{
  "prompt": "A playful orange tabby cat starts in an ordinary cozy living room in a normal house, afternoon light, sofa and rug. The cat crouches, jumps, and does one clean athletic backflip in slow motion. As the backflip completes there is a sharp cinematic cut: the cat lands perfectly on a glowing neon disco dance floor wearing oversized black sunglasses. Mirror ball reflections, colorful lights, joyful party energy, stylish and funny, clear before-and-after transformation.",
  "model_path": "/models/minimax_h3_fl2va_pruned_nvfp4.safetensors",
  "width": 1344,
  "height": 768,
  "frames": 124,
  "steps": 12,
  "sampler_step": 1,
  "seed": 440420,
  "text_tokens": 100,
  "block_index": 49,
  "attention": "sage2",
  "warmup": 2,
  "iterations": 3,
  "hidden_shape": [
    37810,
    5376
  ],
  "output_shape": [
    37810,
    5376
  ],
  "segments": [
    [
      0,
      100,
      1
    ],
    [
      100,
      514,
      2
    ],
    [
      514,
      37810,
      0
    ]
  ],
  "timings": {
    "norm1": {
      "count": 3,
      "mean_s": 0.003877114000109335,
      "p50_s": 0.003574107016902417,
      "p90_s": 0.004309529426973313,
      "p95_s": 0.0044014572282321755,
      "p99_s": 0.004474999469239265,
      "min_s": 0.0035638499539345503,
      "max_s": 0.004493385029491037
    },
    "modulate_msa": {
      "count": 3,
      "mean_s": 0.011260746008095643,
      "p50_s": 0.010931248019915074,
      "p90_s": 0.012151918408926576,
      "p95_s": 0.012304502207553015,
      "p99_s": 0.012426569246454165,
      "min_s": 0.0103939039981924,
      "max_s": 0.012457086006179452
    },
    "linear.attn_qkv_proj.flatten_contiguous": {
      "count": 3,
      "mean_s": 5.386633953700463e-06,
      "p50_s": 4.335946869105101e-06,
      "p90_s": 7.0239766500890255e-06,
      "p95_s": 7.359980372712016e-06,
      "p99_s": 7.628783350810409e-06,
      "min_s": 4.127970896661282e-06,
      "max_s": 7.695984095335007e-06
    },
    "linear.attn_qkv_proj.pre_quant_scale": {
      "count": 3,
      "mean_s": 0.0,
      "p50_s": 0.0,
      "p90_s": 0.0,
      "p95_s": 0.0,
      "p99_s": 0.0,
      "min_s": 0.0,
      "max_s": 0.0
    },
    "linear.attn_qkv_proj.packed_weight_wrapper": {
      "count": 3,
      "mean_s": 1.5333333673576515e-05,
      "p50_s": 8.84797191247344e-06,
      "p90_s": 2.498881658539176e-05,
      "p95_s": 2.700642216950655e-05,
      "p99_s": 2.862050663679838e-05,
      "min_s": 8.128001354634762e-06,
      "max_s": 2.902402775362134e-05
    },
    "linear.attn_qkv_proj.bias_cast": {
      "count": 3,
      "mean_s": 3.562658093869686e-06,
      "p50_s": 3.4239492379128933e-06,
      "p90_s": 3.782415296882391e-06,
      "p95_s": 3.827223554253578e-06,
      "p99_s": 3.8630701601505284e-06,
      "min_s": 3.3919932320713997e-06,
      "max_s": 3.872031811624765e-06
    },
    "linear.attn_qkv_proj.activation_scale": {
      "count": 3,
      "mean_s": 0.0019726533403930566,
      "p50_s": 0.001826349995099008,
      "p90_s": 0.002202194812707603,
      "p95_s": 0.0022491754149086777,
      "p99_s": 0.0022867598966695366,
      "min_s": 0.0017954540089704096,
      "max_s": 0.0022961560171097517
    },
    "linear.attn_qkv_proj.scale_to_device": {
      "count": 3,
      "mean_s": 1.755197687695424e-05,
      "p50_s": 1.7903978005051613e-05,
      "p90_s": 2.3907190188765527e-05,
      "p95_s": 2.4657591711729766e-05,
      "p99_s": 2.5257912930101157e-05,
      "min_s": 9.343959391117096e-06,
      "max_s": 2.5407993234694004e-05
    },
    "linear.attn_qkv_proj.activation_quant_pack": {
      "count": 3,
      "mean_s": 0.002370401343796402,
      "p50_s": 0.0021540760062634945,
      "p90_s": 0.0026746008079499006,
      "p95_s": 0.0027396664081607014,
      "p99_s": 0.0027917188883293423,
      "min_s": 0.00215239601675421,
      "max_s": 0.002804732008371502
    },
    "linear.attn_qkv_proj.activation_quant_wrap": {
      "count": 3,
      "mean_s": 7.146678399294615e-06,
      "p50_s": 6.608024705201387e-06,
      "p90_s": 7.977581117302179e-06,
      "p95_s": 8.148775668814778e-06,
      "p99_s": 8.285731310024857e-06,
      "min_s": 6.51204027235508e-06,
      "max_s": 8.319970220327377e-06
    },
    "linear.attn_qkv_proj.gemm": {
      "count": 3,
      "mean_s": 0.03147544667202359,
      "p50_s": 0.026172921003308147,
      "p90_s": 0.039067967410665004,
      "p95_s": 0.040679848211584616,
      "p99_s": 0.0419693528523203,
      "min_s": 0.025961690000258386,
      "max_s": 0.04229172901250422
    },
    "linear.attn_qkv_proj.slice_reshape": {
      "count": 3,
      "mean_s": 6.479987253745397e-06,
      "p50_s": 5.583977326750755e-06,
      "p90_s": 7.811188697814942e-06,
      "p95_s": 8.089590119197965e-06,
      "p99_s": 8.312311256304384e-06,
      "min_s": 5.4879928939044476e-06,
      "max_s": 8.367991540580988e-06
    },
    "attn_qkv_proj": {
      "count": 3,
      "mean_s": 0.03591976532091697,
      "p50_s": 0.030206882976926863,
      "p90_s": 0.04406598538625986,
      "p95_s": 0.04579837318742648,
      "p99_s": 0.04718428342835978,
      "min_s": 0.030021651997230947,
      "max_s": 0.0475307609885931
    },
    "attn_qkv_split_view": {
      "count": 3,
      "mean_s": 1.6997335478663445e-05,
      "p50_s": 1.4527991879731417e-05,
      "p90_s": 2.1683203522115948e-05,
      "p95_s": 2.257760497741401e-05,
      "p99_s": 2.3293126141652467e-05,
      "min_s": 1.2992008123546839e-05,
      "max_s": 2.3472006432712078e-05
    },
    "attn_qk_rms_rope": {
      "count": 3,
      "mean_s": 0.012158526010656109,
      "p50_s": 0.012062589987181127,
      "p90_s": 0.012462397222407162,
      "p95_s": 0.012512373126810417,
      "p99_s": 0.012552353850333021,
      "min_s": 0.011850639013573527,
      "max_s": 0.012562349031213671
    },
    "attn_q_transpose_contiguous": {
      "count": 3,
      "mean_s": 0.007270698668435216,
      "p50_s": 0.005053576023783535,
      "p90_s": 0.010375015193130822,
      "p95_s": 0.01104019508929923,
      "p99_s": 0.01157233900623396,
      "min_s": 0.0050531449960544705,
      "max_s": 0.011705374985467643
    },
    "attn_k_transpose_contiguous": {
      "count": 3,
      "mean_s": 0.00565119698876515,
      "p50_s": 0.005379768030252308,
      "p90_s": 0.006368463987018912,
      "p95_s": 0.006492050981614738,
      "p99_s": 0.006590920577291399,
      "min_s": 0.004958184959832579,
      "max_s": 0.006615637976210564
    },
    "attn_v_transpose_contiguous": {
      "count": 3,
      "mean_s": 0.005205602676142007,
      "p50_s": 0.005224872031249106,
      "p90_s": 0.005333428806625307,
      "p95_s": 0.005346998403547332,
      "p99_s": 0.005357854081084952,
      "min_s": 0.005031367996707559,
      "max_s": 0.005360568000469357
    },
    "attention_kernel": {
      "count": 3,
      "mean_s": 0.2570537893722455,
      "p50_s": 0.25454208703013137,
      "p90_s": 0.26159199584508314,
      "p95_s": 0.26247323444695214,
      "p99_s": 0.2631782253284473,
      "min_s": 0.25326480803778395,
      "max_s": 0.2633544730488211
    },
    "attn_output_reshape": {
      "count": 3,
      "mean_s": 0.004732787667308003,
      "p50_s": 0.00468229700345546,
      "p90_s": 0.004807135392911732,
      "p95_s": 0.004822740191593766,
      "p99_s": 0.004835224030539393,
      "min_s": 0.004677721008192748,
      "max_s": 0.0048383449902758
    },
    "linear.attn_out_proj.flatten_contiguous": {
      "count": 3,
      "mean_s": 4.373335589965184e-06,
      "p50_s": 4.224013537168503e-06,
      "p90_s": 4.607997834682465e-06,
      "p95_s": 4.6559958718717095e-06,
      "p99_s": 4.694394301623106e-06,
      "min_s": 4.191999323666096e-06,
      "max_s": 4.703993909060955e-06
    },
    "linear.attn_out_proj.pre_quant_scale": {
      "count": 3,
      "mean_s": 0.0,
      "p50_s": 0.0,
      "p90_s": 0.0,
      "p95_s": 0.0,
      "p99_s": 0.0,
      "min_s": 0.0,
      "max_s": 0.0
    },
    "linear.attn_out_proj.packed_weight_wrapper": {
      "count": 3,
      "mean_s": 1.0650682573517164e-05,
      "p50_s": 1.0048039257526398e-05,
      "p90_s": 1.1801626533269882e-05,
      "p95_s": 1.2020824942737818e-05,
      "p99_s": 1.2196183670312166e-05,
      "min_s": 9.66398511081934e-06,
      "max_s": 1.2240023352205753e-05
    },
    "linear.attn_out_proj.bias_cast": {
      "count": 3,
      "mean_s": 3.53597958261768e-06,
      "p50_s": 3.503984771668911e-06,
      "p90_s": 3.6448007449507716e-06,
      "p95_s": 3.662402741611004e-06,
      "p99_s": 3.6764843389391897e-06,
      "min_s": 3.4239492379128933e-06,
      "max_s": 3.6800047382712364e-06
    },
    "linear.attn_out_proj.activation_scale": {
      "count": 3,
      "mean_s": 0.0022405670024454594,
      "p50_s": 0.0022424450144171715,
      "p90_s": 0.002249599387869239,
      "p95_s": 0.0022504936845507474,
      "p99_s": 0.002251209121895954,
      "min_s": 0.002227868011686951,
      "max_s": 0.0022513879812322557
    },
    "linear.attn_out_proj.scale_to_device": {
      "count": 3,
      "mean_s": 1.0266647829363743e-05,
      "p50_s": 1.0079995263367891e-05,
      "p90_s": 1.0655971709638835e-05,
      "p95_s": 1.0727968765422702e-05,
      "p99_s": 1.0785566410049796e-05,
      "min_s": 9.91998240351677e-06,
      "max_s": 1.079996582120657e-05
    },
    "linear.attn_out_proj.activation_quant_pack": {
      "count": 3,
      "mean_s": 0.002841467969119549,
      "p50_s": 0.002840427972842008,
      "p90_s": 0.002845663169864565,
      "p95_s": 0.002846317569492385,
      "p99_s": 0.0028468410891946405,
      "min_s": 0.002837003965396434,
      "max_s": 0.0028469719691202044
    },
    "linear.attn_out_proj.activation_quant_wrap": {
      "count": 3,
      "mean_s": 6.853311788290739e-06,
      "p50_s": 6.783986464142799e-06,
      "p90_s": 7.1551650762557985e-06,
      "p95_s": 7.201562402769923e-06,
      "p99_s": 7.238680263981223e-06,
      "min_s": 6.52798917144537e-06,
      "max_s": 7.247959729284048e-06
    },
    "linear.attn_out_proj.gemm": {
      "count": 3,
      "mean_s": 0.007955518667586148,
      "p50_s": 0.007876515970565379,
      "p90_s": 0.008085207198746502,
      "p95_s": 0.008111293602269143,
      "p99_s": 0.008132162725087256,
      "min_s": 0.007852660026401281,
      "max_s": 0.008137380005791783
    },
    "linear.attn_out_proj.slice_reshape": {
      "count": 3,
      "mean_s": 5.562673322856426e-06,
      "p50_s": 5.4879928939044476e-06,
      "p90_s": 5.718402098864317e-06,
      "p95_s": 5.7472032494843004e-06,
      "p99_s": 5.770244169980288e-06,
      "min_s": 5.424022674560547e-06,
      "max_s": 5.776004400104284e-06
    },
    "attn_out_proj": {
      "count": 3,
      "mean_s": 0.01312111063937967,
      "p50_s": 0.013056731957476586,
      "p90_s": 0.013249423156958075,
      "p95_s": 0.01327350955689326,
      "p99_s": 0.013292778676841407,
      "min_s": 0.01300900400383398,
      "max_s": 0.013297595956828445
    },
    "gate_msa": {
      "count": 3,
      "mean_s": 0.008771182018487403,
      "p50_s": 0.008706916007213295,
      "p90_s": 0.008864918421022595,
      "p95_s": 0.008884668722748757,
      "p99_s": 0.008900468964129686,
      "min_s": 0.008702211023773998,
      "max_s": 0.008904419024474919
    },
    "norm2": {
      "count": 3,
      "mean_s": 0.003586383982716749,
      "p50_s": 0.0035896420013159513,
      "p90_s": 0.0036038891877979043,
      "p95_s": 0.003605670086108148,
      "p99_s": 0.0036070948047563435,
      "min_s": 0.003562058962415904,
      "max_s": 0.003607450984418392
    },
    "modulate_mlp": {
      "count": 3,
      "mean_s": 0.010889343665136645,
      "p50_s": 0.01104473602026701,
      "p90_s": 0.011215244000777603,
      "p95_s": 0.011236557498341427,
      "p99_s": 0.011253608296392485,
      "min_s": 0.010365423979237676,
      "max_s": 0.01125787099590525
    },
    "linear.mlp_fc1.flatten_contiguous": {
      "count": 3,
      "mean_s": 4.560007558514674e-06,
      "p50_s": 4.415982402861118e-06,
      "p90_s": 4.851189441978932e-06,
      "p95_s": 4.905590321868658e-06,
      "p99_s": 4.9491110257804395e-06,
      "min_s": 4.3040490709245205e-06,
      "max_s": 4.959991201758385e-06
    },
    "linear.mlp_fc1.pre_quant_scale": {
      "count": 3,
      "mean_s": 0.0,
      "p50_s": 0.0,
      "p90_s": 0.0,
      "p95_s": 0.0,
      "p99_s": 0.0,
      "min_s": 0.0,
      "max_s": 0.0
    },
    "linear.mlp_fc1.packed_weight_wrapper": {
      "count": 3,
      "mean_s": 8.78933739537994e-06,
      "p50_s": 8.608039934188128e-06,
      "p90_s": 9.23519255593419e-06,
      "p95_s": 9.313586633652448e-06,
      "p99_s": 9.376301895827056e-06,
      "min_s": 8.367991540580988e-06,
      "max_s": 9.391980711370707e-06
    },
    "linear.mlp_fc1.bias_cast": {
      "count": 3,
      "mean_s": 3.509359279026588e-06,
      "p50_s": 3.503984771668911e-06,
      "p90_s": 3.6064302548766136e-06,
      "p95_s": 3.6192359402775764e-06,
      "p99_s": 3.6294804885983467e-06,
      "min_s": 3.392051439732313e-06,
      "max_s": 3.6320416256785393e-06
    },
    "linear.mlp_fc1.activation_scale": {
      "count": 3,
      "mean_s": 0.0017761682975105941,
      "p50_s": 0.0017401569639332592,
      "p90_s": 0.0018271209788508712,
      "p95_s": 0.0018379914807155728,
      "p99_s": 0.001846687882207334,
      "min_s": 0.0017394859460182488,
      "max_s": 0.0018488619825802743
    },
    "linear.mlp_fc1.scale_to_device": {
      "count": 3,
      "mean_s": 1.0149359392623106e-05,
      "p50_s": 9.776034858077765e-06,
      "p90_s": 1.0953645687550305e-05,
      "p95_s": 1.1100847041234373e-05,
      "p99_s": 1.1218608124181627e-05,
      "min_s": 9.423994924873114e-06,
      "max_s": 1.1248048394918442e-05
    },
    "linear.mlp_fc1.activation_quant_pack": {
      "count": 3,
      "mean_s": 0.0022371909581124783,
      "p50_s": 0.002281755965668708,
      "p90_s": 0.0022871831548400222,
      "p95_s": 0.0022878615534864368,
      "p99_s": 0.002288404272403568,
      "min_s": 0.002141276956535876,
      "max_s": 0.002288539952132851
    },
    "linear.mlp_fc1.activation_quant_wrap": {
      "count": 3,
      "mean_s": 6.458683249851068e-06,
      "p50_s": 6.336020305752754e-06,
      "p90_s": 6.720004603266716e-06,
      "p95_s": 6.768002640455961e-06,
      "p99_s": 6.8064010702073575e-06,
      "min_s": 6.224028766155243e-06,
      "max_s": 6.8160006776452065e-06
    },
    "linear.mlp_fc1.gemm": {
      "count": 3,
      "mean_s": 0.03520335033924008,
      "p50_s": 0.03511902003083378,
      "p90_s": 0.03545432800892741,
      "p95_s": 0.0354962415061891,
      "p99_s": 0.035529772303998466,
      "min_s": 0.03495287598343566,
      "max_s": 0.03553815500345081
    },
    "linear.mlp_fc1.slice_reshape": {
      "count": 3,
      "mean_s": 5.594648731251557e-06,
      "p50_s": 5.6800199672579765e-06,
      "p90_s": 5.79517800360918e-06,
      "p95_s": 5.809572758153081e-06,
      "p99_s": 5.821088561788201e-06,
      "min_s": 5.279958713799715e-06,
      "max_s": 5.8239675126969814e-06
    },
    "mlp_fc1": {
      "count": 3,
      "mean_s": 0.039302336323695876,
      "p50_s": 0.03923152602510527,
      "p90_s": 0.039652977988589556,
      "p95_s": 0.03970565948402509,
      "p99_s": 0.03974780468037352,
      "min_s": 0.03891714196652174,
      "max_s": 0.03975834097946063
    },
    "mlp_swiglu": {
      "count": 3,
      "mean_s": 0.02406409801915288,
      "p50_s": 0.023905293026473373,
      "p90_s": 0.024343602627050133,
      "p95_s": 0.024398391327122227,
      "p99_s": 0.0244422222871799,
      "min_s": 0.023833821003790945,
      "max_s": 0.02445318002719432
    },
    "linear.mlp_fc2.flatten_contiguous": {
      "count": 3,
      "mean_s": 4.298674563566844e-06,
      "p50_s": 4.336005076766014e-06,
      "p90_s": 4.591979086399079e-06,
      "p95_s": 4.6239758376032116e-06,
      "p99_s": 4.649573238566518e-06,
      "min_s": 3.9040460251271725e-06,
      "max_s": 4.6559725888073444e-06
    },
    "linear.mlp_fc2.pre_quant_scale": {
      "count": 3,
      "mean_s": 0.0,
      "p50_s": 0.0,
      "p90_s": 0.0,
      "p95_s": 0.0,
      "p99_s": 0.0,
      "min_s": 0.0,
      "max_s": 0.0
    },
    "linear.mlp_fc2.packed_weight_wrapper": {
      "count": 3,
      "mean_s": 8.506661591430506e-06,
      "p50_s": 8.463975973427296e-06,
      "p90_s": 8.899183012545108e-06,
      "p95_s": 8.953583892434835e-06,
      "p99_s": 8.997104596346618e-06,
      "min_s": 8.048024028539658e-06,
      "max_s": 9.007984772324562e-06
    },
    "linear.mlp_fc2.bias_cast": {
      "count": 3,
      "mean_s": 3.5466509871184826e-06,
      "p50_s": 3.4879776649177074e-06,
      "p90_s": 3.641599323600531e-06,
      "p95_s": 3.6608020309358834e-06,
      "p99_s": 3.6761641968041657e-06,
      "min_s": 3.471970558166504e-06,
      "max_s": 3.6800047382712364e-06
    },
    "linear.mlp_fc2.activation_scale": {
      "count": 3,
      "mean_s": 0.004390478308778256,
      "p50_s": 0.004361385013908148,
      "p90_s": 0.004447554564103484,
      "p95_s": 0.004458325757877901,
      "p99_s": 0.004466942712897435,
      "min_s": 0.0043409529607743025,
      "max_s": 0.004469096951652318
    },
    "linear.mlp_fc2.scale_to_device": {
      "count": 3,
      "mean_s": 1.036798736701409e-05,
      "p50_s": 9.519979357719421e-06,
      "p90_s": 1.2169592082500458e-05,
      "p95_s": 1.2500793673098087e-05,
      "p99_s": 1.2765754945576191e-05,
      "min_s": 8.751987479627132e-06,
      "max_s": 1.2831995263695717e-05
    },
    "linear.mlp_fc2.activation_quant_pack": {
      "count": 3,
      "mean_s": 0.0059309463443545,
      "p50_s": 0.00591752800391987,
      "p90_s": 0.0060289128334261475,
      "p95_s": 0.006042835937114432,
      "p99_s": 0.00605397442006506,
      "min_s": 0.005818551988340914,
      "max_s": 0.006056759040802717
    },
    "linear.mlp_fc2.activation_quant_wrap": {
      "count": 3,
      "mean_s": 6.405326227347056e-06,
      "p50_s": 6.52798917144537e-06,
      "p90_s": 6.566406227648258e-06,
      "p95_s": 6.571208359673619e-06,
      "p99_s": 6.575050065293908e-06,
      "min_s": 6.111979018896818e-06,
      "max_s": 6.57601049169898e-06
    },
    "linear.mlp_fc2.gemm": {
      "count": 3,
      "mean_s": 0.05613339397435387,
      "p50_s": 0.056144060974474996,
      "p90_s": 0.05626750335795805,
      "p95_s": 0.05628293365589343,
      "p99_s": 0.056295277894241734,
      "min_s": 0.0559577569947578,
      "max_s": 0.05629836395382881
    },
    "linear.mlp_fc2.slice_reshape": {
      "count": 3,
      "mean_s": 5.423983869453271e-06,
      "p50_s": 5.423964466899633e-06,
      "p90_s": 5.462381523102522e-06,
      "p95_s": 5.467183655127883e-06,
      "p99_s": 5.471025360748172e-06,
      "min_s": 5.376001354306936e-06,
      "max_s": 5.471985787153244e-06
    },
    "mlp_fc2": {
      "count": 3,
      "mean_s": 0.06653381067250545,
      "p50_s": 0.06663658900652081,
      "p90_s": 0.06664222099352628,
      "p95_s": 0.06664292499190197,
      "p99_s": 0.06664348819060252,
      "min_s": 0.06632121402071789,
      "max_s": 0.06664362899027765
    },
    "gate_mlp": {
      "count": 3,
      "mean_s": 0.014599877002183348,
      "p50_s": 0.008974290976766497,
      "p90_s": 0.022819262195844205,
      "p95_s": 0.024549883598228916,
      "p99_s": 0.025934380720136686,
      "min_s": 0.008544835029169917,
      "max_s": 0.02628050500061363
    },
    "block_total": {
      "count": 3,
      "mean_s": 0.5241046730079688,
      "p50_s": 0.5232612389954738,
      "p90_s": 0.5255022222059779,
      "p95_s": 0.5257823451072908,
      "p99_s": 0.5260064434283414,
      "min_s": 0.5229903120198287,
      "max_s": 0.5260624680086039
    }
  },
  "module_forward": {
    "count": 3,
    "mean_s": 0.4765262039921557,
    "p50_s": 0.4681359010282904,
    "p90_s": 0.4901721953880042,
    "p95_s": 0.4929267321829684,
    "p99_s": 0.4951303616189398,
    "min_s": 0.46576144197024405,
    "max_s": 0.49568126897793263
  },
  "module_forward_checksum": -778077863936.0,
  "fused_elementwise": true,
  "profiler_summary": {
    "profiled_iterations": 1,
    "runtime_kernel_launches": 0,
    "runtime_kernel_launches_per_block": 0.0,
    "positive_self_device_allocated_bytes": 7351755264
  },
  "profiler_top_events": [
    {
      "key": "comfy_kitchen::rms_rope_split_half_",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 500.23900000000003,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[1, 37810, 56, 128], [1, 37810, 56, 128], [1, 37810, 1, 48, 2, 2], [128], [128], [], []]"
    },
    {
      "key": "aten::linear",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 416.1590000000001,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 7168], [5376, 7168], []]"
    },
    {
      "key": "comfy_kitchen::scaled_mm_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 323.05600000000004,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": -512,
      "input_shapes": "[[37824, 3584], [5376, 3584], [], [], [37888, 448], [5376, 448], [], [], []]"
    },
    {
      "key": "aten::_to_copy",
      "count": 8,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 217.82299999999896,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[], [], [], [], [], [], []]"
    },
    {
      "key": "comfy_kitchen::quantize_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 201.44000000000005,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 7168], [], [], [], []]"
    },
    {
      "key": "aten::to",
      "count": 13,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 188.896000000002,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[], [], [], [], []]"
    },
    {
      "key": "aten::rms_norm",
      "count": 2,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 186.01600000000002,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 5376], [], [5376], []]"
    },
    {
      "key": "aten::_fused_rms_norm",
      "count": 2,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 175.42400000000006,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 5376], [], [5376], []]"
    },
    {
      "key": "aten::copy_",
      "count": 8,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 147.2470000000003,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[], [], []]"
    },
    {
      "key": "aten::linear",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 133.7439999999997,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 5376], [28672, 5376], []]"
    },
    {
      "key": "aten::linear",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 129.29600000000028,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 14336], [5376, 14336], []]"
    },
    {
      "key": "aten::empty",
      "count": 29,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 120.14400000000344,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 6267507200,
      "input_shapes": "[[], [], [], [], [], []]"
    },
    {
      "key": "comfy_kitchen::scaled_mm_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 112.3680000000004,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": -512,
      "input_shapes": "[[37824, 7168], [5376, 7168], [], [], [37888, 896], [5376, 896], [], [], []]"
    },
    {
      "key": "comfy_kitchen::scaled_mm_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 109.55199999999968,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": -512,
      "input_shapes": "[[37824, 2688], [28672, 2688], [], [], [37888, 336], [28672, 336], [], [], []]"
    },
    {
      "key": "aten::zeros",
      "count": 3,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 107.92000000000007,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[], [], [], [], []]"
    },
    {
      "key": "aten::mean",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 86.096,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 14336,
      "input_shapes": "[[1, 37810, 56, 128], [], [], []]"
    },
    {
      "key": "sageattention_sm89::qk_int8_sv_f8_accum_f16_fuse_v_scale_attn_inst_buf",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 77.87199999999984,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[1, 37810, 56, 128], [1, 37810, 56, 128], [1, 128, 56, 37824], [1, 37810, 56, 128], [1, 56, 1184], [1, 56, 591], [1, 56, 128], [], [], [], [], []]"
    },
    {
      "key": "aten::mul",
      "count": 4,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 70.12799999999925,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 2048,
      "input_shapes": "[[], []]"
    },
    {
      "key": "comfy_kitchen::quantize_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 65.21600000000035,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 5376], [], [], [], []]"
    },
    {
      "key": "aten::slice",
      "count": 19,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 63.07199999999966,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 5376], [], [], [], []]"
    }
  ]
}