{
  "reference_variant": "L016",
  "aggregation": "repeat mean/median/P95 per sample; mean columns use mean across samples, median/P95 columns use median across samples",
  "bootstrap": {
    "unit": "sample_id",
    "iterations": 5000,
    "seed": 7
  },
  "target_control_hz": 10,
  "plot": {
    "x_scale": "linear",
    "log_base": null,
    "x_range": [
      30,
      137
    ]
  },
  "slopes": [
    {
      "metric": "prefill_ms_median",
      "per_token": 0.10440920155768525,
      "per_100_tokens": 10.440920155768525,
      "n": 1024
    },
    {
      "metric": "e2e_ms_median",
      "per_token": 0.14053770125841658,
      "per_100_tokens": 14.053770125841659,
      "n": 1024
    },
    {
      "metric": "peak_gpu_memory_mb_median",
      "per_token": 1.0031333419325994,
      "per_100_tokens": 100.31333419325993,
      "n": 1024
    }
  ],
  "summary": [
    {
      "variant_id": "L016",
      "target_L_instr": 16,
      "L_instr": 17,
      "L_input": 32,
      "multimodal_seq_len": 288,
      "n_samples": 128,
      "tokenize_ms_mean": 0.4053340885416667,
      "tokenize_ms_median": 0.36951100000000003,
      "tokenize_ms_p95": 0.41849309999999995,
      "tokenize_ms_p99": 0.42313301999999997,
      "image_preprocess_ms_mean": 4.066320434895832,
      "image_preprocess_ms_median": 3.731224,
      "image_preprocess_ms_p95": 4.19835255,
      "image_preprocess_ms_p99": 4.222664910000001,
      "host_to_device_ms_mean": 0.271160749757352,
      "host_to_device_ms_median": 0.23399999737739563,
      "host_to_device_ms_p95": 0.27926401048898697,
      "host_to_device_ms_p99": 0.2825152114033699,
      "prefill_ms_mean": 126.93774495522182,
      "prefill_ms_median": 126.3360481262207,
      "prefill_ms_p95": 127.51419296264649,
      "prefill_ms_p99": 127.59803970336914,
      "decode_ms_mean": 173.72042616208395,
      "decode_ms_median": 171.71148681640625,
      "decode_ms_p95": 177.46780319213866,
      "decode_ms_p99": 177.76532196044923,
      "postprocess_ms_mean": 0.20253742447916656,
      "postprocess_ms_median": 0.1891155,
      "postprocess_ms_p95": 0.20739195,
      "postprocess_ms_p99": 0.20861998999999998,
      "e2e_ms_mean": 305.6035238149797,
      "e2e_ms_median": 303.41158052872845,
      "e2e_ms_p95": 309.2498652454747,
      "e2e_ms_p99": 309.4819992185613,
      "peak_gpu_memory_mb_mean": 14731.883972167969,
      "peak_gpu_memory_mb_median": 14731.576171875,
      "peak_gpu_memory_mb_p95": 14731.576171875,
      "peak_gpu_memory_mb_p99": 14731.576171875,
      "cuda_memory_reserved_mb_mean": 14912,
      "cuda_memory_reserved_mb_median": 14912,
      "cuda_memory_reserved_mb_p95": 14912,
      "cuda_memory_reserved_mb_p99": 14912,
      "tokenize_us_per_token_mean": 12.553924658151606,
      "tokenize_us_per_token_median": 11.539080645161292,
      "tokenize_us_per_token_p95": 13.012364155844157,
      "tokenize_us_per_token_p99": 13.088116949308755,
      "decode_tokens_per_sec_mean": 34.601112771892986,
      "decode_tokens_per_sec_median": 34.942333279239975,
      "decode_tokens_per_sec_p95": 35.71059310565818,
      "decode_tokens_per_sec_p99": 35.83460673653984,
      "action_chunks_per_sec_mean": 3.274669399891757,
      "action_chunks_per_sec_median": 3.295853122957005,
      "action_chunks_per_sec_p95": 3.3375535662874682,
      "action_chunks_per_sec_p99": 3.341789377874865,
      "e2e_ms_median_ci95_low": 302.669009642664,
      "e2e_ms_median_ci95_high": 305.0530629622365,
      "e2e_ms_p95_ci95_low": 307.9516776696744,
      "e2e_ms_p95_ci95_high": 311.5550919788594,
      "prefill_ms_median_ci95_low": 126.2138900756836,
      "prefill_ms_median_ci95_high": 126.52611541748047,
      "prefill_ms_p95_ci95_low": 127.23916816711426,
      "prefill_ms_p95_ci95_high": 127.8188591003418,
      "peak_gpu_memory_mb_median_ci95_low": 14730.615234375,
      "peak_gpu_memory_mb_median_ci95_high": 14731.576171875,
      "peak_gpu_memory_mb_p95_ci95_low": 14730.615234375,
      "peak_gpu_memory_mb_p95_ci95_high": 14731.576171875,
      "target_control_hz": 10,
      "meets_target_control_hz_median": false,
      "meets_target_control_hz_p95": false,
      "delta_e2e_ms": 0,
      "delta_e2e_ms_ci95_low": 0,
      "delta_e2e_ms_ci95_high": 0,
      "delta_e2e_ms_percent": 0,
      "delta_prefill_ms": 0,
      "delta_prefill_ms_ci95_low": 0,
      "delta_prefill_ms_ci95_high": 0,
      "delta_prefill_ms_percent": 0,
      "delta_peak_gpu_memory_mb": 0,
      "delta_peak_gpu_memory_mb_ci95_low": 0,
      "delta_peak_gpu_memory_mb_ci95_high": 0,
      "delta_peak_gpu_memory_mb_percent": 0
    },
    {
      "variant_id": "L032",
      "target_L_instr": 32,
      "L_instr": 32,
      "L_input": 46,
      "multimodal_seq_len": 302,
      "n_samples": 128,
      "tokenize_ms_mean": 0.42364745052083336,
      "tokenize_ms_median": 0.393736,
      "tokenize_ms_p95": 0.44912215,
      "tokenize_ms_p99": 0.45326730000000004,
      "image_preprocess_ms_mean": 4.100199453124994,
      "image_preprocess_ms_median": 3.6629775,
      "image_preprocess_ms_p95": 4.2576034499999995,
      "image_preprocess_ms_p99": 4.29172133,
      "host_to_device_ms_mean": 0.2624956659274176,
      "host_to_device_ms_median": 0.2303999960422516,
      "host_to_device_ms_p95": 0.2733936034142971,
      "host_to_device_ms_p99": 0.2758735993504524,
      "prefill_ms_mean": 127.21572550137837,
      "prefill_ms_median": 126.81566619873047,
      "prefill_ms_p95": 127.94468612670899,
      "prefill_ms_p99": 128.04141792297364,
      "decode_ms_mean": 173.32261713345846,
      "decode_ms_median": 172.29209899902344,
      "decode_ms_p95": 177.19085617065429,
      "decode_ms_p99": 177.60669570922852,
      "postprocess_ms_mean": 0.19515809375000007,
      "postprocess_ms_median": 0.185281,
      "postprocess_ms_p95": 0.2022256,
      "postprocess_ms_p99": 0.20319865999999998,
      "e2e_ms_mean": 305.5198432981603,
      "e2e_ms_median": 304.1564531519909,
      "e2e_ms_p95": 309.48653193560824,
      "e2e_ms_p99": 310.0222870137753,
      "peak_gpu_memory_mb_mean": 14757.03759765625,
      "peak_gpu_memory_mb_median": 14757.005859375,
      "peak_gpu_memory_mb_p95": 14757.005859375,
      "peak_gpu_memory_mb_p99": 14757.005859375,
      "cuda_memory_reserved_mb_mean": 14912,
      "cuda_memory_reserved_mb_median": 14912,
      "cuda_memory_reserved_mb_p95": 14912,
      "cuda_memory_reserved_mb_p99": 14912,
      "tokenize_us_per_token_mean": 9.129580871299725,
      "tokenize_us_per_token_median": 8.477905874190565,
      "tokenize_us_per_token_p95": 9.659198242368177,
      "tokenize_us_per_token_p99": 9.740199347826087,
      "decode_tokens_per_sec_mean": 34.66852106593537,
      "decode_tokens_per_sec_median": 34.82458009220164,
      "decode_tokens_per_sec_p95": 35.67511870972719,
      "decode_tokens_per_sec_p99": 35.739491619807126,
      "action_chunks_per_sec_mean": 3.274945424477639,
      "action_chunks_per_sec_median": 3.28778163590202,
      "action_chunks_per_sec_p95": 3.3319076899135545,
      "action_chunks_per_sec_p99": 3.3366369770221267,
      "e2e_ms_median_ci95_low": 303.0236870966412,
      "e2e_ms_median_ci95_high": 305.4492928716307,
      "e2e_ms_p95_ci95_low": 308.2655084495365,
      "e2e_ms_p95_ci95_high": 311.56973979600696,
      "prefill_ms_median_ci95_low": 126.65414428710938,
      "prefill_ms_median_ci95_high": 127.01799774169922,
      "prefill_ms_p95_ci95_low": 127.66786766052246,
      "prefill_ms_p95_ci95_high": 128.1455421447754,
      "peak_gpu_memory_mb_median_ci95_low": 14757.005859375,
      "peak_gpu_memory_mb_median_ci95_high": 14757.005859375,
      "peak_gpu_memory_mb_p95_ci95_low": 14757.005859375,
      "peak_gpu_memory_mb_p95_ci95_high": 14757.005859375,
      "target_control_hz": 10,
      "meets_target_control_hz_median": false,
      "meets_target_control_hz_p95": false,
      "delta_e2e_ms": -0.38827927973889587,
      "delta_e2e_ms_ci95_low": -1.669828207365498,
      "delta_e2e_ms_ci95_high": 0.9536886933421442,
      "delta_e2e_ms_percent": -0.12797114700179738,
      "delta_prefill_ms": 0.4276313781738281,
      "delta_prefill_ms_ci95_low": 0.0840301513671875,
      "delta_prefill_ms_ci95_high": 0.6836410522460936,
      "delta_prefill_ms_percent": 0.3384872208022426,
      "delta_peak_gpu_memory_mb": 25.5078125,
      "delta_peak_gpu_memory_mb_ci95_low": 25.5078125,
      "delta_peak_gpu_memory_mb_ci95_high": 26.390625,
      "delta_peak_gpu_memory_mb_percent": 0.1731505997891699
    },
    {
      "variant_id": "L048",
      "target_L_instr": 48,
      "L_instr": 48,
      "L_input": 62,
      "multimodal_seq_len": 318,
      "n_samples": 128,
      "tokenize_ms_mean": 0.4497659661458333,
      "tokenize_ms_median": 0.411922,
      "tokenize_ms_p95": 0.463597,
      "tokenize_ms_p99": 0.46796355,
      "image_preprocess_ms_mean": 4.011469539062499,
      "image_preprocess_ms_median": 3.6730229999999997,
      "image_preprocess_ms_p95": 4.0078229,
      "image_preprocess_ms_p99": 4.04246889,
      "host_to_device_ms_mean": 0.2573349999729544,
      "host_to_device_ms_median": 0.22836799919605255,
      "host_to_device_ms_p95": 0.27114400193095206,
      "host_to_device_ms_p99": 0.2747321598231792,
      "prefill_ms_mean": 127.95866839090984,
      "prefill_ms_median": 127.51649475097656,
      "prefill_ms_p95": 128.48462142944334,
      "prefill_ms_p99": 128.53004661560058,
      "decode_ms_mean": 173.8758623600006,
      "decode_ms_median": 172.0248260498047,
      "decode_ms_p95": 179.04608383178712,
      "decode_ms_p99": 179.59393432617188,
      "postprocess_ms_mean": 0.19589202083333346,
      "postprocess_ms_median": 0.1859565,
      "postprocess_ms_p95": 0.20582245,
      "postprocess_ms_p99": 0.20755813,
      "e2e_ms_mean": 306.7489932769249,
      "e2e_ms_median": 304.69490107748743,
      "e2e_ms_p95": 311.58149563091183,
      "e2e_ms_p99": 312.0809853206979,
      "peak_gpu_memory_mb_mean": 14762.43603515625,
      "peak_gpu_memory_mb_median": 14761.974609375,
      "peak_gpu_memory_mb_p95": 14761.974609375,
      "peak_gpu_memory_mb_p99": 14761.974609375,
      "cuda_memory_reserved_mb_mean": 14912,
      "cuda_memory_reserved_mb_median": 14912,
      "cuda_memory_reserved_mb_p95": 14912,
      "cuda_memory_reserved_mb_p99": 14912,
      "tokenize_us_per_token_mean": 7.1987496606449435,
      "tokenize_us_per_token_median": 6.599129032258064,
      "tokenize_us_per_token_p95": 7.459737903225806,
      "tokenize_us_per_token_p99": 7.525487780337941,
      "decode_tokens_per_sec_mean": 34.5838042116036,
      "decode_tokens_per_sec_median": 34.87868662784102,
      "decode_tokens_per_sec_p95": 35.75910530549374,
      "decode_tokens_per_sec_p99": 35.885915063596286,
      "action_chunks_per_sec_mean": 3.2629201654021376,
      "action_chunks_per_sec_median": 3.2819720484913155,
      "action_chunks_per_sec_p95": 3.328496852714955,
      "action_chunks_per_sec_p99": 3.335256758238481,
      "e2e_ms_median_ci95_low": 303.4420755372779,
      "e2e_ms_median_ci95_high": 305.52531320897344,
      "e2e_ms_p95_ci95_low": 309.85864693176194,
      "e2e_ms_p95_ci95_high": 314.0487452145727,
      "prefill_ms_median_ci95_low": 127.32877082824707,
      "prefill_ms_median_ci95_high": 127.70660781860352,
      "prefill_ms_p95_ci95_low": 128.2796600341797,
      "prefill_ms_p95_ci95_high": 128.8871711730957,
      "peak_gpu_memory_mb_median_ci95_low": 14761.974609375,
      "peak_gpu_memory_mb_median_ci95_high": 14762.912109375,
      "peak_gpu_memory_mb_p95_ci95_low": 14761.974609375,
      "peak_gpu_memory_mb_p95_ci95_high": 14762.912109375,
      "target_control_hz": 10,
      "meets_target_control_hz_median": false,
      "meets_target_control_hz_p95": false,
      "delta_e2e_ms": 0.595313437077067,
      "delta_e2e_ms_ci95_low": -1.2329793018946589,
      "delta_e2e_ms_ci95_high": 2.061735765159162,
      "delta_e2e_ms_percent": 0.1962065640473139,
      "delta_prefill_ms": 1.1956825256347656,
      "delta_prefill_ms_ci95_low": 1.0014877319335938,
      "delta_prefill_ms_ci95_high": 1.3696632385253906,
      "delta_prefill_ms_percent": 0.9464302100380525,
      "delta_peak_gpu_memory_mb": 30.3984375,
      "delta_peak_gpu_memory_mb_ci95_low": 30.3984375,
      "delta_peak_gpu_memory_mb_ci95_high": 31.359375,
      "delta_peak_gpu_memory_mb_percent": 0.20634884648687907
    },
    {
      "variant_id": "L064",
      "target_L_instr": 64,
      "L_instr": 64,
      "L_input": 78,
      "multimodal_seq_len": 334,
      "n_samples": 128,
      "tokenize_ms_mean": 0.46312760677083326,
      "tokenize_ms_median": 0.427196,
      "tokenize_ms_p95": 0.47629984999999997,
      "tokenize_ms_p99": 0.48173077,
      "image_preprocess_ms_mean": 3.9513943958333324,
      "image_preprocess_ms_median": 3.6819889999999997,
      "image_preprocess_ms_p95": 4.07923575,
      "image_preprocess_ms_p99": 4.10010526,
      "host_to_device_ms_mean": 0.2619468334208553,
      "host_to_device_ms_median": 0.2337760031223297,
      "host_to_device_ms_p95": 0.26383999139070513,
      "host_to_device_ms_p99": 0.26560351356863976,
      "prefill_ms_mean": 129.13675743341446,
      "prefill_ms_median": 128.69175720214844,
      "prefill_ms_p95": 129.7557312011719,
      "prefill_ms_p99": 129.87894775390623,
      "decode_ms_mean": 173.81736449400583,
      "decode_ms_median": 171.87891387939453,
      "decode_ms_p95": 177.9135986328125,
      "decode_ms_p99": 178.33821166992186,
      "postprocess_ms_mean": 0.1978785286458333,
      "postprocess_ms_median": 0.18766549999999999,
      "postprocess_ms_p95": 0.20165575,
      "postprocess_ms_p99": 0.20257904,
      "e2e_ms_mean": 307.82846929209103,
      "e2e_ms_median": 306.1836927257614,
      "e2e_ms_p95": 312.5059172171384,
      "e2e_ms_p99": 312.7877301354555,
      "peak_gpu_memory_mb_mean": 14777.3896484375,
      "peak_gpu_memory_mb_median": 14777.1865234375,
      "peak_gpu_memory_mb_p95": 14777.1865234375,
      "peak_gpu_memory_mb_p99": 14777.1865234375,
      "cuda_memory_reserved_mb_mean": 14912,
      "cuda_memory_reserved_mb_median": 14912,
      "cuda_memory_reserved_mb_p95": 14912,
      "cuda_memory_reserved_mb_p99": 14912,
      "tokenize_us_per_token_mean": 5.922534443152224,
      "tokenize_us_per_token_median": 5.467972360972361,
      "tokenize_us_per_token_p95": 6.139971203796204,
      "tokenize_us_per_token_p99": 6.179007179487179,
      "decode_tokens_per_sec_mean": 34.59018642339857,
      "decode_tokens_per_sec_median": 34.9082960787197,
      "decode_tokens_per_sec_p95": 35.693786987177575,
      "decode_tokens_per_sec_p99": 35.7661066848485,
      "action_chunks_per_sec_mean": 3.251043416580028,
      "action_chunks_per_sec_median": 3.2660132780169446,
      "action_chunks_per_sec_p95": 3.3180704336141096,
      "action_chunks_per_sec_p99": 3.3228485760914,
      "e2e_ms_median_ci95_low": 304.9687383076529,
      "e2e_ms_median_ci95_high": 306.93680533368445,
      "e2e_ms_p95_ci95_low": 311.0583818131728,
      "e2e_ms_p95_ci95_high": 313.7860440225352,
      "prefill_ms_median_ci95_low": 128.60080032348634,
      "prefill_ms_median_ci95_high": 128.8337631225586,
      "prefill_ms_p95_ci95_low": 129.48661199569702,
      "prefill_ms_p95_ci95_high": 130.01752166748048,
      "peak_gpu_memory_mb_median_ci95_low": 14777.1865234375,
      "peak_gpu_memory_mb_median_ci95_high": 14777.1865234375,
      "peak_gpu_memory_mb_p95_ci95_low": 14777.1865234375,
      "peak_gpu_memory_mb_p95_ci95_high": 14777.1865234375,
      "target_control_hz": 10,
      "meets_target_control_hz_median": false,
      "meets_target_control_hz_p95": false,
      "delta_e2e_ms": 1.9582719007630374,
      "delta_e2e_ms_ci95_low": 0.11751784055519465,
      "delta_e2e_ms_ci95_high": 3.7147626358890022,
      "delta_e2e_ms_percent": 0.6454176525993275,
      "delta_prefill_ms": 2.366809844970703,
      "delta_prefill_ms_ci95_low": 2.1575851440429688,
      "delta_prefill_ms_ci95_high": 2.488475799560547,
      "delta_prefill_ms_percent": 1.873423998988835,
      "delta_peak_gpu_memory_mb": 46.5712890625,
      "delta_peak_gpu_memory_mb_ci95_low": 46.5712890625,
      "delta_peak_gpu_memory_mb_ci95_high": 46.5712890625,
      "delta_peak_gpu_memory_mb_percent": 0.3161324254726541
    },
    {
      "variant_id": "L080",
      "target_L_instr": 80,
      "L_instr": 80,
      "L_input": 94,
      "multimodal_seq_len": 350,
      "n_samples": 128,
      "tokenize_ms_mean": 0.4842347760416668,
      "tokenize_ms_median": 0.444969,
      "tokenize_ms_p95": 0.4997545,
      "tokenize_ms_p99": 0.50353366,
      "image_preprocess_ms_mean": 4.080658460937502,
      "image_preprocess_ms_median": 3.6898405,
      "image_preprocess_ms_p95": 4.18712065,
      "image_preprocess_ms_p99": 4.23607549,
      "host_to_device_ms_mean": 0.2650541668214525,
      "host_to_device_ms_median": 0.23585599660873413,
      "host_to_device_ms_p95": 0.26801119595766065,
      "host_to_device_ms_p99": 0.26999903589487073,
      "prefill_ms_mean": 129.89998334646225,
      "prefill_ms_median": 129.30335998535156,
      "prefill_ms_p95": 130.44280319213868,
      "prefill_ms_p99": 130.56405502319336,
      "decode_ms_mean": 173.1332858800888,
      "decode_ms_median": 171.61062622070312,
      "decode_ms_p95": 175.83702621459963,
      "decode_ms_p99": 176.2802703857422,
      "postprocess_ms_mean": 0.19748389062500013,
      "postprocess_ms_median": 0.186565,
      "postprocess_ms_p95": 0.20065870000000002,
      "postprocess_ms_p99": 0.20205182,
      "e2e_ms_mean": 308.06070052097687,
      "e2e_ms_median": 305.9766272969413,
      "e2e_ms_p95": 311.6314695693742,
      "e2e_ms_p99": 312.10627594377354,
      "peak_gpu_memory_mb_mean": 14803.702453613281,
      "peak_gpu_memory_mb_median": 14804.0615234375,
      "peak_gpu_memory_mb_p95": 14804.0615234375,
      "peak_gpu_memory_mb_p99": 14804.0615234375,
      "cuda_memory_reserved_mb_mean": 14912,
      "cuda_memory_reserved_mb_median": 14912,
      "cuda_memory_reserved_mb_p95": 14912,
      "cuda_memory_reserved_mb_p99": 14912,
      "tokenize_us_per_token_mean": 5.156207842693641,
      "tokenize_us_per_token_median": 4.727652081903454,
      "tokenize_us_per_token_p95": 5.335885769846717,
      "tokenize_us_per_token_p99": 5.38415440860215,
      "decode_tokens_per_sec_mean": 34.71686254078771,
      "decode_tokens_per_sec_median": 34.96287185416196,
      "decode_tokens_per_sec_p95": 35.67607907598463,
      "decode_tokens_per_sec_p99": 35.754212322209696,
      "action_chunks_per_sec_mean": 3.2484556993958513,
      "action_chunks_per_sec_median": 3.268223494798351,
      "action_chunks_per_sec_p95": 3.3088863761398644,
      "action_chunks_per_sec_p99": 3.312923173912208,
      "e2e_ms_median_ci95_low": 305.38294797900346,
      "e2e_ms_median_ci95_high": 307.0178788963127,
      "e2e_ms_p95_ci95_low": 309.61821069696026,
      "e2e_ms_p95_ci95_high": 314.05527567408353,
      "prefill_ms_median_ci95_low": 129.19187927246094,
      "prefill_ms_median_ci95_high": 129.4865264892578,
      "prefill_ms_p95_ci95_low": 130.24709396362306,
      "prefill_ms_p95_ci95_high": 130.67102966308596,
      "peak_gpu_memory_mb_median_ci95_low": 14804.0615234375,
      "peak_gpu_memory_mb_median_ci95_high": 14804.0615234375,
      "peak_gpu_memory_mb_p95_ci95_low": 14804.0615234375,
      "peak_gpu_memory_mb_p95_ci95_high": 14804.0615234375,
      "target_control_hz": 10,
      "meets_target_control_hz_median": false,
      "meets_target_control_hz_p95": false,
      "delta_e2e_ms": 2.8693500348524594,
      "delta_e2e_ms_ci95_low": 0.7992617737383512,
      "delta_e2e_ms_ci95_high": 4.011569434416742,
      "delta_e2e_ms_percent": 0.9456956223794417,
      "delta_prefill_ms": 3.005584716796875,
      "delta_prefill_ms_ci95_low": 2.7654457092285156,
      "delta_prefill_ms_ci95_high": 3.1633071899414062,
      "delta_prefill_ms_percent": 2.3790396813695125,
      "delta_peak_gpu_memory_mb": 72.4853515625,
      "delta_peak_gpu_memory_mb_ci95_low": 72.4853515625,
      "delta_peak_gpu_memory_mb_ci95_high": 72.5087890625,
      "delta_peak_gpu_memory_mb_percent": 0.4920407070961385
    },
    {
      "variant_id": "L096",
      "target_L_instr": 96,
      "L_instr": 96,
      "L_input": 110,
      "multimodal_seq_len": 366,
      "n_samples": 128,
      "tokenize_ms_mean": 0.5044586614583334,
      "tokenize_ms_median": 0.46571399999999996,
      "tokenize_ms_p95": 0.523871,
      "tokenize_ms_p99": 0.5277238,
      "image_preprocess_ms_mean": 3.9924818489583322,
      "image_preprocess_ms_median": 3.667445,
      "image_preprocess_ms_p95": 4.10462725,
      "image_preprocess_ms_p99": 4.1313588999999995,
      "host_to_device_ms_mean": 0.26088866707868874,
      "host_to_device_ms_median": 0.23043199628591537,
      "host_to_device_ms_p95": 0.2681871935725212,
      "host_to_device_ms_p99": 0.2713807965815067,
      "prefill_ms_mean": 131.58172460397086,
      "prefill_ms_median": 131.1648941040039,
      "prefill_ms_p95": 132.3068878173828,
      "prefill_ms_p99": 132.42980575561523,
      "decode_ms_mean": 173.80566112200418,
      "decode_ms_median": 172.16614532470703,
      "decode_ms_p95": 177.68566131591797,
      "decode_ms_p99": 178.14551361083983,
      "postprocess_ms_mean": 0.19716464322916652,
      "postprocess_ms_median": 0.1864265,
      "postprocess_ms_p95": 0.2021199,
      "postprocess_ms_p99": 0.20325678,
      "e2e_ms_mean": 310.3423795466995,
      "e2e_ms_median": 308.3481399628172,
      "e2e_ms_p95": 314.66026624956305,
      "e2e_ms_p99": 315.062865160543,
      "peak_gpu_memory_mb_mean": 14810.287963867188,
      "peak_gpu_memory_mb_median": 14810.2958984375,
      "peak_gpu_memory_mb_p95": 14810.2958984375,
      "peak_gpu_memory_mb_p99": 14810.2958984375,
      "cuda_memory_reserved_mb_mean": 14912,
      "cuda_memory_reserved_mb_median": 14912,
      "cuda_memory_reserved_mb_p95": 14912,
      "cuda_memory_reserved_mb_p99": 14912,
      "tokenize_us_per_token_mean": 4.586509718901687,
      "tokenize_us_per_token_median": 4.229109090909091,
      "tokenize_us_per_token_p95": 4.792289033579033,
      "tokenize_us_per_token_p99": 4.835515289693363,
      "decode_tokens_per_sec_mean": 34.58946733956871,
      "decode_tokens_per_sec_median": 34.850072593630664,
      "decode_tokens_per_sec_p95": 35.707157653007556,
      "decode_tokens_per_sec_p99": 35.76862892954428,
      "action_chunks_per_sec_mean": 3.224603857759021,
      "action_chunks_per_sec_median": 3.2430876396547395,
      "action_chunks_per_sec_p95": 3.28564225849977,
      "action_chunks_per_sec_p99": 3.2920327257238458,
      "e2e_ms_median_ci95_low": 307.2569610036201,
      "e2e_ms_median_ci95_high": 309.204082885716,
      "e2e_ms_p95_ci95_low": 313.1660112915565,
      "e2e_ms_p95_ci95_high": 316.81860697779894,
      "prefill_ms_median_ci95_low": 131.04156494140625,
      "prefill_ms_median_ci95_high": 131.29492950439453,
      "prefill_ms_p95_ci95_low": 132.02224655151366,
      "prefill_ms_p95_ci95_high": 132.56070709228516,
      "peak_gpu_memory_mb_median_ci95_low": 14810.2958984375,
      "peak_gpu_memory_mb_median_ci95_high": 14810.2958984375,
      "peak_gpu_memory_mb_p95_ci95_low": 14810.2958984375,
      "peak_gpu_memory_mb_p95_ci95_high": 14810.2958984375,
      "target_control_hz": 10,
      "meets_target_control_hz_median": false,
      "meets_target_control_hz_p95": false,
      "delta_e2e_ms": 4.611273667407943,
      "delta_e2e_ms_ci95_low": 2.976128114552978,
      "delta_e2e_ms_ci95_high": 5.598866320846042,
      "delta_e2e_ms_percent": 1.5198080638096558,
      "delta_prefill_ms": 4.725677490234375,
      "delta_prefill_ms_ci95_low": 4.512508392333984,
      "delta_prefill_ms_ci95_high": 4.873279571533203,
      "delta_prefill_ms_percent": 3.7405614314554243,
      "delta_peak_gpu_memory_mb": 77.8291015625,
      "delta_peak_gpu_memory_mb_ci95_low": 77.8291015625,
      "delta_peak_gpu_memory_mb_ci95_high": 79.6806640625,
      "delta_peak_gpu_memory_mb_percent": 0.5283148296859677
    },
    {
      "variant_id": "L112",
      "target_L_instr": 112,
      "L_instr": 112,
      "L_input": 126,
      "multimodal_seq_len": 382,
      "n_samples": 128,
      "tokenize_ms_mean": 0.5135195156249994,
      "tokenize_ms_median": 0.4805875,
      "tokenize_ms_p95": 0.52942555,
      "tokenize_ms_p99": 0.53463193,
      "image_preprocess_ms_mean": 4.025619747395832,
      "image_preprocess_ms_median": 3.6744855000000003,
      "image_preprocess_ms_p95": 4.203846199999999,
      "image_preprocess_ms_p99": 4.24137324,
      "host_to_device_ms_mean": 0.2683580835194637,
      "host_to_device_ms_median": 0.23043199628591537,
      "host_to_device_ms_p95": 0.2715216040611267,
      "host_to_device_ms_p99": 0.2759824050962925,
      "prefill_ms_mean": 132.3844921986262,
      "prefill_ms_median": 131.89691162109375,
      "prefill_ms_p95": 133.103475189209,
      "prefill_ms_p99": 133.16767990112305,
      "decode_ms_mean": 174.38516430060068,
      "decode_ms_median": 172.01254272460938,
      "decode_ms_p95": 177.19510726928712,
      "decode_ms_p99": 177.58974914550782,
      "postprocess_ms_mean": 0.19621305468749997,
      "postprocess_ms_median": 0.18573050000000002,
      "postprocess_ms_p95": 0.20145475000000002,
      "postprocess_ms_p99": 0.20271982,
      "e2e_ms_mean": 311.7733669004549,
      "e2e_ms_median": 309.7978535967512,
      "e2e_ms_p95": 315.21596732714346,
      "e2e_ms_p99": 315.8016952109655,
      "peak_gpu_memory_mb_mean": 14828.489868164062,
      "peak_gpu_memory_mb_median": 14828.1083984375,
      "peak_gpu_memory_mb_p95": 14828.1083984375,
      "peak_gpu_memory_mb_p99": 14828.1083984375,
      "cuda_memory_reserved_mb_mean": 14912,
      "cuda_memory_reserved_mb_median": 14912,
      "cuda_memory_reserved_mb_p95": 14912,
      "cuda_memory_reserved_mb_p99": 14912,
      "tokenize_us_per_token_mean": 4.063212673832437,
      "tokenize_us_per_token_median": 3.8024677227846517,
      "tokenize_us_per_token_p95": 4.1687051181102355,
      "tokenize_us_per_token_p99": 4.209700236220472,
      "decode_tokens_per_sec_mean": 34.49543374110424,
      "decode_tokens_per_sec_median": 34.88118335461633,
      "decode_tokens_per_sec_p95": 35.68295884557662,
      "decode_tokens_per_sec_p99": 35.76568182288885,
      "action_chunks_per_sec_mean": 3.2106012476827037,
      "action_chunks_per_sec_median": 3.227911667512983,
      "action_chunks_per_sec_p95": 3.2742115609526206,
      "action_chunks_per_sec_p99": 3.276314716125805,
      "e2e_ms_median_ci95_low": 308.5100724317875,
      "e2e_ms_median_ci95_high": 310.64394068701176,
      "e2e_ms_p95_ci95_low": 313.49499813994146,
      "e2e_ms_p95_ci95_high": 317.53883998217816,
      "prefill_ms_median_ci95_low": 131.7450408935547,
      "prefill_ms_median_ci95_high": 132.0551986694336,
      "prefill_ms_p95_ci95_low": 132.75580368041992,
      "prefill_ms_p95_ci95_high": 133.29821166992187,
      "peak_gpu_memory_mb_median_ci95_low": 14828.1083984375,
      "peak_gpu_memory_mb_median_ci95_high": 14829.0849609375,
      "peak_gpu_memory_mb_p95_ci95_low": 14828.1083984375,
      "peak_gpu_memory_mb_p95_ci95_high": 14829.0849609375,
      "target_control_hz": 10,
      "meets_target_control_hz_median": false,
      "meets_target_control_hz_p95": false,
      "delta_e2e_ms": 5.610223204815895,
      "delta_e2e_ms_ci95_low": 3.5664060513448703,
      "delta_e2e_ms_ci95_high": 7.052933989406142,
      "delta_e2e_ms_percent": 1.8490471573429916,
      "delta_prefill_ms": 5.579475402832031,
      "delta_prefill_ms_ci95_low": 5.3685150146484375,
      "delta_prefill_ms_ci95_high": 5.816715240478516,
      "delta_prefill_ms_percent": 4.416376391050043,
      "delta_peak_gpu_memory_mb": 96.5322265625,
      "delta_peak_gpu_memory_mb_ci95_low": 96.5322265625,
      "delta_peak_gpu_memory_mb_ci95_high": 97.4931640625,
      "delta_peak_gpu_memory_mb_percent": 0.6552742587503697
    },
    {
      "variant_id": "L128",
      "target_L_instr": 128,
      "L_instr": 128,
      "L_input": 142,
      "multimodal_seq_len": 398,
      "n_samples": 128,
      "tokenize_ms_mean": 0.5382953098958334,
      "tokenize_ms_median": 0.4986775,
      "tokenize_ms_p95": 0.5591272,
      "tokenize_ms_p99": 0.56268287,
      "image_preprocess_ms_mean": 4.037534213541666,
      "image_preprocess_ms_median": 3.722735,
      "image_preprocess_ms_p95": 4.1559579499999995,
      "image_preprocess_ms_p99": 4.19593916,
      "host_to_device_ms_mean": 0.2655677498939137,
      "host_to_device_ms_median": 0.23518399894237518,
      "host_to_device_ms_p95": 0.2663391903042793,
      "host_to_device_ms_p99": 0.26901439636945723,
      "prefill_ms_mean": 149.51009333133698,
      "prefill_ms_median": 149.00234985351562,
      "prefill_ms_p95": 150.0235939025879,
      "prefill_ms_p99": 150.1263882446289,
      "decode_ms_mean": 173.64276150862375,
      "decode_ms_median": 171.95982360839844,
      "decode_ms_p95": 178.04303665161132,
      "decode_ms_p99": 178.57703155517578,
      "postprocess_ms_mean": 0.20016386197916658,
      "postprocess_ms_median": 0.1860415,
      "postprocess_ms_p95": 0.20539,
      "postprocess_ms_p99": 0.2070708,
      "e2e_ms_mean": 328.1944159752713,
      "e2e_ms_median": 326.4524417601533,
      "e2e_ms_p95": 332.0470597680786,
      "e2e_ms_p99": 332.6158772260153,
      "peak_gpu_memory_mb_mean": 14844.918212890625,
      "peak_gpu_memory_mb_median": 14844.546875,
      "peak_gpu_memory_mb_p95": 14844.546875,
      "peak_gpu_memory_mb_p99": 14844.546875,
      "cuda_memory_reserved_mb_mean": 14912,
      "cuda_memory_reserved_mb_median": 14912,
      "cuda_memory_reserved_mb_p95": 14912,
      "cuda_memory_reserved_mb_p99": 14912,
      "tokenize_us_per_token_mean": 3.7800927199267877,
      "tokenize_us_per_token_median": 3.5031398601398602,
      "tokenize_us_per_token_p95": 3.9255376859056437,
      "tokenize_us_per_token_p99": 3.9482034965034964,
      "decode_tokens_per_sec_mean": 34.622011068418765,
      "decode_tokens_per_sec_median": 34.89189027503147,
      "decode_tokens_per_sec_p95": 35.67726409413112,
      "decode_tokens_per_sec_p99": 35.73528698668595,
      "action_chunks_per_sec_mean": 3.0489937735629624,
      "action_chunks_per_sec_median": 3.063233332675411,
      "action_chunks_per_sec_p95": 3.1020637776062907,
      "action_chunks_per_sec_p99": 3.105610044901656,
      "e2e_ms_median_ci95_low": 325.56352118772554,
      "e2e_ms_median_ci95_high": 327.36809750416467,
      "e2e_ms_p95_ci95_low": 330.87785136867296,
      "e2e_ms_p95_ci95_high": 333.5336060166759,
      "prefill_ms_median_ci95_low": 148.83786010742188,
      "prefill_ms_median_ci95_high": 149.1504135131836,
      "prefill_ms_p95_ci95_low": 149.82537536621095,
      "prefill_ms_p95_ci95_high": 150.2451599121094,
      "peak_gpu_memory_mb_median_ci95_low": 14844.546875,
      "peak_gpu_memory_mb_median_ci95_high": 14844.546875,
      "peak_gpu_memory_mb_p95_ci95_low": 14844.546875,
      "peak_gpu_memory_mb_p95_ci95_high": 14844.546875,
      "target_control_hz": 10,
      "meets_target_control_hz_median": false,
      "meets_target_control_hz_p95": false,
      "delta_e2e_ms": 22.709256701001692,
      "delta_e2e_ms_ci95_low": 21.31639232563259,
      "delta_e2e_ms_ci95_high": 23.877760987800347,
      "delta_e2e_ms_percent": 7.4846374226811925,
      "delta_prefill_ms": 22.60342788696289,
      "delta_prefill_ms_ci95_low": 22.402843475341797,
      "delta_prefill_ms_ci95_high": 22.880924224853516,
      "delta_prefill_ms_percent": 17.891510952107748,
      "delta_peak_gpu_memory_mb": 113.884765625,
      "delta_peak_gpu_memory_mb_ci95_low": 113.884765625,
      "delta_peak_gpu_memory_mb_ci95_high": 113.931640625,
      "delta_peak_gpu_memory_mb_percent": 0.7730657215242503
    }
  ]
}
