu_single_us": 20520.08337341249,
                "e2e_batch_us": 21599.959,
                "e2e_single_us": 21596.5
              },
              "system": {
                "gpu_batch_us": 22038.333292599997,
                "gpu_single_us": 21820.1249954,
                "e2e_batch_us": 23031.417,
                "e2e_single_us": 22206.125
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "e52a83a1f1f76a0899f66da0318074300ab08897414961667365224692b38199",
            "paths": {
              "native": {
                "gpu_batch_us": 21885.9165907,
                "gpu_single_us": 21653.7083616,
                "e2e_batch_us": 21749.667,
                "e2e_single_us": 21855.958
              },
              "torch": {
                "gpu_batch_us": 2754.6527853701264,
                "gpu_single_us": 2466.375008225441,
                "e2e_batch_us": 2771.5555,
                "e2e_single_us": 3100.0
              },
              "system": {
                "gpu_batch_us": 2813.805554366667,
                "gpu_single_us": 2606.29167315,
                "e2e_batch_us": 3027.84033333,
                "e2e_single_us": 3226.792
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "8bebfd4aad6def6763a60e05b848663c49d8becfd6b288e18b890406b7f5c0c4",
            "paths": {
              "native": {
                "gpu_batch_us": 10649.93749605,
                "gpu_single_us": 11049.708293299998,
                "e2e_batch_us": 11314.917,
                "e2e_single_us": 11321.959
              },
              "torch": {
                "gpu_batch_us": 2682.845241257123,
                "gpu_single_us": 2473.8750071264803,
                "e2e_batch_us": 2795.964285714286,
                "e2e_single_us": 2948.375
              },
              "system": {
                "gpu_batch_us": 2831.319444033333,
                "gpu_single_us": 2704.3749578300003,
                "e2e_batch_us": 2998.22916667,
                "e2e_single_us": 2875.375
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "7fad2e25424286e493540a6f1dd7ba3f48d5aa2fbd0fcacfb9676a1c5f704e10",
            "paths": {
              "native": {
                "gpu_batch_us": 21375.2916316,
                "gpu_single_us": 22570.0832671,
                "e2e_batch_us": 22158.875,
                "e2e_single_us": 22790.208
              },
              "torch": {
                "gpu_batch_us": 2701.5624994722507,
                "gpu_single_us": 2515.1250301860273,
                "e2e_batch_us": 2773.7569999999996,
                "e2e_single_us": 2972.166
              },
              "system": {
                "gpu_batch_us": 2750.3124923333335,
                "gpu_single_us": 2622.8750357400004,
                "e2e_batch_us": 2855.22916667,
                "e2e_single_us": 2980.291
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "8bebfd4aad6def6763a60e05b848663c49d8becfd6b288e18b890406b7f5c0c4",
            "paths": {
              "native": {
                "gpu_batch_us": 10570.81250475,
                "gpu_single_us": 10540.5000504,
                "e2e_batch_us": 10880.6875,
                "e2e_single_us": 11012.125
              },
              "torch": {
                "gpu_batch_us": 2580.9523898975126,
                "gpu_single_us": 2517.708344385028,
                "e2e_batch_us": 2682.0774285714288,
                "e2e_single_us": 2934.9159999999997
              },
              "system": {
                "gpu_batch_us": 2807.2500038666667,
                "gpu_single_us": 2621.12502242,
                "e2e_batch_us": 2883.83333333,
                "e2e_single_us": 2892.083
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "acb1daf6a7178124176490e0436dd3a3eb439c11b777a267a95acb6adc1b9198",
            "paths": {
              "native": {
                "gpu_batch_us": 3801.9333384,
                "gpu_single_us": 3727.7083029,
                "e2e_batch_us": 3887.725,
                "e2e_single_us": 3986.458
              },
              "torch": {
                "gpu_batch_us": 388.3519502198125,
                "gpu_single_us": 371.375004760921,
                "e2e_batch_us": 408.2659574468085,
                "e2e_single_us": 680.7090000000001
              },
              "system": {
                "gpu_batch_us": 487.12499992749997,
                "gpu_single_us": 465.58334725,
                "e2e_batch_us": 502.69165,
                "e2e_single_us": 687.916
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "acb1daf6a7178124176490e0436dd3a3eb439c11b777a267a95acb6adc1b9198",
            "paths": {
              "native": {
                "gpu_batch_us": 3815.48332516,
                "gpu_single_us": 3722.25005412,
                "e2e_batch_us": 3857.8332,
                "e2e_single_us": 3998.958
              },
              "torch": {
                "gpu_batch_us": 396.6269848336066,
                "gpu_single_us": 388.2916644215584,
                "e2e_batch_us": 436.2400714285714,
                "e2e_single_us": 628.3330000000001
              },
              "system": {
                "gpu_batch_us": 487.73245591842107,
                "gpu_single_us": 469.041639008,
                "e2e_batch_us": 510.108552632,
                "e2e_single_us": 719.875
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "12a7ebdf4ccc7b30c32eff03f1a8177fc1cf1dd15160901272710b403c52471f",
            "paths": {
              "native": {
                "gpu_batch_us": 4625.937508525,
                "gpu_single_us": 4359.6249888699995,
                "e2e_batch_us": 4766.802,
                "e2e_single_us": 4820.333
              },
              "torch": {
                "gpu_batch_us": 388.21916677989066,
                "gpu_single_us": 372.12501047179103,
                "e2e_batch_us": 409.73084,
                "e2e_single_us": 742.459
              },
              "system": {
                "gpu_batch_us": 493.1875002131579,
                "gpu_single_us": 470.16667667800004,
                "e2e_batch_us": 499.561394737,
                "e2e_single_us": 708.166
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "acb1daf6a7178124176490e0436dd3a3eb439c11b777a267a95acb6adc1b9198",
            "paths": {
              "native": {
                "gpu_batch_us": 3801.489569025,
                "gpu_single_us": 3737.7499975299997,
                "e2e_batch_us": 4139.875,
                "e2e_single_us": 4255.459
              },
              "torch": {
                "gpu_batch_us": 392.48437678907067,
                "gpu_single_us": 388.7083730660379,
                "e2e_batch_us": 408.5138958333334,
                "e2e_single_us": 718.542
              },
              "system": {
                "gpu_batch_us": 486.92948752051285,
                "gpu_single_us": 474.50000420199996,
                "e2e_batch_us": 500.050205128,
                "e2e_single_us": 718.292
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "a3a4bcd315f1e4284cefb6015c9990704da35c56661864e5ea2ff65cff2d5121",
            "paths": {
              "native": {
                "gpu_batch_us": 296.35546798125,
                "gpu_single_us": 297.833292279,
                "e2e_batch_us": 307.1532,
                "e2e_single_us": 531.625
              },
              "torch": {
                "gpu_batch_us": 10.921224202320445,
                "gpu_single_us": 17.37498678267002,
                "e2e_batch_us": 29.13802931596091,
                "e2e_single_us": 304.58299999999997
              },
              "system": {
                "gpu_batch_us": 13.92252579534375,
                "gpu_single_us": 17.2500149347,
                "e2e_batch_us": 16.412807377,
                "e2e_single_us": 269.167
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "b98cc1fee0af0db51a07709382313ed8f0dd1b46f1a8f71a1cb5c5217be5a1ae",
            "paths": {
              "native": {
                "gpu_batch_us": 298.6917979936508,
                "gpu_single_us": 301.749969367,
                "e2e_batch_us": 303.136904762,
                "e2e_single_us": 546.084
              },
              "torch": {
                "gpu_batch_us": 15.224609342112672,
                "gpu_single_us": 20.37500962615013,
                "e2e_batch_us": 29.497794117647057,
                "e2e_single_us": 292.834
              },
              "system": {
                "gpu_batch_us": 13.833984667140625,
                "gpu_single_us": 17.2500149347,
                "e2e_batch_us": 16.0652159848,
                "e2e_single_us": 305.917
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "a3a4bcd315f1e4284cefb6015c9990704da35c56661864e5ea2ff65cff2d5121",
            "paths": {
              "native": {
                "gpu_batch_us": 294.7994789796875,
                "gpu_single_us": 297.33334668000003,
                "e2e_batch_us": 301.8111875,
                "e2e_single_us": 572.917
              },
              "torch": {
                "gpu_batch_us": 14.109374205872882,
                "gpu_single_us": 20.249979570508003,
                "e2e_batch_us": 28.818674074074075,
                "e2e_single_us": 327.708
              },
              "system": {
                "gpu_batch_us": 13.7740880745625,
                "gpu_single_us": 17.0000130311,
                "e2e_batch_us": 16.0088461538,
                "e2e_single_us": 333.833
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "a3a4bcd315f1e4284cefb6015c9990704da35c56661864e5ea2ff65cff2d5121",
            "paths": {
              "native": {
                "gpu_batch_us": 294.50781130468755,
                "gpu_single_us": 297.249993309,
                "e2e_batch_us": 300.512359375,
                "e2e_single_us": 544.417
              },
              "torch": {
                "gpu_batch_us": 12.917316780658439,
                "gpu_single_us": 18.624996300786734,
                "e2e_batch_us": 29.612810283687942,
                "e2e_single_us": 296.58299999999997
              },
              "system": {
                "gpu_batch_us": 14.009765436640624,
                "gpu_single_us": 18.124992493500002,
                "e2e_batch_us": 16.4395667939,
                "e2e_single_us": 260.917
              }
            }
          }
        ],
        "rejections": [],
        "outputs": 60,
        "elements": 327750192
      },
      "old-reverse": {
        "measurements": [
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 278.7128905765625,
                "gpu_single_us": 270.291697234,
                "e2e_batch_us": 283.585144928,
                "e2e_single_us": 497.583
              },
              "torch": {
                "gpu_batch_us": 289.7636722991592,
                "gpu_single_us": 289.54166918992996,
                "e2e_batch_us": 301.5057692307692,
                "e2e_single_us": 637.792
              },
              "system": {
                "gpu_batch_us": 281.89583372343753,
                "gpu_single_us": 269.791635219,
                "e2e_batch_us": 294.537434783,
                "e2e_single_us": 534.875
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 289.6503910953125,
                "gpu_single_us": 295.749981888,
                "e2e_batch_us": 283.423507463,
                "e2e_single_us": 486.584
              },
              "torch": {
                "gpu_batch_us": 286.4082052838057,
                "gpu_single_us": 276.9583370536566,
                "e2e_batch_us": 303.25384615384615,
                "e2e_single_us": 565.041
              },
              "system": {
                "gpu_batch_us": 280.45312501562495,
                "gpu_single_us": 271.749973763,
                "e2e_batch_us": 288.945043478,
                "e2e_single_us": 485.291
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 984.7181328764706,
                "gpu_single_us": 983.708247077,
                "e2e_batch_us": 975.549,
                "e2e_single_us": 1203.209
              },
              "torch": {
                "gpu_batch_us": 331.6550569028639,
                "gpu_single_us": 363.49997390061617,
                "e2e_batch_us": 337.2247213114754,
                "e2e_single_us": 629.417
              },
              "system": {
                "gpu_batch_us": 369.73636385818185,
                "gpu_single_us": 357.749988325,
                "e2e_batch_us": 366.296218182,
                "e2e_single_us": 572.833
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 431.0918567090909,
                "gpu_single_us": 442.000047769,
                "e2e_batch_us": 446.060590909,
                "e2e_single_us": 623.042
              },
              "torch": {
                "gpu_batch_us": 460.45528213697963,
                "gpu_single_us": 365.9166395664215,
                "e2e_batch_us": 487.96546341463414,
                "e2e_single_us": 785.541
              },
              "system": {
                "gpu_batch_us": 429.5460530552632,
                "gpu_single_us": 359.50000165,
                "e2e_batch_us": 435.572368421,
                "e2e_single_us": 586.666
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 140827.08332,
                "gpu_single_us": 140191.541635,
                "e2e_batch_us": 142438.125,
                "e2e_single_us": 142240.417
              },
              "torch": {
                "gpu_batch_us": 20688.750024419278,
                "gpu_single_us": 19569.875032175332,
                "e2e_batch_us": 21128.917,
                "e2e_single_us": 20689.834000000003
              },
              "system": {
                "gpu_batch_us": 22633.916698399997,
                "gpu_single_us": 22586.2499792,
                "e2e_batch_us": 22937.708,
                "e2e_single_us": 23072.791
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 141526.583408,
                "gpu_single_us": 142312.249984,
                "e2e_batch_us": 136774.167,
                "e2e_single_us": 139589.708
              },
              "torch": {
                "gpu_batch_us": 19698.833348229527,
                "gpu_single_us": 20099.416666198522,
                "e2e_batch_us": 20885.0,
                "e2e_single_us": 20655.083000000002
              },
              "system": {
                "gpu_batch_us": 21683.4166786,
                "gpu_single_us": 21650.7916921,
                "e2e_batch_us": 23502.875,
                "e2e_single_us": 23072.084
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15367.625019300001,
                "gpu_single_us": 14765.250030899999,
                "e2e_batch_us": 16244.583,
                "e2e_single_us": 16042.417
              },
              "torch": {
                "gpu_batch_us": 2623.0654745761835,
                "gpu_single_us": 2452.8333451598883,
                "e2e_batch_us": 2754.9582857142855,
                "e2e_single_us": 2881.708
              },
              "system": {
                "gpu_batch_us": 2803.8333306000004,
                "gpu_single_us": 2593.91666623,
                "e2e_batch_us": 2869.6875,
                "e2e_single_us": 2986.875
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15406.64583445,
                "gpu_single_us": 14874.833286699999,
                "e2e_batch_us": 15834.458,
                "e2e_single_us": 15818.75
              },
              "torch": {
                "gpu_batch_us": 2617.0000007065632,
                "gpu_single_us": 2455.374982673675,
                "e2e_batch_us": 2728.4513333333334,
                "e2e_single_us": 2963.708
              },
              "system": {
                "gpu_batch_us": 2815.965272016667,
                "gpu_single_us": 2625.8333819,
                "e2e_batch_us": 2900.13883333,
                "e2e_single_us": 3076.375
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2083.3072922,
                "gpu_single_us": 2194.70838783,
                "e2e_batch_us": 2137.302125,
                "e2e_single_us": 2334.25
              },
              "torch": {
                "gpu_batch_us": 385.1718768904296,
                "gpu_single_us": 377.2500203922391,
                "e2e_batch_us": 404.4548750000001,
                "e2e_single_us": 661.417
              },
              "system": {
                "gpu_batch_us": 489.32500067250004,
                "gpu_single_us": 465.458317194,
                "e2e_batch_us": 493.803125,
                "e2e_single_us": 696.459
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2080.5694447444444,
                "gpu_single_us": 2236.54159345,
                "e2e_batch_us": 2207.412,
                "e2e_single_us": 2332.416
              },
              "torch": {
                "gpu_batch_us": 383.73554648109234,
                "gpu_single_us": 372.04165710136294,
                "e2e_batch_us": 400.81632653061223,
                "e2e_single_us": 659.208
              },
              "system": {
                "gpu_batch_us": 486.87286340769225,
                "gpu_single_us": 467.12497714899996,
                "e2e_batch_us": 502.369666667,
                "e2e_single_us": 705.792
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.3365890178125,
                "gpu_single_us": 37.333404179700004,
                "e2e_batch_us": 33.5884878049,
                "e2e_single_us": 262.5
              },
              "torch": {
                "gpu_batch_us": 15.501953384955414,
                "gpu_single_us": 19.875005818903446,
                "e2e_batch_us": 29.557682196339435,
                "e2e_single_us": 287.125
              },
              "system": {
                "gpu_batch_us": 13.701822354046875,
                "gpu_single_us": 17.0000130311,
                "e2e_batch_us": 15.690038014,
                "e2e_single_us": 302.125
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.014974410937498,
                "gpu_single_us": 38.708385545800006,
                "e2e_batch_us": 32.8038537415,
                "e2e_single_us": 255.959
              },
              "torch": {
                "gpu_batch_us": 17.38086029945407,
                "gpu_single_us": 31.25000512227416,
                "e2e_batch_us": 30.122620626151015,
                "e2e_single_us": 282.375
              },
              "system": {
                "gpu_batch_us": 13.6289063448125,
                "gpu_single_us": 19.7499757633,
                "e2e_batch_us": 16.1658752969,
                "e2e_single_us": 250.042
              }
            }
          }
        ],
        "rejections": [
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          }
        ],
        "outputs": 36,
        "elements": 170166552
      }
    },
    "comparisons": [
      {
        "shape": [
          129,
          257,
          61
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 31.296875931718752,
            "gpu_single_us": 37.916644942,
            "e2e_batch_us": 32.9410699301,
            "e2e_single_us": 320.292
          },
          "torch": {
            "gpu_batch_us": 15.362630620074924,
            "gpu_single_us": 24.333305191248655,
            "e2e_batch_us": 29.368130824372756,
            "e2e_single_us": 259.75
          },
          "system": {
            "gpu_batch_us": 15.21549438621875,
            "gpu_single_us": 17.0833081938,
            "e2e_batch_us": 16.0693724066,
            "e2e_single_us": 257.75
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 295.137369903125,
            "gpu_single_us": 298.49994462,
            "e2e_batch_us": 301.5611875,
            "e2e_single_us": 531
          },
          "torch": {
            "gpu_batch_us": 12.98372353630839,
            "gpu_single_us": 20.33333294093609,
            "e2e_batch_us": 31.27985996705107,
            "e2e_single_us": 309.458
          },
          "system": {
            "gpu_batch_us": 13.74218754790625,
            "gpu_single_us": 17.2083382495,
            "e2e_batch_us": 16.3465167785,
            "e2e_single_us": 272.125
          }
        },
        "new_over_old": {
          "gpu_batch_us": 9.43025018046639,
          "gpu_single_us": 7.872530522587291,
          "e2e_batch_us": 9.154565657396805,
          "e2e_single_us": 1.6578622007418233
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 9.419751101473395,
          "gpu_single_us": 7.851561199323241,
          "e2e_batch_us": 9.08173918874961,
          "e2e_single_us": 2.1568820917612235
        },
        "new_over_mps": {
          "gpu_batch_us": 21.476738610519977,
          "gpu_single_us": 17.346238799593166,
          "e2e_batch_us": 18.4480395173015,
          "e2e_single_us": 1.951309141019752
        },
        "new_over_torch": {
          "gpu_batch_us": 22.73133505021012,
          "gpu_single_us": 14.680325428550127,
          "e2e_batch_us": 9.640746084466244,
          "e2e_single_us": 1.7159032889762098
        }
      },
      {
        "shape": [
          129,
          257,
          61
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 31.014974410937498,
            "gpu_single_us": 38.708385545800006,
            "e2e_batch_us": 32.8038537415,
            "e2e_single_us": 255.959
          },
          "torch": {
            "gpu_batch_us": 17.38086029945407,
            "gpu_single_us": 31.25000512227416,
            "e2e_batch_us": 30.122620626151015,
            "e2e_single_us": 282.375
          },
          "system": {
            "gpu_batch_us": 13.6289063448125,
            "gpu_single_us": 19.7499757633,
            "e2e_batch_us": 16.1658752969,
            "e2e_single_us": 250.042
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 294.50781130468755,
            "gpu_single_us": 297.249993309,
            "e2e_batch_us": 300.512359375,
            "e2e_single_us": 544.417
          },
          "torch": {
            "gpu_batch_us": 12.917316780658439,
            "gpu_single_us": 18.624996300786734,
            "e2e_batch_us": 29.612810283687942,
            "e2e_single_us": 296.58299999999997
          },
          "system": {
            "gpu_batch_us": 14.009765436640624,
            "gpu_single_us": 18.124992493500002,
            "e2e_batch_us": 16.4395667939,
            "e2e_single_us": 260.917
          }
        },
        "new_over_old": {
          "gpu_batch_us": 9.495665139121597,
          "gpu_single_us": 7.6792144419790365,
          "e2e_batch_us": 9.160885844178217,
          "e2e_single_us": 2.1269695537175877
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 9.531726564874912,
          "gpu_single_us": 8.082573127126624,
          "e2e_batch_us": 9.025023886838316,
          "e2e_single_us": 2.08032
        },
        "new_over_mps": {
          "gpu_batch_us": 21.021609008131048,
          "gpu_single_us": 16.40000642293231,
          "e2e_batch_us": 18.279822281357617,
          "e2e_single_us": 2.0865524285500756
        },
        "new_over_torch": {
          "gpu_batch_us": 22.799457217435794,
          "gpu_single_us": 15.959734354226097,
          "e2e_batch_us": 10.148052700710261,
          "e2e_single_us": 1.8356311723868195
        }
      },
      {
        "shape": [
          1025,
          1025,
          1024
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 2065.9404856714286,
            "gpu_single_us": 2240.75000733,
            "e2e_batch_us": 2360.45828571,
            "e2e_single_us": 2650.75
          },
          "torch": {
            "gpu_batch_us": 388.55706604764515,
            "gpu_single_us": 374.8750314116478,
            "e2e_batch_us": 407.5009130434783,
            "e2e_single_us": 638.958
          },
          "system": {
            "gpu_batch_us": 484.42414472307695,
            "gpu_single_us": 469.45834765200004,
            "e2e_batch_us": 501.884615385,
            "e2e_single_us": 707.542
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 3762.3749958500002,
            "gpu_single_us": 3738.3333547,
            "e2e_batch_us": 3885.8125,
            "e2e_single_us": 4038.084
          },
          "torch": {
            "gpu_batch_us": 382.9496514905865,
            "gpu_single_us": 369.208340998739,
            "e2e_batch_us": 404.28820833333333,
            "e2e_single_us": 651.2090000000001
          },
          "system": {
            "gpu_batch_us": 482.8643167205128,
            "gpu_single_us": 470.875005703,
            "e2e_batch_us": 499.682692308,
            "e2e_single_us": 706
          }
        },
        "new_over_old": {
          "gpu_batch_us": 1.821143939984908,
          "gpu_single_us": 1.668340217548172,
          "e2e_batch_us": 1.6462110444926545,
          "e2e_single_us": 1.5233741393945108
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 2.2011009280670213,
          "gpu_single_us": 1.9616348920429576,
          "e2e_batch_us": 2.094150017530231,
          "e2e_single_us": 2.049692770661564
        },
        "new_over_mps": {
          "gpu_batch_us": 7.791785115543554,
          "gpu_single_us": 7.939120381042096,
          "e2e_batch_us": 7.776560124689729,
          "e2e_single_us": 5.719665722379603
        },
        "new_over_torch": {
          "gpu_batch_us": 9.824724950670142,
          "gpu_single_us": 10.12526787609267,
          "e2e_batch_us": 9.611491059853444,
          "e2e_single_us": 6.200903243044859
        }
      },
      {
        "shape": [
          1025,
          1025,
          1024
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 2080.5694447444444,
            "gpu_single_us": 2236.54159345,
            "e2e_batch_us": 2207.412,
            "e2e_single_us": 2332.416
          },
          "torch": {
            "gpu_batch_us": 383.73554648109234,
            "gpu_single_us": 372.04165710136294,
            "e2e_batch_us": 400.81632653061223,
            "e2e_single_us": 659.208
          },
          "system": {
            "gpu_batch_us": 486.87286340769225,
            "gpu_single_us": 467.12497714899996,
            "e2e_batch_us": 502.369666667,
            "e2e_single_us": 705.792
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 3801.489569025,
            "gpu_single_us": 3737.7499975299997,
            "e2e_batch_us": 4139.875,
            "e2e_single_us": 4255.459
          },
          "torch": {
            "gpu_batch_us": 392.48437678907067,
            "gpu_single_us": 388.7083730660379,
            "e2e_batch_us": 408.5138958333334,
            "e2e_single_us": 718.542
          },
          "system": {
            "gpu_batch_us": 486.92948752051285,
            "gpu_single_us": 474.50000420199996,
            "e2e_batch_us": 500.050205128,
            "e2e_single_us": 718.292
          }
        },
        "new_over_old": {
          "gpu_batch_us": 1.8271389972719396,
          "gpu_single_us": 1.6712186388469061,
          "e2e_batch_us": 1.875442826259892,
          "e2e_single_us": 1.8244854262704422
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 2.2204777594955516,
          "gpu_single_us": 1.986425628591388,
          "e2e_batch_us": 2.230289271808027,
          "e2e_single_us": 2.065045732033844
        },
        "new_over_mps": {
          "gpu_batch_us": 7.8070637873719955,
          "gpu_single_us": 7.8772391241935535,
          "e2e_batch_us": 8.278918711652759,
          "e2e_single_us": 5.924413748169267
        },
        "new_over_torch": {
          "gpu_batch_us": 9.685709276188591,
          "gpu_single_us": 9.615820642214443,
          "e2e_batch_us": 10.13398820021779,
          "e2e_single_us": 5.9223524860063845
        }
      },
      {
        "shape": [
          2049,
          4097,
          1025
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 15769.604156950001,
            "gpu_single_us": 15373.875037799999,
            "e2e_batch_us": 15974.1875,
            "e2e_single_us": 15895.958
          },
          "torch": {
            "gpu_batch_us": 2665.6607182563416,
            "gpu_single_us": 2455.500012729317,
            "e2e_batch_us": 2810.3095714285714,
            "e2e_single_us": 3682.9170000000004
          },
          "system": {
            "gpu_batch_us": 2769.0535727285715,
            "gpu_single_us": 2597.12501429,
            "e2e_batch_us": 2854.07142857,
            "e2e_single_us": 3034.583
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 10573.68749755,
            "gpu_single_us": 10843.9583564,
            "e2e_batch_us": 11267,
            "e2e_single_us": 11445.833
          },
          "torch": {
            "gpu_batch_us": 2633.8571499634004,
            "gpu_single_us": 2452.8750218451023,
            "e2e_batch_us": 2744.5832857142855,
            "e2e_single_us": 3448.291
          },
          "system": {
            "gpu_batch_us": 2774.2291664833333,
            "gpu_single_us": 2640.16666915,
            "e2e_batch_us": 2868.35416667,
            "e2e_single_us": 3014
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.6705106477190773,
          "gpu_single_us": 0.705349713701834,
          "e2e_batch_us": 0.7053253882239707,
          "e2e_single_us": 0.7200467565402475
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 1.369314872805531,
          "gpu_single_us": 1.3776002489789614,
          "e2e_batch_us": 1.2922181107992277,
          "e2e_single_us": 1.2960906267309975
        },
        "new_over_mps": {
          "gpu_batch_us": 3.8113965584730014,
          "gpu_single_us": 4.107300680335914,
          "e2e_batch_us": 3.9280365482482806,
          "e2e_single_us": 3.7975557398805577
        },
        "new_over_torch": {
          "gpu_batch_us": 4.014525805887738,
          "gpu_single_us": 4.420917600703095,
          "e2e_batch_us": 4.1051769347446605,
          "e2e_single_us": 3.319276998373977
        }
      },
      {
        "shape": [
          2049,
          4097,
          1025
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 15406.64583445,
            "gpu_single_us": 14874.833286699999,
            "e2e_batch_us": 15834.458,
            "e2e_single_us": 15818.75
          },
          "torch": {
            "gpu_batch_us": 2617.0000007065632,
            "gpu_single_us": 2455.374982673675,
            "e2e_batch_us": 2728.4513333333334,
            "e2e_single_us": 2963.708
          },
          "system": {
            "gpu_batch_us": 2815.965272016667,
            "gpu_single_us": 2625.8333819,
            "e2e_batch_us": 2900.13883333,
            "e2e_single_us": 3076.375
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 10570.81250475,
            "gpu_single_us": 10540.5000504,
            "e2e_batch_us": 10880.6875,
            "e2e_single_us": 11012.125
          },
          "torch": {
            "gpu_batch_us": 2580.9523898975126,
            "gpu_single_us": 2517.708344385028,
            "e2e_batch_us": 2682.0774285714288,
            "e2e_single_us": 2934.9159999999997
          },
          "system": {
            "gpu_batch_us": 2807.2500038666667,
            "gpu_single_us": 2621.12502242,
            "e2e_batch_us": 2883.83333333,
            "e2e_single_us": 2892.083
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.6861203027795417,
          "gpu_single_us": 0.7086129872678676,
          "e2e_batch_us": 0.6871525062619762,
          "e2e_single_us": 0.6961438166732516
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 1.4241573804158911,
          "gpu_single_us": 1.4665317767246862,
          "e2e_batch_us": 1.3388873694080052,
          "e2e_single_us": 1.3623856056104264
        },
        "new_over_mps": {
          "gpu_batch_us": 3.765540115839313,
          "gpu_single_us": 4.021364856785159,
          "e2e_batch_us": 3.7729945674204157,
          "e2e_single_us": 3.807679447650707
        },
        "new_over_torch": {
          "gpu_batch_us": 4.0957022477930165,
          "gpu_single_us": 4.186545305736995,
          "e2e_batch_us": 4.056813343302862,
          "e2e_single_us": 3.75210908932317
        }
      },
      {
        "shape": [
          4097,
          4097,
          4096
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 139695.833321,
            "gpu_single_us": 139235.04168599998,
            "e2e_batch_us": 141215.25,
            "e2e_single_us": 141584.833
          },
          "torch": {
            "gpu_batch_us": 20450.458338018507,
            "gpu_single_us": 20675.37500988692,
            "e2e_batch_us": 20900.959,
            "e2e_single_us": 21502.666999999998
          },
          "system": {
            "gpu_batch_us": 22180.7499765,
            "gpu_single_us": 21256.625011999997,
            "e2e_batch_us": 22479.917,
            "e2e_single_us": 23011.625
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 52429.3750059,
            "gpu_single_us": 53069.333313,
            "e2e_batch_us": 53076.208,
            "e2e_single_us": 53793.791
          },
          "torch": {
            "gpu_batch_us": 20634.41666541621,
            "gpu_single_us": 20720.375003293157,
            "e2e_batch_us": 21483.666,
            "e2e_single_us": 21081.0
          },
          "system": {
            "gpu_batch_us": 21778.7500005,
            "gpu_single_us": 21760.458359499997,
            "e2e_batch_us": 23073.416,
            "e2e_single_us": 22506.458
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.3753109434941068,
          "gpu_single_us": 0.38114926149611733,
          "e2e_batch_us": 0.3758532311489021,
          "e2e_single_us": 0.3799403499667227
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.5888876807462976,
          "gpu_single_us": 0.5899602128019151,
          "e2e_batch_us": 0.6165509627218013,
          "e2e_single_us": 0.600017375056424
        },
        "new_over_mps": {
          "gpu_batch_us": 2.4073638296365183,
          "gpu_single_us": 2.4387966666994148,
          "e2e_batch_us": 2.3003186004187675,
          "e2e_single_us": 2.3901491296409234
        },
        "new_over_torch": {
          "gpu_batch_us": 2.5408702293858845,
          "gpu_single_us": 2.561214905838602,
          "e2e_batch_us": 2.4705377564518085,
          "e2e_single_us": 2.551766567050899
        }
      },
      {
        "shape": [
          4097,
          4097,
          4096
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 141526.583408,
            "gpu_single_us": 142312.249984,
            "e2e_batch_us": 136774.167,
            "e2e_single_us": 139589.708
          },
          "torch": {
            "gpu_batch_us": 19698.833348229527,
            "gpu_single_us": 20099.416666198522,
            "e2e_batch_us": 20885.0,
            "e2e_single_us": 20655.083000000002
          },
          "system": {
            "gpu_batch_us": 21683.4166786,
            "gpu_single_us": 21650.7916921,
            "e2e_batch_us": 23502.875,
            "e2e_single_us": 23072.084
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 51361.5000527,
            "gpu_single_us": 53130.8749923,
            "e2e_batch_us": 53494.5,
            "e2e_single_us": 53262.125
          },
          "torch": {
            "gpu_batch_us": 20354.458305519074,
            "gpu_single_us": 20520.08337341249,
            "e2e_batch_us": 21599.959,
            "e2e_single_us": 21596.5
          },
          "system": {
            "gpu_batch_us": 22038.333292599997,
            "gpu_single_us": 21820.1249954,
            "e2e_batch_us": 23031.417,
            "e2e_single_us": 22206.125
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.3629106194461889,
          "gpu_single_us": 0.3733401376077846,
          "e2e_batch_us": 0.3911155240302067,
          "e2e_single_us": 0.38156197733431746
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.5872074207614854,
          "gpu_single_us": 0.5911361745002042,
          "e2e_batch_us": 0.5868732475943502,
          "e2e_single_us": 0.5824170917609164
        },
        "new_over_mps": {
          "gpu_batch_us": 2.3305528313225983,
          "gpu_single_us": 2.4349482417493373,
          "e2e_batch_us": 2.3226751528140888,
          "e2e_single_us": 2.3985330623870666
        },
        "new_over_torch": {
          "gpu_batch_us": 2.5233538167298426,
          "gpu_single_us": 2.5892134074435944,
          "e2e_batch_us": 2.4766019231795764,
          "e2e_single_us": 2.466238742388813
        }
      },
      {
        "shape": [
          1024,
          1024,
          1024
        ],
        "order": "forward",
        "old_bk": 4096,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 279.0690114125,
            "gpu_single_us": 270.541699138,
            "e2e_batch_us": 285.620656716,
            "e2e_single_us": 499
          },
          "torch": {
            "gpu_batch_us": 286.82682204816956,
            "gpu_single_us": 275.49994410946965,
            "e2e_batch_us": 303.18229687499996,
            "e2e_single_us": 573.25
          },
          "system": {
            "gpu_batch_us": 279.30012947812503,
            "gpu_single_us": 269.250012934,
            "e2e_batch_us": 288.423606061,
            "e2e_single_us": 501.416
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 281.4121098703125,
            "gpu_single_us": 269.166601356,
            "e2e_batch_us": 287.75065625,
            "e2e_single_us": 484.625
          },
          "torch": {
            "gpu_batch_us": 285.9960941350437,
            "gpu_single_us": 276.00000612437725,
            "e2e_batch_us": 301.2416769230769,
            "e2e_single_us": 601.25
          },
          "system": {
            "gpu_batch_us": 283.34221264426225,
            "gpu_single_us": 269.791635219,
            "e2e_batch_us": 288.120901639,
            "e2e_single_us": 502.083
          }
        },
        "new_over_old": {
          "gpu_batch_us": 1.008396125553149,
          "gpu_single_us": 0.9949172427526649,
          "e2e_batch_us": 1.0074574421839453,
          "e2e_single_us": 0.9711923847695391
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.9971448186053099,
          "gpu_single_us": 0.9260561455330794,
          "e2e_batch_us": 0.995983612030895,
          "e2e_single_us": 1.0274390148296115
        },
        "new_over_mps": {
          "gpu_batch_us": 0.9931880860393613,
          "gpu_single_us": 0.9976832718979867,
          "e2e_batch_us": 0.998714965186858,
          "e2e_single_us": 0.9652288565834732
        },
        "new_over_torch": {
          "gpu_batch_us": 0.9839718640962745,
          "gpu_single_us": 0.9752412876204871,
          "e2e_batch_us": 0.9552152915530281,
          "e2e_single_us": 0.806029106029106
        }
      },
      {
        "shape": [
          1024,
          1024,
          1024
        ],
        "order": "reverse",
        "old_bk": 4096,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 431.0918567090909,
            "gpu_single_us": 442.000047769,
            "e2e_batch_us": 446.060590909,
            "e2e_single_us": 623.042
          },
          "torch": {
            "gpu_batch_us": 460.45528213697963,
            "gpu_single_us": 365.9166395664215,
            "e2e_batch_us": 487.96546341463414,
            "e2e_single_us": 785.541
          },
          "system": {
            "gpu_batch_us": 429.5460530552632,
            "gpu_single_us": 359.50000165,
            "e2e_batch_us": 435.572368421,
            "e2e_single_us": 586.666
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 278.8203128140625,
            "gpu_single_us": 269.41666146700004,
            "e2e_batch_us": 283.556768116,
            "e2e_single_us": 512.625
          },
          "torch": {
            "gpu_batch_us": 287.73763096978655,
            "gpu_single_us": 278.45834847539663,
            "e2e_batch_us": 300.7108923076923,
            "e2e_single_us": 687.5
          },
          "system": {
            "gpu_batch_us": 279.83919244375,
            "gpu_single_us": 269.458338153,
            "e2e_batch_us": 288.134191176,
            "e2e_single_us": 505.666
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.6467770348123644,
          "gpu_single_us": 0.6095398921943188,
          "e2e_batch_us": 0.6356911457659973,
          "e2e_single_us": 0.8227775976579428
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.6859423017997981,
          "gpu_single_us": 0.6799949869756865,
          "e2e_batch_us": 0.7104394185079376,
          "e2e_single_us": 0.7549253704053077
        },
        "new_over_mps": {
          "gpu_batch_us": 0.9963590531376613,
          "gpu_single_us": 0.9998453316149516,
          "e2e_batch_us": 0.984113572077935,
          "e2e_single_us": 1.0137620484667744
        },
        "new_over_torch": {
          "gpu_batch_us": 0.9690088566946587,
          "gpu_single_us": 0.9675294813105757,
          "e2e_batch_us": 0.9429547627621686,
          "e2e_single_us": 0.7456363636363637
        }
      }
    ],
    "artifacts": 35,
    "validated_outputs": 192,
    "checked_elements": 995833488,
    "rejected_requests": 16,
    "current_artifacts_checked": false
  },
  "cooperative-replay": {
    "runs": {
      "old-forward": {
        "measurements": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 30.87955701625,
                "gpu_single_us": 38.374972064000005,
                "e2e_batch_us": 32.9826689775,
                "e2e_single_us": 252.5
              },
              "torch": {
                "gpu_batch_us": 19.49674333445728,
                "gpu_single_us": 44.458312913775444,
                "e2e_batch_us": 31.04622648752399,
                "e2e_single_us": 367.08299999999997
              },
              "system": {
                "gpu_batch_us": 14.546874808734376,
                "gpu_single_us": 17.124984879,
                "e2e_batch_us": 16.0165401015,
                "e2e_single_us": 236.833
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.085286536875,
                "gpu_single_us": 38.458325434500004,
                "e2e_batch_us": 33.3077116935,
                "e2e_single_us": 280.25
              },
              "torch": {
                "gpu_batch_us": 14.813801499258261,
                "gpu_single_us": 20.666630007326603,
                "e2e_batch_us": 30.267805128205122,
                "e2e_single_us": 268.25
              },
              "system": {
                "gpu_batch_us": 13.917969226890625,
                "gpu_single_us": 17.0833664015,
                "e2e_batch_us": 15.7748377425,
                "e2e_single_us": 243.708
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2080.8437475375,
                "gpu_single_us": 2190.08338172,
                "e2e_batch_us": 2186.578125,
                "e2e_single_us": 2573.75
              },
              "torch": {
                "gpu_batch_us": 391.1301027983427,
                "gpu_single_us": 374.54167613759637,
                "e2e_batch_us": 403.84522448979595,
                "e2e_single_us": 651.375
              },
              "system": {
                "gpu_batch_us": 482.4615376153846,
                "gpu_single_us": 468.666665256,
                "e2e_batch_us": 498.641025641,
                "e2e_single_us": 691.708
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2118.9843828375,
                "gpu_single_us": 2180.5000142199997,
                "e2e_batch_us": 2211.328125,
                "e2e_single_us": 2322.459
              },
              "torch": {
                "gpu_batch_us": 386.2056740004807,
                "gpu_single_us": 373.50005004554987,
                "e2e_batch_us": 418.5035531914894,
                "e2e_single_us": 668.375
              },
              "system": {
                "gpu_batch_us": 486.58653843076917,
                "gpu_single_us": 466.416706331,
                "e2e_batch_us": 504.939102564,
                "e2e_single_us": 695.959
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15297.83333535,
                "gpu_single_us": 14586.5000086,
                "e2e_batch_us": 16033.083,
                "e2e_single_us": 16155.167
              },
              "torch": {
                "gpu_batch_us": 2549.756948913758,
                "gpu_single_us": 2493.3749809861183,
                "e2e_batch_us": 2815.382,
                "e2e_single_us": 2858.9159999999997
              },
              "system": {
                "gpu_batch_us": 2777.755950642857,
                "gpu_single_us": 2635.75004647,
                "e2e_batch_us": 2858.02971429,
                "e2e_single_us": 3091.75
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15551.58331755,
                "gpu_single_us": 14502.4583326,
                "e2e_batch_us": 15948.0625,
                "e2e_single_us": 16643.208
              },
              "torch": {
                "gpu_batch_us": 2658.4107129435456,
                "gpu_single_us": 2501.6666622832417,
                "e2e_batch_us": 2768.041571428571,
                "e2e_single_us": 2888.417
              },
              "system": {
                "gpu_batch_us": 2781.1874946,
                "gpu_single_us": 2627.58333702,
                "e2e_batch_us": 2915.82633333,
                "e2e_single_us": 2976.208
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 140971.291694,
                "gpu_single_us": 142314.666591,
                "e2e_batch_us": 141691.875,
                "e2e_single_us": 139680.417
              },
              "torch": {
                "gpu_batch_us": 20969.7499522008,
                "gpu_single_us": 20167.37504163757,
                "e2e_batch_us": 20322.958,
                "e2e_single_us": 21905.916999999998
              },
              "system": {
                "gpu_batch_us": 22102.4583443,
                "gpu_single_us": 22124.291688700003,
                "e2e_batch_us": 23064.584,
                "e2e_single_us": 22638.25
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 141349.458368,
                "gpu_single_us": 142939.916637,
                "e2e_batch_us": 142139.875,
                "e2e_single_us": 141180.583
              },
              "torch": {
                "gpu_batch_us": 20251.37498276308,
                "gpu_single_us": 20470.33334383741,
                "e2e_batch_us": 21237.75,
                "e2e_single_us": 20479.625
              },
              "system": {
                "gpu_batch_us": 21973.250026299997,
                "gpu_single_us": 21936.3750075,
                "e2e_batch_us": 22632.375,
                "e2e_single_us": 22640.75
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 279.205729384375,
                "gpu_single_us": 271.41667669599997,
                "e2e_batch_us": 285.661621212,
                "e2e_single_us": 485.792
              },
              "torch": {
                "gpu_batch_us": 287.49088414770085,
                "gpu_single_us": 275.08335188031197,
                "e2e_batch_us": 303.7058181818182,
                "e2e_single_us": 682.083
              },
              "system": {
                "gpu_batch_us": 279.29752559375,
                "gpu_single_us": 269.374984782,
                "e2e_batch_us": 286.295955882,
                "e2e_single_us": 513.5
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 279.6588532875,
                "gpu_single_us": 269.583368208,
                "e2e_batch_us": 288.067857143,
                "e2e_single_us": 495.875
              },
              "torch": {
                "gpu_batch_us": 287.3353500505008,
                "gpu_single_us": 275.37497226148844,
                "e2e_batch_us": 303.2452903225806,
                "e2e_single_us": 537.292
              },
              "system": {
                "gpu_batch_us": 280.3691404546875,
                "gpu_single_us": 269.91666527499996,
                "e2e_batch_us": 286.923913043,
                "e2e_single_us": 503.875
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 677.8452370785715,
                "gpu_single_us": 670.999987051,
                "e2e_batch_us": 698.872035714,
                "e2e_single_us": 929.958
              },
              "torch": {
                "gpu_batch_us": 287.24479216180043,
                "gpu_single_us": 277.83337282016873,
                "e2e_batch_us": 301.7288461538462,
                "e2e_single_us": 572.5409999999999
              },
              "system": {
                "gpu_batch_us": 281.07096295625,
                "gpu_single_us": 273.916695733,
                "e2e_batch_us": 285.226102941,
                "e2e_single_us": 485.791
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 278.5169281203125,
                "gpu_single_us": 268.66665575699994,
                "e2e_batch_us": 282.870228571,
                "e2e_single_us": 533.292
              },
              "torch": {
                "gpu_batch_us": 287.17063740634967,
                "gpu_single_us": 275.749946013093,
                "e2e_batch_us": 302.24007936507934,
                "e2e_single_us": 536.25
              },
              "system": {
                "gpu_batch_us": 279.7792967609375,
                "gpu_single_us": 270.12499049300004,
                "e2e_batch_us": 287.019230769,
                "e2e_single_us": 524.75
              }
            }
          }
        ],
        "rejections": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          }
        ],
        "outputs": 36,
        "elements": 170166552
      },
      "new-forward": {
        "measurements": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "8ae10c1f3d2f687b2c8251424f37f46ac76b4c4fdd832cf2170fb8fa2409b53e",
            "paths": {
              "native": {
                "gpu_batch_us": 534.4398151555556,
                "gpu_single_us": 520.458386745,
                "e2e_batch_us": 544.802083333,
                "e2e_single_us": 782.834
              },
              "torch": {
                "gpu_batch_us": 16.423176930402406,
                "gpu_single_us": 24.54168861731887,
                "e2e_batch_us": 29.203730486008833,
                "e2e_single_us": 344.04200000000003
              },
              "system": {
                "gpu_batch_us": 14.567708603859375,
                "gpu_single_us": 17.2500149347,
                "e2e_batch_us": 15.8111114155,
                "e2e_single_us": 237.125
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "f396b61f109fae66974d61e38ef04418e2088523fcfce0d02cea835170c85dcd",
            "paths": {
              "native": {
                "gpu_batch_us": 148.28320308874999,
                "gpu_single_us": 153.750064783,
                "e2e_batch_us": 153.125,
                "e2e_single_us": 415.75
              },
              "torch": {
                "gpu_batch_us": 15.057942619023379,
                "gpu_single_us": 23.500004317611456,
                "e2e_batch_us": 30.004531194295904,
                "e2e_single_us": 309.0
              },
              "system": {
                "gpu_batch_us": 13.746744116359375,
                "gpu_single_us": 17.1666615643,
                "e2e_batch_us": 15.8211586271,
                "e2e_single_us": 233.208
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "f396b61f109fae66974d61e38ef04418e2088523fcfce0d02cea835170c85dcd",
            "paths": {
              "native": {
                "gpu_batch_us": 148.35611909796873,
                "gpu_single_us": 152.249995153,
                "e2e_batch_us": 151.084645669,
                "e2e_single_us": 439.792
              },
              "torch": {
                "gpu_batch_us": 13.575521734310314,
                "gpu_single_us": 18.208345863968134,
                "e2e_batch_us": 31.363579734219268,
                "e2e_single_us": 282.5
              },
              "system": {
                "gpu_batch_us": 15.15234362159375,
                "gpu_single_us": 18.333317711899998,
                "e2e_batch_us": 16.515125,
                "e2e_single_us": 220.208
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "f396b61f109fae66974d61e38ef04418e2088523fcfce0d02cea835170c85dcd",
            "paths": {
              "native": {
                "gpu_batch_us": 148.51432297296876,
                "gpu_single_us": 153.458327986,
                "e2e_batch_us": 152.268047244,
                "e2e_single_us": 441.667
              },
              "torch": {
                "gpu_batch_us": 13.980467883811798,
                "gpu_single_us": 24.749955628067255,
                "e2e_batch_us": 36.875563685636855,
                "e2e_single_us": 245.583
              },
              "system": {
                "gpu_batch_us": 15.185546544671876,
                "gpu_single_us": 17.374986782700002,
                "e2e_batch_us": 16.028045264,
                "e2e_single_us": 247.916
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "a5eb38354235954c2a39d05e73ab678252a2706b59d7ae77104e5ba6acb565bf",
            "paths": {
              "native": {
                "gpu_batch_us": 781.4236111375001,
                "gpu_single_us": 767.208344769,
                "e2e_batch_us": 791.21875,
                "e2e_single_us": 1038.5
              },
              "torch": {
                "gpu_batch_us": 384.991135240156,
                "gpu_single_us": 368.12503822147846,
                "e2e_batch_us": 402.37323404255324,
                "e2e_single_us": 735.916
              },
              "system": {
                "gpu_batch_us": 484.69166649749997,
                "gpu_single_us": 466.125027742,
                "e2e_batch_us": 494.87605,
                "e2e_single_us": 695.541
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "a5eb38354235954c2a39d05e73ab678252a2706b59d7ae77104e5ba6acb565bf",
            "paths": {
              "native": {
                "gpu_batch_us": 777.654514725,
                "gpu_single_us": 781.624985393,
                "e2e_batch_us": 808.0625,
                "e2e_single_us": 1007.709
              },
              "torch": {
                "gpu_batch_us": 386.65712175619166,
                "gpu_single_us": 373.8749655894935,
                "e2e_batch_us": 403.1631875,
                "e2e_single_us": 642.833
              },
              "system": {
                "gpu_batch_us": 485.0921043394736,
                "gpu_single_us": 467.250007205,
                "e2e_batch_us": 498.355263158,
                "e2e_single_us": 711.541
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "cce0d09dc56b885a4e0e4303998af47133878c48812b49dd9048d6bcc7dd80a0",
            "paths": {
              "native": {
                "gpu_batch_us": 9565.31250265,
                "gpu_single_us": 9676.041721829999,
                "e2e_batch_us": 9809.0625,
                "e2e_single_us": 9936.917
              },
              "torch": {
                "gpu_batch_us": 388.0462932607366,
                "gpu_single_us": 375.54168375208974,
                "e2e_batch_us": 449.81202222222225,
                "e2e_single_us": 1020.9580000000001
              },
              "system": {
                "gpu_batch_us": 480.289584085,
                "gpu_single_us": 468.874990474,
                "e2e_batch_us": 504.696875,
                "e2e_single_us": 696.167
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "a5eb38354235954c2a39d05e73ab678252a2706b59d7ae77104e5ba6acb565bf",
            "paths": {
              "native": {
                "gpu_batch_us": 772.2916658791667,
                "gpu_single_us": 784.374948125,
                "e2e_batch_us": 796.387166667,
                "e2e_single_us": 1011.416
              },
              "torch": {
                "gpu_batch_us": 388.6422112017222,
                "gpu_single_us": 374.8750314116478,
                "e2e_batch_us": 404.5507173913043,
                "e2e_single_us": 711.5409999999999
              },
              "system": {
                "gpu_batch_us": 484.5106830974359,
                "gpu_single_us": 468.374986667,
                "e2e_batch_us": 497.743589744,
                "e2e_single_us": 728.333
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "fc87f86e71fbe4e9d8cad9957928b0fd0130e711a1126e98d5706cdc250c1a71",
            "paths": {
              "native": {
                "gpu_batch_us": 3846.2333264799995,
                "gpu_single_us": 3654.58329907,
                "e2e_batch_us": 3889.1834,
                "e2e_single_us": 4672.041
              },
              "torch": {
                "gpu_batch_us": 2654.934526487653,
                "gpu_single_us": 2474.8332798480988,
                "e2e_batch_us": 2670.1369999999997,
                "e2e_single_us": 3135.041
              },
              "system": {
                "gpu_batch_us": 2771.9375017666666,
                "gpu_single_us": 2604.75004325,
                "e2e_batch_us": 2840.882,
                "e2e_single_us": 3011.125
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "91593521ac7abf8881f1f26fc3a3ee7b670f32f8c819054adce85e13737c9816",
            "paths": {
              "native": {
                "gpu_batch_us": 29593.9166681,
                "gpu_single_us": 29151.7083533,
                "e2e_batch_us": 30867.959,
                "e2e_single_us": 29887.291
              },
              "torch": {
                "gpu_batch_us": 2590.5972191443047,
                "gpu_single_us": 2503.25002707541,
                "e2e_batch_us": 2725.2500000000005,
                "e2e_single_us": 3010.75
              },
              "system": {
                "gpu_batch_us": 2840.5208334666663,
                "gpu_single_us": 2603.74997742,
                "e2e_batch_us": 2883.9515,
                "e2e_single_us": 3062.25
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "dadf52980273db52590a98cb00b9405de2c411d885ab6ee578880395fcbb6063",
            "paths": {
              "native": {
                "gpu_batch_us": 4136.854186075,
                "gpu_single_us": 4000.08331053,
                "e2e_batch_us": 4229.698,
                "e2e_single_us": 4290.5
              },
              "torch": {
                "gpu_batch_us": 2632.345230917313,
                "gpu_single_us": 2474.8333380557597,
                "e2e_batch_us": 2725.880857142857,
                "e2e_single_us": 2911.4170000000004
              },
              "system": {
                "gpu_batch_us": 2779.4999963,
                "gpu_single_us": 2689.99999389,
                "e2e_batch_us": 2843.4235,
                "e2e_single_us": 2989.542
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "fc87f86e71fbe4e9d8cad9957928b0fd0130e711a1126e98d5706cdc250c1a71",
            "paths": {
              "native": {
                "gpu_batch_us": 3833.77498948,
                "gpu_single_us": 3657.25002484,
                "e2e_batch_us": 3884.6584,
                "e2e_single_us": 4247.25
              },
              "torch": {
                "gpu_batch_us": 2631.904756916421,
                "gpu_single_us": 2443.333331029862,
                "e2e_batch_us": 2691.345285714286,
                "e2e_single_us": 3000.458
              },
              "system": {
                "gpu_batch_us": 2805.2291696999996,
                "gpu_single_us": 2596.1250066800003,
                "e2e_batch_us": 2815.5,
                "e2e_single_us": 3090.416
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "9bc5696f0635db48ae870c73e82b590397c811672f48a96017861586feb90943",
            "paths": {
              "native": {
                "gpu_batch_us": 124387.083284,
                "gpu_single_us": 125573.374971,
                "e2e_batch_us": 124359.542,
                "e2e_single_us": 124324.166
              },
              "torch": {
                "gpu_batch_us": 20065.541670192033,
                "gpu_single_us": 20793.624978978187,
                "e2e_batch_us": 20866.917,
                "e2e_single_us": 21110.583
              },
              "system": {
                "gpu_batch_us": 22072.2499653,
                "gpu_single_us": 22251.708316599997,
                "e2e_batch_us": 23519.125,
                "e2e_single_us": 23243.208
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "dd8e064aacb4f1e6bbc81d7f098bf2fe86d35a0c8a026fcf1132540127d53e46",
            "paths": {
              "native": {
                "gpu_batch_us": 24988.7916725,
                "gpu_single_us": 26152.0833592,
                "e2e_batch_us": 26098.5,
                "e2e_single_us": 26072.083
              },
              "torch": {
                "gpu_batch_us": 20437.29170691222,
                "gpu_single_us": 20454.749988857657,
                "e2e_batch_us": 21492.208000000002,
                "e2e_single_us": 21018.208000000002
              },
              "system": {
                "gpu_batch_us": 21720.6249945,
                "gpu_single_us": 21763.0416155,
                "e2e_batch_us": 23226.125,
                "e2e_single_us": 22582.208
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "28522d4512e17182f06a4fddc6fc68123d19f816534fc6136bad647dcec42c44",
            "paths": {
              "native": {
                "gpu_batch_us": 25386.7083229,
                "gpu_single_us": 26032.291643800003,
                "e2e_batch_us": 26508.416,
                "e2e_single_us": 25245.5
              },
              "torch": {
                "gpu_batch_us": 19829.750002827495,
                "gpu_single_us": 20184.583321679384,
                "e2e_batch_us": 20820.333000000002,
                "e2e_single_us": 20898.75
              },
              "system": {
                "gpu_batch_us": 22444.6250359,
                "gpu_single_us": 21402.1666907,
                "e2e_batch_us": 23096.458,
                "e2e_single_us": 23266.125
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "dd8e064aacb4f1e6bbc81d7f098bf2fe86d35a0c8a026fcf1132540127d53e46",
            "paths": {
              "native": {
                "gpu_batch_us": 25484.5833406,
                "gpu_single_us": 25578.166707400003,
                "e2e_batch_us": 26021.333,
                "e2e_single_us": 26643.208
              },
              "torch": {
                "gpu_batch_us": 20286.0000426881,
                "gpu_single_us": 20623.625023290515,
                "e2e_batch_us": 21299.291,
                "e2e_single_us": 21074.958
              },
              "system": {
                "gpu_batch_us": 22143.4999839,
                "gpu_single_us": 22166.5416611,
                "e2e_batch_us": 22366.208,
                "e2e_single_us": 22908.791
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 278.7532539515625,
                "gpu_single_us": 272.24997757,
                "e2e_batch_us": 289.648984848,
                "e2e_single_us": 492.625
              },
              "torch": {
                "gpu_batch_us": 287.48565819114447,
                "gpu_single_us": 275.87503427639604,
                "e2e_batch_us": 302.01162295081974,
                "e2e_single_us": 557.584
              },
              "system": {
                "gpu_batch_us": 282.17578164843746,
                "gpu_single_us": 269.500014838,
                "e2e_batch_us": 287.75675,
                "e2e_single_us": 504.709
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 279.4811198328125,
                "gpu_single_us": 271.6666786,
                "e2e_batch_us": 283.040471429,
                "e2e_single_us": 511.417
              },
              "torch": {
                "gpu_batch_us": 290.2583316123734,
                "gpu_single_us": 276.00000612437725,
                "e2e_batch_us": 302.85,
                "e2e_single_us": 616.0
              },
              "system": {
                "gpu_batch_us": 280.9661464203125,
                "gpu_single_us": 269.166659564,
                "e2e_batch_us": 284.758928571,
                "e2e_single_us": 507.833
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 674.3511891857144,
                "gpu_single_us": 672.0833480360001,
                "e2e_batch_us": 686.035714286,
                "e2e_single_us": 898.5
              },
              "torch": {
                "gpu_batch_us": 288.362629362382,
                "gpu_single_us": 278.4166135825217,
                "e2e_batch_us": 305.653,
                "e2e_single_us": 629.334
              },
              "system": {
                "gpu_batch_us": 280.14192685,
                "gpu_single_us": 269.624986686,
                "e2e_batch_us": 285.493362319,
                "e2e_single_us": 484.292
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 279.764322840625,
                "gpu_single_us": 270.16666717799995,
                "e2e_batch_us": 284.485294118,
                "e2e_single_us": 520.25
              },
              "torch": {
                "gpu_batch_us": 287.32682221743744,
                "gpu_single_us": 276.83330699801445,
                "e2e_batch_us": 303.00641538461537,
                "e2e_single_us": 584.625
              },
              "system": {
                "gpu_batch_us": 280.9654943125,
                "gpu_single_us": 269.291689619,
                "e2e_batch_us": 289.859304348,
                "e2e_single_us": 512.959
              }
            }
          }
        ],
        "rejections": [],
        "outputs": 60,
        "elements": 327750192
      },
      "new-reverse": {
        "measurements": [
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 278.7669282046875,
                "gpu_single_us": 270.374992397,
                "e2e_batch_us": 285.061594203,
                "e2e_single_us": 525.541
              },
              "torch": {
                "gpu_batch_us": 288.8486834994534,
                "gpu_single_us": 276.41665656119585,
                "e2e_batch_us": 302.34796491228076,
                "e2e_single_us": 607.708
              },
              "system": {
                "gpu_batch_us": 279.1972656265625,
                "gpu_single_us": 269.208336249,
                "e2e_batch_us": 286.7718,
                "e2e_single_us": 521.834
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 279.28841063906253,
                "gpu_single_us": 272.416626103,
                "e2e_batch_us": 283.723214286,
                "e2e_single_us": 514.125
              },
              "torch": {
                "gpu_batch_us": 287.0156267817947,
                "gpu_single_us": 290.12502636760473,
                "e2e_batch_us": 302.17179999999996,
                "e2e_single_us": 572.042
              },
              "system": {
                "gpu_batch_us": 279.39713527343747,
                "gpu_single_us": 268.00000341599997,
                "e2e_batch_us": 288.032833333,
                "e2e_single_us": 507.25
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 677.0193452071429,
                "gpu_single_us": 674.541632179,
                "e2e_batch_us": 693.035714286,
                "e2e_single_us": 894.917
              },
              "torch": {
                "gpu_batch_us": 287.30729263770627,
                "gpu_single_us": 277.5833127088845,
                "e2e_batch_us": 301.65,
                "e2e_single_us": 552.083
              },
              "system": {
                "gpu_batch_us": 280.6712236640625,
                "gpu_single_us": 271.833327133,
                "e2e_batch_us": 285.411314286,
                "e2e_single_us": 491.417
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 279.563801996875,
                "gpu_single_us": 272.958364803,
                "e2e_batch_us": 286.30025,
                "e2e_single_us": 486.583
              },
              "torch": {
                "gpu_batch_us": 286.8281235350878,
                "gpu_single_us": 276.2916265055537,
                "e2e_batch_us": 306.6038461538462,
                "e2e_single_us": 567.4169999999999
              },
              "system": {
                "gpu_batch_us": 280.1699220078125,
                "gpu_single_us": 269.95834196,
                "e2e_batch_us": 286.025,
                "e2e_single_us": 469.084
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "dd8e064aacb4f1e6bbc81d7f098bf2fe86d35a0c8a026fcf1132540127d53e46",
            "paths": {
              "native": {
                "gpu_batch_us": 26433.1250219,
                "gpu_single_us": 25683.249929,
                "e2e_batch_us": 26804.417,
                "e2e_single_us": 26378.125
              },
              "torch": {
                "gpu_batch_us": 20454.416691791266,
                "gpu_single_us": 20410.249999258667,
                "e2e_batch_us": 20969.709000000003,
                "e2e_single_us": 21403.25
              },
              "system": {
                "gpu_batch_us": 21977.2916753,
                "gpu_single_us": 22293.5416969,
                "e2e_batch_us": 23019.166,
                "e2e_single_us": 22119.834
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "9bc5696f0635db48ae870c73e82b590397c811672f48a96017861586feb90943",
            "paths": {
              "native": {
                "gpu_batch_us": 126994.249993,
                "gpu_single_us": 123165.708268,
                "e2e_batch_us": 125175.25,
                "e2e_single_us": 125836.166
              },
              "torch": {
                "gpu_batch_us": 20524.541672784835,
                "gpu_single_us": 19996.625022031367,
                "e2e_batch_us": 21301.083,
                "e2e_single_us": 22005.792
              },
              "system": {
                "gpu_batch_us": 21919.250022599997,
                "gpu_single_us": 21907.7499933,
                "e2e_batch_us": 22662.417,
                "e2e_single_us": 21920.291
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "28522d4512e17182f06a4fddc6fc68123d19f816534fc6136bad647dcec42c44",
            "paths": {
              "native": {
                "gpu_batch_us": 25861.6667124,
                "gpu_single_us": 26307.125052,
                "e2e_batch_us": 26648.833,
                "e2e_single_us": 26280.333
              },
              "torch": {
                "gpu_batch_us": 20309.875020757318,
                "gpu_single_us": 20483.00002934411,
                "e2e_batch_us": 21058.084000000003,
                "e2e_single_us": 21482.833
              },
              "system": {
                "gpu_batch_us": 21589.2500128,
                "gpu_single_us": 22006.4999885,
                "e2e_batch_us": 23067.209,
                "e2e_single_us": 22983.25
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "28522d4512e17182f06a4fddc6fc68123d19f816534fc6136bad647dcec42c44",
            "paths": {
              "native": {
                "gpu_batch_us": 25642.7084096,
                "gpu_single_us": 25626.3750489,
                "e2e_batch_us": 26040.5,
                "e2e_single_us": 25238.833
              },
              "torch": {
                "gpu_batch_us": 20657.958288211375,
                "gpu_single_us": 20442.33330525458,
                "e2e_batch_us": 21148.083,
                "e2e_single_us": 21358.540999999997
              },
              "system": {
                "gpu_batch_us": 21990.1250093,
                "gpu_single_us": 22796.041681400002,
                "e2e_batch_us": 23073.667,
                "e2e_single_us": 22546.167
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "91593521ac7abf8881f1f26fc3a3ee7b670f32f8c819054adce85e13737c9816",
            "paths": {
              "native": {
                "gpu_batch_us": 29146.2500463,
                "gpu_single_us": 29930.2500207,
                "e2e_batch_us": 30399.917,
                "e2e_single_us": 30357.208
              },
              "torch": {
                "gpu_batch_us": 2656.5833250060678,
                "gpu_single_us": 2460.9583197161555,
                "e2e_batch_us": 2771.9791666666665,
                "e2e_single_us": 2896.625
              },
              "system": {
                "gpu_batch_us": 2805.0972226333333,
                "gpu_single_us": 2647.62499137,
                "e2e_batch_us": 2792.22916667,
                "e2e_single_us": 2988.5
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "fc87f86e71fbe4e9d8cad9957928b0fd0130e711a1126e98d5706cdc250c1a71",
            "paths": {
              "native": {
                "gpu_batch_us": 3782.03332656,
                "gpu_single_us": 3656.20834054,
                "e2e_batch_us": 3874.8916,
                "e2e_single_us": 4057.625
              },
              "torch": {
                "gpu_batch_us": 2616.5118989800767,
                "gpu_single_us": 2477.124973665923,
                "e2e_batch_us": 2728.851142857143,
                "e2e_single_us": 2919.125
              },
              "system": {
                "gpu_batch_us": 2805.6041725666664,
                "gpu_single_us": 2615.25000678,
                "e2e_batch_us": 2869.0695,
                "e2e_single_us": 3037.041
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "dadf52980273db52590a98cb00b9405de2c411d885ab6ee578880395fcbb6063",
            "paths": {
              "native": {
                "gpu_batch_us": 4169.9999855999995,
                "gpu_single_us": 4014.0000637599996,
                "e2e_batch_us": 4370.6355,
                "e2e_single_us": 4350.458
              },
              "torch": {
                "gpu_batch_us": 2592.1428626004076,
                "gpu_single_us": 2460.4166974313557,
                "e2e_batch_us": 2660.7024285714288,
                "e2e_single_us": 3011.333
              },
              "system": {
                "gpu_batch_us": 2708.6458382333335,
                "gpu_single_us": 2611.12500448,
                "e2e_batch_us": 2827.4515,
                "e2e_single_us": 3081.958
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "fc87f86e71fbe4e9d8cad9957928b0fd0130e711a1126e98d5706cdc250c1a71",
            "paths": {
              "native": {
                "gpu_batch_us": 3789.4416716799997,
                "gpu_single_us": 3654.9999495,
                "e2e_batch_us": 3845.675,
                "e2e_single_us": 4252.875
              },
              "torch": {
                "gpu_batch_us": 2626.9583364150353,
                "gpu_single_us": 2465.749974362552,
                "e2e_batch_us": 2675.5535714285716,
                "e2e_single_us": 3000.167
              },
              "system": {
                "gpu_batch_us": 2772.541668183333,
                "gpu_single_us": 2621.62496801,
                "e2e_batch_us": 2849.65283333,
                "e2e_single_us": 2985.166
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "a5eb38354235954c2a39d05e73ab678252a2706b59d7ae77104e5ba6acb565bf",
            "paths": {
              "native": {
                "gpu_batch_us": 777.9583344625,
                "gpu_single_us": 793.083338067,
                "e2e_batch_us": 792.074666667,
                "e2e_single_us": 988.292
              },
              "torch": {
                "gpu_batch_us": 384.00694393203594,
                "gpu_single_us": 383.50000977516174,
                "e2e_batch_us": 402.67272916666667,
                "e2e_single_us": 643.5
              },
              "system": {
                "gpu_batch_us": 482.294871174359,
                "gpu_single_us": 464.833283331,
                "e2e_batch_us": 497.111102564,
                "e2e_single_us": 727.167
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "a5eb38354235954c2a39d05e73ab678252a2706b59d7ae77104e5ba6acb565bf",
            "paths": {
              "native": {
                "gpu_batch_us": 782.762499875,
                "gpu_single_us": 766.375043895,
                "e2e_batch_us": 804.40415,
                "e2e_single_us": 982.834
              },
              "torch": {
                "gpu_batch_us": 385.503472595398,
                "gpu_single_us": 370.8750009536743,
                "e2e_batch_us": 428.48524999999995,
                "e2e_single_us": 749.417
              },
              "system": {
                "gpu_batch_us": 488.1260683769231,
                "gpu_single_us": 466.541678179,
                "e2e_batch_us": 501.753205128,
                "e2e_single_us": 688.875
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "cce0d09dc56b885a4e0e4303998af47133878c48812b49dd9048d6bcc7dd80a0",
            "paths": {
              "native": {
                "gpu_batch_us": 9499.020816299999,
                "gpu_single_us": 9644.54171481,
                "e2e_batch_us": 9806.1875,
                "e2e_single_us": 10053.916
              },
              "torch": {
                "gpu_batch_us": 384.4591817457457,
                "gpu_single_us": 377.83331936225295,
                "e2e_batch_us": 435.0008571428571,
                "e2e_single_us": 1246.1660000000002
              },
              "system": {
                "gpu_batch_us": 483.3995719820513,
                "gpu_single_us": 467.291625682,
                "e2e_batch_us": 496.228641026,
                "e2e_single_us": 701.833
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "a5eb38354235954c2a39d05e73ab678252a2706b59d7ae77104e5ba6acb565bf",
            "paths": {
              "native": {
                "gpu_batch_us": 777.6215279583333,
                "gpu_single_us": 783.416675404,
                "e2e_batch_us": 793.37325,
                "e2e_single_us": 1048.583
              },
              "torch": {
                "gpu_batch_us": 385.626700317145,
                "gpu_single_us": 379.4166259467602,
                "e2e_batch_us": 430.471081632653,
                "e2e_single_us": 992.042
              },
              "system": {
                "gpu_batch_us": 487.5641017666666,
                "gpu_single_us": 478.99998026,
                "e2e_batch_us": 500.352564103,
                "e2e_single_us": 708.291
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "f396b61f109fae66974d61e38ef04418e2088523fcfce0d02cea835170c85dcd",
            "paths": {
              "native": {
                "gpu_batch_us": 148.29362044109374,
                "gpu_single_us": 151.87502140200002,
                "e2e_batch_us": 151.374015748,
                "e2e_single_us": 424.334
              },
              "torch": {
                "gpu_batch_us": 16.114583559101447,
                "gpu_single_us": 23.79168290644884,
                "e2e_batch_us": 31.478901996370237,
                "e2e_single_us": 268.25
              },
              "system": {
                "gpu_batch_us": 15.65104230390625,
                "gpu_single_us": 17.3750449903,
                "e2e_batch_us": 16.3478347107,
                "e2e_single_us": 230.667
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "8ae10c1f3d2f687b2c8251424f37f46ac76b4c4fdd832cf2170fb8fa2409b53e",
            "paths": {
              "native": {
                "gpu_batch_us": 534.81713015,
                "gpu_single_us": 521.541631315,
                "e2e_batch_us": 544.538194444,
                "e2e_single_us": 793.333
              },
              "torch": {
                "gpu_batch_us": 13.456380656862166,
                "gpu_single_us": 19.12500010803342,
                "e2e_batch_us": 31.713717228464418,
                "e2e_single_us": 296.541
              },
              "system": {
                "gpu_batch_us": 14.754557923875,
                "gpu_single_us": 17.0000130311,
                "e2e_batch_us": 16.008618849,
                "e2e_single_us": 247.542
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "f396b61f109fae66974d61e38ef04418e2088523fcfce0d02cea835170c85dcd",
            "paths": {
              "native": {
                "gpu_batch_us": 147.97981748421876,
                "gpu_single_us": 152.66664559,
                "e2e_batch_us": 150.243907692,
                "e2e_single_us": 382.209
              },
              "torch": {
                "gpu_batch_us": 13.758462955593131,
                "gpu_single_us": 20.33333294093609,
                "e2e_batch_us": 30.820649659863943,
                "e2e_single_us": 243.75
              },
              "system": {
                "gpu_batch_us": 15.003906810312499,
                "gpu_single_us": 17.249956727,
                "e2e_batch_us": 16.9508481481,
                "e2e_single_us": 248.042
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "f396b61f109fae66974d61e38ef04418e2088523fcfce0d02cea835170c85dcd",
            "paths": {
              "native": {
                "gpu_batch_us": 148.7057297709375,
                "gpu_single_us": 151.875079609,
                "e2e_batch_us": 150.900134921,
                "e2e_single_us": 369.75
              },
              "torch": {
                "gpu_batch_us": 15.316406461352017,
                "gpu_single_us": 23.916654754430056,
                "e2e_batch_us": 31.107065466448443,
                "e2e_single_us": 262.542
              },
              "system": {
                "gpu_batch_us": 14.946614101062499,
                "gpu_single_us": 17.2916334122,
                "e2e_batch_us": 15.7816581265,
                "e2e_single_us": 239.833
              }
            }
          }
        ],
        "rejections": [],
        "outputs": 60,
        "elements": 327750192
      },
      "old-reverse": {
        "measurements": [
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 282.272136246875,
                "gpu_single_us": 270.750024356,
                "e2e_batch_us": 284.945470588,
                "e2e_single_us": 482.417
              },
              "torch": {
                "gpu_batch_us": 288.9967463488574,
                "gpu_single_us": 275.7500042207539,
                "e2e_batch_us": 306.324875,
                "e2e_single_us": 582.5
              },
              "system": {
                "gpu_batch_us": 282.54101562187503,
                "gpu_single_us": 269.500014838,
                "e2e_batch_us": 291.197606061,
                "e2e_single_us": 486.416
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 280.1601567625,
                "gpu_single_us": 271.250028163,
                "e2e_batch_us": 285.674242424,
                "e2e_single_us": 577.542
              },
              "torch": {
                "gpu_batch_us": 291.621639213014,
                "gpu_single_us": 289.1666372306645,
                "e2e_batch_us": 304.1565806451613,
                "e2e_single_us": 715.667
              },
              "system": {
                "gpu_batch_us": 282.88867179062504,
                "gpu_single_us": 271.749973763,
                "e2e_batch_us": 291.767953846,
                "e2e_single_us": 600.708
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 675.305555988889,
                "gpu_single_us": 685.666687787,
                "e2e_batch_us": 704.117259259,
                "e2e_single_us": 927.958
              },
              "torch": {
                "gpu_batch_us": 289.2834695483573,
                "gpu_single_us": 280.41668701916933,
                "e2e_batch_us": 323.6454918032787,
                "e2e_single_us": 693.125
              },
              "system": {
                "gpu_batch_us": 284.103515696875,
                "gpu_single_us": 274.124962743,
                "e2e_batch_us": 289.174353846,
                "e2e_single_us": 487.208
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 279.823568234375,
                "gpu_single_us": 270.54164092999997,
                "e2e_batch_us": 283.60261194,
                "e2e_single_us": 499.25
              },
              "torch": {
                "gpu_batch_us": 287.0703137887176,
                "gpu_single_us": 288.125011138618,
                "e2e_batch_us": 302.77243076923077,
                "e2e_single_us": 566.75
              },
              "system": {
                "gpu_batch_us": 278.9707032203125,
                "gpu_single_us": 272.62500952899995,
                "e2e_batch_us": 288.716657143,
                "e2e_single_us": 482.25
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 140384.08332300003,
                "gpu_single_us": 140389.29168599998,
                "e2e_batch_us": 141674.208,
                "e2e_single_us": 141617.792
              },
              "torch": {
                "gpu_batch_us": 20238.666678778827,
                "gpu_single_us": 20113.87498350814,
                "e2e_batch_us": 20746.917,
                "e2e_single_us": 20634.5
              },
              "system": {
                "gpu_batch_us": 22214.7916793,
                "gpu_single_us": 22323.1249838,
                "e2e_batch_us": 23611.875,
                "e2e_single_us": 22455.209
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 142011.25002,
                "gpu_single_us": 143783.291685,
                "e2e_batch_us": 142912,
                "e2e_single_us": 142261.625
              },
              "torch": {
                "gpu_batch_us": 20752.58333934471,
                "gpu_single_us": 20230.250025633723,
                "e2e_batch_us": 20695.708,
                "e2e_single_us": 21454.709000000003
              },
              "system": {
                "gpu_batch_us": 22216.0833655,
                "gpu_single_us": 22212.6250155,
                "e2e_batch_us": 23479.083,
                "e2e_single_us": 23354.791
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15308.0416785,
                "gpu_single_us": 14689.7500381,
                "e2e_batch_us": 15634.1875,
                "e2e_single_us": 16540.833
              },
              "torch": {
                "gpu_batch_us": 2637.138890956218,
                "gpu_single_us": 2447.0833595842123,
                "e2e_batch_us": 2690.5973333333336,
                "e2e_single_us": 2851.125
              },
              "system": {
                "gpu_batch_us": 2774.145832516667,
                "gpu_single_us": 2605.8749645,
                "e2e_batch_us": 2873.84716667,
                "e2e_single_us": 3225.25
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15279.1875007,
                "gpu_single_us": 14884.0416805,
                "e2e_batch_us": 15610.625,
                "e2e_single_us": 16267.916
              },
              "torch": {
                "gpu_batch_us": 2624.0000068875297,
                "gpu_single_us": 2499.750000424683,
                "e2e_batch_us": 2701.029714285715,
                "e2e_single_us": 2989.375
              },
              "system": {
                "gpu_batch_us": 2811.9464238571427,
                "gpu_single_us": 2604.49998314,
                "e2e_batch_us": 2901.04771429,
                "e2e_single_us": 3167.542
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2126.1562506,
                "gpu_single_us": 2208.83334987,
                "e2e_batch_us": 2158.84375,
                "e2e_single_us": 2275.917
              },
              "torch": {
                "gpu_batch_us": 385.6560277277009,
                "gpu_single_us": 376.2500127777457,
                "e2e_batch_us": 403.40425531914894,
                "e2e_single_us": 728.708
              },
              "system": {
                "gpu_batch_us": 488.58881621578945,
                "gpu_single_us": 474.62497605,
                "e2e_batch_us": 499.608552632,
                "e2e_single_us": 665.125
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2143.3101889333334,
                "gpu_single_us": 2221.79165576,
                "e2e_batch_us": 2146.44911111,
                "e2e_single_us": 2271.375
              },
              "torch": {
                "gpu_batch_us": 386.8577908218393,
                "gpu_single_us": 374.4999994523823,
                "e2e_batch_us": 404.6721086956521,
                "e2e_single_us": 646.4159999999999
              },
              "system": {
                "gpu_batch_us": 484.5128203615385,
                "gpu_single_us": 471.500039566,
                "e2e_batch_us": 495.208333333,
                "e2e_single_us": 689.125
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.608723474999998,
                "gpu_single_us": 38.2916769013,
                "e2e_batch_us": 33.3017811966,
                "e2e_single_us": 269.125
              },
              "torch": {
                "gpu_batch_us": 15.324869309552014,
                "gpu_single_us": 23.666711058467627,
                "e2e_batch_us": 30.756215254237283,
                "e2e_single_us": 287.208
              },
              "system": {
                "gpu_batch_us": 15.108723346203126,
                "gpu_single_us": 17.3333100975,
                "e2e_batch_us": 15.9527327935,
                "e2e_single_us": 266.458
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.0618488584375,
                "gpu_single_us": 37.916644942,
                "e2e_batch_us": 33.0719237435,
                "e2e_single_us": 254.167
              },
              "torch": {
                "gpu_batch_us": 16.256511116807815,
                "gpu_single_us": 24.62498378008604,
                "e2e_batch_us": 29.18213696369637,
                "e2e_single_us": 290.833
              },
              "system": {
                "gpu_batch_us": 14.90885460953125,
                "gpu_single_us": 17.0416897163,
                "e2e_batch_us": 15.8002763731,
                "e2e_single_us": 240.416
              }
            }
          }
        ],
        "rejections": [
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          }
        ],
        "outputs": 36,
        "elements": 170166552
      }
    },
    "comparisons": [
      {
        "shape": [
          129,
          257,
          61
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 31.085286536875,
            "gpu_single_us": 38.458325434500004,
            "e2e_batch_us": 33.3077116935,
            "e2e_single_us": 280.25
          },
          "torch": {
            "gpu_batch_us": 14.813801499258261,
            "gpu_single_us": 20.666630007326603,
            "e2e_batch_us": 30.267805128205122,
            "e2e_single_us": 268.25
          },
          "system": {
            "gpu_batch_us": 13.917969226890625,
            "gpu_single_us": 17.0833664015,
            "e2e_batch_us": 15.7748377425,
            "e2e_single_us": 243.708
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 148.51432297296876,
            "gpu_single_us": 153.458327986,
            "e2e_batch_us": 152.268047244,
            "e2e_single_us": 441.667
          },
          "torch": {
            "gpu_batch_us": 13.980467883811798,
            "gpu_single_us": 24.749955628067255,
            "e2e_batch_us": 36.875563685636855,
            "e2e_single_us": 245.583
          },
          "system": {
            "gpu_batch_us": 15.185546544671876,
            "gpu_single_us": 17.374986782700002,
            "e2e_batch_us": 16.028045264,
            "e2e_single_us": 247.916
          }
        },
        "new_over_old": {
          "gpu_batch_us": 4.777640469769943,
          "gpu_single_us": 3.9902498679346126,
          "e2e_batch_us": 4.571555339651721,
          "e2e_single_us": 1.5759750223015163
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 17.307237110762728,
          "gpu_single_us": 13.56244340391971,
          "e2e_batch_us": 16.517828915078134,
          "e2e_single_us": 3.1003326732673266
        },
        "new_over_mps": {
          "gpu_batch_us": 9.779978780222283,
          "gpu_single_us": 8.832140703485083,
          "e2e_batch_us": 9.500100900388873,
          "e2e_single_us": 1.781518740218461
        },
        "new_over_torch": {
          "gpu_batch_us": 10.62298659867713,
          "gpu_single_us": 6.200347600299262,
          "e2e_batch_us": 4.129239854937021,
          "e2e_single_us": 1.7984428889621837
        }
      },
      {
        "shape": [
          129,
          257,
          61
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 31.0618488584375,
            "gpu_single_us": 37.916644942,
            "e2e_batch_us": 33.0719237435,
            "e2e_single_us": 254.167
          },
          "torch": {
            "gpu_batch_us": 16.256511116807815,
            "gpu_single_us": 24.62498378008604,
            "e2e_batch_us": 29.18213696369637,
            "e2e_single_us": 290.833
          },
          "system": {
            "gpu_batch_us": 14.90885460953125,
            "gpu_single_us": 17.0416897163,
            "e2e_batch_us": 15.8002763731,
            "e2e_single_us": 240.416
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 148.7057297709375,
            "gpu_single_us": 151.875079609,
            "e2e_batch_us": 150.900134921,
            "e2e_single_us": 369.75
          },
          "torch": {
            "gpu_batch_us": 15.316406461352017,
            "gpu_single_us": 23.916654754430056,
            "e2e_batch_us": 31.107065466448443,
            "e2e_single_us": 262.542
          },
          "system": {
            "gpu_batch_us": 14.946614101062499,
            "gpu_single_us": 17.2916334122,
            "e2e_batch_us": 15.7816581265,
            "e2e_single_us": 239.833
          }
        },
        "new_over_old": {
          "gpu_batch_us": 4.787407550936678,
          "gpu_single_us": 4.005498900056135,
          "e2e_batch_us": 4.562786733888079,
          "e2e_single_us": 1.4547521904889305
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 16.91992182389137,
          "gpu_single_us": 13.620234826991704,
          "e2e_batch_us": 16.35162369331751,
          "e2e_single_us": 2.947823502090107
        },
        "new_over_mps": {
          "gpu_batch_us": 9.949124849645148,
          "gpu_single_us": 8.783154025336064,
          "e2e_batch_us": 9.56174146667224,
          "e2e_single_us": 1.5416977646946
        },
        "new_over_torch": {
          "gpu_batch_us": 9.708917698558574,
          "gpu_single_us": 6.350180707478263,
          "e2e_batch_us": 4.850992295745748,
          "e2e_single_us": 1.4083460931965173
        }
      },
      {
        "shape": [
          1025,
          1025,
          1024
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 2118.9843828375,
            "gpu_single_us": 2180.5000142199997,
            "e2e_batch_us": 2211.328125,
            "e2e_single_us": 2322.459
          },
          "torch": {
            "gpu_batch_us": 386.2056740004807,
            "gpu_single_us": 373.50005004554987,
            "e2e_batch_us": 418.5035531914894,
            "e2e_single_us": 668.375
          },
          "system": {
            "gpu_batch_us": 486.58653843076917,
            "gpu_single_us": 466.416706331,
            "e2e_batch_us": 504.939102564,
            "e2e_single_us": 695.959
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 772.2916658791667,
            "gpu_single_us": 784.374948125,
            "e2e_batch_us": 796.387166667,
            "e2e_single_us": 1011.416
          },
          "torch": {
            "gpu_batch_us": 388.6422112017222,
            "gpu_single_us": 374.8750314116478,
            "e2e_batch_us": 404.5507173913043,
            "e2e_single_us": 711.5409999999999
          },
          "system": {
            "gpu_batch_us": 484.5106830974359,
            "gpu_single_us": 468.374986667,
            "e2e_batch_us": 497.743589744,
            "e2e_single_us": 728.333
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.3644631230575671,
          "gpu_single_us": 0.3597225145653501,
          "e2e_batch_us": 0.3601397538716693,
          "e2e_single_us": 0.4354935867543841
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 4.596843234370541,
          "gpu_single_us": 4.418115676596222,
          "e2e_batch_us": 4.486033399789911,
          "e2e_single_us": 3.8608711024769304
        },
        "new_over_mps": {
          "gpu_batch_us": 1.5939620999519994,
          "gpu_single_us": 1.6746730087075852,
          "e2e_batch_us": 1.5999948227893777,
          "e2e_single_us": 1.3886724890949609
        },
        "new_over_torch": {
          "gpu_batch_us": 1.987153334402767,
          "gpu_single_us": 2.0923638076703033,
          "e2e_batch_us": 1.9685718809310362,
          "e2e_single_us": 1.421444442414422
        }
      },
      {
        "shape": [
          1025,
          1025,
          1024
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 2143.3101889333334,
            "gpu_single_us": 2221.79165576,
            "e2e_batch_us": 2146.44911111,
            "e2e_single_us": 2271.375
          },
          "torch": {
            "gpu_batch_us": 386.8577908218393,
            "gpu_single_us": 374.4999994523823,
            "e2e_batch_us": 404.6721086956521,
            "e2e_single_us": 646.4159999999999
          },
          "system": {
            "gpu_batch_us": 484.5128203615385,
            "gpu_single_us": 471.500039566,
            "e2e_batch_us": 495.208333333,
            "e2e_single_us": 689.125
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 777.6215279583333,
            "gpu_single_us": 783.416675404,
            "e2e_batch_us": 793.37325,
            "e2e_single_us": 1048.583
          },
          "torch": {
            "gpu_batch_us": 385.626700317145,
            "gpu_single_us": 379.4166259467602,
            "e2e_batch_us": 430.471081632653,
            "e2e_single_us": 992.042
          },
          "system": {
            "gpu_batch_us": 487.5641017666666,
            "gpu_single_us": 478.99998026,
            "e2e_batch_us": 500.352564103,
            "e2e_single_us": 708.291
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.3628133398392205,
          "gpu_single_us": 0.35260582304060345,
          "e2e_batch_us": 0.36962127165908926,
          "e2e_single_us": 0.461651202465467
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 4.467696489201761,
          "gpu_single_us": 4.366351003971226,
          "e2e_batch_us": 4.54233313550367,
          "e2e_single_us": 4.417523134631008
        },
        "new_over_mps": {
          "gpu_batch_us": 1.594911366814449,
          "gpu_single_us": 1.6355254857813635,
          "e2e_batch_us": 1.5856284286707087,
          "e2e_single_us": 1.480440948706111
        },
        "new_over_torch": {
          "gpu_batch_us": 2.016513709550729,
          "gpu_single_us": 2.064792689168896,
          "e2e_batch_us": 1.8430349536860025,
          "e2e_single_us": 1.056994562730207
        }
      },
      {
        "shape": [
          2049,
          4097,
          1025
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 15551.58331755,
            "gpu_single_us": 14502.4583326,
            "e2e_batch_us": 15948.0625,
            "e2e_single_us": 16643.208
          },
          "torch": {
            "gpu_batch_us": 2658.4107129435456,
            "gpu_single_us": 2501.6666622832417,
            "e2e_batch_us": 2768.041571428571,
            "e2e_single_us": 2888.417
          },
          "system": {
            "gpu_batch_us": 2781.1874946,
            "gpu_single_us": 2627.58333702,
            "e2e_batch_us": 2915.82633333,
            "e2e_single_us": 2976.208
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 3833.77498948,
            "gpu_single_us": 3657.25002484,
            "e2e_batch_us": 3884.6584,
            "e2e_single_us": 4247.25
          },
          "torch": {
            "gpu_batch_us": 2631.904756916421,
            "gpu_single_us": 2443.333331029862,
            "e2e_batch_us": 2691.345285714286,
            "e2e_single_us": 3000.458
          },
          "system": {
            "gpu_batch_us": 2805.2291696999996,
            "gpu_single_us": 2596.1250066800003,
            "e2e_batch_us": 2815.5,
            "e2e_single_us": 3090.416
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.2465199144805774,
          "gpu_single_us": 0.25218138476694585,
          "e2e_batch_us": 0.2435818394867715,
          "e2e_single_us": 0.2551941909276145
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 1.9345168704194748,
          "gpu_single_us": 1.9985403171502796,
          "e2e_batch_us": 1.925266587842151,
          "e2e_single_us": 1.850014363825518
        },
        "new_over_mps": {
          "gpu_batch_us": 1.3666530459933854,
          "gpu_single_us": 1.4087341770637605,
          "e2e_batch_us": 1.3797401527259812,
          "e2e_single_us": 1.37432954010075
        },
        "new_over_torch": {
          "gpu_batch_us": 1.4566541511067856,
          "gpu_single_us": 1.496828115261078,
          "e2e_batch_us": 1.443389081519879,
          "e2e_single_us": 1.4155338951586725
        }
      },
      {
        "shape": [
          2049,
          4097,
          1025
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 15279.1875007,
            "gpu_single_us": 14884.0416805,
            "e2e_batch_us": 15610.625,
            "e2e_single_us": 16267.916
          },
          "torch": {
            "gpu_batch_us": 2624.0000068875297,
            "gpu_single_us": 2499.750000424683,
            "e2e_batch_us": 2701.029714285715,
            "e2e_single_us": 2989.375
          },
          "system": {
            "gpu_batch_us": 2811.9464238571427,
            "gpu_single_us": 2604.49998314,
            "e2e_batch_us": 2901.04771429,
            "e2e_single_us": 3167.542
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 3789.4416716799997,
            "gpu_single_us": 3654.9999495,
            "e2e_batch_us": 3845.675,
            "e2e_single_us": 4252.875
          },
          "torch": {
            "gpu_batch_us": 2626.9583364150353,
            "gpu_single_us": 2465.749974362552,
            "e2e_batch_us": 2675.5535714285716,
            "e2e_single_us": 3000.167
          },
          "system": {
            "gpu_batch_us": 2772.541668183333,
            "gpu_single_us": 2621.62496801,
            "e2e_batch_us": 2849.65283333,
            "e2e_single_us": 2985.166
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.24801329727162458,
          "gpu_single_us": 0.24556501708057685,
          "e2e_batch_us": 0.24634984185450615,
          "e2e_single_us": 0.26142715514390413
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 1.9039829299155642,
          "gpu_single_us": 2.0374921249899796,
          "e2e_batch_us": 1.9444513506058438,
          "e2e_single_us": 1.835288948265181
        },
        "new_over_mps": {
          "gpu_batch_us": 1.3667753726359595,
          "gpu_single_us": 1.3941734588660122,
          "e2e_batch_us": 1.3495240385145741,
          "e2e_single_us": 1.4246695158661193
        },
        "new_over_torch": {
          "gpu_batch_us": 1.4425206594069495,
          "gpu_single_us": 1.4823076092477276,
          "e2e_batch_us": 1.4373380675561132,
          "e2e_single_us": 1.417546089934327
        }
      },
      {
        "shape": [
          4097,
          4097,
          4096
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 141349.458368,
            "gpu_single_us": 142939.916637,
            "e2e_batch_us": 142139.875,
            "e2e_single_us": 141180.583
          },
          "torch": {
            "gpu_batch_us": 20251.37498276308,
            "gpu_single_us": 20470.33334383741,
            "e2e_batch_us": 21237.75,
            "e2e_single_us": 20479.625
          },
          "system": {
            "gpu_batch_us": 21973.250026299997,
            "gpu_single_us": 21936.3750075,
            "e2e_batch_us": 22632.375,
            "e2e_single_us": 22640.75
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 25484.5833406,
            "gpu_single_us": 25578.166707400003,
            "e2e_batch_us": 26021.333,
            "e2e_single_us": 26643.208
          },
          "torch": {
            "gpu_batch_us": 20286.0000426881,
            "gpu_single_us": 20623.625023290515,
            "e2e_batch_us": 21299.291,
            "e2e_single_us": 21074.958
          },
          "system": {
            "gpu_batch_us": 22143.4999839,
            "gpu_single_us": 22166.5416611,
            "e2e_batch_us": 22366.208,
            "e2e_single_us": 22908.791
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.18029487792058943,
          "gpu_single_us": 0.17894348415185166,
          "e2e_batch_us": 0.18306849502998365,
          "e2e_single_us": 0.1887172260791698
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.8823575480460333,
          "gpu_single_us": 0.8823642564675852,
          "e2e_batch_us": 0.8776758864966675,
          "e2e_single_us": 0.8900615323907575
        },
        "new_over_mps": {
          "gpu_batch_us": 1.150883255091978,
          "gpu_single_us": 1.153908764770783,
          "e2e_batch_us": 1.1634217566071101,
          "e2e_single_us": 1.1630123999123305
        },
        "new_over_torch": {
          "gpu_batch_us": 1.2562645808425736,
          "gpu_single_us": 1.2402362183425204,
          "e2e_batch_us": 1.2216994922507043,
          "e2e_single_us": 1.264211677195276
        }
      },
      {
        "shape": [
          4097,
          4097,
          4096
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 142011.25002,
            "gpu_single_us": 143783.291685,
            "e2e_batch_us": 142912,
            "e2e_single_us": 142261.625
          },
          "torch": {
            "gpu_batch_us": 20752.58333934471,
            "gpu_single_us": 20230.250025633723,
            "e2e_batch_us": 20695.708,
            "e2e_single_us": 21454.709000000003
          },
          "system": {
            "gpu_batch_us": 22216.0833655,
            "gpu_single_us": 22212.6250155,
            "e2e_batch_us": 23479.083,
            "e2e_single_us": 23354.791
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 25642.7084096,
            "gpu_single_us": 25626.3750489,
            "e2e_batch_us": 26040.5,
            "e2e_single_us": 25238.833
          },
          "torch": {
            "gpu_batch_us": 20657.958288211375,
            "gpu_single_us": 20442.33330525458,
            "e2e_batch_us": 21148.083,
            "e2e_single_us": 21358.540999999997
          },
          "system": {
            "gpu_batch_us": 21990.1250093,
            "gpu_single_us": 22796.041681400002,
            "e2e_batch_us": 23073.667,
            "e2e_single_us": 22546.167
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.18056814798819557,
          "gpu_single_us": 0.17822915826021138,
          "e2e_batch_us": 0.1822135300044783,
          "e2e_single_us": 0.17741139256633684
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.9046200038276969,
          "gpu_single_us": 0.8773155472817478,
          "e2e_batch_us": 0.8835429664092421,
          "e2e_single_us": 0.8885618411562299
        },
        "new_over_mps": {
          "gpu_batch_us": 1.1661010748577034,
          "gpu_single_us": 1.1241589836980057,
          "e2e_batch_us": 1.1285809056705203,
          "e2e_single_us": 1.1194289920765688
        },
        "new_over_torch": {
          "gpu_batch_us": 1.2412992635498354,
          "gpu_single_us": 1.2535934458280698,
          "e2e_batch_us": 1.2313409210659898,
          "e2e_single_us": 1.18167401977504
        }
      },
      {
        "shape": [
          1024,
          1024,
          1024
        ],
        "order": "forward",
        "old_bk": 1024,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 278.5169281203125,
            "gpu_single_us": 268.66665575699994,
            "e2e_batch_us": 282.870228571,
            "e2e_single_us": 533.292
          },
          "torch": {
            "gpu_batch_us": 287.17063740634967,
            "gpu_single_us": 275.749946013093,
            "e2e_batch_us": 302.24007936507934,
            "e2e_single_us": 536.25
          },
          "system": {
            "gpu_batch_us": 279.7792967609375,
            "gpu_single_us": 270.12499049300004,
            "e2e_batch_us": 287.019230769,
            "e2e_single_us": 524.75
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 279.764322840625,
            "gpu_single_us": 270.16666717799995,
            "e2e_batch_us": 284.485294118,
            "e2e_single_us": 520.25
          },
          "torch": {
            "gpu_batch_us": 287.32682221743744,
            "gpu_single_us": 276.83330699801445,
            "e2e_batch_us": 303.00641538461537,
            "e2e_single_us": 584.625
          },
          "system": {
            "gpu_batch_us": 280.9654943125,
            "gpu_single_us": 269.291689619,
            "e2e_batch_us": 289.859304348,
            "e2e_single_us": 512.959
          }
        },
        "new_over_old": {
          "gpu_batch_us": 1.0044787034265066,
          "gpu_single_us": 1.0055831692875081,
          "e2e_batch_us": 1.0057095635520181,
          "e2e_single_us": 0.9755443546874882
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.99484536041307,
          "gpu_single_us": 1.0016145469536615,
          "e2e_batch_us": 0.9816328014686039,
          "e2e_single_us": 0.9661726658623293
        },
        "new_over_mps": {
          "gpu_batch_us": 0.9957248434551931,
          "gpu_single_us": 1.00324918143682,
          "e2e_batch_us": 0.9814599353914544,
          "e2e_single_us": 1.0142136116141838
        },
        "new_over_torch": {
          "gpu_batch_us": 0.9736798001716338,
          "gpu_single_us": 0.9759182163002434,
          "e2e_batch_us": 0.9388754814213887,
          "e2e_single_us": 0.889886679495403
        }
      },
      {
        "shape": [
          1024,
          1024,
          1024
        ],
        "order": "reverse",
        "old_bk": 1024,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 279.823568234375,
            "gpu_single_us": 270.54164092999997,
            "e2e_batch_us": 283.60261194,
            "e2e_single_us": 499.25
          },
          "torch": {
            "gpu_batch_us": 287.0703137887176,
            "gpu_single_us": 288.125011138618,
            "e2e_batch_us": 302.77243076923077,
            "e2e_single_us": 566.75
          },
          "system": {
            "gpu_batch_us": 278.9707032203125,
            "gpu_single_us": 272.62500952899995,
            "e2e_batch_us": 288.716657143,
            "e2e_single_us": 482.25
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 279.563801996875,
            "gpu_single_us": 272.958364803,
            "e2e_batch_us": 286.30025,
            "e2e_single_us": 486.583
          },
          "torch": {
            "gpu_batch_us": 286.8281235350878,
            "gpu_single_us": 276.2916265055537,
            "e2e_batch_us": 306.6038461538462,
            "e2e_single_us": 567.4169999999999
          },
          "system": {
            "gpu_batch_us": 280.1699220078125,
            "gpu_single_us": 269.95834196,
            "e2e_batch_us": 286.025,
            "e2e_single_us": 469.084
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.9990716784896316,
          "gpu_single_us": 1.0089329090512367,
          "e2e_batch_us": 1.0095120353142966,
          "e2e_single_us": 0.9746279419128694
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 1.0025377981908417,
          "gpu_single_us": 0.9837748343237932,
          "e2e_batch_us": 0.9842617904514058,
          "e2e_single_us": 0.9643938626532667
        },
        "new_over_mps": {
          "gpu_batch_us": 0.9978365985663493,
          "gpu_single_us": 1.011112910315046,
          "e2e_batch_us": 1.0009623284677913,
          "e2e_single_us": 1.037304619215322
        },
        "new_over_torch": {
          "gpu_batch_us": 0.9746736078433251,
          "gpu_single_us": 0.9879357121867509,
          "e2e_batch_us": 0.9337790559102825,
          "e2e_single_us": 0.8575403979789116
        }
      }
    ],
    "artifacts": 35,
    "validated_outputs": 192,
    "checked_elements": 995833488,
    "rejected_requests": 16,
    "current_artifacts_checked": false
  },
  "shuffle-replay": {
    "runs": {
      "old-forward": {
        "measurements": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.729166039375,
                "gpu_single_us": 38.0832934752,
                "e2e_batch_us": 33.6676609881,
                "e2e_single_us": 270.042
              },
              "torch": {
                "gpu_batch_us": 15.42578229418723,
                "gpu_single_us": 23.500004317611456,
                "e2e_batch_us": 29.04783464566929,
                "e2e_single_us": 280.583
              },
              "system": {
                "gpu_batch_us": 14.781249774390623,
                "gpu_single_us": 17.541693523499998,
                "e2e_batch_us": 16.6369724026,
                "e2e_single_us": 243.375
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.995442441250002,
                "gpu_single_us": 37.875026464499996,
                "e2e_batch_us": 33.6454501018,
                "e2e_single_us": 264.292
              },
              "torch": {
                "gpu_batch_us": 22.968099074205384,
                "gpu_single_us": 30.125025659799576,
                "e2e_batch_us": 29.399450160771703,
                "e2e_single_us": 367.08299999999997
              },
              "system": {
                "gpu_batch_us": 14.986328096718749,
                "gpu_single_us": 17.5000168383,
                "e2e_batch_us": 16.6094674556,
                "e2e_single_us": 267.75
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2072.3177076375,
                "gpu_single_us": 2047.91669967,
                "e2e_batch_us": 2201.864625,
                "e2e_single_us": 2319.083
              },
              "torch": {
                "gpu_batch_us": 387.72129660679235,
                "gpu_single_us": 374.95832657441497,
                "e2e_batch_us": 406.5703777777778,
                "e2e_single_us": 657.875
              },
              "system": {
                "gpu_batch_us": 483.05592223947366,
                "gpu_single_us": 470.041646622,
                "e2e_batch_us": 508.487921053,
                "e2e_single_us": 678.708
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2122.5833406750003,
                "gpu_single_us": 2236.6665653,
                "e2e_batch_us": 2113.61975,
                "e2e_single_us": 2407.333
              },
              "torch": {
                "gpu_batch_us": 385.9535976186056,
                "gpu_single_us": 375.04167994484305,
                "e2e_batch_us": 404.42709090909096,
                "e2e_single_us": 718.6669999999999
              },
              "system": {
                "gpu_batch_us": 486.6378208717948,
                "gpu_single_us": 466.041616164,
                "e2e_batch_us": 498.787384615,
                "e2e_single_us": 694.5
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 14841.4999712,
                "gpu_single_us": 14883.875032,
                "e2e_batch_us": 16453.208,
                "e2e_single_us": 16413.542
              },
              "torch": {
                "gpu_batch_us": 2624.1428568027914,
                "gpu_single_us": 2469.5833562873304,
                "e2e_batch_us": 2700.4464285714284,
                "e2e_single_us": 3026.916
              },
              "system": {
                "gpu_batch_us": 2816.8263961499997,
                "gpu_single_us": 2678.5833179,
                "e2e_batch_us": 2811.1945,
                "e2e_single_us": 3017.792
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15136.3750338,
                "gpu_single_us": 14775.4166974,
                "e2e_batch_us": 15938.083,
                "e2e_single_us": 16249.5
              },
              "torch": {
                "gpu_batch_us": 2609.839288717402,
                "gpu_single_us": 2448.0416905134916,
                "e2e_batch_us": 2776.226142857143,
                "e2e_single_us": 3061.4159999999997
              },
              "system": {
                "gpu_batch_us": 2797.6249985,
                "gpu_single_us": 2643.5833424300004,
                "e2e_batch_us": 2857.9305,
                "e2e_single_us": 3005.625
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 136471.791659,
                "gpu_single_us": 134053.750022,
                "e2e_batch_us": 139222.333,
                "e2e_single_us": 138383.542
              },
              "torch": {
                "gpu_batch_us": 20320.458337664604,
                "gpu_single_us": 20670.458383392543,
                "e2e_batch_us": 20626.958,
                "e2e_single_us": 21476.165999999997
              },
              "system": {
                "gpu_batch_us": 22336.4999983,
                "gpu_single_us": 21686.4999966,
                "e2e_batch_us": 22772.166,
                "e2e_single_us": 22435.209
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 140672.250011,
                "gpu_single_us": 142828.374985,
                "e2e_batch_us": 144142.375,
                "e2e_single_us": 141140.208
              },
              "torch": {
                "gpu_batch_us": 20189.875038340688,
                "gpu_single_us": 20586.7916229181,
                "e2e_batch_us": 21034.25,
                "e2e_single_us": 21054.208
              },
              "system": {
                "gpu_batch_us": 22332.916618299998,
                "gpu_single_us": 22008.2500018,
                "e2e_batch_us": 22714.25,
                "e2e_single_us": 23061.333
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 278.8014317046875,
                "gpu_single_us": 268.374977168,
                "e2e_batch_us": 283.219537313,
                "e2e_single_us": 485.875
              },
              "torch": {
                "gpu_batch_us": 287.40039215335855,
                "gpu_single_us": 276.50000993162394,
                "e2e_batch_us": 302.6942307692308,
                "e2e_single_us": 559.917
              },
              "system": {
                "gpu_batch_us": 280.328124971875,
                "gpu_single_us": 272.000033874,
                "e2e_batch_us": 286.255597015,
                "e2e_single_us": 533.083
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 280.826171546875,
                "gpu_single_us": 284.66671938099995,
                "e2e_batch_us": 290.362132353,
                "e2e_single_us": 513.667
              },
              "torch": {
                "gpu_batch_us": 294.131289438327,
                "gpu_single_us": 290.6249719671905,
                "e2e_batch_us": 312.7806603773585,
                "e2e_single_us": 621.625
              },
              "system": {
                "gpu_batch_us": 279.83137988437505,
                "gpu_single_us": 272.541656159,
                "e2e_batch_us": 286.987927536,
                "e2e_single_us": 483.792
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 680.230655625,
                "gpu_single_us": 671.708374284,
                "e2e_batch_us": 696.151785714,
                "e2e_single_us": 921.917
              },
              "torch": {
                "gpu_batch_us": 289.58531477237267,
                "gpu_single_us": 290.87497387081385,
                "e2e_batch_us": 302.24736507936507,
                "e2e_single_us": 568.375
              },
              "system": {
                "gpu_batch_us": 282.52994798125,
                "gpu_single_us": 269.166659564,
                "e2e_batch_us": 290.449884058,
                "e2e_single_us": 484.583
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 282.3645836546875,
                "gpu_single_us": 270.291639026,
                "e2e_batch_us": 287.064338235,
                "e2e_single_us": 502.5
              },
              "torch": {
                "gpu_batch_us": 288.31317397435345,
                "gpu_single_us": 275.2500004135072,
                "e2e_batch_us": 304.55914516129036,
                "e2e_single_us": 628.166
              },
              "system": {
                "gpu_batch_us": 284.4264327,
                "gpu_single_us": 273.87501904699997,
                "e2e_batch_us": 293.390926471,
                "e2e_single_us": 568
              }
            }
          }
        ],
        "rejections": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          }
        ],
        "outputs": 36,
        "elements": 170166552
      },
      "new-forward": {
        "measurements": [
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "8f1860ab6c09d536a585ab12cd926e89abdf5488fdd70937ce1f8afaafe7f271",
            "paths": {
              "native": {
                "gpu_batch_us": 53.6796869709375,
                "gpu_single_us": 65.8750068396,
                "e2e_batch_us": 57.0112068966,
                "e2e_single_us": 316.584
              },
              "torch": {
                "gpu_batch_us": 17.969400687434245,
                "gpu_single_us": 24.500011932104826,
                "e2e_batch_us": 30.548894736842108,
                "e2e_single_us": 355.834
              },
              "system": {
                "gpu_batch_us": 14.992839169281249,
                "gpu_single_us": 17.0416897163,
                "e2e_batch_us": 16.1847262895,
                "e2e_single_us": 254.542
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "904f27be631b58b334527ed89cd1a4e5ee27c52c1a94370dd9b8834808ae1cf4",
            "paths": {
              "native": {
                "gpu_batch_us": 24.95312583046875,
                "gpu_single_us": 36.7499887943,
                "e2e_batch_us": 26.5670391061,
                "e2e_single_us": 269.542
              },
              "torch": {
                "gpu_batch_us": 15.719401744718198,
                "gpu_single_us": 24.54168861731887,
                "e2e_batch_us": 29.273819561551434,
                "e2e_single_us": 299.916
              },
              "system": {
                "gpu_batch_us": 15.116536815187501,
                "gpu_single_us": 21.124957129400002,
                "e2e_batch_us": 16.2371134021,
                "e2e_single_us": 272.167
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "904f27be631b58b334527ed89cd1a4e5ee27c52c1a94370dd9b8834808ae1cf4",
            "paths": {
              "native": {
                "gpu_batch_us": 24.5247392740625,
                "gpu_single_us": 37.2916692868,
                "e2e_batch_us": 26.4565217391,
                "e2e_single_us": 270.542
              },
              "torch": {
                "gpu_batch_us": 15.441405594174284,
                "gpu_single_us": 23.708329536020756,
                "e2e_batch_us": 29.054097649186254,
                "e2e_single_us": 276.667
              },
              "system": {
                "gpu_batch_us": 15.468099263671874,
                "gpu_single_us": 17.124984879,
                "e2e_batch_us": 16.4270352832,
                "e2e_single_us": 236.25
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "904f27be631b58b334527ed89cd1a4e5ee27c52c1a94370dd9b8834808ae1cf4",
            "paths": {
              "native": {
                "gpu_batch_us": 24.57552091078125,
                "gpu_single_us": 37.37496444959999,
                "e2e_batch_us": 25.9421542553,
                "e2e_single_us": 261.916
              },
              "torch": {
                "gpu_batch_us": 17.70247399690561,
                "gpu_single_us": 24.666660465300083,
                "e2e_batch_us": 30.257056451612904,
                "e2e_single_us": 394.209
              },
              "system": {
                "gpu_batch_us": 14.67122365284375,
                "gpu_single_us": 17.0000130311,
                "e2e_batch_us": 16.5773128079,
                "e2e_single_us": 224.75
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "c8ce39d5058dcad32e9a1b832a25a8286fa42351caac7b139f5b2cb9583c4ab8",
            "paths": {
              "native": {
                "gpu_batch_us": 470.96849572439027,
                "gpu_single_us": 471.99998516600004,
                "e2e_batch_us": 497.919707317,
                "e2e_single_us": 694.5
              },
              "torch": {
                "gpu_batch_us": 384.971918285136,
                "gpu_single_us": 372.9583113454282,
                "e2e_batch_us": 426.50180434782607,
                "e2e_single_us": 1025.4589999999998
              },
              "system": {
                "gpu_batch_us": 484.74270879499994,
                "gpu_single_us": 468.750018626,
                "e2e_batch_us": 492.389575,
                "e2e_single_us": 692.291
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "c8ce39d5058dcad32e9a1b832a25a8286fa42351caac7b139f5b2cb9583c4ab8",
            "paths": {
              "native": {
                "gpu_batch_us": 468.08028411951216,
                "gpu_single_us": 483.041687403,
                "e2e_batch_us": 479.133121951,
                "e2e_single_us": 768.792
              },
              "torch": {
                "gpu_batch_us": 387.04528782071304,
                "gpu_single_us": 374.00005385279655,
                "e2e_batch_us": 406.6367826086957,
                "e2e_single_us": 747.4159999999999
              },
              "system": {
                "gpu_batch_us": 488.9166662710527,
                "gpu_single_us": 468.624988571,
                "e2e_batch_us": 503.656815789,
                "e2e_single_us": 687.583
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "c5fb98b5a3e9fa7deac8011793868249f2edd5f7a78bd54a236e208c9a92e07e",
            "paths": {
              "native": {
                "gpu_batch_us": 1407.5961528153846,
                "gpu_single_us": 1387.58332469,
                "e2e_batch_us": 1445.04492308,
                "e2e_single_us": 1678.375
              },
              "torch": {
                "gpu_batch_us": 382.58072891039774,
                "gpu_single_us": 368.2916867546737,
                "e2e_batch_us": 419.5546875,
                "e2e_single_us": 924.875
              },
              "system": {
                "gpu_batch_us": 485.7763154789474,
                "gpu_single_us": 469.750026241,
                "e2e_batch_us": 509.885973684,
                "e2e_single_us": 682.792
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "c8ce39d5058dcad32e9a1b832a25a8286fa42351caac7b139f5b2cb9583c4ab8",
            "paths": {
              "native": {
                "gpu_batch_us": 470.91041778999994,
                "gpu_single_us": 475.20833322799996,
                "e2e_batch_us": 489.759375,
                "e2e_single_us": 699.208
              },
              "torch": {
                "gpu_batch_us": 385.7039943492661,
                "gpu_single_us": 365.9999929368496,
                "e2e_batch_us": 412.1545208333334,
                "e2e_single_us": 792.75
              },
              "system": {
                "gpu_batch_us": 484.4294863948718,
                "gpu_single_us": 469.624996185,
                "e2e_batch_us": 493.074794872,
                "e2e_single_us": 691.958
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "b65ed0c63e9a0eb3b969929c3b3dddb862b7d8c43cd967ec520ecab7949bc895",
            "paths": {
              "native": {
                "gpu_batch_us": 3134.8055636,
                "gpu_single_us": 3002.62495875,
                "e2e_batch_us": 3148.49316667,
                "e2e_single_us": 3374.458
              },
              "torch": {
                "gpu_batch_us": 2614.3472253655395,
                "gpu_single_us": 2450.0416475348175,
                "e2e_batch_us": 2707.152833333333,
                "e2e_single_us": 2955.25
              },
              "system": {
                "gpu_batch_us": 2817.7847270833336,
                "gpu_single_us": 2635.1249543900003,
                "e2e_batch_us": 2888.868,
                "e2e_single_us": 3030.334
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "c9759290026e8a0c39158d60e53092d854cbfb78a8dbf799983cddfdae74548b",
            "paths": {
              "native": {
                "gpu_batch_us": 11273.604148299999,
                "gpu_single_us": 11545.041634199999,
                "e2e_batch_us": 11675.271,
                "e2e_single_us": 11273.167
              },
              "torch": {
                "gpu_batch_us": 2664.2142828287824,
                "gpu_single_us": 2503.374998923391,
                "e2e_batch_us": 2744.226142857143,
                "e2e_single_us": 2974.459
              },
              "system": {
                "gpu_batch_us": 2844.1666703,
                "gpu_single_us": 2620.50004676,
                "e2e_batch_us": 2856.40966667,
                "e2e_single_us": 2905.375
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "91d91cdb5dbd3b45d0be58b4f7cc1842c9bc81b71f136c35cd82f202143e1fd2",
            "paths": {
              "native": {
                "gpu_batch_us": 3417.6166635,
                "gpu_single_us": 3308.3749585799997,
                "e2e_batch_us": 3568.55,
                "e2e_single_us": 3838.125
              },
              "torch": {
                "gpu_batch_us": 2614.244053672467,
                "gpu_single_us": 2465.08332202211,
                "e2e_batch_us": 2677.684571428571,
                "e2e_single_us": 2970.8329999999996
              },
              "system": {
                "gpu_batch_us": 2783.5069437666666,
                "gpu_single_us": 2630.2916230600003,
                "e2e_batch_us": 2916.1805,
                "e2e_single_us": 2945.708
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "b65ed0c63e9a0eb3b969929c3b3dddb862b7d8c43cd967ec520ecab7949bc895",
            "paths": {
              "native": {
                "gpu_batch_us": 3121.89166434,
                "gpu_single_us": 2992.66667571,
                "e2e_batch_us": 3289.725,
                "e2e_single_us": 3326.709
              },
              "torch": {
                "gpu_batch_us": 2584.553573147527,
                "gpu_single_us": 2462.9583349451423,
                "e2e_batch_us": 2737.3214285714284,
                "e2e_single_us": 2932.708
              },
              "system": {
                "gpu_batch_us": 2776.187495333333,
                "gpu_single_us": 2605.37496069,
                "e2e_batch_us": 2799.57633333,
                "e2e_single_us": 2936.041
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "e6d039570a126141ec252f922e925939179691eb9057310385fdec66f4a69442",
            "paths": {
              "native": {
                "gpu_batch_us": 54944.3333293,
                "gpu_single_us": 55328.9999953,
                "e2e_batch_us": 56031.167,
                "e2e_single_us": 55559.416
              },
              "torch": {
                "gpu_batch_us": 20508.375018835068,
                "gpu_single_us": 20463.24999537319,
                "e2e_batch_us": 21434.541,
                "e2e_single_us": 21302.0
              },
              "system": {
                "gpu_batch_us": 22339.7500231,
                "gpu_single_us": 22385.2500203,
                "e2e_batch_us": 22992.708,
                "e2e_single_us": 23100
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "7784211defc91c9dcd8562c09c837e9039209e60896f191c6006653ea41529f1",
            "paths": {
              "native": {
                "gpu_batch_us": 21902.7500134,
                "gpu_single_us": 22572.1250172,
                "e2e_batch_us": 21949.5,
                "e2e_single_us": 22382.75
              },
              "torch": {
                "gpu_batch_us": 20320.041687227786,
                "gpu_single_us": 20926.08332168311,
                "e2e_batch_us": 20930.625,
                "e2e_single_us": 21522.375
              },
              "system": {
                "gpu_batch_us": 21914.7500466,
                "gpu_single_us": 22400.1249881,
                "e2e_batch_us": 22935.208,
                "e2e_single_us": 22751.125
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "6944e659768d202473106f722037975663fb4618c1109a577cb11affebf5c1f3",
            "paths": {
              "native": {
                "gpu_batch_us": 21446.541650199997,
                "gpu_single_us": 22179.458348499997,
                "e2e_batch_us": 22245.584,
                "e2e_single_us": 22579.708
              },
              "torch": {
                "gpu_batch_us": 20306.04163883254,
                "gpu_single_us": 20118.083339184523,
                "e2e_batch_us": 21278.625,
                "e2e_single_us": 21119.958
              },
              "system": {
                "gpu_batch_us": 22126.958298,
                "gpu_single_us": 22126.3749991,
                "e2e_batch_us": 22884.625,
                "e2e_single_us": 23311.625
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "6944e659768d202473106f722037975663fb4618c1109a577cb11affebf5c1f3",
            "paths": {
              "native": {
                "gpu_batch_us": 21216.666675200002,
                "gpu_single_us": 21310.2083071,
                "e2e_batch_us": 21841.291,
                "e2e_single_us": 21771.416
              },
              "torch": {
                "gpu_batch_us": 19835.999992210418,
                "gpu_single_us": 20529.29170895368,
                "e2e_batch_us": 20266.125,
                "e2e_single_us": 21399.834
              },
              "system": {
                "gpu_batch_us": 21904.5833801,
                "gpu_single_us": 21707.1250081,
                "e2e_batch_us": 23176.583,
                "e2e_single_us": 22542.792
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 278.80664038125,
                "gpu_single_us": 270.666670986,
                "e2e_batch_us": 284.323214286,
                "e2e_single_us": 521.208
              },
              "torch": {
                "gpu_batch_us": 286.0820304704248,
                "gpu_single_us": 277.24995743483305,
                "e2e_batch_us": 302.3061818181818,
                "e2e_single_us": 532.2090000000001
              },
              "system": {
                "gpu_batch_us": 282.29296913125,
                "gpu_single_us": 269.749958534,
                "e2e_batch_us": 299.032328358,
                "e2e_single_us": 488.458
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 280.3723955359375,
                "gpu_single_us": 270.24996234099996,
                "e2e_batch_us": 283.602553846,
                "e2e_single_us": 564.792
              },
              "torch": {
                "gpu_batch_us": 286.95081948440094,
                "gpu_single_us": 285.1249882951379,
                "e2e_batch_us": 305.8517868852459,
                "e2e_single_us": 578.917
              },
              "system": {
                "gpu_batch_us": 280.14453164375,
                "gpu_single_us": 270.291697234,
                "e2e_batch_us": 286.965272727,
                "e2e_single_us": 536.791
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 676.5684520357144,
                "gpu_single_us": 670.2916580250001,
                "e2e_batch_us": 689.273821429,
                "e2e_single_us": 943.084
              },
              "torch": {
                "gpu_batch_us": 286.9902345992159,
                "gpu_single_us": 276.2500080280006,
                "e2e_batch_us": 301.751953125,
                "e2e_single_us": 568.25
              },
              "system": {
                "gpu_batch_us": 278.9101563390625,
                "gpu_single_us": 270.000018645,
                "e2e_batch_us": 288.234073529,
                "e2e_single_us": 488.542
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 278.88867225625,
                "gpu_single_us": 269.750016741,
                "e2e_batch_us": 287.822294118,
                "e2e_single_us": 503.375
              },
              "torch": {
                "gpu_batch_us": 286.9814800630723,
                "gpu_single_us": 276.83330699801445,
                "e2e_batch_us": 300.60515873015873,
                "e2e_single_us": 562.25
              },
              "system": {
                "gpu_batch_us": 281.6972655640625,
                "gpu_single_us": 271.625001915,
                "e2e_batch_us": 288.486940299,
                "e2e_single_us": 574.291
              }
            }
          }
        ],
        "rejections": [],
        "outputs": 60,
        "elements": 327750192
      },
      "new-reverse": {
        "measurements": [
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 279.04557282499997,
                "gpu_single_us": 270.0832556,
                "e2e_batch_us": 284.249373134,
                "e2e_single_us": 518.5
              },
              "torch": {
                "gpu_batch_us": 287.09439902740996,
                "gpu_single_us": 293.5000229626894,
                "e2e_batch_us": 301.728515625,
                "e2e_single_us": 618.0
              },
              "system": {
                "gpu_batch_us": 285.44466113125,
                "gpu_single_us": 270.12504870099997,
                "e2e_batch_us": 289.096205882,
                "e2e_single_us": 579.542
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 279.7539063984375,
                "gpu_single_us": 270.04169533000004,
                "e2e_batch_us": 284.06841791,
                "e2e_single_us": 487.917
              },
              "torch": {
                "gpu_batch_us": 290.23557726759464,
                "gpu_single_us": 277.9999631457031,
                "e2e_batch_us": 317.21475,
                "e2e_single_us": 857.834
              },
              "system": {
                "gpu_batch_us": 280.49544289374995,
                "gpu_single_us": 270.000018645,
                "e2e_batch_us": 287.249397059,
                "e2e_single_us": 500.334
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 674.2113099392857,
                "gpu_single_us": 672.333349939,
                "e2e_batch_us": 689.726178571,
                "e2e_single_us": 913.584
              },
              "torch": {
                "gpu_batch_us": 288.05040112215903,
                "gpu_single_us": 278.2499650493264,
                "e2e_batch_us": 306.82862903225805,
                "e2e_single_us": 609.917
              },
              "system": {
                "gpu_batch_us": 280.83854158437504,
                "gpu_single_us": 271.625001915,
                "e2e_batch_us": 284.597217391,
                "e2e_single_us": 514.375
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 281.7473960015625,
                "gpu_single_us": 269.708281849,
                "e2e_batch_us": 285.094970588,
                "e2e_single_us": 487.584
              },
              "torch": {
                "gpu_batch_us": 287.64518083335133,
                "gpu_single_us": 275.4583256319165,
                "e2e_batch_us": 303.90832307692307,
                "e2e_single_us": 566.625
              },
              "system": {
                "gpu_batch_us": 280.5332032953125,
                "gpu_single_us": 270.874996204,
                "e2e_batch_us": 289.220405797,
                "e2e_single_us": 511.25
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "7784211defc91c9dcd8562c09c837e9039209e60896f191c6006653ea41529f1",
            "paths": {
              "native": {
                "gpu_batch_us": 21133.5833301,
                "gpu_single_us": 21277.8749992,
                "e2e_batch_us": 22791.25,
                "e2e_single_us": 22644.834
              },
              "torch": {
                "gpu_batch_us": 20295.374968554825,
                "gpu_single_us": 20715.37496522069,
                "e2e_batch_us": 20690.459,
                "e2e_single_us": 21156.208
              },
              "system": {
                "gpu_batch_us": 22020.9999825,
                "gpu_single_us": 22060.6666408,
                "e2e_batch_us": 23010.167,
                "e2e_single_us": 22960.833
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "e6d039570a126141ec252f922e925939179691eb9057310385fdec66f4a69442",
            "paths": {
              "native": {
                "gpu_batch_us": 54715.2917134,
                "gpu_single_us": 53258.833300800005,
                "e2e_batch_us": 56395.333,
                "e2e_single_us": 56060.959
              },
              "torch": {
                "gpu_batch_us": 20610.29168544337,
                "gpu_single_us": 20900.874980725348,
                "e2e_batch_us": 20843.5,
                "e2e_single_us": 21476.708
              },
              "system": {
                "gpu_batch_us": 21960.2499856,
                "gpu_single_us": 21843.2499678,
                "e2e_batch_us": 23301.542,
                "e2e_single_us": 23274.333
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "6944e659768d202473106f722037975663fb4618c1109a577cb11affebf5c1f3",
            "paths": {
              "native": {
                "gpu_batch_us": 21483.6250525,
                "gpu_single_us": 21353.7917007,
                "e2e_batch_us": 22344.584,
                "e2e_single_us": 22161.834
              },
              "torch": {
                "gpu_batch_us": 20099.874993320554,
                "gpu_single_us": 20419.50001148507,
                "e2e_batch_us": 20424.125,
                "e2e_single_us": 20264.083
              },
              "system": {
                "gpu_batch_us": 21147.4166717,
                "gpu_single_us": 21898.6249529,
                "e2e_batch_us": 22852.209,
                "e2e_single_us": 23109.5
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "7784211defc91c9dcd8562c09c837e9039209e60896f191c6006653ea41529f1",
            "paths": {
              "native": {
                "gpu_batch_us": 21629.1666729,
                "gpu_single_us": 21773.6249906,
                "e2e_batch_us": 22422.542,
                "e2e_single_us": 21972.5
              },
              "torch": {
                "gpu_batch_us": 20390.12499153614,
                "gpu_single_us": 20345.124998129904,
                "e2e_batch_us": 20476.625,
                "e2e_single_us": 20740.875
              },
              "system": {
                "gpu_batch_us": 22066.833335,
                "gpu_single_us": 22371.4583553,
                "e2e_batch_us": 22310,
                "e2e_single_us": 22235.917
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "c9759290026e8a0c39158d60e53092d854cbfb78a8dbf799983cddfdae74548b",
            "paths": {
              "native": {
                "gpu_batch_us": 11866.3125031,
                "gpu_single_us": 11292.3749839,
                "e2e_batch_us": 11843.229,
                "e2e_single_us": 12231.417
              },
              "torch": {
                "gpu_batch_us": 2597.9523862978176,
                "gpu_single_us": 2501.041686628014,
                "e2e_batch_us": 2745.8928571428573,
                "e2e_single_us": 3029.25
              },
              "system": {
                "gpu_batch_us": 2796.4861074,
                "gpu_single_us": 2674.37496223,
                "e2e_batch_us": 2859.34033333,
                "e2e_single_us": 2974.833
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "b65ed0c63e9a0eb3b969929c3b3dddb862b7d8c43cd967ec520ecab7949bc895",
            "paths": {
              "native": {
                "gpu_batch_us": 3144.020845266667,
                "gpu_single_us": 2999.62499412,
                "e2e_batch_us": 3182.78483333,
                "e2e_single_us": 3390.959
              },
              "torch": {
                "gpu_batch_us": 2665.3472353549055,
                "gpu_single_us": 2445.375022944063,
                "e2e_batch_us": 2741.8333333333335,
                "e2e_single_us": 2983.875
              },
              "system": {
                "gpu_batch_us": 2757.52083205,
                "gpu_single_us": 2627.6250137,
                "e2e_batch_us": 2868.73616667,
                "e2e_single_us": 3124.75
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "91d91cdb5dbd3b45d0be58b4f7cc1842c9bc81b71f136c35cd82f202143e1fd2",
            "paths": {
              "native": {
                "gpu_batch_us": 3421.77500716,
                "gpu_single_us": 3306.99997721,
                "e2e_batch_us": 3521.6582,
                "e2e_single_us": 3960.333
              },
              "torch": {
                "gpu_batch_us": 2616.298608093833,
                "gpu_single_us": 2459.749986883253,
                "e2e_batch_us": 2688.326333333333,
                "e2e_single_us": 2931.083
              },
              "system": {
                "gpu_batch_us": 2824.6041620166666,
                "gpu_single_us": 2596.24997852,
                "e2e_batch_us": 2882.65983333,
                "e2e_single_us": 3209.75
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "b65ed0c63e9a0eb3b969929c3b3dddb862b7d8c43cd967ec520ecab7949bc895",
            "paths": {
              "native": {
                "gpu_batch_us": 3102.3263873,
                "gpu_single_us": 2981.70826863,
                "e2e_batch_us": 3166.27083333,
                "e2e_single_us": 3340.083
              },
              "torch": {
                "gpu_batch_us": 2588.9166669609644,
                "gpu_single_us": 2513.0833382718265,
                "e2e_batch_us": 2752.5555,
                "e2e_single_us": 2977.083
              },
              "system": {
                "gpu_batch_us": 2754.160710271429,
                "gpu_single_us": 2606.49999836,
                "e2e_batch_us": 2869.54757143,
                "e2e_single_us": 2958.792
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "c8ce39d5058dcad32e9a1b832a25a8286fa42351caac7b139f5b2cb9583c4ab8",
            "paths": {
              "native": {
                "gpu_batch_us": 468.33536570731707,
                "gpu_single_us": 475.99995741600003,
                "e2e_batch_us": 481.62804878,
                "e2e_single_us": 734.792
              },
              "torch": {
                "gpu_batch_us": 384.5994895305104,
                "gpu_single_us": 369.6250496432185,
                "e2e_batch_us": 405.0935306122449,
                "e2e_single_us": 710.666
              },
              "system": {
                "gpu_batch_us": 482.9198722,
                "gpu_single_us": 467.12497714899996,
                "e2e_batch_us": 495.125,
                "e2e_single_us": 696
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "c8ce39d5058dcad32e9a1b832a25a8286fa42351caac7b139f5b2cb9583c4ab8",
            "paths": {
              "native": {
                "gpu_batch_us": 470.06503900243905,
                "gpu_single_us": 463.95836398,
                "e2e_batch_us": 487.769317073,
                "e2e_single_us": 672.125
              },
              "torch": {
                "gpu_batch_us": 383.32313094858307,
                "gpu_single_us": 370.9583543241024,
                "e2e_batch_us": 403.88604081632656,
                "e2e_single_us": 819.416
              },
              "system": {
                "gpu_batch_us": 485.2499994923077,
                "gpu_single_us": 468.500016723,
                "e2e_batch_us": 492.121794872,
                "e2e_single_us": 675.084
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "c5fb98b5a3e9fa7deac8011793868249f2edd5f7a78bd54a236e208c9a92e07e",
            "paths": {
              "native": {
                "gpu_batch_us": 1414.878204323077,
                "gpu_single_us": 1384.20826988,
                "e2e_batch_us": 1448.57369231,
                "e2e_single_us": 1712.625
              },
              "torch": {
                "gpu_batch_us": 388.391490560025,
                "gpu_single_us": 379.0416521951556,
                "e2e_batch_us": 413.90279166666664,
                "e2e_single_us": 834.0840000000001
              },
              "system": {
                "gpu_batch_us": 486.78952943846156,
                "gpu_single_us": 468.458340038,
                "e2e_batch_us": 501.023512821,
                "e2e_single_us": 723.291
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "c8ce39d5058dcad32e9a1b832a25a8286fa42351caac7b139f5b2cb9583c4ab8",
            "paths": {
              "native": {
                "gpu_batch_us": 467.19230699487184,
                "gpu_single_us": 475.500011817,
                "e2e_batch_us": 480.783128205,
                "e2e_single_us": 720.208
              },
              "torch": {
                "gpu_batch_us": 385.10544477410764,
                "gpu_single_us": 378.8333269767463,
                "e2e_batch_us": 404.986387755102,
                "e2e_single_us": 674.1669999999999
              },
              "system": {
                "gpu_batch_us": 485.9935904205128,
                "gpu_single_us": 467.50000910800003,
                "e2e_batch_us": 495.834410256,
                "e2e_single_us": 699.416
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "904f27be631b58b334527ed89cd1a4e5ee27c52c1a94370dd9b8834808ae1cf4",
            "paths": {
              "native": {
                "gpu_batch_us": 24.856119125625,
                "gpu_single_us": 37.50005271289999,
                "e2e_batch_us": 26.1292331903,
                "e2e_single_us": 238.792
              },
              "torch": {
                "gpu_batch_us": 15.451822946488392,
                "gpu_single_us": 20.833336748182774,
                "e2e_batch_us": 29.405230769230773,
                "e2e_single_us": 403.792
              },
              "system": {
                "gpu_batch_us": 14.42382836103125,
                "gpu_single_us": 17.2083382495,
                "e2e_batch_us": 16.7751602374,
                "e2e_single_us": 262.25
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "8f1860ab6c09d536a585ab12cd926e89abdf5488fdd70937ce1f8afaafe7f271",
            "paths": {
              "native": {
                "gpu_batch_us": 54.91080628421875,
                "gpu_single_us": 66.1250087433,
                "e2e_batch_us": 56.9919556962,
                "e2e_single_us": 279.875
              },
              "torch": {
                "gpu_batch_us": 13.701172065339051,
                "gpu_single_us": 17.624988686293364,
                "e2e_batch_us": 29.90623020257827,
                "e2e_single_us": 289.209
              },
              "system": {
                "gpu_batch_us": 15.6666665134375,
                "gpu_single_us": 17.333368305100002,
                "e2e_batch_us": 16.0752685009,
                "e2e_single_us": 235.542
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "904f27be631b58b334527ed89cd1a4e5ee27c52c1a94370dd9b8834808ae1cf4",
            "paths": {
              "native": {
                "gpu_batch_us": 24.595052309375003,
                "gpu_single_us": 37.3332877643,
                "e2e_batch_us": 26.0574229692,
                "e2e_single_us": 281.792
              },
              "torch": {
                "gpu_batch_us": 14.744139662070666,
                "gpu_single_us": 20.291656255722046,
                "e2e_batch_us": 29.41936544850498,
                "e2e_single_us": 289.666
              },
              "system": {
                "gpu_batch_us": 15.21158810646875,
                "gpu_single_us": 17.4166634679,
                "e2e_batch_us": 16.7132699029,
                "e2e_single_us": 250.75
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "fresh",
            "source_sha256": "904f27be631b58b334527ed89cd1a4e5ee27c52c1a94370dd9b8834808ae1cf4",
            "paths": {
              "native": {
                "gpu_batch_us": 24.854166440468752,
                "gpu_single_us": 38.708327338100005,
                "e2e_batch_us": 26.0205823293,
                "e2e_single_us": 256.709
              },
              "torch": {
                "gpu_batch_us": 15.923827049846295,
                "gpu_single_us": 23.874978069216013,
                "e2e_batch_us": 30.174560131795715,
                "e2e_single_us": 294.333
              },
              "system": {
                "gpu_batch_us": 14.757812095921876,
                "gpu_single_us": 17.333368305100002,
                "e2e_batch_us": 16.3591916933,
                "e2e_single_us": 290.875
              }
            }
          }
        ],
        "rejections": [],
        "outputs": 60,
        "elements": 327750192
      },
      "old-reverse": {
        "measurements": [
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "phase": "trial",
            "source_sha256": "9a3882584d67f153f07bb7141774a116e61191216407241e0974c48611e65a5e",
            "paths": {
              "native": {
                "gpu_batch_us": 281.58007808125,
                "gpu_single_us": 270.708347671,
                "e2e_batch_us": 296.670338235,
                "e2e_single_us": 1590.042
              },
              "torch": {
                "gpu_batch_us": 289.3756722070036,
                "gpu_single_us": 291.50000773370266,
                "e2e_batch_us": 306.52082258064513,
                "e2e_single_us": 558.5830000000001
              },
              "system": {
                "gpu_batch_us": 280.98763050312505,
                "gpu_single_us": 269.874988589,
                "e2e_batch_us": 288.21809375,
                "e2e_single_us": 474.291
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "trial",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 280.179036053125,
                "gpu_single_us": 269.750016741,
                "e2e_batch_us": 284.960714286,
                "e2e_single_us": 485
              },
              "torch": {
                "gpu_batch_us": 290.0551077008488,
                "gpu_single_us": 280.7083656080067,
                "e2e_batch_us": 304.8535,
                "e2e_single_us": 557.25
              },
              "system": {
                "gpu_batch_us": 282.10221353281247,
                "gpu_single_us": 273.250043392,
                "e2e_batch_us": 292.938720588,
                "e2e_single_us": 521.875
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4348457736954f054432a3bb20131e46eae0a1fff6df71a9dd5a648cfae508f9",
            "paths": {
              "native": {
                "gpu_batch_us": 680.40624735,
                "gpu_single_us": 680.166704115,
                "e2e_batch_us": 704.988107143,
                "e2e_single_us": 959.375
              },
              "torch": {
                "gpu_batch_us": 288.09457961913375,
                "gpu_single_us": 282.0416702888906,
                "e2e_batch_us": 300.71361904761903,
                "e2e_single_us": 583.625
              },
              "system": {
                "gpu_batch_us": 280.335286334375,
                "gpu_single_us": 268.91665766,
                "e2e_batch_us": 289.567164179,
                "e2e_single_us": 493.917
              }
            }
          },
          {
            "shape": [
              1024,
              1024,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "phase": "fresh",
            "source_sha256": "bfa52e1a27b05b7647359d31c8a6f9e305318e1c6899c3b2aff0f099b617f42b",
            "paths": {
              "native": {
                "gpu_batch_us": 278.56966153125,
                "gpu_single_us": 268.33335869,
                "e2e_batch_us": 286.058462687,
                "e2e_single_us": 499.167
              },
              "torch": {
                "gpu_batch_us": 286.3997397071216,
                "gpu_single_us": 277.87499129772186,
                "e2e_batch_us": 300.53333846153845,
                "e2e_single_us": 531.2090000000001
              },
              "system": {
                "gpu_batch_us": 279.8893228828125,
                "gpu_single_us": 268.416653853,
                "e2e_batch_us": 287.865530303,
                "e2e_single_us": 656.375
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 141174.499993,
                "gpu_single_us": 140550.95828400002,
                "e2e_batch_us": 142737.5,
                "e2e_single_us": 143006.167
              },
              "torch": {
                "gpu_batch_us": 19950.3333424218,
                "gpu_single_us": 20301.249984186143,
                "e2e_batch_us": 20519.541,
                "e2e_single_us": 21229.790999999997
              },
              "system": {
                "gpu_batch_us": 22427.7499947,
                "gpu_single_us": 22875.416674699998,
                "e2e_batch_us": 23211.166,
                "e2e_single_us": 23244.792
              }
            }
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "0f4e4ecdb4df311a41a4b5ee6bcad9ae8309b9e2550e7690995a73220aaa00df",
            "paths": {
              "native": {
                "gpu_batch_us": 140243.29168,
                "gpu_single_us": 141116.249957,
                "e2e_batch_us": 140979.5,
                "e2e_single_us": 141980.542
              },
              "torch": {
                "gpu_batch_us": 20690.49997953698,
                "gpu_single_us": 20839.833363424987,
                "e2e_batch_us": 21220.0,
                "e2e_single_us": 20821.75
              },
              "system": {
                "gpu_batch_us": 21939.2083236,
                "gpu_single_us": 22218.708356400002,
                "e2e_batch_us": 23098.125,
                "e2e_single_us": 22399.042
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15312.77084725,
                "gpu_single_us": 14861.9583342,
                "e2e_batch_us": 15517.521,
                "e2e_single_us": 16622.25
              },
              "torch": {
                "gpu_batch_us": 2607.354176385949,
                "gpu_single_us": 2449.124993290752,
                "e2e_batch_us": 2778.3195,
                "e2e_single_us": 3348.875
              },
              "system": {
                "gpu_batch_us": 2817.2916693999996,
                "gpu_single_us": 2681.1250136199997,
                "e2e_batch_us": 2825.27766667,
                "e2e_single_us": 2979.166
              }
            }
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "81418a3cd198adc2e432216b317e38be52c4d6431b8fed802a265fa4a5abe5ff",
            "paths": {
              "native": {
                "gpu_batch_us": 15928.8750147,
                "gpu_single_us": 15046.208282,
                "e2e_batch_us": 16792.917,
                "e2e_single_us": 15900.875
              },
              "torch": {
                "gpu_batch_us": 2731.284718417252,
                "gpu_single_us": 2471.999963745475,
                "e2e_batch_us": 2839.6458333333335,
                "e2e_single_us": 3132.959
              },
              "system": {
                "gpu_batch_us": 2826.5694466666664,
                "gpu_single_us": 2617.12499196,
                "e2e_batch_us": 2975.5,
                "e2e_single_us": 3023.625
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2115.194439988889,
                "gpu_single_us": 2129.24997322,
                "e2e_batch_us": 2141.05088889,
                "e2e_single_us": 2408.875
              },
              "torch": {
                "gpu_batch_us": 385.01992695155025,
                "gpu_single_us": 373.6666985787451,
                "e2e_batch_us": 401.51447826086957,
                "e2e_single_us": 775.25
              },
              "system": {
                "gpu_batch_us": 482.5801275179487,
                "gpu_single_us": 466.24999959,
                "e2e_batch_us": 494.170948718,
                "e2e_single_us": 698.208
              }
            }
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "4fe59cc8e1786f0776bb91acc70fdfa567c9539bf8943ade44f1b7b12c97ff13",
            "paths": {
              "native": {
                "gpu_batch_us": 2156.7222235444447,
                "gpu_single_us": 2216.6250273600003,
                "e2e_batch_us": 2096.86577778,
                "e2e_single_us": 2285.458
              },
              "torch": {
                "gpu_batch_us": 383.5697282979987,
                "gpu_single_us": 387.6666887663305,
                "e2e_batch_us": 399.9472857142857,
                "e2e_single_us": 639.75
              },
              "system": {
                "gpu_batch_us": 485.26458413000006,
                "gpu_single_us": 467.291625682,
                "e2e_batch_us": 497.407275,
                "e2e_single_us": 716.792
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "trial",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.699870305625,
                "gpu_single_us": 37.45831782,
                "e2e_batch_us": 33.6038439108,
                "e2e_single_us": 282.417
              },
              "torch": {
                "gpu_batch_us": 13.816405953548383,
                "gpu_single_us": 20.291656255722046,
                "e2e_batch_us": 29.945231958762886,
                "e2e_single_us": 302.79200000000003
              },
              "system": {
                "gpu_batch_us": 15.828124560312501,
                "gpu_single_us": 17.0416315086,
                "e2e_batch_us": 16.1634770346,
                "e2e_single_us": 237.291
              }
            }
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              16
            ],
            "phase": "fresh",
            "source_sha256": "129528e6144d894b940109c4cb9f4c2773b77b9e018b57a2bb39a4254771863e",
            "paths": {
              "native": {
                "gpu_batch_us": 31.23632723140625,
                "gpu_single_us": 37.9999401048,
                "e2e_batch_us": 32.9919393443,
                "e2e_single_us": 322.375
              },
              "torch": {
                "gpu_batch_us": 15.479819012398366,
                "gpu_single_us": 20.250037778168917,
                "e2e_batch_us": 30.13367505720824,
                "e2e_single_us": 366.58299999999997
              },
              "system": {
                "gpu_batch_us": 14.25195296178125,
                "gpu_single_us": 17.0416315086,
                "e2e_batch_us": 15.9316617502,
                "e2e_single_us": 269.625
              }
            }
          }
        ],
        "rejections": [
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              4097,
              4097,
              4096
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              2049,
              4097,
              1025
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              1025,
              1025,
              1024
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              1024
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          },
          {
            "shape": [
              129,
              257,
              61
            ],
            "block": [
              128,
              32,
              4096
            ],
            "error": "native benchmark failed (2):\nno legal Metal MPP group plan: each local matrix requires M or N to be a multiple of 16 and must fit shared-memory capacity and cost bounds\n\n"
          }
        ],
        "outputs": 36,
        "elements": 170166552
      }
    },
    "comparisons": [
      {
        "shape": [
          129,
          257,
          61
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 31.995442441250002,
            "gpu_single_us": 37.875026464499996,
            "e2e_batch_us": 33.6454501018,
            "e2e_single_us": 264.292
          },
          "torch": {
            "gpu_batch_us": 22.968099074205384,
            "gpu_single_us": 30.125025659799576,
            "e2e_batch_us": 29.399450160771703,
            "e2e_single_us": 367.08299999999997
          },
          "system": {
            "gpu_batch_us": 14.986328096718749,
            "gpu_single_us": 17.5000168383,
            "e2e_batch_us": 16.6094674556,
            "e2e_single_us": 267.75
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 24.57552091078125,
            "gpu_single_us": 37.37496444959999,
            "e2e_batch_us": 25.9421542553,
            "e2e_single_us": 261.916
          },
          "torch": {
            "gpu_batch_us": 17.70247399690561,
            "gpu_single_us": 24.666660465300083,
            "e2e_batch_us": 30.257056451612904,
            "e2e_single_us": 394.209
          },
          "system": {
            "gpu_batch_us": 14.67122365284375,
            "gpu_single_us": 17.0000130311,
            "e2e_batch_us": 16.5773128079,
            "e2e_single_us": 224.75
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.7680944233200336,
          "gpu_single_us": 0.9867970517362223,
          "e2e_batch_us": 0.7710449459527997,
          "e2e_single_us": 0.9910099435472887
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 1.6918089465169848,
          "gpu_single_us": 1.72976128975027,
          "e2e_batch_us": 1.6933521730764394,
          "e2e_single_us": 1.172350967627258
        },
        "new_over_mps": {
          "gpu_batch_us": 1.6750832440631311,
          "gpu_single_us": 2.198525635317211,
          "e2e_batch_us": 1.5649191491963124,
          "e2e_single_us": 1.1653659621802002
        },
        "new_over_torch": {
          "gpu_batch_us": 1.3882534675705243,
          "gpu_single_us": 1.5152016424021957,
          "e2e_batch_us": 0.8573918714395335,
          "e2e_single_us": 0.6644089810227569
        }
      },
      {
        "shape": [
          129,
          257,
          61
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 31.23632723140625,
            "gpu_single_us": 37.9999401048,
            "e2e_batch_us": 32.9919393443,
            "e2e_single_us": 322.375
          },
          "torch": {
            "gpu_batch_us": 15.479819012398366,
            "gpu_single_us": 20.250037778168917,
            "e2e_batch_us": 30.13367505720824,
            "e2e_single_us": 366.58299999999997
          },
          "system": {
            "gpu_batch_us": 14.25195296178125,
            "gpu_single_us": 17.0416315086,
            "e2e_batch_us": 15.9316617502,
            "e2e_single_us": 269.625
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 24.854166440468752,
            "gpu_single_us": 38.708327338100005,
            "e2e_batch_us": 26.0205823293,
            "e2e_single_us": 256.709
          },
          "torch": {
            "gpu_batch_us": 15.923827049846295,
            "gpu_single_us": 23.874978069216013,
            "e2e_batch_us": 30.174560131795715,
            "e2e_single_us": 294.333
          },
          "system": {
            "gpu_batch_us": 14.757812095921876,
            "gpu_single_us": 17.333368305100002,
            "e2e_batch_us": 16.3591916933,
            "e2e_single_us": 290.875
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.7956814594860365,
          "gpu_single_us": 1.0186417986803755,
          "e2e_batch_us": 0.788695143312197,
          "e2e_single_us": 0.7963055447848003
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 1.7322091779812447,
          "gpu_single_us": 1.76529573647843,
          "e2e_batch_us": 1.6959951322081712,
          "e2e_single_us": 0.9909991254067567
        },
        "new_over_mps": {
          "gpu_batch_us": 1.6841362580661172,
          "gpu_single_us": 2.2331682253997247,
          "e2e_batch_us": 1.5905787288962985,
          "e2e_single_us": 0.8825406102277611
        },
        "new_over_torch": {
          "gpu_batch_us": 1.5608161507072296,
          "gpu_single_us": 1.6212926866730764,
          "e2e_batch_us": 0.8623351000196169,
          "e2e_single_us": 0.8721719956647742
        }
      },
      {
        "shape": [
          1025,
          1025,
          1024
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 2122.5833406750003,
            "gpu_single_us": 2236.6665653,
            "e2e_batch_us": 2113.61975,
            "e2e_single_us": 2407.333
          },
          "torch": {
            "gpu_batch_us": 385.9535976186056,
            "gpu_single_us": 375.04167994484305,
            "e2e_batch_us": 404.42709090909096,
            "e2e_single_us": 718.6669999999999
          },
          "system": {
            "gpu_batch_us": 486.6378208717948,
            "gpu_single_us": 466.041616164,
            "e2e_batch_us": 498.787384615,
            "e2e_single_us": 694.5
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 470.91041778999994,
            "gpu_single_us": 475.20833322799996,
            "e2e_batch_us": 489.759375,
            "e2e_single_us": 699.208
          },
          "torch": {
            "gpu_batch_us": 385.7039943492661,
            "gpu_single_us": 365.9999929368496,
            "e2e_batch_us": 412.1545208333334,
            "e2e_single_us": 792.75
          },
          "system": {
            "gpu_batch_us": 484.4294863948718,
            "gpu_single_us": 469.624996185,
            "e2e_batch_us": 493.074794872,
            "e2e_single_us": 691.958
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.22185720992243643,
          "gpu_single_us": 0.21246275175766358,
          "e2e_batch_us": 0.23171593424030032,
          "e2e_single_us": 0.29044922326906997
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.679237622507257,
          "gpu_single_us": 0.6775584792650962,
          "e2e_batch_us": 0.6562823648070553,
          "e2e_single_us": 0.7237235579752859
        },
        "new_over_mps": {
          "gpu_batch_us": 0.9720928040415523,
          "gpu_single_us": 1.0118889264591029,
          "e2e_batch_us": 0.9932760305201553,
          "e2e_single_us": 1.0104775145312288
        },
        "new_over_torch": {
          "gpu_batch_us": 1.220911436461757,
          "gpu_single_us": 1.2983834491767146,
          "e2e_batch_us": 1.188290677995616,
          "e2e_single_us": 0.8820031535793125
        }
      },
      {
        "shape": [
          1025,
          1025,
          1024
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 2156.7222235444447,
            "gpu_single_us": 2216.6250273600003,
            "e2e_batch_us": 2096.86577778,
            "e2e_single_us": 2285.458
          },
          "torch": {
            "gpu_batch_us": 383.5697282979987,
            "gpu_single_us": 387.6666887663305,
            "e2e_batch_us": 399.9472857142857,
            "e2e_single_us": 639.75
          },
          "system": {
            "gpu_batch_us": 485.26458413000006,
            "gpu_single_us": 467.291625682,
            "e2e_batch_us": 497.407275,
            "e2e_single_us": 716.792
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 467.19230699487184,
            "gpu_single_us": 475.500011817,
            "e2e_batch_us": 480.783128205,
            "e2e_single_us": 720.208
          },
          "torch": {
            "gpu_batch_us": 385.10544477410764,
            "gpu_single_us": 378.8333269767463,
            "e2e_batch_us": 404.986387755102,
            "e2e_single_us": 674.1669999999999
          },
          "system": {
            "gpu_batch_us": 485.9935904205128,
            "gpu_single_us": 467.50000910800003,
            "e2e_batch_us": 495.834410256,
            "e2e_single_us": 699.416
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.21662145541722524,
          "gpu_single_us": 0.2145153131214621,
          "e2e_batch_us": 0.22928655391286715,
          "e2e_single_us": 0.3151263335401482
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.6689116506615388,
          "gpu_single_us": 0.6500919512924562,
          "e2e_batch_us": 0.6765713509317821,
          "e2e_single_us": 0.7109646619272482
        },
        "new_over_mps": {
          "gpu_batch_us": 0.9613137214230071,
          "gpu_single_us": 1.017112304926505,
          "e2e_batch_us": 0.9696445391048415,
          "e2e_single_us": 1.0297276585036659
        },
        "new_over_torch": {
          "gpu_batch_us": 1.213154249919562,
          "gpu_single_us": 1.2551694319285358,
          "e2e_batch_us": 1.1871587360504887,
          "e2e_single_us": 1.0682931677166045
        }
      },
      {
        "shape": [
          2049,
          4097,
          1025
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 15136.3750338,
            "gpu_single_us": 14775.4166974,
            "e2e_batch_us": 15938.083,
            "e2e_single_us": 16249.5
          },
          "torch": {
            "gpu_batch_us": 2609.839288717402,
            "gpu_single_us": 2448.0416905134916,
            "e2e_batch_us": 2776.226142857143,
            "e2e_single_us": 3061.4159999999997
          },
          "system": {
            "gpu_batch_us": 2797.6249985,
            "gpu_single_us": 2643.5833424300004,
            "e2e_batch_us": 2857.9305,
            "e2e_single_us": 3005.625
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 3121.89166434,
            "gpu_single_us": 2992.66667571,
            "e2e_batch_us": 3289.725,
            "e2e_single_us": 3326.709
          },
          "torch": {
            "gpu_batch_us": 2584.553573147527,
            "gpu_single_us": 2462.9583349451423,
            "e2e_batch_us": 2737.3214285714284,
            "e2e_single_us": 2932.708
          },
          "system": {
            "gpu_batch_us": 2776.187495333333,
            "gpu_single_us": 2605.37496069,
            "e2e_batch_us": 2799.57633333,
            "e2e_single_us": 2936.041
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.2062509456437699,
          "gpu_single_us": 0.20254363968202763,
          "e2e_batch_us": 0.20640656721388637,
          "e2e_single_us": 0.20472685313394257
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.7596000518934394,
          "gpu_single_us": 0.7756744536875254,
          "e2e_batch_us": 0.7096045342646857,
          "e2e_single_us": 0.6868211017463506
        },
        "new_over_mps": {
          "gpu_batch_us": 1.1245247914947325,
          "gpu_single_us": 1.1486510467258928,
          "e2e_batch_us": 1.1750795864483483,
          "e2e_single_us": 1.1330594497828879
        },
        "new_over_torch": {
          "gpu_batch_us": 1.2079036382821389,
          "gpu_single_us": 1.21506995601558,
          "e2e_batch_us": 1.2018044229891056,
          "e2e_single_us": 1.13434716309977
        }
      },
      {
        "shape": [
          2049,
          4097,
          1025
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 15928.8750147,
            "gpu_single_us": 15046.208282,
            "e2e_batch_us": 16792.917,
            "e2e_single_us": 15900.875
          },
          "torch": {
            "gpu_batch_us": 2731.284718417252,
            "gpu_single_us": 2471.999963745475,
            "e2e_batch_us": 2839.6458333333335,
            "e2e_single_us": 3132.959
          },
          "system": {
            "gpu_batch_us": 2826.5694466666664,
            "gpu_single_us": 2617.12499196,
            "e2e_batch_us": 2975.5,
            "e2e_single_us": 3023.625
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 3102.3263873,
            "gpu_single_us": 2981.70826863,
            "e2e_batch_us": 3166.27083333,
            "e2e_single_us": 3340.083
          },
          "torch": {
            "gpu_batch_us": 2588.9166669609644,
            "gpu_single_us": 2513.0833382718265,
            "e2e_batch_us": 2752.5555,
            "e2e_single_us": 2977.083
          },
          "system": {
            "gpu_batch_us": 2754.160710271429,
            "gpu_single_us": 2606.49999836,
            "e2e_batch_us": 2869.54757143,
            "e2e_single_us": 2958.792
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.1947611733055229,
          "gpu_single_us": 0.19817007798549893,
          "e2e_batch_us": 0.18854799516546172,
          "e2e_single_us": 0.21005655349155314
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.7749291504111455,
          "gpu_single_us": 0.7598174298412775,
          "e2e_batch_us": 0.763216560170919,
          "e2e_single_us": 0.7358460497225104
        },
        "new_over_mps": {
          "gpu_batch_us": 1.1264144375199727,
          "gpu_single_us": 1.1439509957821137,
          "e2e_batch_us": 1.1034041968337651,
          "e2e_single_us": 1.1288671187430546
        },
        "new_over_torch": {
          "gpu_batch_us": 1.1983106396938257,
          "gpu_single_us": 1.1864740907002445,
          "e2e_batch_us": 1.1503022675946044,
          "e2e_single_us": 1.121931434226053
        }
      },
      {
        "shape": [
          4097,
          4097,
          4096
        ],
        "order": "forward",
        "old_bk": 16,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 140672.250011,
            "gpu_single_us": 142828.374985,
            "e2e_batch_us": 144142.375,
            "e2e_single_us": 141140.208
          },
          "torch": {
            "gpu_batch_us": 20189.875038340688,
            "gpu_single_us": 20586.7916229181,
            "e2e_batch_us": 21034.25,
            "e2e_single_us": 21054.208
          },
          "system": {
            "gpu_batch_us": 22332.916618299998,
            "gpu_single_us": 22008.2500018,
            "e2e_batch_us": 22714.25,
            "e2e_single_us": 23061.333
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 21216.666675200002,
            "gpu_single_us": 21310.2083071,
            "e2e_batch_us": 21841.291,
            "e2e_single_us": 21771.416
          },
          "torch": {
            "gpu_batch_us": 19835.999992210418,
            "gpu_single_us": 20529.29170895368,
            "e2e_batch_us": 20266.125,
            "e2e_single_us": 21399.834
          },
          "system": {
            "gpu_batch_us": 21904.5833801,
            "gpu_single_us": 21707.1250081,
            "e2e_batch_us": 23176.583,
            "e2e_single_us": 22542.792
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.1508233974614108,
          "gpu_single_us": 0.14920150361815726,
          "e2e_batch_us": 0.151525816055133,
          "e2e_single_us": 0.1542538183024358
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.4026057887961827,
          "gpu_single_us": 0.41273742798108803,
          "e2e_batch_us": 0.4024581817631227,
          "e2e_single_us": 0.4014886105458986
        },
        "new_over_mps": {
          "gpu_batch_us": 0.968594850997031,
          "gpu_single_us": 0.9817149115393269,
          "e2e_batch_us": 0.9423861576143473,
          "e2e_single_us": 0.9657817008647376
        },
        "new_over_torch": {
          "gpu_batch_us": 1.0696040876956932,
          "gpu_single_us": 1.038039139840647,
          "e2e_batch_us": 1.0777240839084927,
          "e2e_single_us": 1.0173637795508135
        }
      },
      {
        "shape": [
          4097,
          4097,
          4096
        ],
        "order": "reverse",
        "old_bk": 16,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 140243.29168,
            "gpu_single_us": 141116.249957,
            "e2e_batch_us": 140979.5,
            "e2e_single_us": 141980.542
          },
          "torch": {
            "gpu_batch_us": 20690.49997953698,
            "gpu_single_us": 20839.833363424987,
            "e2e_batch_us": 21220.0,
            "e2e_single_us": 20821.75
          },
          "system": {
            "gpu_batch_us": 21939.2083236,
            "gpu_single_us": 22218.708356400002,
            "e2e_batch_us": 23098.125,
            "e2e_single_us": 22399.042
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 21629.1666729,
            "gpu_single_us": 21773.6249906,
            "e2e_batch_us": 22422.542,
            "e2e_single_us": 21972.5
          },
          "torch": {
            "gpu_batch_us": 20390.12499153614,
            "gpu_single_us": 20345.124998129904,
            "e2e_batch_us": 20476.625,
            "e2e_single_us": 20740.875
          },
          "system": {
            "gpu_batch_us": 22066.833335,
            "gpu_single_us": 22371.4583553,
            "e2e_batch_us": 22310,
            "e2e_single_us": 22235.917
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.15422603401417823,
          "gpu_single_us": 0.15429566047308313,
          "e2e_batch_us": 0.15904824460293873,
          "e2e_single_us": 0.15475712157796948
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.3875720595157979,
          "gpu_single_us": 0.3789289945158834,
          "e2e_batch_us": 0.3950982257640774,
          "e2e_single_us": 0.39201777221257883
        },
        "new_over_mps": {
          "gpu_batch_us": 0.9801663131517008,
          "gpu_single_us": 0.9732769605268774,
          "e2e_batch_us": 1.0050444643657552,
          "e2e_single_us": 0.9881535355614072
        },
        "new_over_torch": {
          "gpu_batch_us": 1.0607667526255076,
          "gpu_single_us": 1.0702133799915903,
          "e2e_batch_us": 1.0950311391647793,
          "e2e_single_us": 1.0593815352534548
        }
      },
      {
        "shape": [
          1024,
          1024,
          1024
        ],
        "order": "forward",
        "old_bk": 1024,
        "new_bk": 1024,
        "old": {
          "native": {
            "gpu_batch_us": 282.3645836546875,
            "gpu_single_us": 270.291639026,
            "e2e_batch_us": 287.064338235,
            "e2e_single_us": 502.5
          },
          "torch": {
            "gpu_batch_us": 288.31317397435345,
            "gpu_single_us": 275.2500004135072,
            "e2e_batch_us": 304.55914516129036,
            "e2e_single_us": 628.166
          },
          "system": {
            "gpu_batch_us": 284.4264327,
            "gpu_single_us": 273.87501904699997,
            "e2e_batch_us": 293.390926471,
            "e2e_single_us": 568
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 278.88867225625,
            "gpu_single_us": 269.750016741,
            "e2e_batch_us": 287.822294118,
            "e2e_single_us": 503.375
          },
          "torch": {
            "gpu_batch_us": 286.9814800630723,
            "gpu_single_us": 276.83330699801445,
            "e2e_batch_us": 300.60515873015873,
            "e2e_single_us": 562.25
          },
          "system": {
            "gpu_batch_us": 281.6972655640625,
            "gpu_single_us": 271.625001915,
            "e2e_batch_us": 288.486940299,
            "e2e_single_us": 574.291
          }
        },
        "new_over_old": {
          "gpu_batch_us": 0.9876899880521549,
          "gpu_single_us": 0.9979961559782178,
          "e2e_batch_us": 1.002640369359915,
          "e2e_single_us": 1.0017412935323382
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.9946162326572582,
          "gpu_single_us": 0.9978908759913704,
          "e2e_batch_us": 0.9901200220610715,
          "e2e_single_us": 1.0229597675278794
        },
        "new_over_mps": {
          "gpu_batch_us": 0.9900297459324334,
          "gpu_single_us": 0.9930971554135996,
          "e2e_batch_us": 0.9976960961202918,
          "e2e_single_us": 0.8765155644089843
        },
        "new_over_torch": {
          "gpu_batch_us": 0.9718002436775932,
          "gpu_single_us": 0.9744131573840382,
          "e2e_batch_us": 0.9574762300615293,
          "e2e_single_us": 0.8952867941307248
        }
      },
      {
        "shape": [
          1024,
          1024,
          1024
        ],
        "order": "reverse",
        "old_bk": 1024,
        "new_bk": 4096,
        "old": {
          "native": {
            "gpu_batch_us": 278.56966153125,
            "gpu_single_us": 268.33335869,
            "e2e_batch_us": 286.058462687,
            "e2e_single_us": 499.167
          },
          "torch": {
            "gpu_batch_us": 286.3997397071216,
            "gpu_single_us": 277.87499129772186,
            "e2e_batch_us": 300.53333846153845,
            "e2e_single_us": 531.2090000000001
          },
          "system": {
            "gpu_batch_us": 279.8893228828125,
            "gpu_single_us": 268.416653853,
            "e2e_batch_us": 287.865530303,
            "e2e_single_us": 656.375
          }
        },
        "new": {
          "native": {
            "gpu_batch_us": 281.7473960015625,
            "gpu_single_us": 269.708281849,
            "e2e_batch_us": 285.094970588,
            "e2e_single_us": 487.584
          },
          "torch": {
            "gpu_batch_us": 287.64518083335133,
            "gpu_single_us": 275.4583256319165,
            "e2e_batch_us": 303.90832307692307,
            "e2e_single_us": 566.625
          },
          "system": {
            "gpu_batch_us": 280.5332032953125,
            "gpu_single_us": 270.874996204,
            "e2e_batch_us": 289.220405797,
            "e2e_single_us": 511.25
          }
        },
        "new_over_old": {
          "gpu_batch_us": 1.0114073243038924,
          "gpu_single_us": 1.0051239367543132,
          "e2e_batch_us": 0.9966318350104042,
          "e2e_single_us": 0.9767953410381697
        },
        "same_bk16_new_over_old": {
          "gpu_batch_us": 0.9908952373161741,
          "gpu_single_us": 0.9884831848889274,
          "e2e_batch_us": 0.9783515091710557,
          "e2e_single_us": 0.95226996742671
        },
        "new_over_mps": {
          "gpu_batch_us": 1.0043281604173315,
          "gpu_single_us": 0.9956927941989842,
          "e2e_batch_us": 0.9857360161098191,
          "e2e_single_us": 0.9537095354523227
        },
        "new_over_torch": {
          "gpu_batch_us": 0.979496319685586,
          "gpu_single_us": 0.979125540062274,
          "e2e_batch_us": 0.938095303549284,
          "e2e_single_us": 0.8605056254136334
        }
      }
    ],
    "artifacts": 35,
    "validated_outputs": 192,
    "checked_elements": 995833488,
    "rejected_requests": 16,
    "current_artifacts_checked": true
  },
  "negative_probes_rejected": 5
}
