{
  "title": "FLUX.3 Action · RTX 5090",
  "date": "2026-09-28",
  "hardware": "NVIDIA GeForce RTX 5090 / SM120 / single GPU",
  "runtime": {
    "torch": "2.12.1+cu130",
    "triton": "3.7.1",
    "sageattention": "2.2.0"
  },
  "timing_scope": "warm predict_action_chunk including image/VAE conditioning and all four video/action denoising steps; excludes checkpoint load, Qwen text encoding, first compilation and warmups",
  "statistic": "median",
  "warmup": 5,
  "repeats_per_path_per_input": 30,
  "abba_groups": 15,
  "exact_requests": 120,
  "model_revisions": {
    "droid": {
      "repo_id": "black-forest-labs/flux-3-action-droid",
      "revision": "3d0887bdc7acee1686b19afac267125d519ff4f1"
    },
    "encoders": {
      "repo_id": "black-forest-labs/flux-3-action-base",
      "revision": "62878e2925e59b7a89ec14463ce89932624c490d"
    }
  },
  "config": {
    "sampler": "cosmos_unipc",
    "num_inference_steps": 4,
    "sampler_shift": 5.0,
    "guidance_scale": 4.0,
    "guidance_scale_action": 1.0,
    "inference_seed": 0,
    "chunk_size": 32,
    "action_dim": 8,
    "fps": 15.0,
    "canvas_hw": [
      544,
      736
    ]
  },
  "cases": [
    {
      "id": 0,
      "task": "Pour the contents of the yellow cup into the pink bowl",
      "observation_sha256": "722f9211bfaae0526eecda317cff6ba015ddfebe887bbf92450221b3838ba07f",
      "original_ms": 1396.405591396615,
      "sage_baseline_ms": 722.625227062963,
      "final_ms": 684.6034820191562,
      "total_speedup": 2.039729022817826,
      "latency_reduction_pct": 50.97387992163149,
      "lossless_speedup": 1.0555383459804588,
      "lossless_reduction_pct": 5.261613298271117,
      "quality": {
        "joint_mae_rad": 0.009817952103081293,
        "joint_max_abs_rad": 0.04699385166168213,
        "gripper_mae": 0.002278873696923256,
        "gripper_max_abs": 0.007245779037475586
      },
      "recorded_mae_degradation_rad": -0.008592513851389025,
      "all_quality_pass": true,
      "exact": true
    },
    {
      "id": 1,
      "task": "Pour the contents of the cup into the bowl",
      "observation_sha256": "aa63d0eeb61347dbfb2808c6dd832c8d00f4533759231edab6943fc4895007e7",
      "original_ms": 1408.6075149243698,
      "sage_baseline_ms": 724.1225234465674,
      "final_ms": 685.8274585101753,
      "total_speedup": 2.0538803126726526,
      "latency_reduction_pct": 51.31167118989859,
      "lossless_speedup": 1.055837754031576,
      "lossless_reduction_pct": 5.2884786339915335,
      "quality": {
        "joint_mae_rad": 0.015379066013598017,
        "joint_max_abs_rad": 0.0453333854675293,
        "gripper_mae": 0.002805577591061592,
        "gripper_max_abs": 0.010859251022338867
      },
      "recorded_mae_degradation_rad": -0.007676871393674185,
      "all_quality_pass": true,
      "exact": true
    }
  ],
  "history": [
    {
      "id": "original-v1",
      "label": "原始 BF16",
      "detail": "原始 PyTorch 路径",
      "kind": "baseline",
      "cases": [
        {
          "id": 0,
          "median_ms": 1396.405591396615,
          "seconds": [
            1.381904570851475,
            1.3823811470065266,
            1.3835190190002322,
            1.384274504845962,
            1.3872690789867193,
            1.3898364559281617,
            1.3898142939433455,
            1.3900694691110402,
            1.3914941870607436,
            1.3923688249196857,
            1.3926732998806983,
            1.393239316996187,
            1.39420221792534,
            1.3945417830254883,
            1.396204066928476,
            1.3966071158647537,
            1.3980501149781048,
            1.3979450820479542,
            1.4022394558414817,
            1.3986742771230638,
            1.399807906942442,
            1.3996493748854846,
            1.4003535558003932,
            1.4010304759722203,
            1.4015033501200378,
            1.401030495064333,
            1.4022455080412328,
            1.4056538008153439,
            1.4028159650042653,
            1.4028324319515377
          ]
        },
        {
          "id": 1,
          "median_ms": 1408.6075149243698,
          "seconds": [
            1.405564164975658,
            1.407197569962591,
            1.4047191909048706,
            1.4090528008528054,
            1.4029856929555535,
            1.405674556037411,
            1.4063881929032505,
            1.4073772400151938,
            1.4071643080096692,
            1.4072683060076088,
            1.4071268800180405,
            1.4094544500112534,
            1.4084794430527836,
            1.4084400490392,
            1.4084894431289285,
            1.4089460279792547,
            1.40849255095236,
            1.4126053720247,
            1.4111325331032276,
            1.4123897759709507,
            1.407148095080629,
            1.4098139212001115,
            1.4112632770556957,
            1.409819349180907,
            1.4094535470940173,
            1.4090065830387175,
            1.409109738888219,
            1.4087432380765676,
            1.4087224788963795,
            1.4097991108428687
          ]
        }
      ]
    },
    {
      "id": "optimized-v1",
      "label": "编译与执行优化",
      "detail": "编译 / 共享 CFG / 融合 UniPC",
      "kind": "execution",
      "cases": [
        {
          "id": 0,
          "median_ms": 1105.726829962805,
          "seconds": [
            1.0939354179427028,
            1.0950800881255418,
            1.095570659963414,
            1.0971358560491353,
            1.1001360081136227,
            1.1005526389926672,
            1.101168611086905,
            1.1015601688995957,
            1.1027813260443509,
            1.1028800609055907,
            1.104188428958878,
            1.1057327119633555,
            1.1049868501722813,
            1.1052685719914734,
            1.1054474129341543,
            1.1057209479622543,
            1.1067480009514838,
            1.1067145089618862,
            1.106929527129978,
            1.1069671148434281,
            1.1073070678394288,
            1.108028369024396,
            1.1088999500498176,
            1.109519376885146,
            1.109954441897571,
            1.1100162561051548,
            1.109944449039176,
            1.1108431718312204,
            1.1105548560153693,
            1.1105069699697196
          ]
        },
        {
          "id": 1,
          "median_ms": 1115.74458447285,
          "seconds": [
            1.112592165125534,
            1.1125846428330988,
            1.113525175023824,
            1.1139120161533356,
            1.1142586760688573,
            1.1139136960264295,
            1.1139425560832024,
            1.113907593069598,
            1.114756850991398,
            1.114386186003685,
            1.1148598240688443,
            1.114664763910696,
            1.1149549309629947,
            1.1150471579749137,
            1.115875778021291,
            1.1167280869558454,
            1.1158945939969271,
            1.1162209569010884,
            1.116134142037481,
            1.1156974020414054,
            1.1157917669042945,
            1.1158420711290091,
            1.1168441488407552,
            1.1167204279918224,
            1.116568208904937,
            1.116884137969464,
            1.1174702809657902,
            1.1172106221783906,
            1.1169800360221416,
            1.1172712589614093
          ]
        }
      ]
    },
    {
      "id": "rowwise-fp8-qkbf16-v3",
      "label": "混合精度投影",
      "detail": "保留 BF16 Q/K；其他选定投影 FP8",
      "kind": "precision",
      "cases": [
        {
          "id": 0,
          "median_ms": 777.765735052526,
          "seconds": [
            0.77187519101426,
            0.7719290389213711,
            0.7720805930439383,
            0.772506375098601,
            0.7726934109814465,
            0.7733639311045408,
            0.7783755280543119,
            0.7772313428577036,
            0.7740706850308925,
            0.7745594850275666,
            0.7735607258509845,
            0.7744390738662332,
            0.7762669958174229,
            0.7767264801077545,
            0.7775059519335628,
            0.7782417458947748,
            0.77767852996476,
            0.777852940140292,
            0.7778786199633032,
            0.7795840729959309,
            0.7801161350216717,
            0.7802730789408088,
            0.7787462549749762,
            0.7800344449933618,
            0.779274526052177,
            0.7795446920208633,
            0.7796726578380913,
            0.7794401370920241,
            0.7796120920684189,
            0.7798830349929631
          ]
        },
        {
          "id": 1,
          "median_ms": 784.4128310680389,
          "seconds": [
            0.7819512649439275,
            0.7858534760307521,
            0.7826038890052587,
            0.7843325608409941,
            0.7836677909363061,
            0.7824699589982629,
            0.783658463973552,
            0.7829139940440655,
            0.7832127250730991,
            0.7833697120659053,
            0.7866980000399053,
            0.7831063650082797,
            0.7833628219086677,
            0.7835651049390435,
            0.7846796528901905,
            0.7841983931139112,
            0.7844528339337558,
            0.7843951319810003,
            0.7842397268395871,
            0.7844305301550776,
            0.7853629479650408,
            0.7863788141403347,
            0.7865908748935908,
            0.7854677538853139,
            0.7849521010648459,
            0.7856644811108708,
            0.7855948850046843,
            0.7851674810517579,
            0.7856167980935425,
            0.7849965011700988
          ]
        }
      ]
    },
    {
      "id": "sage2-smooth-thread-v3-v8",
      "label": "SageAttention2",
      "detail": "INT8 QK / FP8 PV / K、V smooth",
      "kind": "precision",
      "cases": [
        {
          "id": 0,
          "median_ms": 711.7815556703135,
          "seconds": [
            0.7070603989996016,
            0.7084723520092666,
            0.7090139030478895,
            0.7103253770619631,
            0.7089668491389602,
            0.7082902421243489,
            0.709116626996547,
            0.708626707084477,
            0.7087468160316348,
            0.7092719411011785,
            0.7096243631094694,
            0.7107974591199309,
            0.711419414030388,
            0.7116676061414182,
            0.711255727801472,
            0.7157650021836162,
            0.7118955051992089,
            0.7131657260470092,
            0.7133188520092517,
            0.7130064801312983,
            0.7137848499696702,
            0.7135631800629199,
            0.713997723069042,
            0.7139290149789304,
            0.714752787956968,
            0.7143818931654096,
            0.7146923071704805,
            0.7146690050140023,
            0.7152265170589089,
            0.717653329949826
          ]
        },
        {
          "id": 1,
          "median_ms": 718.0130195338279,
          "seconds": [
            0.715362663147971,
            0.7150260568596423,
            0.7157730159815401,
            0.7159747669938952,
            0.7159235149156302,
            0.71656369487755,
            0.7166073801927269,
            0.7165130679495633,
            0.7167868439573795,
            0.7174572390504181,
            0.7176454800646752,
            0.7179852838162333,
            0.7185234068892896,
            0.7180336250457913,
            0.7182956060860306,
            0.7177435311023146,
            0.7184870031196624,
            0.7185444550123066,
            0.7183321320917457,
            0.718570914119482,
            0.7184261330403388,
            0.7179924140218645,
            0.7179150430019945,
            0.7181572718545794,
            0.718563802074641,
            0.71898004389368,
            0.7186842069495469,
            0.7187076800037175,
            0.7191525478847325,
            0.7190727861598134
          ]
        }
      ]
    }
  ],
  "thresholds": {
    "joint_mae_rad": 0.02,
    "joint_max_abs_rad": 0.05,
    "gripper_max_abs": 0.05,
    "recorded_joint_mae_degradation_rad": 0.02
  },
  "limitations": [
    "Overall speedup compares historical original BF16 with final ABBA median on the same GPU and workload. Historical stages ran in independent processes and are not isolated ablations.",
    "Only the last optimization batch is lossless relative to the pinned mixed-precision Sage2 baseline; the full route is not bit-identical to original BF16.",
    "Quality gate covers two fixed offline observations; no robot rollout, success rate or generalization claim.",
    "Launcher pinning is specific to the validated generated code, environment and workload; failed restart attempts remain disclosed in the report.",
    "This static website summarizes completed RTX 5090 experiments; it does not run inference or claim B200 VC-Attn results."
  ]
}