{
  "status": "pass",
  "audit_sha256": "094cebbe0ab088003cbc09b22d97086784003cfee46ba1f1aa7dc19669e279c9",
  "source_reports_sha256": {
    "reference": "4af0528c8887e71a49030ffd133b7a557d4c137adea3adf014b7fd5b114f45b8",
    "candidate": "7bbd1fb3cf0cce65ebefede283dc38fc4e474ca026a7113dba2b0130febacce3",
    "automatic": "fd5af778c262733d6e2426f5c4f57a232a013336ee1554af62317ab3b22396be"
  },
  "executed_output_validations": 456,
  "automatic_candidates": {
    "softmax_37x769": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 145.0,
        "waves": 1,
        "kernel_score": 145.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 137.0,
        "waves": 1,
        "kernel_score": 137.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 134.33333333333334,
        "waves": 1,
        "kernel_score": 134.33333333333334
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 133.0,
        "waves": 1,
        "kernel_score": 133.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 132.2,
        "waves": 1,
        "kernel_score": 132.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 131.66666666666666,
        "waves": 1,
        "kernel_score": 131.66666666666666
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 131.28571428571428,
        "waves": 1,
        "kernel_score": 131.28571428571428
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 131.0,
        "waves": 1,
        "kernel_score": 131.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 13,
        "rounds": 65,
        "storage": 26,
        "reductions": 2,
        "score": 89.0,
        "waves": 1,
        "kernel_score": 89.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 73.0,
        "waves": 1,
        "kernel_score": 73.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 72.0,
        "waves": 1,
        "kernel_score": 72.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 76.0,
        "waves": 1,
        "kernel_score": 76.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 5,
        "rounds": 25,
        "storage": 10,
        "reductions": 2,
        "score": 65.0,
        "waves": 1,
        "kernel_score": 65.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 64.0,
        "waves": 1,
        "kernel_score": 64.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 68.0,
        "waves": 1,
        "kernel_score": 68.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 72.0,
        "waves": 1,
        "kernel_score": 72.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 76.0,
        "waves": 1,
        "kernel_score": 76.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 80.0,
        "waves": 1,
        "kernel_score": 80.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 84.0,
        "waves": 1,
        "kernel_score": 84.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 88.0,
        "waves": 1,
        "kernel_score": 88.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 92.0,
        "waves": 1,
        "kernel_score": 92.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 96.0,
        "waves": 1,
        "kernel_score": 96.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 100.0,
        "waves": 1,
        "kernel_score": 100.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 104.0,
        "waves": 1,
        "kernel_score": 104.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 108.0,
        "waves": 1,
        "kernel_score": 108.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 112.0,
        "waves": 1,
        "kernel_score": 112.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 116.0,
        "waves": 1,
        "kernel_score": 116.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 120.0,
        "waves": 1,
        "kernel_score": 120.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 124.0,
        "waves": 1,
        "kernel_score": 124.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 128.0,
        "waves": 1,
        "kernel_score": 128.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 132.0,
        "waves": 1,
        "kernel_score": 132.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 136.0,
        "waves": 1,
        "kernel_score": 136.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 140.0,
        "waves": 1,
        "kernel_score": 140.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 144.0,
        "waves": 1,
        "kernel_score": 144.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 148.0,
        "waves": 1,
        "kernel_score": 148.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 152.0,
        "waves": 1,
        "kernel_score": 152.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 156.0,
        "waves": 1,
        "kernel_score": 156.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 160.0,
        "waves": 1,
        "kernel_score": 160.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 164.0,
        "waves": 1,
        "kernel_score": 164.0
      }
    ],
    "softmax_1024x1024": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 180.0,
        "waves": 16,
        "kernel_score": 2880.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 172.0,
        "waves": 16,
        "kernel_score": 2752.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 169.33333333333334,
        "waves": 16,
        "kernel_score": 2709.3333333333335
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 168.0,
        "waves": 16,
        "kernel_score": 2688.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 167.2,
        "waves": 16,
        "kernel_score": 2675.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.66666666666666,
        "waves": 16,
        "kernel_score": 2666.6666666666665
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.28571428571428,
        "waves": 16,
        "kernel_score": 2660.5714285714284
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.0,
        "waves": 16,
        "kernel_score": 2656.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 16,
        "rounds": 80,
        "storage": 32,
        "reductions": 2,
        "score": 104.0,
        "waves": 16,
        "kernel_score": 1664.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 12,
        "rounds": 60,
        "storage": 24,
        "reductions": 2,
        "score": 88.0,
        "waves": 16,
        "kernel_score": 1408.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 72.0,
        "waves": 16,
        "kernel_score": 1152.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 76.0,
        "waves": 16,
        "kernel_score": 1216.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 80.0,
        "waves": 16,
        "kernel_score": 1280.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 84.0,
        "waves": 16,
        "kernel_score": 1344.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 68.0,
        "waves": 16,
        "kernel_score": 1088.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 72.0,
        "waves": 16,
        "kernel_score": 1152.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 76.0,
        "waves": 16,
        "kernel_score": 1216.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 80.0,
        "waves": 16,
        "kernel_score": 1280.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 84.0,
        "waves": 16,
        "kernel_score": 1344.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 88.0,
        "waves": 16,
        "kernel_score": 1408.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 92.0,
        "waves": 16,
        "kernel_score": 1472.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 96.0,
        "waves": 16,
        "kernel_score": 1536.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 100.0,
        "waves": 16,
        "kernel_score": 1600.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 104.0,
        "waves": 16,
        "kernel_score": 1664.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 108.0,
        "waves": 16,
        "kernel_score": 1728.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 112.0,
        "waves": 16,
        "kernel_score": 1792.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 116.0,
        "waves": 16,
        "kernel_score": 1856.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 120.0,
        "waves": 16,
        "kernel_score": 1920.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 124.0,
        "waves": 16,
        "kernel_score": 1984.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 128.0,
        "waves": 16,
        "kernel_score": 2048.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 132.0,
        "waves": 16,
        "kernel_score": 2112.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 136.0,
        "waves": 16,
        "kernel_score": 2176.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 140.0,
        "waves": 16,
        "kernel_score": 2240.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 144.0,
        "waves": 16,
        "kernel_score": 2304.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 148.0,
        "waves": 16,
        "kernel_score": 2368.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 152.0,
        "waves": 16,
        "kernel_score": 2432.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 156.0,
        "waves": 16,
        "kernel_score": 2496.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 160.0,
        "waves": 16,
        "kernel_score": 2560.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 164.0,
        "waves": 16,
        "kernel_score": 2624.0
      }
    ],
    "softmax_16384x257": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 65.0,
        "waves": 256,
        "kernel_score": 16640.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 57.0,
        "waves": 256,
        "kernel_score": 14592.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 54.333333333333336,
        "waves": 256,
        "kernel_score": 13909.333333333334
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 53.0,
        "waves": 256,
        "kernel_score": 13568.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 52.2,
        "waves": 256,
        "kernel_score": 13363.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 51.666666666666664,
        "waves": 256,
        "kernel_score": 13226.666666666666
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 51.285714285714285,
        "waves": 256,
        "kernel_score": 13129.142857142857
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 51.0,
        "waves": 256,
        "kernel_score": 13056.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 5,
        "rounds": 25,
        "storage": 10,
        "reductions": 2,
        "score": 49.0,
        "waves": 256,
        "kernel_score": 12544.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 48.0,
        "waves": 256,
        "kernel_score": 12288.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 52.0,
        "waves": 256,
        "kernel_score": 13312.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 56.0,
        "waves": 256,
        "kernel_score": 14336.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 60.0,
        "waves": 256,
        "kernel_score": 15360.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 64.0,
        "waves": 256,
        "kernel_score": 16384.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 68.0,
        "waves": 256,
        "kernel_score": 17408.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 72.0,
        "waves": 256,
        "kernel_score": 18432.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 76.0,
        "waves": 256,
        "kernel_score": 19456.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 80.0,
        "waves": 256,
        "kernel_score": 20480.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 84.0,
        "waves": 256,
        "kernel_score": 21504.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 88.0,
        "waves": 256,
        "kernel_score": 22528.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 92.0,
        "waves": 256,
        "kernel_score": 23552.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 96.0,
        "waves": 256,
        "kernel_score": 24576.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 100.0,
        "waves": 256,
        "kernel_score": 25600.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 104.0,
        "waves": 256,
        "kernel_score": 26624.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 108.0,
        "waves": 256,
        "kernel_score": 27648.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 112.0,
        "waves": 256,
        "kernel_score": 28672.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 116.0,
        "waves": 256,
        "kernel_score": 29696.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 120.0,
        "waves": 256,
        "kernel_score": 30720.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 124.0,
        "waves": 256,
        "kernel_score": 31744.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 128.0,
        "waves": 256,
        "kernel_score": 32768.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 132.0,
        "waves": 256,
        "kernel_score": 33792.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 136.0,
        "waves": 256,
        "kernel_score": 34816.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 140.0,
        "waves": 256,
        "kernel_score": 35840.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 144.0,
        "waves": 256,
        "kernel_score": 36864.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 148.0,
        "waves": 256,
        "kernel_score": 37888.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 152.0,
        "waves": 256,
        "kernel_score": 38912.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 156.0,
        "waves": 256,
        "kernel_score": 39936.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 160.0,
        "waves": 256,
        "kernel_score": 40960.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 164.0,
        "waves": 256,
        "kernel_score": 41984.0
      }
    ],
    "softmax_4096x1024": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 180.0,
        "waves": 64,
        "kernel_score": 11520.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 172.0,
        "waves": 64,
        "kernel_score": 11008.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 169.33333333333334,
        "waves": 64,
        "kernel_score": 10837.333333333334
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 168.0,
        "waves": 64,
        "kernel_score": 10752.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 167.2,
        "waves": 64,
        "kernel_score": 10700.8
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.66666666666666,
        "waves": 64,
        "kernel_score": 10666.666666666666
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.28571428571428,
        "waves": 64,
        "kernel_score": 10642.285714285714
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.0,
        "waves": 64,
        "kernel_score": 10624.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 16,
        "rounds": 80,
        "storage": 32,
        "reductions": 2,
        "score": 104.0,
        "waves": 64,
        "kernel_score": 6656.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 12,
        "rounds": 60,
        "storage": 24,
        "reductions": 2,
        "score": 88.0,
        "waves": 64,
        "kernel_score": 5632.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 72.0,
        "waves": 64,
        "kernel_score": 4608.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 76.0,
        "waves": 64,
        "kernel_score": 4864.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 80.0,
        "waves": 64,
        "kernel_score": 5120.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 84.0,
        "waves": 64,
        "kernel_score": 5376.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 68.0,
        "waves": 64,
        "kernel_score": 4352.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 72.0,
        "waves": 64,
        "kernel_score": 4608.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 76.0,
        "waves": 64,
        "kernel_score": 4864.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 80.0,
        "waves": 64,
        "kernel_score": 5120.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 84.0,
        "waves": 64,
        "kernel_score": 5376.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 88.0,
        "waves": 64,
        "kernel_score": 5632.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 92.0,
        "waves": 64,
        "kernel_score": 5888.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 96.0,
        "waves": 64,
        "kernel_score": 6144.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 100.0,
        "waves": 64,
        "kernel_score": 6400.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 104.0,
        "waves": 64,
        "kernel_score": 6656.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 108.0,
        "waves": 64,
        "kernel_score": 6912.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 112.0,
        "waves": 64,
        "kernel_score": 7168.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 116.0,
        "waves": 64,
        "kernel_score": 7424.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 120.0,
        "waves": 64,
        "kernel_score": 7680.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 124.0,
        "waves": 64,
        "kernel_score": 7936.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 128.0,
        "waves": 64,
        "kernel_score": 8192.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 132.0,
        "waves": 64,
        "kernel_score": 8448.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 136.0,
        "waves": 64,
        "kernel_score": 8704.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 140.0,
        "waves": 64,
        "kernel_score": 8960.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 144.0,
        "waves": 64,
        "kernel_score": 9216.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 148.0,
        "waves": 64,
        "kernel_score": 9472.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 152.0,
        "waves": 64,
        "kernel_score": 9728.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 156.0,
        "waves": 64,
        "kernel_score": 9984.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 160.0,
        "waves": 64,
        "kernel_score": 10240.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 164.0,
        "waves": 64,
        "kernel_score": 10496.0
      }
    ],
    "rmsnorm_37x769": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 25,
        "rounds": 75,
        "storage": 25,
        "reductions": 1,
        "score": 93.0,
        "waves": 1,
        "kernel_score": 93.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 25,
        "rounds": 75,
        "storage": 25,
        "reductions": 1,
        "score": 85.0,
        "waves": 1,
        "kernel_score": 85.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 25,
        "rounds": 75,
        "storage": 25,
        "reductions": 1,
        "score": 82.33333333333333,
        "waves": 1,
        "kernel_score": 82.33333333333333
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 25,
        "rounds": 75,
        "storage": 25,
        "reductions": 1,
        "score": 81.0,
        "waves": 1,
        "kernel_score": 81.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 25,
        "rounds": 75,
        "storage": 25,
        "reductions": 1,
        "score": 80.2,
        "waves": 1,
        "kernel_score": 80.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 25,
        "rounds": 75,
        "storage": 25,
        "reductions": 1,
        "score": 79.66666666666667,
        "waves": 1,
        "kernel_score": 79.66666666666667
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 25,
        "rounds": 75,
        "storage": 25,
        "reductions": 1,
        "score": 79.28571428571429,
        "waves": 1,
        "kernel_score": 79.28571428571429
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 25,
        "rounds": 75,
        "storage": 25,
        "reductions": 1,
        "score": 79.0,
        "waves": 1,
        "kernel_score": 79.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 13,
        "rounds": 39,
        "storage": 13,
        "reductions": 1,
        "score": 59.0,
        "waves": 1,
        "kernel_score": 59.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 49.0,
        "waves": 1,
        "kernel_score": 49.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 48.0,
        "waves": 1,
        "kernel_score": 48.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 50.0,
        "waves": 1,
        "kernel_score": 50.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 5,
        "rounds": 15,
        "storage": 5,
        "reductions": 1,
        "score": 43.0,
        "waves": 1,
        "kernel_score": 43.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 42.0,
        "waves": 1,
        "kernel_score": 42.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 44.0,
        "waves": 1,
        "kernel_score": 44.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 46.0,
        "waves": 1,
        "kernel_score": 46.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 48.0,
        "waves": 1,
        "kernel_score": 48.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 50.0,
        "waves": 1,
        "kernel_score": 50.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 52.0,
        "waves": 1,
        "kernel_score": 52.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 54.0,
        "waves": 1,
        "kernel_score": 54.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 56.0,
        "waves": 1,
        "kernel_score": 56.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 58.0,
        "waves": 1,
        "kernel_score": 58.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 60.0,
        "waves": 1,
        "kernel_score": 60.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 62.0,
        "waves": 1,
        "kernel_score": 62.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 64.0,
        "waves": 1,
        "kernel_score": 64.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 66.0,
        "waves": 1,
        "kernel_score": 66.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 68.0,
        "waves": 1,
        "kernel_score": 68.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 70.0,
        "waves": 1,
        "kernel_score": 70.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 72.0,
        "waves": 1,
        "kernel_score": 72.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 74.0,
        "waves": 1,
        "kernel_score": 74.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 76.0,
        "waves": 1,
        "kernel_score": 76.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 78.0,
        "waves": 1,
        "kernel_score": 78.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 80.0,
        "waves": 1,
        "kernel_score": 80.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 82.0,
        "waves": 1,
        "kernel_score": 82.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 84.0,
        "waves": 1,
        "kernel_score": 84.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 86.0,
        "waves": 1,
        "kernel_score": 86.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 88.0,
        "waves": 1,
        "kernel_score": 88.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 90.0,
        "waves": 1,
        "kernel_score": 90.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 92.0,
        "waves": 1,
        "kernel_score": 92.0
      }
    ],
    "rmsnorm_1024x1024": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 114.0,
        "waves": 16,
        "kernel_score": 1824.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 106.0,
        "waves": 16,
        "kernel_score": 1696.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 103.33333333333333,
        "waves": 16,
        "kernel_score": 1653.3333333333333
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 102.0,
        "waves": 16,
        "kernel_score": 1632.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 101.2,
        "waves": 16,
        "kernel_score": 1619.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 100.66666666666667,
        "waves": 16,
        "kernel_score": 1610.6666666666667
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 100.28571428571429,
        "waves": 16,
        "kernel_score": 1604.5714285714287
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 100.0,
        "waves": 16,
        "kernel_score": 1600.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 16,
        "rounds": 48,
        "storage": 16,
        "reductions": 1,
        "score": 68.0,
        "waves": 16,
        "kernel_score": 1088.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 12,
        "rounds": 36,
        "storage": 12,
        "reductions": 1,
        "score": 58.0,
        "waves": 16,
        "kernel_score": 928.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 48.0,
        "waves": 16,
        "kernel_score": 768.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 50.0,
        "waves": 16,
        "kernel_score": 800.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 52.0,
        "waves": 16,
        "kernel_score": 832.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 54.0,
        "waves": 16,
        "kernel_score": 864.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 44.0,
        "waves": 16,
        "kernel_score": 704.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 46.0,
        "waves": 16,
        "kernel_score": 736.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 48.0,
        "waves": 16,
        "kernel_score": 768.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 50.0,
        "waves": 16,
        "kernel_score": 800.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 52.0,
        "waves": 16,
        "kernel_score": 832.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 54.0,
        "waves": 16,
        "kernel_score": 864.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 56.0,
        "waves": 16,
        "kernel_score": 896.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 58.0,
        "waves": 16,
        "kernel_score": 928.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 60.0,
        "waves": 16,
        "kernel_score": 960.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 62.0,
        "waves": 16,
        "kernel_score": 992.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 64.0,
        "waves": 16,
        "kernel_score": 1024.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 66.0,
        "waves": 16,
        "kernel_score": 1056.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 68.0,
        "waves": 16,
        "kernel_score": 1088.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 70.0,
        "waves": 16,
        "kernel_score": 1120.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 72.0,
        "waves": 16,
        "kernel_score": 1152.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 74.0,
        "waves": 16,
        "kernel_score": 1184.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 76.0,
        "waves": 16,
        "kernel_score": 1216.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 78.0,
        "waves": 16,
        "kernel_score": 1248.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 80.0,
        "waves": 16,
        "kernel_score": 1280.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 82.0,
        "waves": 16,
        "kernel_score": 1312.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 84.0,
        "waves": 16,
        "kernel_score": 1344.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 86.0,
        "waves": 16,
        "kernel_score": 1376.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 88.0,
        "waves": 16,
        "kernel_score": 1408.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 90.0,
        "waves": 16,
        "kernel_score": 1440.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 92.0,
        "waves": 16,
        "kernel_score": 1472.0
      }
    ],
    "rmsnorm_16384x257": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 45.0,
        "waves": 256,
        "kernel_score": 11520.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 37.0,
        "waves": 256,
        "kernel_score": 9472.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 34.333333333333336,
        "waves": 256,
        "kernel_score": 8789.333333333334
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 33.0,
        "waves": 256,
        "kernel_score": 8448.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 32.2,
        "waves": 256,
        "kernel_score": 8243.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 31.666666666666668,
        "waves": 256,
        "kernel_score": 8106.666666666667
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 31.285714285714285,
        "waves": 256,
        "kernel_score": 8009.142857142857
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 9,
        "rounds": 27,
        "storage": 9,
        "reductions": 1,
        "score": 31.0,
        "waves": 256,
        "kernel_score": 7936.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 5,
        "rounds": 15,
        "storage": 5,
        "reductions": 1,
        "score": 35.0,
        "waves": 256,
        "kernel_score": 8960.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 34.0,
        "waves": 256,
        "kernel_score": 8704.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 36.0,
        "waves": 256,
        "kernel_score": 9216.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 38.0,
        "waves": 256,
        "kernel_score": 9728.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 40.0,
        "waves": 256,
        "kernel_score": 10240.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 42.0,
        "waves": 256,
        "kernel_score": 10752.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 44.0,
        "waves": 256,
        "kernel_score": 11264.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 46.0,
        "waves": 256,
        "kernel_score": 11776.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 48.0,
        "waves": 256,
        "kernel_score": 12288.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 50.0,
        "waves": 256,
        "kernel_score": 12800.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 52.0,
        "waves": 256,
        "kernel_score": 13312.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 54.0,
        "waves": 256,
        "kernel_score": 13824.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 56.0,
        "waves": 256,
        "kernel_score": 14336.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 58.0,
        "waves": 256,
        "kernel_score": 14848.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 60.0,
        "waves": 256,
        "kernel_score": 15360.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 62.0,
        "waves": 256,
        "kernel_score": 15872.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 64.0,
        "waves": 256,
        "kernel_score": 16384.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 66.0,
        "waves": 256,
        "kernel_score": 16896.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 68.0,
        "waves": 256,
        "kernel_score": 17408.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 70.0,
        "waves": 256,
        "kernel_score": 17920.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 72.0,
        "waves": 256,
        "kernel_score": 18432.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 74.0,
        "waves": 256,
        "kernel_score": 18944.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 76.0,
        "waves": 256,
        "kernel_score": 19456.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 78.0,
        "waves": 256,
        "kernel_score": 19968.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 80.0,
        "waves": 256,
        "kernel_score": 20480.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 82.0,
        "waves": 256,
        "kernel_score": 20992.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 84.0,
        "waves": 256,
        "kernel_score": 21504.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 86.0,
        "waves": 256,
        "kernel_score": 22016.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 88.0,
        "waves": 256,
        "kernel_score": 22528.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 90.0,
        "waves": 256,
        "kernel_score": 23040.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 92.0,
        "waves": 256,
        "kernel_score": 23552.0
      }
    ],
    "rmsnorm_4096x1024": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 114.0,
        "waves": 64,
        "kernel_score": 7296.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 106.0,
        "waves": 64,
        "kernel_score": 6784.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 103.33333333333333,
        "waves": 64,
        "kernel_score": 6613.333333333333
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 102.0,
        "waves": 64,
        "kernel_score": 6528.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 101.2,
        "waves": 64,
        "kernel_score": 6476.8
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 100.66666666666667,
        "waves": 64,
        "kernel_score": 6442.666666666667
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 100.28571428571429,
        "waves": 64,
        "kernel_score": 6418.285714285715
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 32,
        "rounds": 96,
        "storage": 32,
        "reductions": 1,
        "score": 100.0,
        "waves": 64,
        "kernel_score": 6400.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 16,
        "rounds": 48,
        "storage": 16,
        "reductions": 1,
        "score": 68.0,
        "waves": 64,
        "kernel_score": 4352.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 12,
        "rounds": 36,
        "storage": 12,
        "reductions": 1,
        "score": 58.0,
        "waves": 64,
        "kernel_score": 3712.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 48.0,
        "waves": 64,
        "kernel_score": 3072.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 50.0,
        "waves": 64,
        "kernel_score": 3200.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 52.0,
        "waves": 64,
        "kernel_score": 3328.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 8,
        "rounds": 24,
        "storage": 8,
        "reductions": 1,
        "score": 54.0,
        "waves": 64,
        "kernel_score": 3456.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 44.0,
        "waves": 64,
        "kernel_score": 2816.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 46.0,
        "waves": 64,
        "kernel_score": 2944.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 48.0,
        "waves": 64,
        "kernel_score": 3072.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 50.0,
        "waves": 64,
        "kernel_score": 3200.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 52.0,
        "waves": 64,
        "kernel_score": 3328.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 54.0,
        "waves": 64,
        "kernel_score": 3456.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 56.0,
        "waves": 64,
        "kernel_score": 3584.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 58.0,
        "waves": 64,
        "kernel_score": 3712.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 60.0,
        "waves": 64,
        "kernel_score": 3840.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 62.0,
        "waves": 64,
        "kernel_score": 3968.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 64.0,
        "waves": 64,
        "kernel_score": 4096.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 66.0,
        "waves": 64,
        "kernel_score": 4224.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 68.0,
        "waves": 64,
        "kernel_score": 4352.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 70.0,
        "waves": 64,
        "kernel_score": 4480.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 72.0,
        "waves": 64,
        "kernel_score": 4608.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 74.0,
        "waves": 64,
        "kernel_score": 4736.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 76.0,
        "waves": 64,
        "kernel_score": 4864.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 78.0,
        "waves": 64,
        "kernel_score": 4992.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 80.0,
        "waves": 64,
        "kernel_score": 5120.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 82.0,
        "waves": 64,
        "kernel_score": 5248.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 84.0,
        "waves": 64,
        "kernel_score": 5376.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 86.0,
        "waves": 64,
        "kernel_score": 5504.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 88.0,
        "waves": 64,
        "kernel_score": 5632.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 90.0,
        "waves": 64,
        "kernel_score": 5760.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 12,
        "storage": 4,
        "reductions": 1,
        "score": 92.0,
        "waves": 64,
        "kernel_score": 5888.0
      }
    ],
    "layernorm_37x769": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 145.0,
        "waves": 1,
        "kernel_score": 145.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 137.0,
        "waves": 1,
        "kernel_score": 137.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 134.33333333333334,
        "waves": 1,
        "kernel_score": 134.33333333333334
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 133.0,
        "waves": 1,
        "kernel_score": 133.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 132.2,
        "waves": 1,
        "kernel_score": 132.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 131.66666666666666,
        "waves": 1,
        "kernel_score": 131.66666666666666
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 131.28571428571428,
        "waves": 1,
        "kernel_score": 131.28571428571428
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 25,
        "rounds": 125,
        "storage": 50,
        "reductions": 2,
        "score": 131.0,
        "waves": 1,
        "kernel_score": 131.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 13,
        "rounds": 65,
        "storage": 26,
        "reductions": 2,
        "score": 89.0,
        "waves": 1,
        "kernel_score": 89.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 73.0,
        "waves": 1,
        "kernel_score": 73.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 72.0,
        "waves": 1,
        "kernel_score": 72.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 76.0,
        "waves": 1,
        "kernel_score": 76.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 5,
        "rounds": 25,
        "storage": 10,
        "reductions": 2,
        "score": 65.0,
        "waves": 1,
        "kernel_score": 65.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 64.0,
        "waves": 1,
        "kernel_score": 64.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 68.0,
        "waves": 1,
        "kernel_score": 68.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 72.0,
        "waves": 1,
        "kernel_score": 72.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 76.0,
        "waves": 1,
        "kernel_score": 76.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 80.0,
        "waves": 1,
        "kernel_score": 80.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 84.0,
        "waves": 1,
        "kernel_score": 84.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 88.0,
        "waves": 1,
        "kernel_score": 88.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 92.0,
        "waves": 1,
        "kernel_score": 92.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 96.0,
        "waves": 1,
        "kernel_score": 96.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 100.0,
        "waves": 1,
        "kernel_score": 100.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 104.0,
        "waves": 1,
        "kernel_score": 104.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 108.0,
        "waves": 1,
        "kernel_score": 108.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 112.0,
        "waves": 1,
        "kernel_score": 112.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 116.0,
        "waves": 1,
        "kernel_score": 116.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 120.0,
        "waves": 1,
        "kernel_score": 120.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 124.0,
        "waves": 1,
        "kernel_score": 124.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 128.0,
        "waves": 1,
        "kernel_score": 128.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 132.0,
        "waves": 1,
        "kernel_score": 132.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 136.0,
        "waves": 1,
        "kernel_score": 136.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 140.0,
        "waves": 1,
        "kernel_score": 140.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 144.0,
        "waves": 1,
        "kernel_score": 144.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 148.0,
        "waves": 1,
        "kernel_score": 148.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 152.0,
        "waves": 1,
        "kernel_score": 152.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 156.0,
        "waves": 1,
        "kernel_score": 156.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 160.0,
        "waves": 1,
        "kernel_score": 160.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 164.0,
        "waves": 1,
        "kernel_score": 164.0
      }
    ],
    "layernorm_1024x1024": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 180.0,
        "waves": 16,
        "kernel_score": 2880.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 172.0,
        "waves": 16,
        "kernel_score": 2752.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 169.33333333333334,
        "waves": 16,
        "kernel_score": 2709.3333333333335
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 168.0,
        "waves": 16,
        "kernel_score": 2688.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 167.2,
        "waves": 16,
        "kernel_score": 2675.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.66666666666666,
        "waves": 16,
        "kernel_score": 2666.6666666666665
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.28571428571428,
        "waves": 16,
        "kernel_score": 2660.5714285714284
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.0,
        "waves": 16,
        "kernel_score": 2656.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 16,
        "rounds": 80,
        "storage": 32,
        "reductions": 2,
        "score": 104.0,
        "waves": 16,
        "kernel_score": 1664.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 12,
        "rounds": 60,
        "storage": 24,
        "reductions": 2,
        "score": 88.0,
        "waves": 16,
        "kernel_score": 1408.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 72.0,
        "waves": 16,
        "kernel_score": 1152.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 76.0,
        "waves": 16,
        "kernel_score": 1216.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 80.0,
        "waves": 16,
        "kernel_score": 1280.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 84.0,
        "waves": 16,
        "kernel_score": 1344.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 68.0,
        "waves": 16,
        "kernel_score": 1088.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 72.0,
        "waves": 16,
        "kernel_score": 1152.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 76.0,
        "waves": 16,
        "kernel_score": 1216.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 80.0,
        "waves": 16,
        "kernel_score": 1280.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 84.0,
        "waves": 16,
        "kernel_score": 1344.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 88.0,
        "waves": 16,
        "kernel_score": 1408.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 92.0,
        "waves": 16,
        "kernel_score": 1472.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 96.0,
        "waves": 16,
        "kernel_score": 1536.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 100.0,
        "waves": 16,
        "kernel_score": 1600.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 104.0,
        "waves": 16,
        "kernel_score": 1664.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 108.0,
        "waves": 16,
        "kernel_score": 1728.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 112.0,
        "waves": 16,
        "kernel_score": 1792.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 116.0,
        "waves": 16,
        "kernel_score": 1856.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 120.0,
        "waves": 16,
        "kernel_score": 1920.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 124.0,
        "waves": 16,
        "kernel_score": 1984.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 128.0,
        "waves": 16,
        "kernel_score": 2048.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 132.0,
        "waves": 16,
        "kernel_score": 2112.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 136.0,
        "waves": 16,
        "kernel_score": 2176.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 140.0,
        "waves": 16,
        "kernel_score": 2240.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 144.0,
        "waves": 16,
        "kernel_score": 2304.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 148.0,
        "waves": 16,
        "kernel_score": 2368.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 152.0,
        "waves": 16,
        "kernel_score": 2432.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 156.0,
        "waves": 16,
        "kernel_score": 2496.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 160.0,
        "waves": 16,
        "kernel_score": 2560.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 164.0,
        "waves": 16,
        "kernel_score": 2624.0
      }
    ],
    "layernorm_16384x257": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 65.0,
        "waves": 256,
        "kernel_score": 16640.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 57.0,
        "waves": 256,
        "kernel_score": 14592.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 54.333333333333336,
        "waves": 256,
        "kernel_score": 13909.333333333334
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 53.0,
        "waves": 256,
        "kernel_score": 13568.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 52.2,
        "waves": 256,
        "kernel_score": 13363.2
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 51.666666666666664,
        "waves": 256,
        "kernel_score": 13226.666666666666
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 51.285714285714285,
        "waves": 256,
        "kernel_score": 13129.142857142857
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 9,
        "rounds": 45,
        "storage": 18,
        "reductions": 2,
        "score": 51.0,
        "waves": 256,
        "kernel_score": 13056.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 5,
        "rounds": 25,
        "storage": 10,
        "reductions": 2,
        "score": 49.0,
        "waves": 256,
        "kernel_score": 12544.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 48.0,
        "waves": 256,
        "kernel_score": 12288.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 52.0,
        "waves": 256,
        "kernel_score": 13312.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 56.0,
        "waves": 256,
        "kernel_score": 14336.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 60.0,
        "waves": 256,
        "kernel_score": 15360.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 64.0,
        "waves": 256,
        "kernel_score": 16384.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 68.0,
        "waves": 256,
        "kernel_score": 17408.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 72.0,
        "waves": 256,
        "kernel_score": 18432.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 76.0,
        "waves": 256,
        "kernel_score": 19456.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 80.0,
        "waves": 256,
        "kernel_score": 20480.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 84.0,
        "waves": 256,
        "kernel_score": 21504.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 88.0,
        "waves": 256,
        "kernel_score": 22528.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 92.0,
        "waves": 256,
        "kernel_score": 23552.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 96.0,
        "waves": 256,
        "kernel_score": 24576.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 100.0,
        "waves": 256,
        "kernel_score": 25600.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 104.0,
        "waves": 256,
        "kernel_score": 26624.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 108.0,
        "waves": 256,
        "kernel_score": 27648.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 112.0,
        "waves": 256,
        "kernel_score": 28672.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 116.0,
        "waves": 256,
        "kernel_score": 29696.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 120.0,
        "waves": 256,
        "kernel_score": 30720.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 124.0,
        "waves": 256,
        "kernel_score": 31744.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 128.0,
        "waves": 256,
        "kernel_score": 32768.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 132.0,
        "waves": 256,
        "kernel_score": 33792.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 136.0,
        "waves": 256,
        "kernel_score": 34816.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 140.0,
        "waves": 256,
        "kernel_score": 35840.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 144.0,
        "waves": 256,
        "kernel_score": 36864.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 148.0,
        "waves": 256,
        "kernel_score": 37888.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 152.0,
        "waves": 256,
        "kernel_score": 38912.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 156.0,
        "waves": 256,
        "kernel_score": 39936.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 160.0,
        "waves": 256,
        "kernel_score": 40960.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 164.0,
        "waves": 256,
        "kernel_score": 41984.0
      }
    ],
    "layernorm_4096x1024": [
      {
        "subgroups": 1,
        "packing": 1,
        "width": 32,
        "threads": 32,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 180.0,
        "waves": 64,
        "kernel_score": 11520.0
      },
      {
        "subgroups": 1,
        "packing": 2,
        "width": 32,
        "threads": 64,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 172.0,
        "waves": 64,
        "kernel_score": 11008.0
      },
      {
        "subgroups": 1,
        "packing": 3,
        "width": 32,
        "threads": 96,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 169.33333333333334,
        "waves": 64,
        "kernel_score": 10837.333333333334
      },
      {
        "subgroups": 1,
        "packing": 4,
        "width": 32,
        "threads": 128,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 168.0,
        "waves": 64,
        "kernel_score": 10752.0
      },
      {
        "subgroups": 1,
        "packing": 5,
        "width": 32,
        "threads": 160,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 167.2,
        "waves": 64,
        "kernel_score": 10700.8
      },
      {
        "subgroups": 1,
        "packing": 6,
        "width": 32,
        "threads": 192,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.66666666666666,
        "waves": 64,
        "kernel_score": 10666.666666666666
      },
      {
        "subgroups": 1,
        "packing": 7,
        "width": 32,
        "threads": 224,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.28571428571428,
        "waves": 64,
        "kernel_score": 10642.285714285714
      },
      {
        "subgroups": 1,
        "packing": 8,
        "width": 32,
        "threads": 256,
        "slots": 32,
        "rounds": 160,
        "storage": 64,
        "reductions": 2,
        "score": 166.0,
        "waves": 64,
        "kernel_score": 10624.0
      },
      {
        "subgroups": 2,
        "packing": 1,
        "width": 64,
        "threads": 64,
        "slots": 16,
        "rounds": 80,
        "storage": 32,
        "reductions": 2,
        "score": 104.0,
        "waves": 64,
        "kernel_score": 6656.0
      },
      {
        "subgroups": 3,
        "packing": 1,
        "width": 96,
        "threads": 96,
        "slots": 12,
        "rounds": 60,
        "storage": 24,
        "reductions": 2,
        "score": 88.0,
        "waves": 64,
        "kernel_score": 5632.0
      },
      {
        "subgroups": 4,
        "packing": 1,
        "width": 128,
        "threads": 128,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 72.0,
        "waves": 64,
        "kernel_score": 4608.0
      },
      {
        "subgroups": 5,
        "packing": 1,
        "width": 160,
        "threads": 160,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 76.0,
        "waves": 64,
        "kernel_score": 4864.0
      },
      {
        "subgroups": 6,
        "packing": 1,
        "width": 192,
        "threads": 192,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 80.0,
        "waves": 64,
        "kernel_score": 5120.0
      },
      {
        "subgroups": 7,
        "packing": 1,
        "width": 224,
        "threads": 224,
        "slots": 8,
        "rounds": 40,
        "storage": 16,
        "reductions": 2,
        "score": 84.0,
        "waves": 64,
        "kernel_score": 5376.0
      },
      {
        "subgroups": 8,
        "packing": 1,
        "width": 256,
        "threads": 256,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 68.0,
        "waves": 64,
        "kernel_score": 4352.0
      },
      {
        "subgroups": 9,
        "packing": 1,
        "width": 288,
        "threads": 288,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 72.0,
        "waves": 64,
        "kernel_score": 4608.0
      },
      {
        "subgroups": 10,
        "packing": 1,
        "width": 320,
        "threads": 320,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 76.0,
        "waves": 64,
        "kernel_score": 4864.0
      },
      {
        "subgroups": 11,
        "packing": 1,
        "width": 352,
        "threads": 352,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 80.0,
        "waves": 64,
        "kernel_score": 5120.0
      },
      {
        "subgroups": 12,
        "packing": 1,
        "width": 384,
        "threads": 384,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 84.0,
        "waves": 64,
        "kernel_score": 5376.0
      },
      {
        "subgroups": 13,
        "packing": 1,
        "width": 416,
        "threads": 416,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 88.0,
        "waves": 64,
        "kernel_score": 5632.0
      },
      {
        "subgroups": 14,
        "packing": 1,
        "width": 448,
        "threads": 448,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 92.0,
        "waves": 64,
        "kernel_score": 5888.0
      },
      {
        "subgroups": 15,
        "packing": 1,
        "width": 480,
        "threads": 480,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 96.0,
        "waves": 64,
        "kernel_score": 6144.0
      },
      {
        "subgroups": 16,
        "packing": 1,
        "width": 512,
        "threads": 512,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 100.0,
        "waves": 64,
        "kernel_score": 6400.0
      },
      {
        "subgroups": 17,
        "packing": 1,
        "width": 544,
        "threads": 544,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 104.0,
        "waves": 64,
        "kernel_score": 6656.0
      },
      {
        "subgroups": 18,
        "packing": 1,
        "width": 576,
        "threads": 576,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 108.0,
        "waves": 64,
        "kernel_score": 6912.0
      },
      {
        "subgroups": 19,
        "packing": 1,
        "width": 608,
        "threads": 608,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 112.0,
        "waves": 64,
        "kernel_score": 7168.0
      },
      {
        "subgroups": 20,
        "packing": 1,
        "width": 640,
        "threads": 640,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 116.0,
        "waves": 64,
        "kernel_score": 7424.0
      },
      {
        "subgroups": 21,
        "packing": 1,
        "width": 672,
        "threads": 672,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 120.0,
        "waves": 64,
        "kernel_score": 7680.0
      },
      {
        "subgroups": 22,
        "packing": 1,
        "width": 704,
        "threads": 704,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 124.0,
        "waves": 64,
        "kernel_score": 7936.0
      },
      {
        "subgroups": 23,
        "packing": 1,
        "width": 736,
        "threads": 736,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 128.0,
        "waves": 64,
        "kernel_score": 8192.0
      },
      {
        "subgroups": 24,
        "packing": 1,
        "width": 768,
        "threads": 768,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 132.0,
        "waves": 64,
        "kernel_score": 8448.0
      },
      {
        "subgroups": 25,
        "packing": 1,
        "width": 800,
        "threads": 800,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 136.0,
        "waves": 64,
        "kernel_score": 8704.0
      },
      {
        "subgroups": 26,
        "packing": 1,
        "width": 832,
        "threads": 832,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 140.0,
        "waves": 64,
        "kernel_score": 8960.0
      },
      {
        "subgroups": 27,
        "packing": 1,
        "width": 864,
        "threads": 864,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 144.0,
        "waves": 64,
        "kernel_score": 9216.0
      },
      {
        "subgroups": 28,
        "packing": 1,
        "width": 896,
        "threads": 896,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 148.0,
        "waves": 64,
        "kernel_score": 9472.0
      },
      {
        "subgroups": 29,
        "packing": 1,
        "width": 928,
        "threads": 928,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 152.0,
        "waves": 64,
        "kernel_score": 9728.0
      },
      {
        "subgroups": 30,
        "packing": 1,
        "width": 960,
        "threads": 960,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 156.0,
        "waves": 64,
        "kernel_score": 9984.0
      },
      {
        "subgroups": 31,
        "packing": 1,
        "width": 992,
        "threads": 992,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 160.0,
        "waves": 64,
        "kernel_score": 10240.0
      },
      {
        "subgroups": 32,
        "packing": 1,
        "width": 1024,
        "threads": 1024,
        "slots": 4,
        "rounds": 20,
        "storage": 8,
        "reductions": 2,
        "score": 164.0,
        "waves": 64,
        "kernel_score": 10496.0
      }
    ]
  },
  "replays": {
    "fixed-replay": {
      "baseline": "reference",
      "replay_sha256": "ee435e3554d0dbac024fec4d2a3db226e8a8528a0ff666c1ffa477b93c0643f7",
      "replay_pairs": 48,
      "recorded_unchanged_artifacts": 21,
      "current_artifacts_checked": true,
      "torch_probe_control_ratio_range": [
        0.5312412858736668,
        6.777120248426219
      ],
      "modeled_candidate_better": 12,
      "summaries": {
        "gpu_throughput": {
          "positive_medians": 7,
          "all_pairs_positive": 4,
          "all_pairs_negative": 2,
          "faster_torch_all_pairs": 10,
          "slower_torch_all_pairs": 0,
          "baseline_faster_torch_all_pairs": 10,
          "baseline_slower_torch_all_pairs": 1
        },
        "e2e_throughput": {
          "positive_medians": 6,
          "all_pairs_positive": 3,
          "all_pairs_negative": 3,
          "faster_torch_all_pairs": 12,
          "slower_torch_all_pairs": 0,
          "baseline_faster_torch_all_pairs": 12,
          "baseline_slower_torch_all_pairs": 0
        },
        "gpu_latency": {
          "positive_medians": 6,
          "all_pairs_positive": 2,
          "all_pairs_negative": 1,
          "faster_torch_all_pairs": 9,
          "slower_torch_all_pairs": 0,
          "baseline_faster_torch_all_pairs": 10,
          "baseline_slower_torch_all_pairs": 0
        },
        "e2e_latency": {
          "positive_medians": 11,
          "all_pairs_positive": 2,
          "all_pairs_negative": 0,
          "faster_torch_all_pairs": 7,
          "slower_torch_all_pairs": 1,
          "baseline_faster_torch_all_pairs": 8,
          "baseline_slower_torch_all_pairs": 0
        }
      },
      "records": [
        {
          "name": "softmax_37x769",
          "baseline_kernel_score": 131,
          "candidate_kernel_score": 77,
          "gpu_throughput": {
            "reference_us": 5.235026037554688,
            "candidate_us": 4.719075604953125,
            "torch_us": 15.605793578288285,
            "reference_rounds_us": [
              5.228514964984376,
              5.241537110125,
              5.115234671390626,
              5.532551767828125
            ],
            "candidate_rounds_us": [
              4.7975263441875,
              4.46289141109375,
              4.867188181375,
              4.64062486571875
            ],
            "torch_rounds_us": [
              14.270181054598652,
              16.05403576832032,
              15.157551388256252,
              21.067707166366745
            ],
            "reference_torch_us": 14.40136702512973,
            "reference_torch_rounds_us": [
              12.790364053216763,
              16.012369997042697,
              12.736977623717394,
              19.764974240388256
            ],
            "reference_native_torch_ratio": 0.3644740265337195,
            "reference_ratio_min": 0.2799169733559666,
            "reference_ratio_max": 0.40878546875054816,
            "gain": 1.1321533602039713,
            "gain_min": 1.050962995629553,
            "gain_max": 1.1921997420429784,
            "native_torch_ratio": 0.2995491815056975,
            "ratio_min": 0.22027194649483323,
            "ratio_max": 0.33619239488495967
          },
          "e2e_throughput": {
            "reference_us": 7.375395606425,
            "candidate_us": 6.399175420095,
            "torch_us": 33.2233991824404,
            "reference_rounds_us": [
              7.5673140977,
              7.41886431227,
              7.33192690058,
              5.76154223744
            ],
            "candidate_rounds_us": [
              6.57450878754,
              6.22384205265,
              6.70937741833,
              5.11850996016
            ],
            "torch_rounds_us": [
              30.630219206680586,
              38.17735419847328,
              33.529815699658705,
              32.9169826652221
            ],
            "reference_torch_us": 34.64488693943239,
            "reference_torch_rounds_us": [
              28.78749482163406,
              36.48180565805658,
              35.76017974683544,
              33.52959413202934
            ],
            "reference_native_torch_ratio": 0.2041942038328996,
            "reference_ratio_min": 0.17183453562702847,
            "reference_ratio_max": 0.2628681010482752,
            "gain": 1.138318544810987,
            "gain_min": 1.0927879657729787,
            "gain_max": 1.19200716366367,
            "native_torch_ratio": 0.181563137518931,
            "ratio_min": 0.15549754399475618,
            "ratio_max": 0.21464125813719515
          },
          "gpu_latency": {
            "reference_us": 10.020856279870001,
            "candidate_us": 7.812457624825,
            "torch_us": 71.60418317653239,
            "reference_rounds_us": [
              11.958356481000001,
              10.333373211300001,
              9.708339348440001,
              8.00000270829
            ],
            "candidate_rounds_us": [
              7.70826591179,
              7.91664933786,
              9.87498788163,
              7.24999699742
            ],
            "torch_rounds_us": [
              70.95834007486701,
              69.16665006428957,
              72.25002627819777,
              102.16666851192713
            ],
            "reference_torch_us": 78.64582585170865,
            "reference_torch_rounds_us": [
              71.8749943189323,
              85.416657384485,
              57.708355598151684,
              131.00000796839595
            ],
            "reference_native_torch_ratio": 0.143676618282902,
            "reference_ratio_min": 0.06106871924939133,
            "reference_ratio_max": 0.16823108625800012,
            "gain": 1.2043600793167373,
            "gain_min": 0.9831241784610179,
            "gain_max": 1.5513679234533637,
            "native_torch_ratio": 0.11154423698684199,
            "ratio_min": 0.07096244893777291,
            "ratio_max": 0.1366779832523035
          },
          "e2e_latency": {
            "reference_us": 223.7295,
            "candidate_us": 212.6255,
            "torch_us": 374.79200000000003,
            "reference_rounds_us": [
              244.958,
              203.333,
              223.709,
              223.75
            ],
            "candidate_rounds_us": [
              201.458,
              256.625,
              219.792,
              205.459
            ],
            "torch_rounds_us": [
              485.25,
              346.45799999999997,
              377.375,
              372.209
            ],
            "reference_torch_us": 385.08299999999997,
            "reference_torch_rounds_us": [
              439.0,
              331.625,
              373.791,
              396.375
            ],
            "reference_native_torch_ratio": 0.5814887759473661,
            "reference_ratio_min": 0.557990888382688,
            "reference_ratio_max": 0.6131413494157557,
            "gain": 1.053423227867154,
            "gain_min": 0.7923351193375547,
            "gain_max": 1.2159259001876321,
            "native_torch_ratio": 0.5672111651439787,
            "ratio_min": 0.4151633178773828,
            "ratio_max": 0.7407102736839676
          }
        },
        {
          "name": "softmax_1024x1024",
          "baseline_kernel_score": 2656,
          "candidate_kernel_score": 1472,
          "gpu_throughput": {
            "reference_us": 16.790038898828122,
            "candidate_us": 21.430013021015625,
            "torch_us": 42.98111934986082,
            "reference_rounds_us": [
              20.593750377812498,
              16.75716157484375,
              16.72395865171875,
              16.8229162228125
            ],
            "candidate_rounds_us": [
              23.758463612499998,
              23.22330692546875,
              18.123697373125,
              19.6367191165625
            ],
            "torch_rounds_us": [
              43.58528713055421,
              42.77734387869714,
              42.80794382793829,
              43.15429487178335
            ],
            "reference_torch_us": 43.71549630377558,
            "reference_torch_rounds_us": [
              43.896486204175744,
              43.53450640337542,
              42.88216405257117,
              45.425130338117015
            ],
            "reference_native_torch_ratio": 0.3874573899234732,
            "reference_ratio_min": 0.3703438184457139,
            "reference_ratio_max": 0.469143481827333,
            "gain": 0.861751730771058,
            "gain_min": 0.7215665550398574,
            "gain_max": 0.9227674854314288,
            "native_torch_ratio": 0.4989615557735749,
            "ratio_min": 0.42337229384273073,
            "ratio_max": 0.545102835765072
          },
          "e2e_throughput": {
            "reference_us": 20.3076440698,
            "candidate_us": 23.5222229044,
            "torch_us": 62.479121796885636,
            "reference_rounds_us": [
              20.3209818044,
              20.2943063352,
              21.1153218504,
              17.0616703085
            ],
            "candidate_rounds_us": [
              24.8967924883,
              23.0354146101,
              24.0090311987,
              19.6845462687
            ],
            "torch_rounds_us": [
              61.77847721822542,
              63.17976637554586,
              64.38997395833333,
              50.57826475849732
            ],
            "reference_torch_us": 62.981858447716775,
            "reference_torch_rounds_us": [
              62.60666145833333,
              64.86940051679588,
              63.35705543710022,
              49.22646808510638
            ],
            "reference_native_torch_ratio": 0.32892838518802836,
            "reference_ratio_min": 0.31284868017156153,
            "reference_ratio_max": 0.34659545915426054,
            "gain": 0.8731143472578705,
            "gain_min": 0.8162088274603904,
            "gain_max": 0.8810046043756407,
            "native_torch_ratio": 0.3810294616378297,
            "ratio_min": 0.3646011362747933,
            "ratio_max": 0.4030010710745576
          },
          "gpu_latency": {
            "reference_us": 27.18751784415,
            "candidate_us": 30.583323678,
            "torch_us": 67.35415081493556,
            "reference_rounds_us": [
              30.8750313707,
              59.666635934300004,
              23.5000043176,
              22.3333481699
            ],
            "candidate_rounds_us": [
              32.2500127368,
              30.958326533399998,
              30.208320822599998,
              24.0833032876
            ],
            "torch_rounds_us": [
              67.87496386095881,
              66.83333776891232,
              68.1666424497962,
              66.16662722080946
            ],
            "reference_torch_us": 68.77083797007799,
            "reference_torch_rounds_us": [
              69.2916801199317,
              74.74998710677028,
              67.58334347978234,
              68.24999582022429
            ],
            "reference_native_torch_ratio": 0.3966497600555131,
            "reference_ratio_min": 0.32722856465409533,
            "reference_ratio_max": 0.7982160030218368,
            "gain": 0.9423511744151409,
            "gain_min": 0.7779314995893035,
            "gain_max": 1.927321099540942,
            "native_torch_ratio": 0.45318541662586775,
            "ratio_min": 0.363979611764551,
            "ratio_max": 0.47513856217829936
          },
          "e2e_latency": {
            "reference_us": 283.20849999999996,
            "candidate_us": 297.25,
            "torch_us": 398.08349999999996,
            "reference_rounds_us": [
              270.417,
              306.166,
              296,
              251.209
            ],
            "candidate_rounds_us": [
              245.416,
              305.708,
              300.25,
              294.25
            ],
            "torch_rounds_us": [
              472.75,
              426.542,
              369.625,
              363.875
            ],
            "reference_torch_us": 405.375,
            "reference_torch_rounds_us": [
              513.583,
              414.375,
              392.292,
              396.375
            ],
            "reference_native_torch_ratio": 0.6863140730974,
            "reference_ratio_min": 0.5265302784554784,
            "reference_ratio_max": 0.7545399855209896,
            "gain": 0.9936716453517811,
            "gain_min": 0.8537264231096007,
            "gain_max": 1.1018719235909638,
            "native_torch_ratio": 0.7626846792038375,
            "ratio_min": 0.5191242728714965,
            "ratio_max": 0.8123097734190058
          }
        },
        {
          "name": "softmax_16384x257",
          "baseline_kernel_score": 13056,
          "candidate_kernel_score": 9472,
          "gpu_throughput": {
            "reference_us": 52.32063813304688,
            "candidate_us": 51.146810164921874,
            "torch_us": 205.2311215265945,
            "reference_rounds_us": [
              52.629557103437506,
              54.94661490953125,
              52.01171916265625,
              47.194661419843754
            ],
            "candidate_rounds_us": [
              54.67578102984375,
              53.715494686875,
              47.923177589843746,
              48.57812564296875
            ],
            "torch_rounds_us": [
              207.32226676045684,
              203.13997629273217,
              233.3535176148871,
              175.96158613741864
            ],
            "reference_torch_us": 204.00195307956892,
            "reference_torch_rounds_us": [
              204.43033827177715,
              203.5735678873607,
              205.95247224264313,
              171.75911398226162
            ],
            "reference_native_torch_ratio": 0.26367765207183014,
            "reference_ratio_min": 0.25254233948393,
            "reference_ratio_max": 0.27477238514817776,
            "gain": 0.9972200576547412,
            "gain_min": 0.9625753141909514,
            "gain_max": 1.085314492453,
            "native_torch_ratio": 0.26407482378802116,
            "ratio_min": 0.20536728170917629,
            "ratio_max": 0.276072333225226
          },
          "e2e_throughput": {
            "reference_us": 63.14329230535,
            "candidate_us": 61.694425893100004,
            "torch_us": 239.98125603228544,
            "reference_rounds_us": [
              63.5139578164,
              66.7658678756,
              62.7726267943,
              49.8562651007
            ],
            "candidate_rounds_us": [
              62.7239010989,
              62.346862069,
              61.0419897172,
              50.5661529851
            ],
            "torch_rounds_us": [
              241.62577570093455,
              244.2909752066116,
              238.33673636363633,
              185.66370967741935
            ],
            "reference_torch_us": 238.28295901639342,
            "reference_torch_rounds_us": [
              240.986,
              249.1053,
              235.57991803278688,
              185.03316447368422
            ],
            "reference_native_torch_ratio": 0.26724133941400086,
            "reference_ratio_min": 0.2635587038931722,
            "reference_ratio_max": 0.2694450221532619,
            "gain": 1.0204736838814112,
            "gain_min": 0.9859612044323567,
            "gain_max": 1.0708777580772137,
            "native_torch_ratio": 0.2578538402203864,
            "ratio_min": 0.2552155764913931,
            "ratio_max": 0.27235345600363126
          },
          "gpu_latency": {
            "reference_us": 101.41669190479999,
            "candidate_us": 91.66667587125,
            "torch_us": 193.31247312948108,
            "reference_rounds_us": [
              171.00002150999998,
              106.666702777,
              96.1666810326,
              59.7083126195
            ],
            "candidate_rounds_us": [
              101.916666608,
              92.45832916350001,
              90.875022579,
              80.4166775197
            ],
            "torch_rounds_us": [
              194.24996571615338,
              198.0833476409316,
              192.37498054280877,
              191.41670782119036
            ],
            "reference_torch_us": 193.6874759849161,
            "reference_torch_rounds_us": [
              194.49996761977673,
              192.87498435005546,
              195.33332670107484,
              190.333288628608
            ],
            "reference_native_torch_ratio": 0.5226781771291289,
            "reference_ratio_min": 0.31370399287329687,
            "reference_ratio_max": 0.879177634848165,
            "gain": 1.1059516664932891,
            "gain_min": 0.7424866888447736,
            "gain_max": 1.6778415856919053,
            "native_torch_ratio": 0.4695748026832672,
            "ratio_min": 0.420113157493129,
            "ratio_max": 0.5246676169658898
          },
          "e2e_latency": {
            "reference_us": 360.4375,
            "candidate_us": 317.0,
            "torch_us": 554.7915,
            "reference_rounds_us": [
              268.375,
              361.458,
              473.958,
              359.417
            ],
            "candidate_rounds_us": [
              322.5,
              298.125,
              327.083,
              311.5
            ],
            "torch_rounds_us": [
              564.0,
              545.5830000000001,
              498.0,
              593.375
            ],
            "reference_torch_us": 532.5,
            "reference_torch_rounds_us": [
              606.75,
              527.5,
              537.5,
              520.9580000000001
            ],
            "reference_native_torch_ratio": 0.6875719689222202,
            "reference_ratio_min": 0.442315615986815,
            "reference_ratio_max": 0.8817823255813954,
            "gain": 1.183132190556952,
            "gain_min": 0.8321705426356589,
            "gain_max": 1.4490450436127833,
            "native_torch_ratio": 0.5591211627374519,
            "ratio_min": 0.5249631346113335,
            "ratio_max": 0.6567931726907631
          }
        },
        {
          "name": "softmax_4096x1024",
          "baseline_kernel_score": 10624,
          "candidate_kernel_score": 5888,
          "gpu_throughput": {
            "reference_us": 52.72688804328125,
            "candidate_us": 55.66829395320312,
            "torch_us": 131.80436189941247,
            "reference_rounds_us": [
              53.44791588875,
              52.684896218125004,
              48.5240880153125,
              52.7688798684375
            ],
            "candidate_rounds_us": [
              55.0944014321875,
              58.041667216406246,
              55.813150538593746,
              55.5234373678125
            ],
            "torch_rounds_us": [
              128.8235689571593,
              135.97005272458773,
              129.65429505129578,
              133.95442874752916
            ],
            "reference_torch_us": 129.77506366951275,
            "reference_torch_rounds_us": [
              128.53841326432303,
              130.27669137954945,
              129.27343595947605,
              132.34830657893326
            ],
            "reference_native_torch_ratio": 0.4015599290906907,
            "reference_ratio_min": 0.3753600858147189,
            "reference_ratio_max": 0.41581278725481935,
            "gain": 0.9290487327530595,
            "gain_min": 0.8694024176570897,
            "gain_max": 0.9701151931841194,
            "native_torch_ratio": 0.4272721332698555,
            "ratio_min": 0.4144949733051409,
            "ratio_max": 0.43047668044095344
          },
          "e2e_throughput": {
            "reference_us": 65.24289198505,
            "candidate_us": 71.10903478505,
            "torch_us": 149.18984411118817,
            "reference_rounds_us": [
              61.8199645833,
              66.5805322997,
              63.9052516704,
              66.7586722488
            ],
            "candidate_rounds_us": [
              77.3021122715,
              69.7787814815,
              69.6225502846,
              72.4392880886
            ],
            "torch_rounds_us": [
              143.00338578680203,
              150.65699404761904,
              147.7226941747573,
              155.9055101010101
            ],
            "reference_torch_us": 143.0001098390354,
            "reference_torch_rounds_us": [
              144.77209302325582,
              143.16779891304347,
              138.7847258064516,
              142.8324207650273
            ],
            "reference_native_torch_ratio": 0.46275778706699133,
            "reference_ratio_min": 0.4270157548483422,
            "reference_ratio_max": 0.46739159002719877,
            "gain": 0.9197312603933723,
            "gain_min": 0.7997189567883515,
            "gain_max": 0.9541658780234227,
            "native_torch_ratio": 0.467970779882233,
            "ratio_min": 0.4631632399319252,
            "ratio_max": 0.5405614129077096
          },
          "gpu_latency": {
            "reference_us": 93.81250129085001,
            "candidate_us": 99.64578202909999,
            "torch_us": 128.81250586360693,
            "reference_rounds_us": [
              88.6250054464,
              98.9999971353,
              75.6249646656,
              103.41667803
            ],
            "candidate_rounds_us": [
              98.9165855572,
              88.87500735,
              100.374978501,
              100.500066765
            ],
            "torch_rounds_us": [
              125.75002619996667,
              131.8749855272472,
              121.54167052358389,
              135.04165690392256
            ],
            "reference_torch_us": 125.00002048909664,
            "reference_torch_rounds_us": [
              126.87500566244125,
              125.75002619996667,
              121.99999764561653,
              124.25001477822661
            ],
            "reference_native_torch_ratio": 0.7428991671892855,
            "reference_ratio_min": 0.6198767715166199,
            "reference_ratio_max": 0.8323272895748789,
            "gain": 0.9624889850693188,
            "gain_min": 0.7534244668839115,
            "gain_max": 1.1139239262780214,
            "native_torch_ratio": 0.7654140696342414,
            "ratio_min": 0.6739337789852284,
            "ratio_max": 0.8258482713673356
          },
          "e2e_latency": {
            "reference_us": 383.16650000000004,
            "candidate_us": 365.292,
            "torch_us": 486.97850000000005,
            "reference_rounds_us": [
              369.375,
              443.583,
              278.792,
              396.958
            ],
            "candidate_rounds_us": [
              345.917,
              384.667,
              522.5,
              263.417
            ],
            "torch_rounds_us": [
              722.583,
              559.041,
              414.916,
              406.75
            ],
            "reference_torch_us": 501.813,
            "reference_torch_rounds_us": [
              494.125,
              505.79200000000003,
              569.875,
              497.834
            ],
            "reference_native_torch_ratio": 0.7724518632931876,
            "reference_ratio_min": 0.489216056152665,
            "reference_ratio_max": 0.8770067537643933,
            "gain": 1.1104875005792263,
            "gain_min": 0.5335732057416267,
            "gain_max": 1.5069566504819358,
            "native_torch_ratio": 0.667848857000874,
            "ratio_min": 0.4787228595192524,
            "ratio_max": 1.2592910372219919
          }
        },
        {
          "name": "rmsnorm_37x769",
          "baseline_kernel_score": 79,
          "candidate_kernel_score": 47,
          "gpu_throughput": {
            "reference_us": 5.893229172210937,
            "candidate_us": 5.059570412414063,
            "torch_us": 4.138346412219107,
            "reference_rounds_us": [
              5.4296879170625,
              8.977864126796874,
              5.90364561503125,
              5.882812729390625
            ],
            "candidate_rounds_us": [
              5.0390626711249995,
              5.867838808625,
              5.080078153703125,
              4.847005584453125
            ],
            "torch_rounds_us": [
              7.012369678704999,
              3.7597665141220205,
              4.304687536205165,
              3.972005288233049
            ],
            "reference_torch_us": 4.0423174141324125,
            "reference_torch_rounds_us": [
              4.111979251320008,
              3.771484443859663,
              3.972655576944817,
              4.4882817746838555
            ],
            "reference_native_torch_ratio": 1.4032631984008312,
            "reference_ratio_min": 1.3107048587217982,
            "reference_ratio_max": 2.3804590103543166,
            "gain": 1.1879087540781992,
            "gain_min": 1.0775194260186272,
            "gain_max": 1.5300120571820275,
            "native_torch_ratio": 1.2002094320400514,
            "ratio_min": 0.7185962665983665,
            "ratio_max": 1.560692342618849
          },
          "e2e_throughput": {
            "reference_us": 8.02045296643,
            "candidate_us": 6.7747268922099995,
            "torch_us": 8.368697307116314,
            "reference_rounds_us": [
              8.02759960938,
              8.1535112,
              8.01330632348,
              6.27764395887
            ],
            "candidate_rounds_us": [
              6.73945830145,
              7.06792393321,
              6.80999548297,
              6.72952793834
            ],
            "torch_rounds_us": [
              8.149497894395854,
              8.354026845637584,
              8.413909954058193,
              8.383367768595042
            ],
            "reference_torch_us": 8.20953805506864,
            "reference_torch_rounds_us": [
              8.525959245283019,
              8.239853364755172,
              8.179222745382107,
              7.984169344333479
            ],
            "reference_native_torch_ratio": 0.9606314300796632,
            "reference_ratio_min": 0.7862613739931965,
            "reference_ratio_max": 0.989521395474768,
            "gain": 1.165145628667445,
            "gain_min": 0.9328505678837454,
            "gain_max": 1.1911342500112887,
            "native_torch_ratio": 0.8181759079755255,
            "ratio_min": 0.8027236934003426,
            "ratio_max": 0.8460499425975417
          },
          "gpu_latency": {
            "reference_us": 10.39583003145,
            "candidate_us": 8.29162308946,
            "torch_us": 13.479177141562104,
            "reference_rounds_us": [
              10.708346962899999,
              10.0833131,
              12.583273928599999,
              9.624985978009999
            ],
            "candidate_rounds_us": [
              7.95838423073,
              8.16659303382,
              9.37498407438,
              8.4166531451
            ],
            "torch_rounds_us": [
              8.874980267137289,
              13.500044587999582,
              13.458309695124626,
              14.208315405994654
            ],
            "reference_torch_us": 9.895855328068137,
            "reference_torch_rounds_us": [
              8.750008419156075,
              10.750023648142815,
              9.04168700799346,
              18.624996300786734
            ],
            "reference_native_torch_ratio": 1.0808952272479284,
            "reference_ratio_min": 0.5167778732714934,
            "reference_ratio_max": 1.3916953680740705,
            "gain": 1.2884603490974893,
            "gain_min": 1.1435645276190887,
            "gain_max": 1.3455428454373273,
            "native_torch_ratio": 0.6507626433802187,
            "ratio_min": 0.59237516233268,
            "ratio_max": 0.8967213437306103
          },
          "e2e_latency": {
            "reference_us": 254.312,
            "candidate_us": 235.1045,
            "torch_us": 230.9375,
            "reference_rounds_us": [
              201.417,
              274,
              242.416,
              266.208
            ],
            "candidate_rounds_us": [
              262.583,
              235.334,
              234.708,
              234.875
            ],
            "torch_rounds_us": [
              245.833,
              217.25,
              229.958,
              231.917
            ],
            "reference_torch_us": 293.063,
            "reference_torch_rounds_us": [
              274.41700000000003,
              233.375,
              311.709,
              354.291
            ],
            "reference_native_torch_ratio": 0.764541025881111,
            "reference_ratio_min": 0.7339814953155234,
            "reference_ratio_max": 1.174076057846813,
            "gain": 1.0831218401589717,
            "gain_min": 0.7670603199750173,
            "gain_max": 1.1643026507007062,
            "native_torch_ratio": 1.0443958157380426,
            "ratio_min": 1.012754563054886,
            "ratio_max": 1.083240506329114
          }
        },
        {
          "name": "rmsnorm_1024x1024",
          "baseline_kernel_score": 1600,
          "candidate_kernel_score": 896,
          "gpu_throughput": {
            "reference_us": 19.479166439843752,
            "candidate_us": 19.18522093546875,
            "torch_us": 20.64062528006616,
            "reference_rounds_us": [
              19.413411791875003,
              20.867838429687502,
              19.37369779625,
              19.5449210878125
            ],
            "candidate_rounds_us": [
              19.072264876718748,
              19.29817699421875,
              19.5026050278125,
              18.912760424374998
            ],
            "torch_rounds_us": [
              19.76757812371943,
              20.585937818395905,
              20.695312741736416,
              23.58658821322024
            ],
            "reference_torch_us": 22.55403660456068,
            "reference_torch_rounds_us": [
              21.16080759151373,
              24.097004825307522,
              20.64648379018763,
              23.947265617607627
            ],
            "reference_native_torch_ratio": 0.891708024231836,
            "reference_ratio_min": 0.8161650436382921,
            "reference_ratio_max": 0.9383533774141953,
            "gain": 1.02565607795202,
            "gain_min": 0.9933902557438523,
            "gain_max": 1.081337290871515,
            "native_torch_ratio": 0.9399064281551702,
            "ratio_min": 0.8018438382612043,
            "ratio_max": 0.9648255723261128
          },
          "e2e_throughput": {
            "reference_us": 26.4438432735,
            "candidate_us": 24.22651211665,
            "torch_us": 30.4984408005007,
            "reference_rounds_us": [
              26.4849822917,
              26.4027042553,
              27.0738748466,
              20.8237089066
            ],
            "candidate_rounds_us": [
              24.169743513,
              26.0306910569,
              24.2832807203,
              20.2053908507
            ],
            "torch_rounds_us": [
              30.166301969365428,
              31.386299035369777,
              30.83057963163597,
              24.310228331780056
            ],
            "reference_torch_us": 31.120603472621323,
            "reference_torch_rounds_us": [
              33.49564096916299,
              29.69513209494324,
              32.546074850299405,
              24.67166313213703
            ],
            "reference_native_torch_ratio": 0.8379482146345237,
            "reference_ratio_min": 0.7906993723775223,
            "reference_ratio_max": 0.8891256712003681,
            "gain": 1.0631962130239971,
            "gain_min": 1.0142913300913459,
            "gain_max": 1.1149183324297347,
            "native_torch_ratio": 0.8152907177548424,
            "ratio_min": 0.7876362043930684,
            "ratio_max": 0.831147720002534
          },
          "gpu_latency": {
            "reference_us": 29.833347070949998,
            "candidate_us": 27.541682356949998,
            "torch_us": 28.583337552845478,
            "reference_rounds_us": [
              35.0833288394,
              29.9167004414,
              29.7499937005,
              23.5416810028
            ],
            "candidate_rounds_us": [
              30.9166498482,
              24.1667148657,
              34.6249435097,
              23.0833538808
            ],
            "torch_rounds_us": [
              31.791685614734888,
              32.91666507720947,
              25.374989490956068,
              25.20834095776081
            ],
            "reference_torch_us": 32.18752681277692,
            "reference_torch_rounds_us": [
              31.37503517791629,
              35.91668792068958,
              33.00001844763756,
              30.291674192994833
            ],
            "reference_native_torch_ratio": 0.8672307699238311,
            "reference_ratio_min": 0.7771667175875073,
            "reference_ratio_max": 1.1181924941424077,
            "gain": 1.077313335175295,
            "gain_min": 0.8592069960248654,
            "gain_max": 1.2379299630774805,
            "native_torch_ratio": 0.9440894865403047,
            "ratio_min": 0.734178714915817,
            "ratio_max": 1.3645303586053
          },
          "e2e_latency": {
            "reference_us": 262.9585,
            "candidate_us": 252.6245,
            "torch_us": 292.8125,
            "reference_rounds_us": [
              256.375,
              276.625,
              261.959,
              263.958
            ],
            "candidate_rounds_us": [
              296.75,
              259.208,
              245.792,
              246.041
            ],
            "torch_rounds_us": [
              282.541,
              293.25,
              297.042,
              292.375
            ],
            "reference_torch_us": 281.6875,
            "reference_torch_rounds_us": [
              274.375,
              351.25,
              289.0,
              264.25
            ],
            "reference_native_torch_ratio": 0.9204144406523161,
            "reference_ratio_min": 0.7875444839857652,
            "reference_ratio_max": 0.9988949858088932,
            "gain": 1.066484134561783,
            "gain_min": 0.8639427127211458,
            "gain_max": 1.0728211964672556,
            "native_torch_ratio": 0.8627200933647806,
            "ratio_min": 0.8274654762626161,
            "ratio_max": 1.0502900464003455
          }
        },
        {
          "name": "rmsnorm_16384x257",
          "baseline_kernel_score": 7936,
          "candidate_kernel_score": 5888,
          "gpu_throughput": {
            "reference_us": 52.830729600781254,
            "candidate_us": 49.28808584734375,
            "torch_us": 60.84863298383425,
            "reference_rounds_us": [
              48.74218666375,
              55.3404943275,
              55.147135753906255,
              50.51432344765625
            ],
            "candidate_rounds_us": [
              48.9114590890625,
              53.0364586665625,
              48.014322601406256,
              49.664712605625
            ],
            "torch_rounds_us": [
              59.089844398840796,
              62.25585912034148,
              59.441406847327016,
              64.19726560125127
            ],
            "reference_torch_us": 61.06510409153998,
            "reference_torch_rounds_us": [
              63.6276035947958,
              59.229167163721286,
              61.62890622363193,
              60.50130195944803
            ],
            "reference_native_torch_ratio": 0.8648776672268099,
            "reference_ratio_min": 0.7660541008924104,
            "reference_ratio_max": 0.934345306165251,
            "gain": 1.0302747092564455,
            "gain_min": 0.9965392072028709,
            "gain_max": 1.148555946768498,
            "native_torch_ratio": 0.8177530757916677,
            "ratio_min": 0.7736266045053636,
            "ratio_max": 0.8519111199484414
          },
          "e2e_throughput": {
            "reference_us": 60.9913358568,
            "candidate_us": 62.5409731978,
            "torch_us": 84.16139931063739,
            "reference_rounds_us": [
              60.7471421569,
              61.2355295567,
              61.526112381,
              50.9068136273
            ],
            "candidate_rounds_us": [
              63.3319463956,
              74.0920163934,
              61.75,
              53.3385106796
            ],
            "torch_rounds_us": [
              85.96195100222717,
              87.8522435897436,
              82.36084761904762,
              66.28466917293234
            ],
            "reference_torch_us": 87.7205952963064,
            "reference_torch_rounds_us": [
              86.83828963414635,
              90.91680528052805,
              88.60290095846645,
              66.4799140811456
            ],
            "reference_native_torch_ratio": 0.6969730968150321,
            "reference_ratio_min": 0.6735336703457069,
            "reference_ratio_max": 0.7657472836857607,
            "gain": 0.956798256818032,
            "gain_min": 0.8264794580776825,
            "gain_max": 0.9963742895708502,
            "native_torch_ratio": 0.7772189712715707,
            "ratio_min": 0.7367439391174259,
            "ratio_max": 0.8433707935724245
          },
          "gpu_latency": {
            "reference_us": 79.6666718088,
            "candidate_us": 90.68750659935,
            "torch_us": 89.93750088848174,
            "reference_rounds_us": [
              92.2083272599,
              100.916658994,
              67.1250163577,
              59.4166922383
            ],
            "candidate_rounds_us": [
              92.5000058487,
              96.79159848020001,
              88.87500735,
              65.45835640279999
            ],
            "torch_rounds_us": [
              91.75000013783574,
              95.99997429177165,
              82.04166078940034,
              88.12500163912773
            ],
            "reference_torch_us": 97.708310931921,
            "reference_torch_rounds_us": [
              96.66662663221359,
              102.58331894874573,
              98.74999523162842,
              79.87499702721834
            ],
            "reference_native_torch_ratio": 0.8488753107952751,
            "reference_ratio_min": 0.6797470339137868,
            "reference_ratio_max": 0.9837531094545845,
            "gain": 0.9522744447425102,
            "gain_min": 0.7552743831947486,
            "gain_max": 1.0426179604280823,
            "native_torch_ratio": 1.0082102683284173,
            "ratio_min": 0.7427898460740149,
            "ratio_max": 1.0832911778582928
          },
          "e2e_latency": {
            "reference_us": 343.83299999999997,
            "candidate_us": 317.104,
            "torch_us": 370.0,
            "reference_rounds_us": [
              364.541,
              323.125,
              388.084,
              321.125
            ],
            "candidate_rounds_us": [
              339.667,
              294.541,
              370.083,
              289.709
            ],
            "torch_rounds_us": [
              370.25,
              557.292,
              322.542,
              369.75
            ],
            "reference_torch_us": 341.1875,
            "reference_torch_rounds_us": [
              342.167,
              409.5,
              323.0,
              340.208
            ],
            "reference_native_torch_ratio": 1.0046484837339609,
            "reference_ratio_min": 0.789072039072039,
            "reference_ratio_max": 1.201498452012384,
            "gain": 1.0851382292510916,
            "gain_min": 1.0486404401174871,
            "gain_max": 1.1084398482615314,
            "native_torch_ratio": 0.8504628809636965,
            "ratio_min": 0.5285218520990791,
            "ratio_max": 1.1473947578919956
          }
        },
        {
          "name": "rmsnorm_4096x1024",
          "baseline_kernel_score": 6400,
          "candidate_kernel_score": 3584,
          "gpu_throughput": {
            "reference_us": 58.21223976453125,
            "candidate_us": 55.14290387515625,
            "torch_us": 69.6409501870221,
            "reference_rounds_us": [
              59.99544373484375,
              57.9498700971875,
              58.474609431874995,
              56.539062825312506
            ],
            "candidate_rounds_us": [
              57.1074224353125,
              56.934895837968746,
              51.74153648109375,
              53.35091191234375
            ],
            "torch_rounds_us": [
              70.14322909526527,
              74.1998692319612,
              68.38737044745358,
              69.13867127877893
            ],
            "reference_torch_us": 85.75976517022355,
            "reference_torch_rounds_us": [
              99.11458346323343,
              71.73372341640061,
              98.96484334603883,
              72.55468699440826
            ],
            "reference_native_torch_ratio": 0.6922876652265826,
            "reference_ratio_min": 0.5908624462468317,
            "reference_ratio_max": 0.8078469558982675,
            "gain": 1.0551649330410664,
            "gain_min": 1.017826927480596,
            "gain_max": 1.1301289719767307,
            "native_torch_ratio": 0.7694844104739254,
            "ratio_min": 0.7565949113491666,
            "ratio_max": 0.814154454705127
          },
          "e2e_throughput": {
            "reference_us": 73.00759849185,
            "candidate_us": 70.77720055520001,
            "torch_us": 96.19635099851689,
            "reference_rounds_us": [
              76.6413389356,
              75.5322734864,
              70.4829234973,
              58.0812139831
            ],
            "candidate_rounds_us": [
              71.9739010989,
              70.7179852941,
              70.8364158163,
              55.5390827068
            ],
            "torch_rounds_us": [
              96.81108174386921,
              98.09867441860465,
              95.58162025316456,
              74.50533536585365
            ],
            "reference_torch_us": 99.60489663865548,
            "reference_torch_rounds_us": [
              101.49087581699347,
              97.71891746031747,
              104.77083333333333,
              76.22854464285714
            ],
            "reference_native_torch_ratio": 0.758545100178805,
            "reference_ratio_min": 0.6727342071724797,
            "reference_ratio_max": 0.7729544641862492,
            "gain": 1.0553104835812284,
            "gain_min": 0.995009737365641,
            "gain_max": 1.0680772815045347,
            "native_torch_ratio": 0.742278050810383,
            "ratio_min": 0.7208862475789802,
            "ratio_max": 0.745437658042594
          },
          "gpu_latency": {
            "reference_us": 90.99999442695,
            "candidate_us": 92.79168443755,
            "torch_us": 99.31251406669617,
            "reference_rounds_us": [
              91.458321549,
              90.54166730489999,
              92.33335731550001,
              71.3332556188
            ],
            "candidate_rounds_us": [
              95.04164336250001,
              94.5416977629,
              91.04167111219999,
              65.24997297670001
            ],
            "torch_rounds_us": [
              96.99998190626502,
              101.62504622712731,
              102.041638456285,
              91.04167111217976
            ],
            "reference_torch_us": 94.12498911842704,
            "reference_torch_rounds_us": [
              97.0833352766931,
              91.83335350826383,
              96.41662472859025,
              88.5000335983932
            ],
            "reference_native_torch_ratio": 0.9498548516784631,
            "reference_ratio_min": 0.8060251812163732,
            "reference_ratio_max": 0.9859344545960896,
            "gain": 0.9882426034674435,
            "gain_min": 0.9576903043561621,
            "gain_max": 1.093230423930938,
            "native_torch_ratio": 0.9112501871790816,
            "ratio_min": 0.7167044736722844,
            "ratio_max": 0.9798109390818502
          },
          "e2e_latency": {
            "reference_us": 325.875,
            "candidate_us": 320.167,
            "torch_us": 352.04200000000003,
            "reference_rounds_us": [
              316.25,
              344.333,
              335.5,
              307.375
            ],
            "candidate_rounds_us": [
              569.417,
              340.042,
              300.292,
              295.625
            ],
            "torch_rounds_us": [
              311.584,
              473.333,
              356.75,
              347.334
            ],
            "reference_torch_us": 375.79150000000004,
            "reference_torch_rounds_us": [
              385.458,
              331.41700000000003,
              474.708,
              366.125
            ],
            "reference_native_torch_ratio": 0.8299941151267898,
            "reference_ratio_min": 0.7067502548935345,
            "reference_ratio_max": 1.038972050317274,
            "gain": 1.026182664812715,
            "gain_min": 0.5553926208736304,
            "gain_max": 1.117245880676142,
            "native_torch_ratio": 0.8464347616929804,
            "ratio_min": 0.7183990974641531,
            "ratio_max": 1.8274911420355346
          }
        },
        {
          "name": "layernorm_37x769",
          "baseline_kernel_score": 131,
          "candidate_kernel_score": 77,
          "gpu_throughput": {
            "reference_us": 6.185546681085937,
            "candidate_us": 5.407552635,
            "torch_us": 5.661458544636844,
            "reference_rounds_us": [
              6.4160158217343755,
              6.048828254281251,
              6.117838893265625,
              6.25325446890625
            ],
            "candidate_rounds_us": [
              5.673176929125,
              5.60612079425,
              5.1373699534375,
              5.2089844757499995
            ],
            "torch_rounds_us": [
              5.407551725511439,
              7.983073373907246,
              5.82552092964761,
              5.497396159626078
            ],
            "reference_torch_us": 5.432291800389066,
            "reference_torch_rounds_us": [
              5.344400960893836,
              7.031901077425573,
              5.342448275769129,
              5.520182639884297
            ],
            "reference_native_torch_ratio": 1.1389680833203166,
            "reference_ratio_min": 0.8601981438134463,
            "reference_ratio_max": 1.2005116885281966,
            "gain": 1.1608945757439129,
            "gain_min": 1.0789685909881428,
            "gain_max": 1.2004747754610834,
            "native_torch_ratio": 0.9147048659950741,
            "ratio_min": 0.7022509416703674,
            "ratio_max": 1.049121158168568
          },
          "e2e_throughput": {
            "reference_us": 8.32125210373,
            "candidate_us": 7.195627165755001,
            "torch_us": 12.192381140973815,
            "reference_rounds_us": [
              8.39428003697,
              8.24822417049,
              8.58929735936,
              6.58130406613
            ],
            "candidate_rounds_us": [
              7.0482310757,
              7.34302325581,
              7.75618300206,
              5.66197656788
            ],
            "torch_rounds_us": [
              12.034968043264504,
              12.69552813163482,
              12.349794238683128,
              10.518465909090908
            ],
            "reference_torch_us": 11.95908212190345,
            "reference_torch_rounds_us": [
              11.819169111302847,
              12.098995132504054,
              12.136655272026964,
              10.220271520342612
            ],
            "reference_native_torch_ratio": 0.6947216974025363,
            "reference_ratio_min": 0.6439461077947347,
            "reference_ratio_max": 0.7102259014927222,
            "gain": 1.1428211257899574,
            "gain_min": 1.107412931989708,
            "gain_max": 1.1909768489161399,
            "native_torch_ratio": 0.5820202421566021,
            "ratio_min": 0.5382891970003404,
            "ratio_max": 0.6280414760081906
          },
          "gpu_latency": {
            "reference_us": 10.062503861269999,
            "candidate_us": 9.104202035815,
            "torch_us": 10.500021744519472,
            "reference_rounds_us": [
              9.91666456684,
              10.2083431557,
              11.5832663141,
              9.125040378420001
            ],
            "candidate_rounds_us": [
              9.20839374885,
              9.00001032278,
              9.37504228204,
              8.25000461191
            ],
            "torch_rounds_us": [
              10.374991688877344,
              10.6250518001616,
              10.25001984089613,
              11.2500274553895
            ],
            "reference_torch_us": 10.270829079672694,
            "reference_torch_rounds_us": [
              10.124989785254002,
              10.416668374091387,
              11.8750031106174,
              9.624985978007317
            ],
            "reference_native_torch_ratio": 0.9774286731417283,
            "reference_ratio_min": 0.9480575243714984,
            "reference_ratio_max": 0.980000782312554,
            "gain": 1.1201619658843776,
            "gain_min": 1.0769157832850549,
            "gain_max": 1.2355428344350348,
            "native_torch_ratio": 0.8673062002046186,
            "ratio_min": 0.7333319535996072,
            "ratio_max": 0.914636500959238
          },
          "e2e_latency": {
            "reference_us": 239.625,
            "candidate_us": 246.584,
            "torch_us": 339.166,
            "reference_rounds_us": [
              233.166,
              265.208,
              246.084,
              225.75
            ],
            "candidate_rounds_us": [
              236.959,
              256.209,
              259.083,
              197.709
            ],
            "torch_rounds_us": [
              403.45799999999997,
              332.541,
              345.791,
              257.458
            ],
            "reference_torch_us": 267.4995,
            "reference_torch_rounds_us": [
              233.5,
              285.416,
              360.125,
              249.583
            ],
            "reference_native_torch_ratio": 0.9168534016721561,
            "reference_ratio_min": 0.6833293995140576,
            "reference_ratio_max": 0.9985695931477516,
            "gain": 1.0095583400083483,
            "gain_min": 0.9498268894524148,
            "gain_max": 1.1418296587408767,
            "native_torch_ratio": 0.758587286821723,
            "ratio_min": 0.5873201175834908,
            "ratio_max": 0.7704583795682337
          }
        },
        {
          "name": "layernorm_1024x1024",
          "baseline_kernel_score": 2656,
          "candidate_kernel_score": 1472,
          "gpu_throughput": {
            "reference_us": 18.986327631875,
            "candidate_us": 19.192382751484374,
            "torch_us": 42.21451854391489,
            "reference_rounds_us": [
              18.98762911875,
              18.73046858234375,
              21.253906197656253,
              18.985026145
            ],
            "candidate_rounds_us": [
              23.97330808890625,
              19.02473923109375,
              19.360026271874997,
              18.755208657187502
            ],
            "torch_rounds_us": [
              41.27278680243762,
              44.40429711394245,
              40.40950534545118,
              43.15625028539216
            ],
            "reference_torch_us": 41.92089863863657,
            "reference_torch_rounds_us": [
              41.06705728190718,
              43.49283881310839,
              42.49934954714263,
              41.342447730130516
            ],
            "reference_native_torch_ratio": 0.460785294227639,
            "reference_ratio_min": 0.4306563814523539,
            "reference_ratio_max": 0.5000995644434568,
            "gain": 0.9983928699477,
            "gain_min": 0.7920320820277867,
            "gain_max": 1.0978242435823842,
            "native_torch_ratio": 0.4568421580575349,
            "ratio_min": 0.4284436522500476,
            "ratio_max": 0.5808502392547517
          },
          "e2e_throughput": {
            "reference_us": 25.022317980399997,
            "candidate_us": 26.323218880200002,
            "torch_us": 60.72101360475749,
            "reference_rounds_us": [
              25.7447671861,
              26.7317557058,
              24.2998687747,
              20.2511344884
            ],
            "candidate_rounds_us": [
              27.8629958246,
              26.6075966587,
              26.0388411017,
              20.0493797468
            ],
            "torch_rounds_us": [
              60.63155690072639,
              66.227416,
              60.810470308788595,
              47.34138229376257
            ],
            "reference_torch_us": 61.493547269816546,
            "reference_torch_rounds_us": [
              61.98121028037384,
              62.831768774703555,
              61.005884259259254,
              46.89616720257234
            ],
            "reference_native_torch_ratio": 0.4204068686284964,
            "reference_ratio_min": 0.39832008124711105,
            "reference_ratio_max": 0.4318292026067578,
            "gain": 0.9689412585982534,
            "gain_min": 0.9239769961624215,
            "gain_max": 1.010062891927228,
            "native_torch_ratio": 0.42585155113189377,
            "ratio_min": 0.40176105706283327,
            "ratio_max": 0.45954610517788286
          },
          "gpu_latency": {
            "reference_us": 27.50000567175,
            "candidate_us": 35.2083297912,
            "torch_us": 46.750035835430026,
            "reference_rounds_us": [
              27.3333280347,
              27.6666833088,
              34.91668030620001,
              23.375032469599997
            ],
            "candidate_rounds_us": [
              35.8333345503,
              34.5833250321,
              54.2916823179,
              23.666652850800002
            ],
            "torch_rounds_us": [
              48.37504820898175,
              48.62499190494418,
              45.1250234618783,
              43.625012040138245
            ],
            "reference_torch_us": 50.81246490590274,
            "reference_torch_rounds_us": [
              56.95829167962074,
              44.666638132184744,
              58.500037994235754,
              43.666630517691374
            ],
            "reference_native_torch_ratio": 0.5660862363046505,
            "reference_ratio_min": 0.47988321328955275,
            "reference_ratio_max": 0.6194037533544449,
            "gain": 0.7813955985734005,
            "gain_min": 0.6431313014348785,
            "gain_max": 0.9876780048687727,
            "native_torch_ratio": 0.7259826697831473,
            "ratio_min": 0.5425019213524784,
            "ratio_max": 1.2031391488087693
          },
          "e2e_latency": {
            "reference_us": 283.16700000000003,
            "candidate_us": 263.02099999999996,
            "torch_us": 349.104,
            "reference_rounds_us": [
              271.583,
              285.084,
              295.583,
              281.25
            ],
            "candidate_rounds_us": [
              263,
              263.042,
              287.375,
              226
            ],
            "torch_rounds_us": [
              422.916,
              359.25,
              338.95799999999997,
              322.459
            ],
            "reference_torch_us": 337.83299999999997,
            "reference_torch_rounds_us": [
              346.54200000000003,
              278.5,
              342.791,
              332.875
            ],
            "reference_native_torch_ratio": 0.8535975929200377,
            "reference_ratio_min": 0.7836943285373779,
            "reference_ratio_max": 1.023640933572711,
            "gain": 1.0582157424844731,
            "gain_min": 1.0285619834710744,
            "gain_max": 1.2444690265486726,
            "native_torch_ratio": 0.7165309649428159,
            "ratio_min": 0.6218729014745245,
            "ratio_max": 0.8478189038169921
          }
        },
        {
          "name": "layernorm_16384x257",
          "baseline_kernel_score": 13056,
          "candidate_kernel_score": 9472,
          "gpu_throughput": {
            "reference_us": 53.59342458179687,
            "candidate_us": 53.8457024959375,
            "torch_us": 142.50260437620454,
            "reference_rounds_us": [
              52.3971357325,
              54.789713431093745,
              55.01953182828125,
              49.662760829999996
            ],
            "candidate_rounds_us": [
              54.358071793125,
              54.50390654,
              51.17708315078125,
              53.33333319875
            ],
            "torch_rounds_us": [
              142.8059895260958,
              145.90755199606065,
              142.1992192263133,
              116.15625044214539
            ],
            "reference_torch_us": 143.4583332411421,
            "reference_torch_rounds_us": [
              145.25455753755523,
              144.96549465548014,
              141.95117182680406,
              115.76367251109332
            ],
            "reference_native_torch_ratio": 0.38277240141629637,
            "reference_ratio_min": 0.36072627682579145,
            "reference_ratio_max": 0.4290012553397608,
            "gain": 0.9845846822685503,
            "gain_min": 0.9311767679122663,
            "gain_max": 1.0750814317842057,
            "native_torch_ratio": 0.3770968817324962,
            "ratio_min": 0.35989707559035017,
            "ratio_max": 0.45915164268593567
          },
          "e2e_throughput": {
            "reference_us": 67.58330198195,
            "candidate_us": 68.52260869915,
            "torch_us": 159.04628642205142,
            "reference_rounds_us": [
              69.8394953052,
              65.3271086587,
              70.3241,
              51.7906953405
            ],
            "candidate_rounds_us": [
              68.6247832898,
              68.4204341085,
              68.7661556122,
              53.2111541353
            ],
            "torch_rounds_us": [
              158.01177173913044,
              160.53372164948453,
              160.08080110497238,
              120.974
            ],
            "reference_torch_us": 163.0699732142857,
            "reference_torch_rounds_us": [
              167.41692592592594,
              162.63690476190476,
              163.50304166666666,
              120.67531779661017
            ],
            "reference_native_torch_ratio": 0.423166497534591,
            "reference_ratio_min": 0.40167456921500566,
            "reference_ratio_max": 0.43010881805715645,
            "gain": 0.9955030105409706,
            "gain_min": 0.9547894501093831,
            "gain_max": 1.0226556854012006,
            "native_torch_ratio": 0.43193662496449214,
            "ratio_min": 0.42620599214594795,
            "ratio_max": 0.43985611896192567
          },
          "gpu_latency": {
            "reference_us": 96.41665383265,
            "candidate_us": 99.20835145769999,
            "torch_us": 115.91665679588914,
            "reference_rounds_us": [
              92.75000775230001,
              102.791644167,
              100.08329991299999,
              60.166639741500006
            ],
            "candidate_rounds_us": [
              104.125065263,
              102.625053842,
              95.7916490734,
              82.166690845
            ],
            "torch_rounds_us": [
              115.49994815140963,
              116.33336544036865,
              117.91667202487588,
              113.62502118572593
            ],
            "reference_torch_us": 114.75000064820051,
            "reference_torch_rounds_us": [
              114.87497249618173,
              114.6250288002193,
              122.66664998605847,
              114.20832015573978
            ],
            "reference_native_torch_ratio": 0.8116480927710717,
            "reference_ratio_min": 0.5268148560407331,
            "reference_ratio_max": 0.8967643911885617,
            "gain": 0.9461895548282504,
            "gain_min": 0.7322509781366138,
            "gain_max": 1.044801930868854,
            "native_torch_ratio": 0.8472654240941819,
            "ratio_min": 0.7231390585238645,
            "ratio_max": 0.9015161212583557
          },
          "e2e_latency": {
            "reference_us": 332.083,
            "candidate_us": 299.5625,
            "torch_us": 386.2915,
            "reference_rounds_us": [
              351.416,
              317.583,
              346.583,
              271.208
            ],
            "candidate_rounds_us": [
              320.875,
              283.125,
              277,
              316
            ],
            "torch_rounds_us": [
              377.20799999999997,
              421.58299999999997,
              375.834,
              395.375
            ],
            "reference_torch_us": 381.54150000000004,
            "reference_torch_rounds_us": [
              372.625,
              392.625,
              361.25,
              390.458
            ],
            "reference_native_torch_ratio": 0.875976623678714,
            "reference_ratio_min": 0.694589430873487,
            "reference_ratio_max": 0.9593993079584776,
            "gain": 1.1084431632255551,
            "gain_min": 0.8582531645569621,
            "gain_max": 1.2512021660649821,
            "native_torch_ratio": 0.7681343720756733,
            "ratio_min": 0.6715759411551225,
            "ratio_max": 0.8506579924073722
          }
        },
        {
          "name": "layernorm_4096x1024",
          "baseline_kernel_score": 10624,
          "candidate_kernel_score": 5888,
          "gpu_throughput": {
            "reference_us": 56.37630147242187,
            "candidate_us": 53.817708248984374,
            "torch_us": 167.92480482763494,
            "reference_rounds_us": [
              54.214193369218755,
              60.43489611328125,
              56.692056205,
              56.060546739843744
            ],
            "candidate_rounds_us": [
              58.763670495,
              53.5501303603125,
              53.494140956875,
              54.08528613765625
            ],
            "torch_rounds_us": [
              181.8535156417056,
              175.49414133100072,
              160.35546832426917,
              135.16015678760596
            ],
            "reference_torch_us": 170.88769482143107,
            "reference_torch_rounds_us": [
              175.50585926073836,
              172.81835880567087,
              168.95703083719127,
              140.12369774718536
            ],
            "reference_native_torch_ratio": 0.3426215478949877,
            "reference_ratio_min": 0.30890246968151663,
            "reference_ratio_max": 0.40007898479092074,
            "gain": 1.0481509415359314,
            "gain_min": 0.9225801062551335,
            "gain_max": 1.128566741231899,
            "native_torch_ratio": 0.32836731307545697,
            "ratio_min": 0.3051391342991401,
            "ratio_max": 0.4001570242526961
          },
          "e2e_throughput": {
            "reference_us": 68.8513083413,
            "candidate_us": 71.9538032981,
            "torch_us": 187.99964535044148,
            "reference_rounds_us": [
              67.8981737968,
              69.8044428858,
              73.0657926078,
              57.2048958743
            ],
            "candidate_rounds_us": [
              69.7748091603,
              74.1968838527,
              74.1327974359,
              55.9506965377
            ],
            "torch_rounds_us": [
              204.43088235294118,
              188.29427083333334,
              187.70501986754965,
              142.7099656862745
            ],
            "reference_torch_us": 186.5198645320197,
            "reference_torch_rounds_us": [
              203.12853107344634,
              191.03714285714287,
              182.00258620689655,
              147.08412322274881
            ],
            "reference_native_torch_ratio": 0.37716180608436556,
            "reference_ratio_min": 0.3342621218102033,
            "reference_ratio_max": 0.4014546943016536,
            "gain": 0.9793556225629952,
            "gain_min": 0.9408001961966471,
            "gain_max": 1.0224161523307385,
            "native_torch_ratio": 0.39305313772042244,
            "ratio_min": 0.34131246882668526,
            "ratio_max": 0.394943073382962
          },
          "gpu_latency": {
            "reference_us": 101.22917592515,
            "candidate_us": 91.4791598916,
            "torch_us": 136.56250666826963,
            "reference_rounds_us": [
              104.249978904,
              93.0833630264,
              102.666672319,
              99.7916795313
            ],
            "candidate_rounds_us": [
              136.166694574,
              90.4583139345,
              92.5000058487,
              62.3333617114
            ],
            "torch_rounds_us": [
              138.87498062103987,
              139.49995627626777,
              134.2500327154994,
              132.00001558288932
            ],
            "reference_torch_us": 138.3333292324096,
            "reference_torch_rounds_us": [
              140.99996769800782,
              139.1249825246632,
              137.54167594015598,
              135.04165690392256
            ],
            "reference_native_torch_ratio": 0.7391656622261683,
            "reference_ratio_min": 0.669062891058396,
            "reference_ratio_max": 0.7464404633521406,
            "gain": 1.069464667534426,
            "gain_min": 0.7656055633144946,
            "gain_max": 1.6009353064147243,
            "native_torch_ratio": 0.668729904578079,
            "ratio_min": 0.4722223814606886,
            "ratio_max": 0.9804983875790398
          },
          "e2e_latency": {
            "reference_us": 337.7505,
            "candidate_us": 325.2295,
            "torch_us": 457.77099999999996,
            "reference_rounds_us": [
              351.209,
              324.292,
              416.291,
              313
            ],
            "candidate_rounds_us": [
              369,
              349.167,
              301.292,
              271.084
            ],
            "torch_rounds_us": [
              417.0,
              512.75,
              428.334,
              487.20799999999997
            ],
            "reference_torch_us": 447.64599999999996,
            "reference_torch_rounds_us": [
              439.5,
              506.708,
              455.792,
              371.25
            ],
            "reference_native_torch_ratio": 0.8211039978855679,
            "reference_ratio_min": 0.6399977896540018,
            "reference_ratio_max": 0.9133354688103346,
            "gain": 1.0532047834731493,
            "gain_min": 0.9287590178911523,
            "gain_max": 1.3816862047448988,
            "native_torch_ratio": 0.6921868179772503,
            "ratio_min": 0.5564030147288223,
            "ratio_max": 0.8848920863309353
          }
        }
      ]
    },
    "automatic-replay": {
      "baseline": "automatic",
      "replay_sha256": "9cf71bf3b758712637b2db0767c6181aa95a96148e5cbba6926d177dd5caac33",
      "replay_pairs": 48,
      "recorded_unchanged_artifacts": 21,
      "current_artifacts_checked": true,
      "torch_probe_control_ratio_range": [
        0.8610229959492322,
        5.915401637957651
      ],
      "modeled_candidate_better": 3,
      "summaries": {
        "gpu_throughput": {
          "positive_medians": 2,
          "all_pairs_positive": 2,
          "all_pairs_negative": 8,
          "faster_torch_all_pairs": 11,
          "slower_torch_all_pairs": 1,
          "baseline_faster_torch_all_pairs": 12,
          "baseline_slower_torch_all_pairs": 0
        },
        "e2e_throughput": {
          "positive_medians": 3,
          "all_pairs_positive": 2,
          "all_pairs_negative": 8,
          "faster_torch_all_pairs": 12,
          "slower_torch_all_pairs": 0,
          "baseline_faster_torch_all_pairs": 12,
          "baseline_slower_torch_all_pairs": 0
        },
        "gpu_latency": {
          "positive_medians": 3,
          "all_pairs_positive": 0,
          "all_pairs_negative": 5,
          "faster_torch_all_pairs": 9,
          "slower_torch_all_pairs": 0,
          "baseline_faster_torch_all_pairs": 11,
          "baseline_slower_torch_all_pairs": 0
        },
        "e2e_latency": {
          "positive_medians": 5,
          "all_pairs_positive": 1,
          "all_pairs_negative": 1,
          "faster_torch_all_pairs": 9,
          "slower_torch_all_pairs": 0,
          "baseline_faster_torch_all_pairs": 8,
          "baseline_slower_torch_all_pairs": 0
        }
      },
      "records": [
        {
          "name": "softmax_37x769",
          "baseline_kernel_score": 64,
          "candidate_kernel_score": 77,
          "gpu_throughput": {
            "reference_us": 3.315103640484375,
            "candidate_us": 4.684895884570312,
            "torch_us": 21.028971787018236,
            "reference_rounds_us": [
              3.2154948712499998,
              3.4114582376718747,
              3.257811840734375,
              3.372395440234375
            ],
            "candidate_rounds_us": [
              4.82682298753125,
              4.700521458296874,
              4.59049351775,
              4.66927031084375
            ],
            "torch_rounds_us": [
              20.94791670970153,
              21.11002686433494,
              24.50130250508664,
              20.0169270101469
            ],
            "reference_torch_us": 21.36523471563123,
            "reference_torch_rounds_us": [
              18.23112052079523,
              24.367838705074973,
              21.791665858472697,
              20.938803572789766
            ],
            "reference_native_torch_ratio": 0.15527883430489753,
            "reference_ratio_min": 0.13999839210037887,
            "reference_ratio_max": 0.17637395724427704,
            "gain": 0.7159699130795729,
            "gain_min": 0.6661721135323034,
            "gain_max": 0.7257616560074035,
            "native_torch_ratio": 0.2265439537389286,
            "ratio_min": 0.18735712180187078,
            "ratio_max": 0.23326609066800424
          },
          "e2e_throughput": {
            "reference_us": 3.478246045975,
            "candidate_us": 5.04430346891,
            "torch_us": 31.943564089818345,
            "reference_rounds_us": [
              3.49872005027,
              3.48184863991,
              3.47464345204,
              3.4026429342
            ],
            "candidate_rounds_us": [
              5.25495889921,
              4.96665440404,
              4.9634177003,
              5.12195253378
            ],
            "torch_rounds_us": [
              30.362863459037712,
              32.62880633373934,
              32.58383461117196,
              31.30329356846473
            ],
            "reference_torch_us": 31.725646407136654,
            "reference_torch_rounds_us": [
              29.171511627906977,
              32.66817934165721,
              32.54679843225084,
              30.90449438202247
            ],
            "reference_native_torch_ratio": 0.10843012968718792,
            "reference_ratio_min": 0.10658226782384779,
            "reference_ratio_max": 0.11993619305360041,
            "gain": 0.6829222859853644,
            "gain_min": 0.6643253547859121,
            "gain_max": 0.7010450811874042,
            "native_torch_ratio": 0.15797552506682322,
            "ratio_min": 0.15221685872413615,
            "ratio_max": 0.17307191419213216
          },
          "gpu_latency": {
            "reference_us": 5.81247149967,
            "candidate_us": 7.2500261012500005,
            "torch_us": 101.1041458696127,
            "reference_rounds_us": [
              5.87495742366,
              5.9999874793,
              5.70830889046,
              5.749985575679999
            ],
            "candidate_rounds_us": [
              7.25005520508,
              7.5000571087,
              7.24999699742,
              7.0416717790100005
            ],
            "torch_rounds_us": [
              102.37499373033643,
              99.83329800888896,
              107.12497169151902,
              83.20831693708897
            ],
            "reference_torch_us": 118.3541608043015,
            "reference_torch_rounds_us": [
              105.62501847743988,
              132.7083446085453,
              105.00004282221198,
              131.08330313116312
            ],
            "reference_native_torch_ratio": 0.0497883285707901,
            "reference_ratio_min": 0.04386512575080987,
            "reference_ratio_max": 0.055620888955534846,
            "gain": 0.8051625118756945,
            "gain_min": 0.7873532764898208,
            "gain_max": 0.8165654060758281,
            "native_torch_ratio": 0.07297221085126401,
            "ratio_min": 0.06767793617997264,
            "ratio_max": 0.08462701852669335
          },
          "e2e_latency": {
            "reference_us": 204.25,
            "candidate_us": 225.83350000000002,
            "torch_us": 372.2505,
            "reference_rounds_us": [
              223.875,
              194.333,
              214.167,
              190.25
            ],
            "candidate_rounds_us": [
              221.083,
              221.25,
              245.458,
              230.417
            ],
            "torch_rounds_us": [
              354.66700000000003,
              405.083,
              373.917,
              370.584
            ],
            "reference_torch_us": 386.938,
            "reference_torch_rounds_us": [
              387.167,
              386.709,
              421.875,
              374.333
            ],
            "reference_native_torch_ratio": 0.5079462146104612,
            "reference_ratio_min": 0.502530326421159,
            "reference_ratio_max": 0.5782388478356885,
            "gain": 0.875430592624888,
            "gain_min": 0.8256769248796747,
            "gain_max": 1.012628741241977,
            "native_torch_ratio": 0.6225605010456248,
            "ratio_min": 0.5461843622171259,
            "ratio_max": 0.6564504956982432
          }
        },
        {
          "name": "softmax_1024x1024",
          "baseline_kernel_score": 1088,
          "candidate_kernel_score": 1472,
          "gpu_throughput": {
            "reference_us": 16.1972657225,
            "candidate_us": 18.473633190312498,
            "torch_us": 42.8740240749903,
            "reference_rounds_us": [
              15.9759119925,
              16.5520832525,
              15.76367230875,
              16.4186194525
            ],
            "candidate_rounds_us": [
              18.73567725859375,
              18.0657552846875,
              18.50911485234375,
              18.43815152828125
            ],
            "torch_rounds_us": [
              42.845052121265326,
              43.85156262287637,
              42.828773985092994,
              42.90299602871528
            ],
            "reference_torch_us": 42.917968585243216,
            "reference_torch_rounds_us": [
              42.852212573052384,
              43.022787394875195,
              42.970052163582295,
              42.865885006904136
            ],
            "reference_native_torch_ratio": 0.3779185792715746,
            "reference_ratio_min": 0.36685252902973964,
            "reference_ratio_max": 0.38472828597971453,
            "gain": 0.8715849635349087,
            "gain_min": 0.8516707813693153,
            "gain_max": 0.9162131885252267,
            "native_torch_ratio": 0.43096457128664334,
            "ratio_min": 0.4119751772600004,
            "ratio_max": 0.43728916948369523
          },
          "e2e_throughput": {
            "reference_us": 17.45557158895,
            "candidate_us": 19.42400509655,
            "torch_us": 49.83920681860697,
            "reference_rounds_us": [
              17.4321114319,
              17.5828016997,
              17.479031746,
              17.1424919817
            ],
            "candidate_rounds_us": [
              19.4774744788,
              19.8690689655,
              19.1854059278,
              19.3705357143
            ],
            "torch_rounds_us": [
              49.71029294755877,
              50.09595019920319,
              48.95776274509804,
              49.96812068965517
            ],
            "reference_torch_us": 49.67436909930274,
            "reference_torch_rounds_us": [
              49.68029848229342,
              49.73486882129277,
              49.66843971631206,
              49.43311904761905
            ],
            "reference_native_torch_ratio": 0.3514000269049178,
            "reference_ratio_min": 0.3467815163592367,
            "reference_ratio_max": 0.3535306740805628,
            "gain": 0.889982987072738,
            "gain_min": 0.8849333469137483,
            "gain_max": 0.911058739740949,
            "native_torch_ratio": 0.39184821834818917,
            "ratio_min": 0.38765787960302966,
            "ratio_max": 0.39662026344428997
          },
          "gpu_latency": {
            "reference_us": 20.3124945983,
            "candidate_us": 23.37497426195,
            "torch_us": 66.54168828390539,
            "reference_rounds_us": [
              20.3333329409,
              20.583334844600003,
              20.2916562557,
              19.916682504100002
            ],
            "candidate_rounds_us": [
              23.041677195600002,
              23.4582694247,
              23.2916790992,
              23.874978069199997
            ],
            "torch_rounds_us": [
              66.45836401730776,
              66.62501255050302,
              65.83333015441895,
              69.500005338341
            ],
            "reference_torch_us": 67.10417801514268,
            "reference_torch_rounds_us": [
              67.66669685021043,
              67.99999391660094,
              66.54165918007493,
              66.12500874325633
            ],
            "reference_native_torch_ratio": 0.30194677744552567,
            "reference_ratio_min": 0.3004924709995914,
            "reference_ratio_max": 0.3049466530551448,
            "gain": 0.8743211926439189,
            "gain_min": 0.8342073633061716,
            "gain_max": 0.8824588925663284,
            "native_torch_ratio": 0.34940124808004736,
            "ratio_min": 0.3435248379186658,
            "ratio_max": 0.35379767428697495
          },
          "e2e_latency": {
            "reference_us": 275.5835,
            "candidate_us": 287.312,
            "torch_us": 350.1245,
            "reference_rounds_us": [
              325.125,
              248.041,
              278.584,
              272.583
            ],
            "candidate_rounds_us": [
              295.625,
              281.333,
              264.875,
              293.291
            ],
            "torch_rounds_us": [
              354.208,
              341.45799999999997,
              346.041,
              375.167
            ],
            "reference_torch_us": 410.313,
            "reference_torch_rounds_us": [
              397.959,
              478.209,
              391.459,
              422.66700000000003
            ],
            "reference_native_torch_ratio": 0.6782837893664013,
            "reference_ratio_min": 0.5186874358282676,
            "reference_ratio_max": 0.8169811462989905,
            "gain": 0.9905754223431041,
            "gain_min": 0.8816633668997237,
            "gain_max": 1.0997885835095138,
            "native_torch_ratio": 0.8028390198493236,
            "ratio_min": 0.7654439791816576,
            "ratio_max": 0.8346084786340229
          }
        },
        {
          "name": "softmax_16384x257",
          "baseline_kernel_score": 12288,
          "candidate_kernel_score": 9472,
          "gpu_throughput": {
            "reference_us": 59.436198171171874,
            "candidate_us": 48.560220875546875,
            "torch_us": 172.97460908594076,
            "reference_rounds_us": [
              59.043619330625,
              61.270833612031254,
              59.82877701171875,
              58.70833410875
            ],
            "candidate_rounds_us": [
              48.88932289765625,
              47.96484336109375,
              48.68033738609375,
              48.440104364999996
            ],
            "torch_rounds_us": [
              170.4824217085843,
              173.88085871061776,
              174.4492183206603,
              172.06835946126375
            ],
            "reference_torch_us": 174.53222744734376,
            "reference_torch_rounds_us": [
              175.16276238893624,
              173.90169250575127,
              170.42513081833022,
              177.5013015503646
            ],
            "reference_native_torch_ratio": 0.3440673377967271,
            "reference_ratio_min": 0.3307487528033251,
            "reference_ratio_max": 0.3523302891948847,
            "gain": 1.2204955286282422,
            "gain_min": 1.2076996741031885,
            "gain_max": 1.2774113145906891,
            "native_torch_ratio": 0.2802841252332662,
            "ratio_min": 0.27584889858934686,
            "ratio_max": 0.28677046236019377
          },
          "e2e_throughput": {
            "reference_us": 63.09219962185,
            "candidate_us": 50.8144973447,
            "torch_us": 184.71495793633838,
            "reference_rounds_us": [
              63.4390889831,
              63.0012752941,
              63.1831239496,
              62.3816207627
            ],
            "candidate_rounds_us": [
              50.4418364312,
              51.0998050089,
              51.0298694779,
              50.5991252115
            ],
            "torch_rounds_us": [
              184.15949006622515,
              187.16752413793105,
              185.2704258064516,
              183.85254609929078
            ],
            "reference_torch_us": 186.62590101878232,
            "reference_torch_rounds_us": [
              185.67494520547947,
              186.18105035971223,
              187.07075167785237,
              187.50965562913908
            ],
            "reference_native_torch_ratio": 0.33806851579942665,
            "reference_ratio_min": 0.3326848452331426,
            "reference_ratio_max": 0.34166747114369334,
            "gain": 1.2355329983392116,
            "gain_min": 1.2328596690545575,
            "gain_max": 1.2576681078934857,
            "native_torch_ratio": 0.2745593980390941,
            "ratio_min": 0.27301640732952454,
            "ratio_max": 0.27543451285209386
          },
          "gpu_latency": {
            "reference_us": 91.81248606185,
            "candidate_us": 81.41668513415,
            "torch_us": 191.70835730619729,
            "reference_rounds_us": [
              66.9583678246,
              87.66667451709999,
              96.8750100583,
              95.9582976066
            ],
            "candidate_rounds_us": [
              56.0416374356,
              89.208362624,
              79.708348494,
              83.1250217743
            ],
            "torch_rounds_us": [
              191.29167776554823,
              191.91671162843704,
              191.50000298395753,
              191.95833010599017
            ],
            "reference_torch_us": 191.66668062098324,
            "reference_torch_rounds_us": [
              190.58334873989224,
              193.1249862536788,
              192.0000067912042,
              191.33335445076227
            ],
            "reference_native_torch_ratio": 0.47773082685195345,
            "reference_ratio_min": 0.35133377741192195,
            "reference_ratio_max": 0.5045573262070217,
            "gain": 1.174591005210589,
            "gain_min": 0.9827181212438794,
            "gain_max": 1.2153684261265583,
            "native_torch_ratio": 0.4246341911801516,
            "ratio_min": 0.29296432594566924,
            "ratio_max": 0.4648285283082229
          },
          "e2e_latency": {
            "reference_us": 315.60400000000004,
            "candidate_us": 291.792,
            "torch_us": 515.125,
            "reference_rounds_us": [
              313,
              343.333,
              311.125,
              318.208
            ],
            "candidate_rounds_us": [
              289.917,
              311.125,
              286.958,
              293.667
            ],
            "torch_rounds_us": [
              512.125,
              484.333,
              518.125,
              533.25
            ],
            "reference_torch_us": 564.1665,
            "reference_torch_rounds_us": [
              579.792,
              469.58299999999997,
              548.5409999999999,
              614.459
            ],
            "reference_native_torch_ratio": 0.5535175911954977,
            "reference_ratio_min": 0.5178669366060227,
            "reference_ratio_max": 0.7311444409188579,
            "gain": 1.0838926707591168,
            "gain_min": 1.0796193393281526,
            "gain_max": 1.1035210928083568,
            "native_torch_ratio": 0.5599726278282411,
            "ratio_min": 0.5507116736990154,
            "ratio_max": 0.6423782810586931
          }
        },
        {
          "name": "softmax_4096x1024",
          "baseline_kernel_score": 4352,
          "candidate_kernel_score": 5888,
          "gpu_throughput": {
            "reference_us": 48.15136662734375,
            "candidate_us": 53.3313796040625,
            "torch_us": 100.85937537951395,
            "reference_rounds_us": [
              47.59440071218749,
              49.86588555765625,
              46.96224004875,
              48.7083325425
            ],
            "candidate_rounds_us": [
              53.026041314375,
              53.44400960890625,
              53.21874959921875,
              53.587239563124996
            ],
            "torch_rounds_us": [
              99.4251304291538,
              102.2936203298741,
              103.18619933968876,
              99.13736903399695
            ],
            "reference_torch_us": 100.6253260129597,
            "reference_torch_rounds_us": [
              99.68294398277067,
              102.4075509121758,
              101.48958517675055,
              99.76106684916886
            ],
            "reference_native_torch_ratio": 0.4821967240878695,
            "reference_ratio_min": 0.4627296482389033,
            "reference_ratio_max": 0.48824991633402737,
            "gain": 0.9032602427112912,
            "gain_min": 0.8824378701569382,
            "gain_max": 0.9330491091998135,
            "native_torch_ratio": 0.5278916310058384,
            "ratio_min": 0.5157545286072872,
            "ratio_max": 0.5405352198195662
          },
          "e2e_throughput": {
            "reference_us": 50.37105194595,
            "candidate_us": 55.3478447736,
            "torch_us": 109.4090596707819,
            "reference_rounds_us": [
              50.5077128028,
              50.2343910891,
              50.0739247312,
              50.5841757576
            ],
            "candidate_rounds_us": [
              54.8040275049,
              55.360460039,
              55.3352295082,
              55.4342605042
            ],
            "torch_rounds_us": [
              109.534,
              110.02215873015874,
              109.17299629629629,
              109.2841193415638
            ],
            "reference_torch_us": 109.41564309545876,
            "reference_torch_rounds_us": [
              109.70043076923078,
              109.13085542168676,
              108.7014705882353,
              109.73435632183909
            ],
            "reference_native_torch_ratio": 0.46053517418348255,
            "reference_ratio_min": 0.4603133632096253,
            "reference_ratio_max": 0.4609693577574014,
            "gain": 0.9099565222302345,
            "gain_min": 0.9049194369706132,
            "gain_max": 0.9216058582242725,
            "native_torch_ratio": 0.5050168792414398,
            "ratio_min": 0.5003380457656983,
            "ratio_max": 0.5072490023087628
          },
          "gpu_latency": {
            "reference_us": 69.93746501395,
            "candidate_us": 77.87495269440001,
            "torch_us": 119.20835822820663,
            "reference_rounds_us": [
              58.2082429901,
              92.5000058487,
              58.916630223400006,
              80.9582998045
            ],
            "candidate_rounds_us": [
              77.2499479353,
              57.875004131299995,
              78.4999574535,
              87.37499592830001
            ],
            "torch_rounds_us": [
              119.16668154299259,
              119.25003491342068,
              121.99999764561653,
              118.99997480213642
            ],
            "reference_torch_us": 119.45833102799952,
            "reference_torch_rounds_us": [
              120.08333578705788,
              123.12503531575203,
              118.83332626894116,
              118.62500105053186
            ],
            "reference_native_torch_ratio": 0.5891323194197465,
            "reference_ratio_min": 0.4847320621848803,
            "reference_ratio_max": 0.7512688675498475,
            "gain": 0.8400333429009941,
            "gain_min": 0.7505307281000717,
            "gain_max": 1.598272125196689,
            "native_torch_ratio": 0.6458467585230996,
            "ratio_min": 0.4853248401421357,
            "ratio_max": 0.7342438187367697
          },
          "e2e_latency": {
            "reference_us": 319.25,
            "candidate_us": 316.2915,
            "torch_us": 484.5625,
            "reference_rounds_us": [
              313.583,
              324.917,
              296.292,
              350.833
            ],
            "candidate_rounds_us": [
              286.5,
              295.042,
              385.25,
              337.541
            ],
            "torch_rounds_us": [
              477.625,
              423.334,
              497.70799999999997,
              491.5
            ],
            "reference_torch_us": 440.9375,
            "reference_torch_rounds_us": [
              461.584,
              447.0,
              434.875,
              419.833
            ],
            "reference_native_torch_ratio": 0.7041052434774835,
            "reference_ratio_min": 0.6793628028701169,
            "reference_ratio_max": 0.8356489366009818,
            "gain": 1.066954730453196,
            "gain_min": 0.7690902011680726,
            "gain_max": 1.1012567702225446,
            "native_torch_ratio": 0.6918526877384961,
            "ratio_min": 0.5998429730437058,
            "ratio_max": 0.7740482371189533
          }
        },
        {
          "name": "rmsnorm_37x769",
          "baseline_kernel_score": 42,
          "candidate_kernel_score": 47,
          "gpu_throughput": {
            "reference_us": 3.15494798996875,
            "candidate_us": 4.935221568304687,
            "torch_us": 3.9098304114304483,
            "reference_rounds_us": [
              3.169921910734375,
              3.139974069203125,
              3.25520795740625,
              3.0996097848437496
            ],
            "candidate_rounds_us": [
              4.9446616685625004,
              5.3769535952656256,
              4.91471291753125,
              4.925781468046875
            ],
            "torch_rounds_us": [
              3.7597656046273187,
              3.9524738895124756,
              4.046874892082997,
              3.867186933348421
            ],
            "reference_torch_us": 4.05989612772828,
            "reference_torch_rounds_us": [
              4.343750333646312,
              4.314453690312803,
              3.694661245390307,
              3.8053385651437566
            ],
            "reference_native_torch_ratio": 0.772154292891899,
            "reference_ratio_min": 0.7277802230797553,
            "reference_ratio_max": 0.8810572177537664,
            "gain": 0.6351710915379012,
            "gain_min": 0.5839689730571327,
            "gain_max": 0.662339390322192,
            "native_torch_ratio": 1.2944445974764225,
            "ratio_min": 1.2144464676054174,
            "ratio_max": 1.360402053390631
          },
          "e2e_throughput": {
            "reference_us": 3.488899586315,
            "candidate_us": 5.39085323728,
            "torch_us": 6.846371526892781,
            "reference_rounds_us": [
              3.47700531423,
              3.5075307377,
              3.5007938584,
              3.45474897758
            ],
            "candidate_rounds_us": [
              5.40598653811,
              5.39924046033,
              5.38246601423,
              5.33811965255
            ],
            "torch_rounds_us": [
              6.742550259965338,
              6.901927384475302,
              7.1908411356085775,
              6.790815669310262
            ],
            "reference_torch_us": 7.0519527668326,
            "reference_torch_rounds_us": [
              6.795834465887469,
              7.225096774193548,
              7.019180581180812,
              7.084724952484388
            ],
            "reference_native_torch_ratio": 0.49319013288024716,
            "reference_ratio_min": 0.4854648798931145,
            "reference_ratio_max": 0.5116377292123959,
            "gain": 0.6484093655325638,
            "gain_min": 0.643176835480172,
            "gain_max": 0.6504070530393888,
            "native_torch_ratio": 0.7841797029705534,
            "ratio_min": 0.7485168859559946,
            "ratio_max": 0.8017717821413453
          },
          "gpu_latency": {
            "reference_us": 6.08334084973,
            "candidate_us": 7.49999890104,
            "torch_us": 8.45832983031869,
            "reference_rounds_us": [
              6.29166606814,
              6.2917242758,
              5.875015631319999,
              5.66669041291
            ],
            "candidate_rounds_us": [
              8.208327926699999,
              7.3749688454,
              7.24999699742,
              7.6250289566800005
            ],
            "torch_rounds_us": [
              9.125040378421545,
              7.791619282215834,
              7.7500008046627045,
              9.166717063635588
            ],
            "reference_torch_us": 8.14584200270474,
            "reference_torch_rounds_us": [
              8.791626896709204,
              7.708324119448662,
              8.583359885960817,
              7.6249707490205765
            ],
            "reference_native_torch_ratio": 0.7294091718807845,
            "reference_ratio_min": 0.684465723140578,
            "reference_ratio_max": 0.8162246654789104,
            "gain": 0.7884226065603521,
            "gain_min": 0.7431696909092557,
            "gain_max": 0.8531187599150803,
            "native_torch_ratio": 0.9175110962042177,
            "ratio_min": 0.8318167675239512,
            "ratio_max": 0.946525821947329
          },
          "e2e_latency": {
            "reference_us": 215.7295,
            "candidate_us": 224.27100000000002,
            "torch_us": 263.0625,
            "reference_rounds_us": [
              221.084,
              235.125,
              206.709,
              210.375
            ],
            "candidate_rounds_us": [
              219.917,
              218.167,
              228.625,
              235.625
            ],
            "torch_rounds_us": [
              265.875,
              228.0,
              260.25,
              281.875
            ],
            "reference_torch_us": 262.4165,
            "reference_torch_rounds_us": [
              259.541,
              232.459,
              287.0,
              265.292
            ],
            "reference_native_torch_ratio": 0.8224105068774648,
            "reference_ratio_min": 0.720240418118467,
            "reference_ratio_max": 1.0114686891021643,
            "gain": 0.9547232573326853,
            "gain_min": 0.8928381962864721,
            "gain_max": 1.0777294457915267,
            "native_torch_ratio": 0.8572012030049565,
            "ratio_min": 0.8271443347437706,
            "ratio_max": 0.9568728070175438
          }
        },
        {
          "name": "rmsnorm_1024x1024",
          "baseline_kernel_score": 704,
          "candidate_kernel_score": 896,
          "gpu_throughput": {
            "reference_us": 17.172525986093753,
            "candidate_us": 19.280924334453125,
            "torch_us": 20.80696640405222,
            "reference_rounds_us": [
              16.90820317890625,
              17.28255210765625,
              17.42708354875,
              17.06249986453125
            ],
            "candidate_rounds_us": [
              18.83007826109375,
              19.32486975421875,
              19.2369789146875,
              19.75976556421875
            ],
            "torch_rounds_us": [
              20.430338736332487,
              21.74674500565743,
              20.1855464183609,
              21.18359407177195
            ],
            "reference_torch_us": 20.34635417658137,
            "reference_torch_rounds_us": [
              22.476562662632205,
              20.33333294093609,
              20.242838218109682,
              20.359375412226655
            ],
            "reference_native_torch_ratio": 0.844013786690053,
            "reference_ratio_min": 0.7522592948349937,
            "reference_ratio_max": 0.8609011918674208,
            "gain": 0.8961262551662897,
            "gain_min": 0.8634970799161836,
            "gain_max": 0.9059158210879132,
            "native_torch_ratio": 0.927229305504371,
            "ratio_min": 0.8886327470704867,
            "ratio_max": 0.953007588498542
          },
          "e2e_throughput": {
            "reference_us": 18.01195236505,
            "candidate_us": 20.5921927085,
            "torch_us": 24.52249226562158,
            "reference_rounds_us": [
              17.6549744352,
              18.601898982,
              17.8813366864,
              18.1425680437
            ],
            "candidate_rounds_us": [
              20.0075193065,
              21.0147606178,
              20.1960802023,
              20.9883052147
            ],
            "torch_rounds_us": [
              24.53544682926829,
              24.669549865229108,
              24.234715242881073,
              24.509537701974864
            ],
            "reference_torch_us": 24.343769468812226,
            "reference_torch_rounds_us": [
              24.404411764705884,
              24.574356223175965,
              24.092238878143135,
              24.283127172918572
            ],
            "reference_native_torch_ratio": 0.7446648525737052,
            "reference_ratio_min": 0.7234337219605905,
            "reference_ratio_max": 0.7569638371424207,
            "gain": 0.8837997470512682,
            "gain_min": 0.8644131985937162,
            "gain_max": 0.8853864961560022,
            "native_torch_ratio": 0.8426017483036745,
            "ratio_min": 0.8154536351313996,
            "ratio_max": 0.8563321540336054
          },
          "gpu_latency": {
            "reference_us": 21.95834531445,
            "candidate_us": 27.1875178441,
            "torch_us": 25.666638975962996,
            "reference_rounds_us": [
              21.3333405554,
              22.5833500735,
              22.8333519772,
              21.1666920222
            ],
            "candidate_rounds_us": [
              23.125030566,
              30.625029467,
              23.7500062212,
              39.0416826122
            ],
            "torch_rounds_us": [
              25.749963242560625,
              25.583314709365368,
              25.500019546598196,
              43.70836541056633
            ],
            "reference_torch_us": 27.020840207114816,
            "reference_torch_rounds_us": [
              45.37496715784073,
              25.75002145022154,
              26.124995201826096,
              27.916685212403536
            ],
            "reference_native_torch_ratio": 0.816106711869699,
            "reference_ratio_min": 0.4701566059803446,
            "reference_ratio_max": 0.877022573249379,
            "gain": 0.8299682092682787,
            "gain_min": 0.5421562444541183,
            "gain_max": 0.9614040419416073,
            "native_torch_ratio": 0.9147164002373177,
            "ratio_min": 0.8932313584703816,
            "ratio_max": 1.197070427147933
          },
          "e2e_latency": {
            "reference_us": 269.083,
            "candidate_us": 273.6465,
            "torch_us": 291.125,
            "reference_rounds_us": [
              266.208,
              287.291,
              271.958,
              262.458
            ],
            "candidate_rounds_us": [
              261.084,
              259.834,
              286.209,
              305.041
            ],
            "torch_rounds_us": [
              265.458,
              322.791,
              260.667,
              316.79200000000003
            ],
            "reference_torch_us": 281.354,
            "reference_torch_rounds_us": [
              277.75,
              301.0,
              284.95799999999997,
              259.583
            ],
            "reference_native_torch_ratio": 0.9564498969830538,
            "reference_ratio_min": 0.9543792418531856,
            "reference_ratio_max": 1.011075455634614,
            "gain": 0.9849167914388861,
            "gain_min": 0.8604023721401386,
            "gain_max": 1.1056713132230578,
            "native_torch_ratio": 0.9732145387139575,
            "ratio_min": 0.804960485267557,
            "ratio_max": 1.0979870869730346
          }
        },
        {
          "name": "rmsnorm_16384x257",
          "baseline_kernel_score": 7936,
          "candidate_kernel_score": 5888,
          "gpu_throughput": {
            "reference_us": 48.78320305578125,
            "candidate_us": 49.46223953085938,
            "torch_us": 59.84342442388879,
            "reference_rounds_us": [
              48.54231792703125,
              48.83007841265625,
              49.6263028253125,
              48.73632769890625
            ],
            "candidate_rounds_us": [
              47.77278718390625,
              49.91341120328125,
              50.15624992671875,
              49.0110678584375
            ],
            "torch_rounds_us": [
              60.01888050377602,
              59.741536460933276,
              59.94531238684431,
              59.20507828705013
            ],
            "reference_torch_us": 59.96419258735841,
            "reference_torch_rounds_us": [
              60.130208112241235,
              59.390625210653525,
              60.394531828933395,
              59.79817706247559
            ],
            "reference_native_torch_ratio": 0.8183577616817042,
            "reference_ratio_min": 0.8072867108063286,
            "reference_ratio_max": 0.8221849532558395,
            "gain": 0.9919142003081585,
            "gain_min": 0.9782957573023223,
            "gain_max": 1.016108139978574,
            "native_torch_ratio": 0.8316539465542208,
            "ratio_min": 0.7959626501347469,
            "ratio_max": 0.8367001176513365
          },
          "e2e_throughput": {
            "reference_us": 51.47326535305,
            "candidate_us": 51.962343337600004,
            "torch_us": 67.36605153357462,
            "reference_rounds_us": [
              50.879639738,
              50.9912247788,
              52.5529126214,
              51.9553059273
            ],
            "candidate_rounds_us": [
              50.4644059041,
              52.9223901099,
              51.4155177165,
              52.5091689587
            ],
            "torch_rounds_us": [
              65.50036865342163,
              67.25366839378239,
              67.75048688046647,
              67.47843467336683
            ],
            "reference_torch_us": 67.32349792601659,
            "reference_torch_rounds_us": [
              65.90789473684211,
              67.1115426356589,
              68.17565957446807,
              67.53545321637426
            ],
            "reference_native_torch_ratio": 0.7700749257608986,
            "reference_ratio_min": 0.7597981327239889,
            "reference_ratio_max": 0.7719809582926732,
            "gain": 0.998840160920216,
            "gain_min": 0.9635094838481465,
            "gain_max": 1.0221216269992939,
            "native_torch_ratio": 0.7743034175997376,
            "ratio_min": 0.7588951767566396,
            "ratio_max": 0.7869071141224571
          },
          "gpu_latency": {
            "reference_us": 67.2083115205,
            "candidate_us": 76.1249975767,
            "torch_us": 82.4583403300494,
            "reference_rounds_us": [
              52.2916670889,
              64.5416439511,
              69.87497908990001,
              90.999994427
            ],
            "candidate_rounds_us": [
              60.041667893500005,
              62.0000064373,
              90.3749605641,
              90.24998871609999
            ],
            "torch_rounds_us": [
              94.33331433683634,
              70.58336632326245,
              95.6250005401671,
              70.12503920122981
            ],
            "reference_torch_us": 94.1875041462481,
            "reference_torch_rounds_us": [
              66.70830771327019,
              96.49997809901834,
              95.04164336249232,
              93.33336493000388
            ],
            "reference_native_torch_ratio": 0.7595445987191234,
            "reference_ratio_min": 0.6688254777102023,
            "reference_ratio_max": 0.9749996102170557,
            "gain": 0.9396166365107679,
            "gain_min": 0.7731674642373977,
            "gain_max": 1.0409941491920705,
            "native_torch_ratio": 0.9117458175321289,
            "ratio_min": 0.6364842401180667,
            "ratio_max": 1.2869866419200124
          },
          "e2e_latency": {
            "reference_us": 332.75,
            "candidate_us": 314.60400000000004,
            "torch_us": 314.021,
            "reference_rounds_us": [
              304.5,
              324.625,
              349.417,
              340.875
            ],
            "candidate_rounds_us": [
              329.833,
              291.5,
              299.375,
              369.833
            ],
            "torch_rounds_us": [
              446.75,
              317.375,
              306.0,
              310.66700000000003
            ],
            "reference_torch_us": 331.3535,
            "reference_torch_rounds_us": [
              317.29200000000003,
              379.542,
              332.916,
              329.791
            ],
            "reference_native_torch_ratio": 0.996646494477899,
            "reference_ratio_min": 0.8553071860294776,
            "reference_ratio_max": 1.0495650554494227,
            "gain": 1.0184154143570727,
            "gain_min": 0.9216997942314503,
            "gain_max": 1.1671549060542796,
            "native_torch_ratio": 0.948410756254714,
            "ratio_min": 0.7382943480693901,
            "ratio_max": 1.1904482935104146
          }
        },
        {
          "name": "rmsnorm_4096x1024",
          "baseline_kernel_score": 2816,
          "candidate_kernel_score": 3584,
          "gpu_throughput": {
            "reference_us": 49.34667958875,
            "candidate_us": 53.390299854062505,
            "torch_us": 68.38671879449976,
            "reference_rounds_us": [
              49.1067703478125,
              49.598958867187505,
              48.71289092984375,
              49.5865888296875
            ],
            "candidate_rounds_us": [
              52.93164122125,
              53.848958486875006,
              52.664713621249994,
              54.0794271728125
            ],
            "torch_rounds_us": [
              68.34114537923597,
              69.57682307984214,
              68.03320320614148,
              68.43229220976355
            ],
            "reference_torch_us": 69.38216120033758,
            "reference_torch_rounds_us": [
              70.37174509605393,
              69.05729151185369,
              69.70703088882146,
              67.79687555535929
            ],
            "reference_native_torch_ratio": 0.7085261722479146,
            "reference_ratio_min": 0.6978194200070525,
            "reference_ratio_max": 0.7313993222179944,
            "gain": 0.9230190779106227,
            "gain_min": 0.9169214879298185,
            "gain_max": 0.9277394241858127,
            "native_torch_ratio": 0.7743119700917094,
            "ratio_min": 0.773949658855237,
            "ratio_max": 0.7902618110035593
          },
          "e2e_throughput": {
            "reference_us": 52.665761982,
            "candidate_us": 55.4781827389,
            "torch_us": 74.90402183466081,
            "reference_rounds_us": [
              52.9666361974,
              52.3602415027,
              52.3683043478,
              52.9632196162
            ],
            "candidate_rounds_us": [
              54.973666,
              58.0723467933,
              55.5342562141,
              55.4221092637
            ],
            "torch_rounds_us": [
              74.79545454545455,
              74.03523936170212,
              75.01258912386707,
              75.58018731117825
            ],
            "reference_torch_us": 74.99115013089857,
            "reference_torch_rounds_us": [
              77.07630813953489,
              74.22639425587468,
              74.81230283911673,
              75.16999742268042
            ],
            "reference_native_torch_ratio": 0.7022875636666155,
            "reference_ratio_min": 0.6871973694109997,
            "reference_ratio_max": 0.7054127043030374,
            "gain": 0.9493122130856213,
            "gain_min": 0.9016381185535449,
            "gain_max": 0.9634910685672663,
            "native_torch_ratio": 0.7376596018003193,
            "ratio_min": 0.7332888583024604,
            "ratio_max": 0.7843879116752122
          },
          "gpu_latency": {
            "reference_us": 74.0416580811,
            "candidate_us": 69.62500629015,
            "torch_us": 85.37500980310142,
            "reference_rounds_us": [
              56.4166693948,
              57.541648857300004,
              90.54166730489999,
              95.9999742918
            ],
            "candidate_rounds_us": [
              60.458318330299996,
              88.2916501723,
              68.49999772379999,
              70.7500148565
            ],
            "torch_rounds_us": [
              82.66669465228915,
              86.91666880622506,
              83.83335079997778,
              92.62497769668698
            ],
            "reference_torch_us": 77.97917351126671,
            "reference_torch_rounds_us": [
              72.66667671501637,
              83.29167030751705,
              87.33331924304366,
              70.33336441963911
            ],
            "reference_native_torch_ratio": 0.906556446385233,
            "reference_ratio_min": 0.6908451786937797,
            "reference_ratio_max": 1.3649279411549666,
            "gain": 1.127463018839799,
            "gain_min": 0.6517224306602972,
            "gain_max": 1.3568898110694914,
            "native_torch_ratio": 0.7904651105008098,
            "ratio_min": 0.7313503773750536,
            "ratio_max": 1.015819535941263
          },
          "e2e_latency": {
            "reference_us": 299.041,
            "candidate_us": 339.14599999999996,
            "torch_us": 344.33299999999997,
            "reference_rounds_us": [
              293.917,
              326.292,
              300.916,
              297.166
            ],
            "candidate_rounds_us": [
              326.667,
              351.625,
              291.083,
              400.667
            ],
            "torch_rounds_us": [
              303.08299999999997,
              371.417,
              337.375,
              351.291
            ],
            "reference_torch_us": 341.85400000000004,
            "reference_torch_rounds_us": [
              431.66700000000003,
              334.041,
              325.83299999999997,
              349.66700000000003
            ],
            "reference_native_torch_ratio": 0.8866913002268044,
            "reference_ratio_min": 0.6808882773063495,
            "reference_ratio_max": 0.9768022488257428,
            "gain": 0.9138497486192594,
            "gain_min": 0.7416782515155977,
            "gain_max": 1.0337807429496053,
            "native_torch_ratio": 1.0122629271182937,
            "ratio_min": 0.8627876991478326,
            "ratio_max": 1.140555835475432
          }
        },
        {
          "name": "layernorm_37x769",
          "baseline_kernel_score": 64,
          "candidate_kernel_score": 77,
          "gpu_throughput": {
            "reference_us": 3.5143234526953124,
            "candidate_us": 5.274088380249999,
            "torch_us": 5.750650871050311,
            "reference_rounds_us": [
              3.76497428078125,
              3.48242247128125,
              3.546224434109375,
              3.481771273078125
            ],
            "candidate_rounds_us": [
              5.20572939421875,
              5.417317879625,
              5.34244736628125,
              5.1901042752499995
            ],
            "torch_rounds_us": [
              5.73632769373944,
              5.498697646544315,
              5.777343176305294,
              5.764974048361182
            ],
            "reference_torch_us": 5.597981726168655,
            "reference_torch_rounds_us": [
              5.659505404764786,
              5.536458047572523,
              5.534505362447817,
              5.92773449170636
            ],
            "reference_native_torch_ratio": 0.6348732370492538,
            "reference_ratio_min": 0.5873696397754584,
            "reference_ratio_max": 0.6652479344946797,
            "gain": 0.6673153970593272,
            "gain_min": 0.6428314802014742,
            "gain_max": 0.7232366486361088,
            "native_torch_ratio": 0.9161130015809854,
            "ratio_min": 0.9002823311451676,
            "ratio_max": 0.9852001742684546
          },
          "e2e_throughput": {
            "reference_us": 3.914741357,
            "candidate_us": 5.76315683136,
            "torch_us": 10.462697957063543,
            "reference_rounds_us": [
              4.01996684151,
              3.9570608153,
              3.8724218987,
              3.82157807309
            ],
            "candidate_rounds_us": [
              5.7103281177,
              5.81598554502,
              5.59024086672,
              5.88709884133
            ],
            "torch_rounds_us": [
              10.348568535262206,
              10.432995798319329,
              10.49240011580776,
              10.523342425514057
            ],
            "reference_torch_us": 10.308788242241832,
            "reference_torch_rounds_us": [
              10.160946909366704,
              10.247441534988713,
              10.37013494949495,
              10.447109712937475
            ],
            "reference_native_torch_ratio": 0.37978584412205035,
            "reference_ratio_min": 0.3658024255605778,
            "reference_ratio_max": 0.3956291551729552,
            "gain": 0.6865438794984435,
            "gain_min": 0.6491445406455308,
            "gain_max": 0.7039817605313297,
            "native_torch_ratio": 0.5546297895777226,
            "ratio_min": 0.5327895243241622,
            "ratio_max": 0.5594324125628194
          },
          "gpu_latency": {
            "reference_us": 6.12498843111,
            "candidate_us": 8.37503466755,
            "torch_us": 10.562478564679623,
            "reference_rounds_us": [
              6.16669422016,
              6.166636012490001,
              6.08334084973,
              5.95836900175
            ],
            "candidate_rounds_us": [
              8.37503466755,
              9.04168700799,
              8.37503466755,
              8.24994640425
            ],
            "torch_rounds_us": [
              11.249969247728586,
              10.416668374091387,
              10.499963536858559,
              10.624993592500687
            ],
            "reference_torch_us": 10.374991688877344,
            "reference_torch_rounds_us": [
              12.87501072511077,
              8.958333637565374,
              10.374991688877344,
              10.374991688877344
            ],
            "reference_native_torch_ratio": 0.5803238312175943,
            "reference_ratio_min": 0.47896614238408297,
            "reference_ratio_max": 0.6883686477841342,
            "gain": 0.7242986437835601,
            "gain_min": 0.6820227250778134,
            "gain_max": 0.7363186500055385,
            "native_torch_ratio": 0.787045565685713,
            "ratio_min": 0.7444495609835514,
            "ratio_max": 0.8680018104904561
          },
          "e2e_latency": {
            "reference_us": 210.75,
            "candidate_us": 214.9375,
            "torch_us": 259.625,
            "reference_rounds_us": [
              208.417,
              227.291,
              213.083,
              201.875
            ],
            "candidate_rounds_us": [
              216.458,
              199.709,
              244.917,
              213.417
            ],
            "torch_rounds_us": [
              289.66700000000003,
              258.75,
              255.917,
              260.5
            ],
            "reference_torch_us": 243.47899999999998,
            "reference_torch_rounds_us": [
              241.083,
              245.875,
              239.667,
              252.042
            ],
            "reference_native_torch_ratio": 0.876791263743935,
            "reference_ratio_min": 0.8009577768784567,
            "reference_ratio_max": 0.9244168784951703,
            "gain": 0.9543850005480234,
            "gain_min": 0.8700212725127288,
            "gain_max": 1.1381109514343368,
            "native_torch_ratio": 0.7955406696523779,
            "ratio_min": 0.7472649628711588,
            "ratio_max": 0.9570173142073407
          }
        },
        {
          "name": "layernorm_1024x1024",
          "baseline_kernel_score": 1088,
          "candidate_kernel_score": 1472,
          "gpu_throughput": {
            "reference_us": 18.1165364665625,
            "candidate_us": 18.96093772296875,
            "torch_us": 41.363931813975796,
            "reference_rounds_us": [
              17.93164028640625,
              18.6217448571875,
              17.766927157968748,
              18.301432646718748
            ],
            "candidate_rounds_us": [
              19.53580704140625,
              19.19335954875,
              18.7285158971875,
              18.422525499843747
            ],
            "torch_rounds_us": [
              42.22851566737518,
              42.22525967634283,
              40.24218742415542,
              40.50260395160876
            ],
            "reference_torch_us": 41.40722603551694,
            "reference_torch_rounds_us": [
              42.427735024830326,
              42.27148383506574,
              40.54296823596815,
              40.164062738767825
            ],
            "reference_native_torch_ratio": 0.4393759900273308,
            "reference_ratio_min": 0.4226395841284474,
            "reference_ratio_max": 0.4556668673125424,
            "gain": 0.9594372690143482,
            "gain_min": 0.917885821066928,
            "gain_max": 0.9934269135289824,
            "native_torch_ratio": 0.4587345694890209,
            "ratio_min": 0.45454686829322904,
            "ratio_max": 0.4653950765595234
          },
          "e2e_throughput": {
            "reference_us": 19.42825565935,
            "candidate_us": 19.9279012547,
            "torch_us": 47.68799348626364,
            "reference_rounds_us": [
              19.2410977814,
              19.4534155937,
              19.4167631173,
              19.4397482014
            ],
            "candidate_rounds_us": [
              20.6480708302,
              20.0235860215,
              19.8322164879,
              19.8140141844
            ],
            "torch_rounds_us": [
              48.02154294975689,
              48.778460122699386,
              47.046318770226534,
              47.354444022770394
            ],
            "reference_torch_us": 48.30531513589503,
            "reference_torch_rounds_us": [
              48.464948453608244,
              49.89417987804878,
              47.11441740674956,
              48.14568181818182
            ],
            "reference_native_torch_ratio": 0.4003899445550321,
            "reference_ratio_min": 0.3898934833932131,
            "reference_ratio_max": 0.4121193508490328,
            "gain": 0.9752883253657996,
            "gain_min": 0.9318593460681978,
            "gain_max": 0.9811110469833686,
            "native_torch_ratio": 0.41998296728799356,
            "ratio_min": 0.410500576917185,
            "ratio_max": 0.42997516451737683
          },
          "gpu_latency": {
            "reference_us": 24.2083042394,
            "candidate_us": 24.520792067050003,
            "torch_us": 44.64582889340818,
            "reference_rounds_us": [
              23.124972358300003,
              25.291636120499998,
              21.7916094698,
              25.2916943282
            ],
            "candidate_rounds_us": [
              26.2083485723,
              25.4582846537,
              23.5832994804,
              23.3333557844
            ],
            "torch_rounds_us": [
              44.79166818782687,
              44.49998959898949,
              55.166659876704216,
              43.458305299282074
            ],
            "reference_torch_us": 43.9999857917428,
            "reference_torch_rounds_us": [
              44.08333916217089,
              55.4583384655416,
              43.916632421314716,
              43.3750101365149
            ],
            "reference_native_torch_ratio": 0.5103889954572205,
            "reference_ratio_min": 0.456047491148958,
            "reference_ratio_max": 0.5830936810988406,
            "gain": 0.9587406122827463,
            "gain_min": 0.8823513734375898,
            "gain_max": 1.0839287139790363,
            "native_torch_ratio": 0.5545050163550052,
            "ratio_min": 0.427491886097653,
            "ratio_max": 0.5851165994175386
          },
          "e2e_latency": {
            "reference_us": 275.0,
            "candidate_us": 250.8535,
            "torch_us": 297.47900000000004,
            "reference_rounds_us": [
              273.875,
              250.667,
              295.083,
              276.125
            ],
            "candidate_rounds_us": [
              263.583,
              253.041,
              223.541,
              248.666
            ],
            "torch_rounds_us": [
              344.45799999999997,
              283.33399999999995,
              297.5,
              297.458
            ],
            "reference_torch_us": 292.8745,
            "reference_torch_rounds_us": [
              281.0,
              299.541,
              286.208,
              305.375
            ],
            "reference_native_torch_ratio": 0.9394301279128543,
            "reference_ratio_min": 0.8368370273184639,
            "reference_ratio_max": 1.0310089165921288,
            "gain": 1.0747358766333983,
            "gain_min": 0.990618121174039,
            "gain_max": 1.3200397242564006,
            "native_torch_ratio": 0.8005904865718267,
            "ratio_min": 0.751398319327731,
            "ratio_max": 0.893083780979339
          }
        },
        {
          "name": "layernorm_16384x257",
          "baseline_kernel_score": 12288,
          "candidate_kernel_score": 9472,
          "gpu_throughput": {
            "reference_us": 60.993163970156246,
            "candidate_us": 50.508788717734376,
            "torch_us": 115.7314454758307,
            "reference_rounds_us": [
              62.3281248409375,
              62.22395859375,
              59.762369346562494,
              59.512370171406246
            ],
            "candidate_rounds_us": [
              50.687499424375,
              50.96744826,
              49.847005357187506,
              50.33007801109375
            ],
            "torch_rounds_us": [
              115.48828115337528,
              115.07617182360264,
              116.90104111039545,
              115.97460979828611
            ],
            "reference_torch_us": 115.65104159672046,
            "reference_torch_rounds_us": [
              116.25195293163415,
              112.2968742492958,
              115.68749960133573,
              115.61458359210519
            ],
            "reference_native_torch_ratio": 0.5263657037309943,
            "reference_ratio_min": 0.5147479523981963,
            "reference_ratio_max": 0.5541023203870716,
            "gain": 1.2098863981635852,
            "gain_min": 1.1824414450199847,
            "gain_max": 1.2296547580519364,
            "native_torch_ratio": 0.43643615598388996,
            "ratio_min": 0.42640343391052016,
            "ratio_max": 0.44290184016658735
          },
          "e2e_throughput": {
            "reference_us": 62.7562586638,
            "candidate_us": 52.8372326203,
            "torch_us": 121.13751545134181,
            "reference_rounds_us": [
              62.8870135135,
              62.7339028698,
              62.7786144578,
              62.7297884615
            ],
            "candidate_rounds_us": [
              54.0515879121,
              53.2058823529,
              51.950075,
              52.4685828877
            ],
            "torch_rounds_us": [
              125.15404564315352,
              121.73406451612902,
              120.08455,
              120.54096638655462
            ],
            "reference_torch_us": 120.85516194673667,
            "reference_torch_rounds_us": [
              120.86976744186046,
              120.8405564516129,
              120.58888888888889,
              125.33176569037656
            ],
            "reference_native_torch_ratio": 0.5197167323715541,
            "reference_ratio_min": 0.5005098916141467,
            "reference_ratio_max": 0.5206003226022299,
            "gain": 1.1873234481709136,
            "gain_min": 1.1634628313930089,
            "gain_max": 1.2084412670780553,
            "native_torch_ratio": 0.43394421186723386,
            "ratio_min": 0.43188046886007203,
            "ratio_max": 0.4370665069336491
          },
          "gpu_latency": {
            "reference_us": 83.5208338685,
            "candidate_us": 62.3333326075,
            "torch_us": 113.54166781529784,
            "reference_rounds_us": [
              87.9166764207,
              61.125028878500004,
              97.8333409876,
              79.1249913163
            ],
            "candidate_rounds_us": [
              65.41667971759999,
              72.04164285210001,
              52.3333437741,
              59.2499854974
            ],
            "torch_rounds_us": [
              116.87498772516847,
              113.50004933774471,
              113.37501928210258,
              113.58328629285097
            ],
            "reference_torch_us": 113.93750901333988,
            "reference_torch_rounds_us": [
              113.99999493733048,
              113.83334640413523,
              113.87502308934927,
              116.70833919197321
            ],
            "reference_native_torch_ratio": 0.7245854972917212,
            "reference_ratio_min": 0.5369694453283639,
            "reference_ratio_max": 0.8591290551119137,
            "gain": 1.33969607192918,
            "gain_min": 0.8484680034849901,
            "gain_max": 1.8694265248921118,
            "native_torch_ratio": 0.5406792437085883,
            "ratio_min": 0.46159501542295533,
            "ratio_max": 0.6347278549432523
          },
          "e2e_latency": {
            "reference_us": 299.417,
            "candidate_us": 316.6875,
            "torch_us": 377.0625,
            "reference_rounds_us": [
              305,
              300.667,
              298.167,
              291.125
            ],
            "candidate_rounds_us": [
              335.417,
              308.583,
              324.792,
              297.792
            ],
            "torch_rounds_us": [
              384.334,
              367.084,
              398.792,
              369.791
            ],
            "reference_torch_us": 374.85450000000003,
            "reference_torch_rounds_us": [
              353.292,
              383.209,
              366.5,
              412.916
            ],
            "reference_native_torch_ratio": 0.799077851931216,
            "reference_ratio_min": 0.7050465469974523,
            "reference_ratio_max": 0.8633085379799147,
            "gain": 0.9461858585023107,
            "gain_min": 0.9093158665183936,
            "gain_max": 0.9776118901783797,
            "native_torch_ratio": 0.8275364070917606,
            "ratio_min": 0.8052981278614135,
            "ratio_max": 0.8727226839155525
          }
        },
        {
          "name": "layernorm_4096x1024",
          "baseline_kernel_score": 4352,
          "candidate_kernel_score": 5888,
          "gpu_throughput": {
            "reference_us": 53.986978400625,
            "candidate_us": 54.39029882835938,
            "torch_us": 137.04101547773462,
            "reference_rounds_us": [
              56.39713526765625,
              54.52408731796875,
              52.93099002296875,
              53.44986948328125
            ],
            "candidate_rounds_us": [
              54.434895901093746,
              54.86653662953125,
              54.345701755625,
              53.57617192203125
            ],
            "torch_rounds_us": [
              139.87369766255142,
              137.3789064018638,
              136.69270811078604,
              136.70312455360545
            ],
            "reference_torch_us": 139.33105446994887,
            "reference_torch_rounds_us": [
              141.62239585857606,
              142.83593736763578,
              137.03971308132168,
              133.3548170805443
            ],
            "reference_native_torch_ratio": 0.39223374877876727,
            "reference_ratio_min": 0.38172527392481687,
            "reference_ratio_max": 0.40080943945953096,
            "gain": 0.9957005332163396,
            "gain_min": 0.9739682865994123,
            "gain_max": 1.0360474532757042,
            "native_torch_ratio": 0.3947459699799402,
            "ratio_min": 0.38917178004701936,
            "ratio_max": 0.39938108452424603
          },
          "e2e_throughput": {
            "reference_us": 56.00499831005,
            "candidate_us": 55.435083404400004,
            "torch_us": 149.49135160783365,
            "reference_rounds_us": [
              55.9553293269,
              56.177943609,
              56.0546672932,
              55.6256290019
            ],
            "candidate_rounds_us": [
              57.2778054475,
              55.0141545802,
              55.6931818182,
              55.1769849906
            ],
            "torch_rounds_us": [
              153.59781818181816,
              147.56226704545455,
              143.07032682926828,
              151.42043617021275
            ],
            "reference_torch_us": 145.0525335948033,
            "reference_torch_rounds_us": [
              151.3793103448276,
              147.23760487804878,
              142.8674623115578,
              142.57671495327102
            ],
            "reference_native_torch_ratio": 0.38584571108525373,
            "reference_ratio_min": 0.36963657186334853,
            "reference_ratio_max": 0.39235432887412075,
            "gain": 1.0073108292219812,
            "gain_min": 0.9769111943052329,
            "gain_max": 1.0211543563230336,
            "native_torch_ratio": 0.37286380482673964,
            "ratio_min": 0.3643958925635056,
            "ratio_max": 0.3892713678124254
          },
          "gpu_latency": {
            "reference_us": 67.85415462219999,
            "candidate_us": 76.93746010774998,
            "torch_us": 136.97915710508823,
            "reference_rounds_us": [
              67.7083153278,
              68.2499958202,
              67.99999391659999,
              61.7083860561
            ],
            "candidate_rounds_us": [
              93.50001346319999,
              78.58331082389999,
              75.29160939159999,
              61.374972574400005
            ],
            "torch_rounds_us": [
              140.99996769800782,
              138.25000496581197,
              135.7083092443645,
              133.91667744144797
            ],
            "reference_torch_us": 136.08334120362997,
            "reference_torch_rounds_us": [
              142.87501107901335,
              139.000010676682,
              132.74996308609843,
              133.16667173057795
            ],
            "reference_native_torch_ratio": 0.48245302981757154,
            "reference_ratio_min": 0.4633921179688868,
            "reference_ratio_max": 0.512241151227265,
            "gain": 0.8858299866937589,
            "gain_min": 0.7241530008383242,
            "gain_max": 1.0054324013146534,
            "native_torch_ratio": 0.5616096180827608,
            "ratio_min": 0.45830716343179,
            "ratio_max": 0.663120814775343
          },
          "e2e_latency": {
            "reference_us": 329.4375,
            "candidate_us": 305.5625,
            "torch_us": 411.89549999999997,
            "reference_rounds_us": [
              343.25,
              308.667,
              323.5,
              335.375
            ],
            "candidate_rounds_us": [
              300.791,
              310.334,
              289.541,
              353.125
            ],
            "torch_rounds_us": [
              406.875,
              412.625,
              411.166,
              938.5
            ],
            "reference_torch_us": 422.6875,
            "reference_torch_rounds_us": [
              472.375,
              428.333,
              417.04200000000003,
              395.708
            ],
            "reference_native_torch_ratio": 0.7511742547478619,
            "reference_ratio_min": 0.7206239071003167,
            "reference_ratio_max": 0.8475315131359487,
            "gain": 1.055957001497465,
            "gain_min": 0.9497345132743363,
            "gain_max": 1.1411578138973573,
            "native_torch_ratio": 0.7217330871726507,
            "ratio_min": 0.3762653169952051,
            "ratio_max": 0.7520969403211148
          }
        }
      ]
    }
  },
  "environment": {
    "platform": "macOS-26.6.2-arm64-arm-64bit-Mach-O",
    "torch_version": "2.14.0",
    "torch_config": "PyTorch built with:\n  - GCC 4.2\n  - C++ Version: 202002\n  - clang 21.0.0\n  - OpenMP 202011\n  - LAPACK is enabled (usually provided by MKL)\n  - NNPACK is enabled\n  - CPU capability usage: DEFAULT\n  - Build settings: BLAS_INFO=accelerate, BUILD_TYPE=Release, COMMIT_SHA=08187d9e0fba026dc8217405802ab5381dc88d90, CUDA_FLAGS= -DLIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS -Xfatbin -compress-all  -Wno-deprecated-gpu-targets --expt-extended-lambda -DCUB_WRAPPED_NAMESPACE=at_cuda_detail -DDISABLE_CUSPARSE_DEPRECATED -DCUDA_HAS_FP16=1 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -DC10_NODEPRECATED, CXX_COMPILER=/usr/bin/c++, CXX_FLAGS= -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO -DUSE_PYTORCH_QNNPACK -DAT_BUILD_ARM_VEC256_WITH_SLEEF -DUSE_PYTORCH_METAL_EXPORT -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -DUSE_COREML_DELEGATE -O2 -fPIC -DC10_NODEPRECATED -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Werror=braced-scalar-init -Werror=range-loop-construct -Werror=bool-operation -Wnarrowing -Wno-missing-field-initializers -Wno-unknown-pragmas -Wno-unused-parameter -Wno-strict-overflow -Wno-strict-aliasing -Wvla-extension -Wsuggest-override -Wnewline-eof -Winconsistent-missing-override -Winconsistent-missing-destructor-override -Wno-pass-failed -Wno-error=old-style-cast -Wconstant-conversion -Qunused-arguments -faligned-new -fno-math-errno -fno-trapping-math -Werror=format -fclang-abi-compat=17 -DUSE_MPS -Wno-missing-braces, LAPACK_INFO=accelerate, TORCH_VERSION=2.14.0, USE_CUDA=OFF, USE_CUDNN=OFF, USE_CUSPARSELT=OFF, USE_EIGEN_FOR_BLAS=ON, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_GLOO=ON, USE_HIPSPARSELT=OFF, USE_MKL=OFF, USE_MKLDNN=OFF, USE_MPI=OFF, USE_NCCL=OFF, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=OFF, USE_ROCM_KERNEL_ASSERT=OFF, USE_XCCL=OFF, USE_XPU=OFF, \n",
    "threads": 8,
    "thread_environment": {
      "TVM_NUM_THREADS": "8",
      "OMP_NUM_THREADS": "8",
      "VECLIB_MAXIMUM_THREADS": "8"
    }
  }
}