{
  "metadata": {
    "width": 1344,
    "height": 768,
    "frames": 124,
    "steps": 12,
    "sampler_step": 1,
    "seed": 440420,
    "text_tokens": 100,
    "hidden_shape": [
      37810,
      5376
    ],
    "segments": [
      [
        0,
        100,
        1
      ],
      [
        100,
        514,
        2
      ],
      [
        514,
        37810,
        0
      ]
    ]
  },
  "results": [
    {
      "name": "attn_qkv_proj",
      "packer": "ck",
      "threads": null,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 237.3554828038592,
      "effective_gbps_p95": 221.54486834351053,
      "timing": {
        "count": 20,
        "mean_s": 0.002216759545262903,
        "p50_s": 0.0021947434870526195,
        "p90_s": 0.0023419696139171716,
        "p95_s": 0.002351371999247931,
        "p99_s": 0.0023890071938512847,
        "min_s": 0.0021435199887491763,
        "max_s": 0.002398415992502123
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "attn_qkv_proj",
      "packer": "vortex_native",
      "threads": 512,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 233.3057458114144,
      "effective_gbps_p95": 212.10780665923392,
      "timing": {
        "count": 20,
        "mean_s": 0.0022636363486526535,
        "p50_s": 0.0022328399936668575,
        "p90_s": 0.0024042752280365675,
        "p95_s": 0.002455988811561838,
        "p99_s": 0.0025236713525373486,
        "min_s": 0.0021525119664147496,
        "max_s": 0.0025405919877812266
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "attn_out_proj",
      "packer": "ck",
      "threads": null,
      "input_shape": [
        37810,
        7168
      ],
      "qdata_shape": [
        37824,
        3584
      ],
      "block_scale_shape": [
        37888,
        448
      ],
      "bytes_touched": 694579200,
      "effective_gbps_p50": 244.5536052248782,
      "effective_gbps_p95": 231.08574698145156,
      "timing": {
        "count": 20,
        "mean_s": 0.0028653202491113914,
        "p50_s": 0.002840192028088495,
        "p90_s": 0.002945154212648049,
        "p95_s": 0.003005720642977394,
        "p99_s": 0.003050372929428704,
        "min_s": 0.0028018239536322653,
        "max_s": 0.0030615360010415316
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "attn_out_proj",
      "packer": "vortex_native",
      "threads": 512,
      "input_shape": [
        37810,
        7168
      ],
      "qdata_shape": [
        37824,
        3584
      ],
      "block_scale_shape": [
        37888,
        448
      ],
      "bytes_touched": 694579200,
      "effective_gbps_p50": 206.63049000273654,
      "effective_gbps_p95": 182.7584609572448,
      "timing": {
        "count": 20,
        "mean_s": 0.003328758548013866,
        "p50_s": 0.0033614555140957236,
        "p90_s": 0.0037920629896689207,
        "p95_s": 0.0038005310198059306,
        "p99_s": 0.003804483005660586,
        "min_s": 0.00287507101893425,
        "max_s": 0.00380547100212425
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc1",
      "packer": "ck",
      "threads": null,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 188.369281060563,
      "effective_gbps_p95": 165.52457219580273,
      "timing": {
        "count": 20,
        "mean_s": 0.0028155810461612417,
        "p50_s": 0.002765495504718274,
        "p90_s": 0.003082494280533865,
        "p95_s": 0.0031471726106246933,
        "p99_s": 0.0031740097067086024,
        "min_s": 0.0026715680141933262,
        "max_s": 0.00318071898072958
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc1",
      "packer": "vortex_native",
      "threads": 512,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 240.72231247788002,
      "effective_gbps_p95": 218.16303897955598,
      "timing": {
        "count": 20,
        "mean_s": 0.0022075875516748057,
        "p50_s": 0.002164047007681802,
        "p90_s": 0.002363825583597645,
        "p95_s": 0.0023878215230070054,
        "p99_s": 0.002408699474763125,
        "min_s": 0.002137008006684482,
        "max_s": 0.002413918962702155
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc2",
      "packer": "ck",
      "threads": null,
      "input_shape": [
        37810,
        14336
      ],
      "qdata_shape": [
        37824,
        7168
      ],
      "block_scale_shape": [
        37888,
        896
      ],
      "bytes_touched": 1389158400,
      "effective_gbps_p50": 198.68331940880012,
      "effective_gbps_p95": 185.7911050344262,
      "timing": {
        "count": 20,
        "mean_s": 0.007080884993774817,
        "p50_s": 0.006991821981500834,
        "p90_s": 0.007437778793973848,
        "p95_s": 0.007476990891154857,
        "p99_s": 0.007967294955160468,
        "min_s": 0.006620269967243075,
        "max_s": 0.008089870971161872
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc2",
      "packer": "vortex_native",
      "threads": 512,
      "input_shape": [
        37810,
        14336
      ],
      "qdata_shape": [
        37824,
        7168
      ],
      "block_scale_shape": [
        37888,
        896
      ],
      "bytes_touched": 1389158400,
      "effective_gbps_p50": 183.27811590655392,
      "effective_gbps_p95": 174.45709655535202,
      "timing": {
        "count": 20,
        "mean_s": 0.00760674009216018,
        "p50_s": 0.007579510478535667,
        "p90_s": 0.007870518049458042,
        "p95_s": 0.007962750885053538,
        "p99_s": 0.008361125348019412,
        "min_s": 0.007152206962928176,
        "max_s": 0.008460718963760883
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    }
  ]
}