{
  "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": 0,
  "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": 0.0039003503237230084,
      "p50_s": 0.00361842499114573,
      "p90_s": 0.004298052983358502,
      "p95_s": 0.004383006482385099,
      "p99_s": 0.004450969281606376,
      "min_s": 0.003614665998611599,
      "max_s": 0.0044679599814116955
    },
    "modulate_msa": {
      "count": 3,
      "mean_s": 0.011204699655839553,
      "p50_s": 0.010762252961285412,
      "p90_s": 0.012003415380604565,
      "p95_s": 0.01215856068301946,
      "p99_s": 0.012282676924951375,
      "min_s": 0.010538140020798892,
      "max_s": 0.012313705985434353
    },
    "linear.attn_qkv_proj.flatten_contiguous": {
      "count": 3,
      "mean_s": 6.149309532095988e-06,
      "p50_s": 4.4959597289562225e-06,
      "p90_s": 8.566351607441903e-06,
      "p95_s": 9.075150592252612e-06,
      "p99_s": 9.48218978010118e-06,
      "min_s": 4.368019290268421e-06,
      "max_s": 9.583949577063322e-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": 1.7290663284560043e-05,
      "p50_s": 8.816015906631947e-06,
      "p90_s": 2.942399587482214e-05,
      "p95_s": 3.199999337084591e-05,
      "p99_s": 3.406079136766494e-05,
      "min_s": 8.4799830801785e-06,
      "max_s": 3.457599086686969e-05
    },
    "linear.attn_qkv_proj.bias_cast": {
      "count": 3,
      "mean_s": 3.695992442468802e-06,
      "p50_s": 3.6479905247688293e-06,
      "p90_s": 3.993557766079903e-06,
      "p95_s": 4.036753671243787e-06,
      "p99_s": 4.071310395374894e-06,
      "min_s": 3.360037226229906e-06,
      "max_s": 4.079949576407671e-06
    },
    "linear.attn_qkv_proj.activation_scale": {
      "count": 3,
      "mean_s": 0.001928791015719374,
      "p50_s": 0.0017829880234785378,
      "p90_s": 0.002139096811879426,
      "p95_s": 0.002183610410429537,
      "p99_s": 0.0022192212892696264,
      "min_s": 0.001775261014699936,
      "max_s": 0.0022281240089796484
    },
    "linear.attn_qkv_proj.scale_to_device": {
      "count": 3,
      "mean_s": 1.439966338997086e-05,
      "p50_s": 1.3343989849090576e-05,
      "p90_s": 1.871921122074127e-05,
      "p95_s": 1.9391113892197608e-05,
      "p99_s": 1.9928636029362678e-05,
      "min_s": 9.791983757168055e-06,
      "max_s": 2.0063016563653946e-05
    },
    "linear.attn_qkv_proj.activation_quant_pack": {
      "count": 3,
      "mean_s": 0.0023788169880087176,
      "p50_s": 0.002249772020149976,
      "p90_s": 0.0026388911646790802,
      "p95_s": 0.0026875310577452183,
      "p99_s": 0.002726442972198129,
      "min_s": 0.0021505079930648208,
      "max_s": 0.0027361709508113563
    },
    "linear.attn_qkv_proj.activation_quant_wrap": {
      "count": 3,
      "mean_s": 7.8986631706357e-06,
      "p50_s": 6.432004738599062e-06,
      "p90_s": 1.0156782809644936e-05,
      "p95_s": 1.062238006852567e-05,
      "p99_s": 1.0994857875630258e-05,
      "min_s": 6.176007445901632e-06,
      "max_s": 1.1087977327406406e-05
    },
    "linear.attn_qkv_proj.gemm": {
      "count": 3,
      "mean_s": 0.03701395931420848,
      "p50_s": 0.025930000992957503,
      "p90_s": 0.05311313216807321,
      "p95_s": 0.05651102356496267,
      "p99_s": 0.05922933668247424,
      "min_s": 0.025202961987815797,
      "max_s": 0.05990891496185213
    },
    "linear.attn_qkv_proj.slice_reshape": {
      "count": 3,
      "mean_s": 2.0917330402880907e-05,
      "p50_s": 5.44002978131175e-06,
      "p90_s": 4.272641381248832e-05,
      "p95_s": 4.738721181638539e-05,
      "p99_s": 5.1115850219503045e-05,
      "min_s": 5.2639516070485115e-06,
      "max_s": 5.204800982028246e-05
    },
    "attn_qkv_proj": {
      "count": 3,
      "mean_s": 0.041458233628266804,
      "p50_s": 0.029947368951980025,
      "p90_s": 0.05808691458078101,
      "p95_s": 0.061604357784381135,
      "p99_s": 0.06441831234726124,
      "min_s": 0.02930553094483912,
      "max_s": 0.06512180098798126
    },
    "attn_qkv_split_view": {
      "count": 3,
      "mean_s": 2.4512002710253e-05,
      "p50_s": 1.7551996279507875e-05,
      "p90_s": 3.592000575736165e-05,
      "p95_s": 3.8216006942093366e-05,
      "p99_s": 4.005280788987874e-05,
      "min_s": 1.547200372442603e-05,
      "max_s": 4.0512008126825094e-05
    },
    "attn_qk_rms_rope": {
      "count": 3,
      "mean_s": 0.012107417666508505,
      "p50_s": 0.012135433964431286,
      "p90_s": 0.012213666830211878,
      "p95_s": 0.012223445938434451,
      "p99_s": 0.012231269225012512,
      "min_s": 0.011953593988437206,
      "max_s": 0.012233225046657026
    },
    "attn_q_transpose_contiguous": {
      "count": 3,
      "mean_s": 0.00539845967432484,
      "p50_s": 0.005506518005859107,
      "p90_s": 0.005652962799649686,
      "p95_s": 0.005671268398873508,
      "p99_s": 0.005685912878252567,
      "min_s": 0.004999287019018084,
      "max_s": 0.00568957399809733
    },
    "attn_k_transpose_contiguous": {
      "count": 3,
      "mean_s": 0.005182119008774559,
      "p50_s": 0.005078374990262091,
      "p90_s": 0.00533373502548784,
      "p95_s": 0.005365655029891059,
      "p99_s": 0.005391191033413634,
      "min_s": 0.0050704070017673075,
      "max_s": 0.005397575034294277
    },
    "attn_v_transpose_contiguous": {
      "count": 3,
      "mean_s": 0.005088572021729003,
      "p50_s": 0.004968855006154627,
      "p90_s": 0.005264303821604699,
      "p95_s": 0.005301234923535958,
      "p99_s": 0.005330779805080965,
      "min_s": 0.004958695033565164,
      "max_s": 0.005338166025467217
    },
    "attention_kernel": {
      "count": 3,
      "mean_s": 0.2596873336394007,
      "p50_s": 0.255025869992096,
      "p90_s": 0.26813469716580585,
      "p95_s": 0.2697733005625196,
      "p99_s": 0.27108418327989053,
      "min_s": 0.2526242269668728,
      "max_s": 0.2714119039592333
    },
    "attn_output_reshape": {
      "count": 3,
      "mean_s": 0.004714257995753239,
      "p50_s": 0.004697238968219608,
      "p90_s": 0.004743498226162046,
      "p95_s": 0.004749280633404851,
      "p99_s": 0.004753906559199095,
      "min_s": 0.004690471978392452,
      "max_s": 0.004755063040647656
    },
    "linear.attn_out_proj.flatten_contiguous": {
      "count": 3,
      "mean_s": 5.0720021439095335e-06,
      "p50_s": 4.799978341907263e-06,
      "p90_s": 5.862431135028601e-06,
      "p95_s": 5.995237734168768e-06,
      "p99_s": 6.101483013480901e-06,
      "min_s": 4.2879837565124035e-06,
      "max_s": 6.128044333308935e-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": 1.8890675467749436e-05,
      "p50_s": 1.5839992556720972e-05,
      "p90_s": 2.770560095086694e-05,
      "p95_s": 2.9188802000135183e-05,
      "p99_s": 3.037536283954978e-05,
      "min_s": 1.0160030797123909e-05,
      "max_s": 3.067200304940343e-05
    },
    "linear.attn_out_proj.bias_cast": {
      "count": 3,
      "mean_s": 3.722670953720808e-06,
      "p50_s": 3.760040272027254e-06,
      "p90_s": 3.7984107621014117e-06,
      "p95_s": 3.8032070733606814e-06,
      "p99_s": 3.8070441223680975e-06,
      "min_s": 3.5999692045152187e-06,
      "max_s": 3.8080033846199512e-06
    },
    "linear.attn_out_proj.activation_scale": {
      "count": 3,
      "mean_s": 0.002283606658844898,
      "p50_s": 0.0022722999565303326,
      "p90_s": 0.002299871202558279,
      "p95_s": 0.002303317608311772,
      "p99_s": 0.002306074732914567,
      "min_s": 0.002271756005939096,
      "max_s": 0.0023067640140652657
    },
    "linear.attn_out_proj.scale_to_device": {
      "count": 3,
      "mean_s": 1.3381339764843384e-05,
      "p50_s": 1.3264012522995472e-05,
      "p90_s": 1.6271998174488544e-05,
      "p95_s": 1.664799638092518e-05,
      "p99_s": 1.6948794946074483e-05,
      "min_s": 9.856012184172869e-06,
      "max_s": 1.7023994587361813e-05
    },
    "linear.attn_out_proj.activation_quant_pack": {
      "count": 3,
      "mean_s": 0.0028784163296222687,
      "p50_s": 0.0028789869975298643,
      "p90_s": 0.0029005677904933693,
      "p95_s": 0.0029032653896138073,
      "p99_s": 0.0029054234689101575,
      "min_s": 0.0028502990026026964,
      "max_s": 0.0029059629887342453
    },
    "linear.attn_out_proj.activation_quant_wrap": {
      "count": 3,
      "mean_s": 8.378662945081791e-06,
      "p50_s": 7.951981388032436e-06,
      "p90_s": 9.014387615025043e-06,
      "p95_s": 9.14718839339912e-06,
      "p99_s": 9.253429016098379e-06,
      "min_s": 7.90401827543974e-06,
      "max_s": 9.279989171773195e-06
    },
    "linear.attn_out_proj.gemm": {
      "count": 3,
      "mean_s": 0.008133968998057147,
      "p50_s": 0.007810881012119353,
      "p90_s": 0.008587405015714466,
      "p95_s": 0.008684470516163856,
      "p99_s": 0.008762122916523367,
      "min_s": 0.007809489965438843,
      "max_s": 0.008781536016613245
    },
    "linear.attn_out_proj.slice_reshape": {
      "count": 3,
      "mean_s": 5.957340666403373e-06,
      "p50_s": 5.856039933860302e-06,
      "p90_s": 6.444775499403477e-06,
      "p95_s": 6.518367445096373e-06,
      "p99_s": 6.577241001650691e-06,
      "min_s": 5.424022674560547e-06,
      "max_s": 6.59195939078927e-06
    },
    "attn_out_proj": {
      "count": 3,
      "mean_s": 0.013398274313658476,
      "p50_s": 0.013059511955361813,
      "p90_s": 0.013909942365717142,
      "p95_s": 0.014016246167011559,
      "p99_s": 0.014101289208047093,
      "min_s": 0.013012761017307639,
      "max_s": 0.014122549968305975
    },
    "gate_msa": {
      "count": 3,
      "mean_s": 0.008729871983329454,
      "p50_s": 0.008632095996290445,
      "p90_s": 0.00904886396601796,
      "p95_s": 0.0091009599622339,
      "p99_s": 0.009142636759206652,
      "min_s": 0.00840446399524808,
      "max_s": 0.00915305595844984
    },
    "norm2": {
      "count": 3,
      "mean_s": 0.0036775400318826237,
      "p50_s": 0.0036539300344884396,
      "p90_s": 0.0037461916450411085,
      "p95_s": 0.003757724346360192,
      "p99_s": 0.003766950507415459,
      "min_s": 0.0036094330134801567,
      "max_s": 0.0037692570476792753
    },
    "modulate_mlp": {
      "count": 3,
      "mean_s": 0.010796177666634321,
      "p50_s": 0.010941468004602939,
      "p90_s": 0.010954472806770353,
      "p95_s": 0.01095609840704128,
      "p99_s": 0.010957398887258022,
      "min_s": 0.010489340987987816,
      "max_s": 0.010957724007312208
    },
    "linear.mlp_fc1.flatten_contiguous": {
      "count": 3,
      "mean_s": 4.4533129160602885e-06,
      "p50_s": 4.399975296109915e-06,
      "p90_s": 4.7199777327477935e-06,
      "p95_s": 4.759978037327528e-06,
      "p99_s": 4.7919782809913166e-06,
      "min_s": 4.159985110163689e-06,
      "max_s": 4.799978341907263e-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": 9.487984546770653e-06,
      "p50_s": 9.295996278524399e-06,
      "p90_s": 1.0576006025075912e-05,
      "p95_s": 1.0736007243394852e-05,
      "p99_s": 1.0864008218050004e-05,
      "min_s": 8.271948900073767e-06,
      "max_s": 1.0896008461713791e-05
    },
    "linear.mlp_fc1.bias_cast": {
      "count": 3,
      "mean_s": 3.477325662970543e-06,
      "p50_s": 3.471970558166504e-06,
      "p90_s": 3.548804670572281e-06,
      "p95_s": 3.558408934623003e-06,
      "p99_s": 3.5660923458635807e-06,
      "min_s": 3.3919932320713997e-06,
      "max_s": 3.568013198673725e-06
    },
    "linear.mlp_fc1.activation_scale": {
      "count": 3,
      "mean_s": 0.0017434633142935734,
      "p50_s": 0.0017429409781470895,
      "p90_s": 0.0017490209778770803,
      "p95_s": 0.0017497809778433291,
      "p99_s": 0.0017503889778163284,
      "min_s": 0.0017369079869240522,
      "max_s": 0.0017505409778095782
    },
    "linear.mlp_fc1.scale_to_device": {
      "count": 3,
      "mean_s": 1.0010650536666313e-05,
      "p50_s": 9.999959729611874e-06,
      "p90_s": 1.058878842741251e-05,
      "p95_s": 1.066239201463759e-05,
      "p99_s": 1.0721274884417652e-05,
      "min_s": 9.295996278524399e-06,
      "max_s": 1.0735995601862669e-05
    },
    "linear.mlp_fc1.activation_quant_pack": {
      "count": 3,
      "mean_s": 0.0022977933404035866,
      "p50_s": 0.002274715981911868,
      "p90_s": 0.0024394648033194246,
      "p95_s": 0.002460058405995369,
      "p99_s": 0.0024765332881361247,
      "min_s": 0.0021380120306275785,
      "max_s": 0.0024806520086713135
    },
    "linear.mlp_fc1.activation_quant_wrap": {
      "count": 3,
      "mean_s": 6.543996278196573e-06,
      "p50_s": 6.464018952101469e-06,
      "p90_s": 6.7327986471354965e-06,
      "p95_s": 6.766396109014749e-06,
      "p99_s": 6.793274078518152e-06,
      "min_s": 6.367976311594248e-06,
      "max_s": 6.799993570894003e-06
    },
    "linear.mlp_fc1.gemm": {
      "count": 3,
      "mean_s": 0.03486676298780367,
      "p50_s": 0.03488327999366447,
      "p90_s": 0.0348947360063903,
      "p95_s": 0.03489616800798103,
      "p99_s": 0.03489731360925361,
      "min_s": 0.03481940896017477,
      "max_s": 0.03489760000957176
    },
    "linear.mlp_fc1.slice_reshape": {
      "count": 3,
      "mean_s": 5.68533626695474e-06,
      "p50_s": 5.6480057537555695e-06,
      "p90_s": 5.852803587913513e-06,
      "p95_s": 5.878403317183256e-06,
      "p99_s": 5.89888310059905e-06,
      "min_s": 5.504000000655651e-06,
      "max_s": 5.904003046452999e-06
    },
    "mlp_fc1": {
      "count": 3,
      "mean_s": 0.038996499322820455,
      "p50_s": 0.03899847296997905,
      "p90_s": 0.03917223219759762,
      "p95_s": 0.03919395210104994,
      "p99_s": 0.039211328023811796,
      "min_s": 0.03877535299398005,
      "max_s": 0.03921567200450227
    },
    "mlp_swiglu": {
      "count": 3,
      "mean_s": 0.02435958234127611,
      "p50_s": 0.02435267501277849,
      "p90_s": 0.024421091785188764,
      "p95_s": 0.024429643881740046,
      "p99_s": 0.024436485558981078,
      "min_s": 0.024287876032758504,
      "max_s": 0.024438195978291333
    },
    "linear.mlp_fc2.flatten_contiguous": {
      "count": 3,
      "mean_s": 4.271666208902995e-06,
      "p50_s": 4.208006430417299e-06,
      "p90_s": 4.579185042530298e-06,
      "p95_s": 4.625582369044423e-06,
      "p99_s": 4.662700230255723e-06,
      "min_s": 3.935012500733137e-06,
      "max_s": 4.671979695558548e-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": 8.533359505236149e-06,
      "p50_s": 8.704024367034435e-06,
      "p90_s": 9.216018952429295e-06,
      "p95_s": 9.280018275603652e-06,
      "p99_s": 9.331217734143137e-06,
      "min_s": 7.552036549896002e-06,
      "max_s": 9.34401759877801e-06
    },
    "linear.mlp_fc2.bias_cast": {
      "count": 3,
      "mean_s": 3.471989960720142e-06,
      "p50_s": 3.5199918784201145e-06,
      "p90_s": 3.5199918784201145e-06,
      "p95_s": 3.5199918784201145e-06,
      "p99_s": 3.5199918784201145e-06,
      "min_s": 3.375986125320196e-06,
      "max_s": 3.5199918784201145e-06
    },
    "linear.mlp_fc2.activation_scale": {
      "count": 3,
      "mean_s": 0.004591644974425435,
      "p50_s": 0.004597686987835914,
      "p90_s": 0.00461162697756663,
      "p95_s": 0.00461336947628297,
      "p99_s": 0.00461476347525604,
      "min_s": 0.004562135960441083,
      "max_s": 0.004615111974999309
    },
    "linear.mlp_fc2.scale_to_device": {
      "count": 3,
      "mean_s": 9.818662268420061e-06,
      "p50_s": 9.311945177614689e-06,
      "p90_s": 1.0924809612333774e-05,
      "p95_s": 1.112641766667366e-05,
      "p99_s": 1.1287704110145569e-05,
      "min_s": 8.816015906631947e-06,
      "max_s": 1.1328025721013546e-05
    },
    "linear.mlp_fc2.activation_quant_pack": {
      "count": 3,
      "mean_s": 0.006046714649225275,
      "p50_s": 0.006064180983230472,
      "p90_s": 0.006119732977822423,
      "p95_s": 0.006126676977146417,
      "p99_s": 0.006132232176605612,
      "min_s": 0.005942341987974942,
      "max_s": 0.006133620976470411
    },
    "linear.mlp_fc2.activation_quant_wrap": {
      "count": 3,
      "mean_s": 6.224009363601605e-06,
      "p50_s": 6.208021659404039e-06,
      "p90_s": 6.40001380816102e-06,
      "p95_s": 6.4240128267556436e-06,
      "p99_s": 6.443212041631341e-06,
      "min_s": 6.01599458605051e-06,
      "max_s": 6.4480118453502655e-06
    },
    "linear.mlp_fc2.gemm": {
      "count": 3,
      "mean_s": 0.05645514999438698,
      "p50_s": 0.05642460100352764,
      "p90_s": 0.05665761139243841,
      "p95_s": 0.05668673769105226,
      "p99_s": 0.05671003872994333,
      "min_s": 0.0562249849899672,
      "max_s": 0.056715863989666104
    },
    "linear.mlp_fc2.slice_reshape": {
      "count": 3,
      "mean_s": 5.6533220534523325e-06,
      "p50_s": 5.71197597309947e-06,
      "p90_s": 5.724781658500433e-06,
      "p95_s": 5.726382369175553e-06,
      "p99_s": 5.72766293771565e-06,
      "min_s": 5.520007107406855e-06,
      "max_s": 5.727983079850674e-06
    },
    "mlp_fc2": {
      "count": 3,
      "mean_s": 0.06717234667545806,
      "p50_s": 0.0670450460165739,
      "p90_s": 0.06738878921605647,
      "p95_s": 0.06743175711599178,
      "p99_s": 0.06746613143594005,
      "min_s": 0.06699726899387315,
      "max_s": 0.0674747250159271
    },
    "gate_mlp": {
      "count": 3,
      "mean_s": 0.018475912996412564,
      "p50_s": 0.008704703999683261,
      "p90_s": 0.032231163978576664,
      "p95_s": 0.035171971475938334,
      "p99_s": 0.037524617473827675,
      "min_s": 0.008610256016254425,
      "max_s": 0.03811277897330001
    },
    "block_total": {
      "count": 3,
      "mean_s": 0.5345096856666108,
      "p50_s": 0.535079502966255,
      "p90_s": 0.5420267941895872,
      "p95_s": 0.5428952055925038,
      "p99_s": 0.5435899347148371,
      "min_s": 0.5246859370381571,
      "max_s": 0.5437636169954203
    }
  },
  "module_forward": {
    "count": 3,
    "mean_s": 0.46528928333039704,
    "p50_s": 0.4646527419681661,
    "p90_s": 0.46618508918909357,
    "p95_s": 0.4663766325917095,
    "p99_s": 0.46652986731380225,
    "min_s": 0.46464693202869967,
    "max_s": 0.46656817599432543
  },
  "module_forward_checksum": 606380672.0,
  "fused_elementwise": true,
  "profiler_summary": {
    "profiled_iterations": 1,
    "runtime_kernel_launches": 0,
    "runtime_kernel_launches_per_block": 0.0,
    "positive_self_device_allocated_bytes": 9093546496
  },
  "profiler_top_events": [
    {
      "key": "aten::rms_norm",
      "count": 2,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 1360.0130000000001,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": -303104,
      "input_shapes": "[[37810, 5376], [], [5376], []]"
    },
    {
      "key": "aten::_fused_rms_norm",
      "count": 2,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 1339.1809999999998,
      "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::linear",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 603.1990000000001,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 5376], [21504, 5376], []]"
    },
    {
      "key": "comfy_kitchen::quantize_nvfp4",
      "count": 2,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 527.5030000000006,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[37810, 5376], [], [], [], []]"
    },
    {
      "key": "comfy_kitchen::scaled_mm_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 445.02400000000034,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": -512,
      "input_shapes": "[[37824, 2688], [21504, 2688], [], [], [37888, 336], [21504, 336], [], [], []]"
    },
    {
      "key": "comfy_kitchen::rms_rope_split_half_",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 265.37599999999975,
      "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::_to_copy",
      "count": 8,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 222.23999999999978,
      "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": 161.48800000000028,
      "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": "aten::to",
      "count": 12,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 160.35200000000077,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[], [], [], [], []]"
    },
    {
      "key": "aten::copy_",
      "count": 8,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 154.1279999999988,
      "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": 148.03099999999995,
      "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": 34,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 141.58399999999915,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 8009298432,
      "input_shapes": "[[], [], [], [], [], []]"
    },
    {
      "key": "comfy_kitchen::scaled_mm_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 138.3199999999997,
      "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": "comfy_kitchen::scaled_mm_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 132.09499999999935,
      "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": "aten::zeros",
      "count": 4,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 122.49499999999898,
      "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": 112.32000000000062,
      "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::mean",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 99.2470000000003,
      "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": "comfy_kitchen::scaled_mm_nvfp4",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 93.92000000000007,
      "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": "sageattention_sm89::qk_int8_sv_f8_accum_f16_fuse_v_scale_attn_inst_buf",
      "count": 1,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 81.16800000000057,
      "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::to",
      "count": 4,
      "device_type": "DeviceType.CPU",
      "cpu_time_total_us": 79.10399999999981,
      "device_time_total_us": 0.0,
      "self_device_time_total_us": 0.0,
      "self_device_memory_usage_bytes": 0,
      "input_shapes": "[[], [], [], [], [], []]"
    }
  ]
}