{
  "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": 245.29710325616531,
      "effective_gbps_p95": 226.7388199302835,
      "timing": {
        "count": 40,
        "mean_s": 0.002157641400117427,
        "p50_s": 0.0021236875327304006,
        "p90_s": 0.002282506227493286,
        "p95_s": 0.002297508649644442,
        "p99_s": 0.00230576561531052,
        "min_s": 0.002105231978930533,
        "max_s": 0.002310976036824286
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "attn_qkv_proj",
      "packer": "vortex_native",
      "threads": 64,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 224.59107678952034,
      "effective_gbps_p95": 207.23756769247012,
      "timing": {
        "count": 40,
        "mean_s": 0.002354832128912676,
        "p50_s": 0.002319479506695643,
        "p90_s": 0.002483506331918761,
        "p95_s": 0.0025137063988950104,
        "p99_s": 0.0026119847997324542,
        "min_s": 0.002309999952558428,
        "max_s": 0.0026610869681462646
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "attn_qkv_proj",
      "packer": "vortex_native",
      "threads": 128,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 233.2105432536058,
      "effective_gbps_p95": 211.95950883121543,
      "timing": {
        "count": 40,
        "mean_s": 0.002263313374714926,
        "p50_s": 0.002233751496532932,
        "p90_s": 0.002288319927174598,
        "p95_s": 0.002457707148278132,
        "p99_s": 0.0026292865216964853,
        "min_s": 0.0022275829687714577,
        "max_s": 0.002655663003679365
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "attn_qkv_proj",
      "packer": "vortex_native",
      "threads": 256,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 238.55814010715088,
      "effective_gbps_p95": 216.5766578813339,
      "timing": {
        "count": 40,
        "mean_s": 0.002226993774820585,
        "p50_s": 0.0021836789965163916,
        "p90_s": 0.0023945495078805832,
        "p95_s": 0.0024053118424490092,
        "p99_s": 0.0024354551278520375,
        "min_s": 0.0021724309772253036,
        "max_s": 0.002446942962706089
      },
      "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": 246.31591915364592,
      "effective_gbps_p95": 245.46406404228864,
      "timing": {
        "count": 40,
        "mean_s": 0.002115138196677435,
        "p50_s": 0.0021149035019334406,
        "p90_s": 0.0021202821750193833,
        "p95_s": 0.0021222430339548734,
        "p99_s": 0.002123958675074391,
        "min_s": 0.0021011190256103873,
        "max_s": 0.002125006983987987
      },
      "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": 244.56265212477584,
      "effective_gbps_p95": 231.34306753591417,
      "timing": {
        "count": 40,
        "mean_s": 0.0028569515954586676,
        "p50_s": 0.0028400869632605463,
        "p90_s": 0.002908607915742323,
        "p95_s": 0.00300237741030287,
        "p99_s": 0.0030247071670601146,
        "min_s": 0.002825614996254444,
        "max_s": 0.0030301110236905515
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "attn_out_proj",
      "packer": "vortex_native",
      "threads": 64,
      "input_shape": [
        37810,
        7168
      ],
      "qdata_shape": [
        37824,
        3584
      ],
      "block_scale_shape": [
        37888,
        448
      ],
      "bytes_touched": 694579200,
      "effective_gbps_p50": 231.12093550926818,
      "effective_gbps_p95": 211.18017623027535,
      "timing": {
        "count": 40,
        "mean_s": 0.003069607949873898,
        "p50_s": 0.0030052630172576755,
        "p90_s": 0.0032192086800932884,
        "p95_s": 0.0032890359900193287,
        "p99_s": 0.0033825889090076085,
        "min_s": 0.0029931350145488977,
        "max_s": 0.003397358988877386
      },
      "qdata_diff": 0,
      "block_scale_diff": 16827811
    },
    {
      "name": "attn_out_proj",
      "packer": "vortex_native",
      "threads": 128,
      "input_shape": [
        37810,
        7168
      ],
      "qdata_shape": [
        37824,
        3584
      ],
      "block_scale_shape": [
        37888,
        448
      ],
      "bytes_touched": 694579200,
      "effective_gbps_p50": 236.64753270567476,
      "effective_gbps_p95": 219.20626346231876,
      "timing": {
        "count": 40,
        "mean_s": 0.002979868765396532,
        "p50_s": 0.00293507898459211,
        "p90_s": 0.0031074534170329573,
        "p95_s": 0.003168610189459286,
        "p99_s": 0.003295062365941703,
        "min_s": 0.002914640004746616,
        "max_s": 0.003295599017292261
      },
      "qdata_diff": 0,
      "block_scale_diff": 16827134
    },
    {
      "name": "attn_out_proj",
      "packer": "vortex_native",
      "threads": 256,
      "input_shape": [
        37810,
        7168
      ],
      "qdata_shape": [
        37824,
        3584
      ],
      "block_scale_shape": [
        37888,
        448
      ],
      "bytes_touched": 694579200,
      "effective_gbps_p50": 242.57225236552964,
      "effective_gbps_p95": 223.35673896586212,
      "timing": {
        "count": 40,
        "mean_s": 0.0029051131932646967,
        "p50_s": 0.002863390982383862,
        "p90_s": 0.0029911637655459346,
        "p95_s": 0.003109730215510353,
        "p99_s": 0.0032386753160972147,
        "min_s": 0.002843327005393803,
        "max_s": 0.0032468310091644526
      },
      "qdata_diff": 0,
      "block_scale_diff": 16827064
    },
    {
      "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": 250.80788235490002,
      "effective_gbps_p95": 229.17894761840023,
      "timing": {
        "count": 40,
        "mean_s": 0.002809190830157604,
        "p50_s": 0.002769367507426068,
        "p90_s": 0.0029243686178233477,
        "p95_s": 0.0030307286389870563,
        "p99_s": 0.003207745099207386,
        "min_s": 0.002747935999650508,
        "max_s": 0.0032453910098411143
      },
      "qdata_diff": 0,
      "block_scale_diff": 16827062
    },
    {
      "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": 243.50477927357494,
      "effective_gbps_p95": 226.67511573686335,
      "timing": {
        "count": 40,
        "mean_s": 0.0021565302042290567,
        "p50_s": 0.002139318996341899,
        "p90_s": 0.002206075092544779,
        "p95_s": 0.002298154335585423,
        "p99_s": 0.0023168863559840246,
        "min_s": 0.002122094971127808,
        "max_s": 0.002320768020581454
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc1",
      "packer": "vortex_native",
      "threads": 64,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 223.27653165470315,
      "effective_gbps_p95": 203.9842303404134,
      "timing": {
        "count": 40,
        "mean_s": 0.002379627717891708,
        "p50_s": 0.00233313548960723,
        "p90_s": 0.0025109710812103005,
        "p95_s": 0.0025537974142935127,
        "p99_s": 0.002575046133133583,
        "min_s": 0.0023052149917930365,
        "max_s": 0.0025864150375127792
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc1",
      "packer": "vortex_native",
      "threads": 128,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 231.32488856945838,
      "effective_gbps_p95": 230.5624261809017,
      "timing": {
        "count": 40,
        "mean_s": 0.0022529397654579954,
        "p50_s": 0.002251960017019883,
        "p90_s": 0.0022583077894523743,
        "p95_s": 0.0022594071750063447,
        "p99_s": 0.002268195869983174,
        "min_s": 0.0022337600239552557,
        "max_s": 0.0022697430104017258
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc1",
      "packer": "vortex_native",
      "threads": 256,
      "input_shape": [
        37810,
        5376
      ],
      "qdata_shape": [
        37824,
        2688
      ],
      "block_scale_shape": [
        37888,
        336
      ],
      "bytes_touched": 520934400,
      "effective_gbps_p50": 237.7793630736635,
      "effective_gbps_p95": 220.94531333027638,
      "timing": {
        "count": 40,
        "mean_s": 0.0022103461771621367,
        "p50_s": 0.00219083100091666,
        "p90_s": 0.0022536566131748257,
        "p95_s": 0.0023577526590088382,
        "p99_s": 0.0024101987341418863,
        "min_s": 0.0021838549873791635,
        "max_s": 0.00241096003446728
      },
      "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": 244.1410580831981,
      "effective_gbps_p95": 241.96521054631256,
      "timing": {
        "count": 40,
        "mean_s": 0.00214694502792554,
        "p50_s": 0.0021337435173336416,
        "p90_s": 0.0021429663756862283,
        "p95_s": 0.0021529309888137504,
        "p99_s": 0.002424711784115061,
        "min_s": 0.002115023962687701,
        "max_s": 0.0025554709718562663
      },
      "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": 244.68703463081846,
      "effective_gbps_p95": 233.4123192225869,
      "timing": {
        "count": 40,
        "mean_s": 0.005744802793196868,
        "p50_s": 0.00567728650639765,
        "p90_s": 0.005920993315521627,
        "p95_s": 0.005951521344832144,
        "p99_s": 0.006018744396278634,
        "min_s": 0.005660655035171658,
        "max_s": 0.006024766014888883
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc2",
      "packer": "vortex_native",
      "threads": 64,
      "input_shape": [
        37810,
        14336
      ],
      "qdata_shape": [
        37824,
        7168
      ],
      "block_scale_shape": [
        37888,
        896
      ],
      "bytes_touched": 1389158400,
      "effective_gbps_p50": 224.47756060169795,
      "effective_gbps_p95": 211.92747342574486,
      "timing": {
        "count": 40,
        "mean_s": 0.006284207080898341,
        "p50_s": 0.006188406521687284,
        "p90_s": 0.00652997883153148,
        "p95_s": 0.006554876427981072,
        "p99_s": 0.006563341356813908,
        "min_s": 0.006155902985483408,
        "max_s": 0.006565663032233715
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc2",
      "packer": "vortex_native",
      "threads": 128,
      "input_shape": [
        37810,
        14336
      ],
      "qdata_shape": [
        37824,
        7168
      ],
      "block_scale_shape": [
        37888,
        896
      ],
      "bytes_touched": 1389158400,
      "effective_gbps_p50": 228.51826075047396,
      "effective_gbps_p95": 213.8415197976548,
      "timing": {
        "count": 40,
        "mean_s": 0.006181065877899528,
        "p50_s": 0.006078982027247548,
        "p90_s": 0.006448514922522008,
        "p95_s": 0.006496205233270302,
        "p99_s": 0.00660812895919662,
        "min_s": 0.006062957982067019,
        "max_s": 0.0066334379953332245
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    },
    {
      "name": "mlp_fc2",
      "packer": "vortex_native",
      "threads": 256,
      "input_shape": [
        37810,
        14336
      ],
      "qdata_shape": [
        37824,
        7168
      ],
      "block_scale_shape": [
        37888,
        896
      ],
      "bytes_touched": 1389158400,
      "effective_gbps_p50": 231.47246817495252,
      "effective_gbps_p95": 214.90532206804656,
      "timing": {
        "count": 40,
        "mean_s": 0.00611898949573515,
        "p50_s": 0.006001397967338562,
        "p90_s": 0.006412335694767534,
        "p95_s": 0.00646404838480521,
        "p99_s": 0.006495686792186461,
        "min_s": 0.005980286980047822,
        "max_s": 0.006505389988888055
      },
      "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": 238.58180014194556,
      "effective_gbps_p95": 223.09015483782161,
      "timing": {
        "count": 40,
        "mean_s": 0.005879163170175161,
        "p50_s": 0.005822566512506455,
        "p90_s": 0.006110190035542473,
        "p95_s": 0.006226892446284182,
        "p99_s": 0.006299175039166584,
        "min_s": 0.005806254979688674,
        "max_s": 0.006312365992926061
      },
      "qdata_diff": 0,
      "block_scale_diff": 0
    }
  ]
}