{
 "generated_at": "2026-09-11T03:57:42.895095+00:00",
 "spec": {
  "bf16_dense_tflops": 989.5,
  "tf32_dense_tflops": 494.5,
  "fp32_tflops": 67.0,
  "dram_tb_s": 4.8
 },
 "spec_source": "raw/h200_official_spec_review_20260910.html; dense = sparsity figure / 2",
 "same_gpu_empirical": {
  "compute": {
   "bf16": {
    "three_launch_max_tflops": [
     787.0565134516788,
     787.7060668090745,
     789.0809256432037
    ],
    "reference_tflops": 787.7060668090745,
    "selection": "Median of three per-launch maxima across measured square GEMM sizes"
   },
   "fp32_ieee": {
    "three_launch_max_tflops": [
     51.286787913396516,
     51.275693333347874,
     51.250146619931485
    ],
    "reference_tflops": 51.275693333347874,
    "selection": "Median of three per-launch maxima across measured square GEMM sizes"
   },
   "fp32_tf32_allowed": {
    "three_launch_max_tflops": [
     397.8479222355699,
     397.44295597962594,
     397.4245633398784
    ],
    "reference_tflops": 397.44295597962594,
    "selection": "Median of three per-launch maxima across measured square GEMM sizes"
   }
  },
  "bandwidth_GB_s": 4349.809747597248,
  "gpu_uuid": "GPU-46890c79-bf10-5b49-3d4a-a6b78f956122"
 },
 "points": [
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 4680,
    "N": 5120,
    "K": 144
   },
   "flops": 6900940800.0,
   "kernel_us": 25.312,
   "dram_bytes": 24836608.0,
   "tflops": 272.63514538558786,
   "intensity": 277.8535941783999,
   "input_sha256": "824a77bcaead8af0dacf908827027966a81c2d964d919eb4200bac0771b4a369",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_cd8cbc742f5c3f37_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 4680,
    "N": 5120,
    "K": 1536
   },
   "flops": 73610035200.0,
   "kernel_us": 95.52,
   "dram_bytes": 79865088.0,
   "tflops": 770.6243216080403,
   "intensity": 921.6797607485263,
   "input_sha256": "787b43c2c5ab957d3bd3c45047d646e6e407e506ea71b3249408d5d20bf00b0c",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_9296b219534b05e0_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 4680,
    "N": 5120,
    "K": 5120
   },
   "flops": 245366784000.0,
   "kernel_us": 292.4479999999999,
   "dram_bytes": 356987648.0,
   "tflops": 839.009957325747,
   "intensity": 687.3256970504481,
   "input_sha256": "4a5b492ee606c1a2c9fe4844a0c3125609fa599f4c027f56b4329c89d2879b7e",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_dit_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 3,
    "N": 5120,
    "K": 256
   },
   "flops": 7864320.0,
   "kernel_us": 4.672,
   "dram_bytes": 2649344.0,
   "tflops": 1.683287671232877,
   "intensity": 2.9684027442264953,
   "input_sha256": "ea1184eaad12555ac80bf5ee33de7aa2251926b289572cadadb4043884848587",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_10a11d3772070763_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 3,
    "N": 5120,
    "K": 5120
   },
   "flops": 157286400.0,
   "kernel_us": 17.344,
   "dram_bytes": 55696128.0,
   "tflops": 9.068634686346863,
   "intensity": 2.8240095972201154,
   "input_sha256": "a6cb8d5375b991c9c5d1a1e5674fa86eb61d10648b2ccf71bb7ec40c56f2ac1e",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_ca590b3a668a9ecf_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 3,
    "N": 30720,
    "K": 5120
   },
   "flops": 943718400.0,
   "kernel_us": 79.808,
   "dram_bytes": 319836416.0,
   "tflops": 11.824859663191658,
   "intensity": 2.950628361218255,
   "input_sha256": "d6a2e382ecf0fdad711df2cd1bcb239a004321ae666d26aa04920718cf49bf81",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_22805ed4eb786e66_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 1170,
    "N": 15360,
    "K": 5120
   },
   "flops": 184025088000.0,
   "kernel_us": 252.832,
   "dram_bytes": 296887040.0,
   "tflops": 727.8552082014935,
   "intensity": 619.8488421724303,
   "input_sha256": "4f1e9988bf86b3f91f274f9b8c3bf33bf42b39945181ce6f09fbce8e0132cb6f",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_fcebe966800b94b1_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 1170,
    "N": 5120,
    "K": 5120
   },
   "flops": 61341696000.0,
   "kernel_us": 87.104,
   "dram_bytes": 84771584.0,
   "tflops": 704.2351212343865,
   "intensity": 723.6115347331483,
   "input_sha256": "a64f7da6eb1fcc9779d40258d974eb529c2397aaba4df6e32363d7c5d4e73065",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_10a9e2fea034c6cf_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     16,
     3,
     60,
     104
    ],
    "weight": [
     16,
     16,
     1,
     1,
     1
    ],
    "output": [
     1,
     16,
     3,
     60,
     104
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 9584640.0,
   "kernel_us": 8.511999999999999,
   "dram_bytes": 2438400.0,
   "tflops": 1.126015037593985,
   "intensity": 3.930708661417323,
   "input_sha256": "32bcd9b4d8be26c876626649d4ff2e623424efe309222bcd7e80afc47310ded4",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_6638f430ac58a438_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     16,
     3,
     62,
     28
    ],
    "weight": [
     384,
     16,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     1,
     60,
     26
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 517570560.0,
   "kernel_us": 26.944,
   "dram_bytes": 4486912.0,
   "tflops": 19.20912114014252,
   "intensity": 115.35117247675015,
   "input_sha256": "2b9e80cee384305d79045df290612b04bcfe041bd67dc69a2bc096e45a6c5872",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_29ef146661136bf4_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     3,
     62,
     28
    ],
    "weight": [
     384,
     384,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     1,
     60,
     26
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 12421693440.0,
   "kernel_us": 122.848,
   "dram_bytes": 54811136.0,
   "tflops": 101.1143318572545,
   "intensity": 226.6271846655395,
   "input_sha256": "6309c12f24013edabd88edced7c2ce22672322d203206b12ba218c88696effb5",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_551fa1ff3095f40c_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     1,
     120,
     52
    ],
    "weight": [
     384,
     192,
     1,
     1,
     1
    ],
    "output": [
     1,
     384,
     1,
     120,
     52
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 920125440.0,
   "kernel_us": 21.152,
   "dram_bytes": 14720512.0,
   "tflops": 43.50063540090772,
   "intensity": 62.50634760530068,
   "input_sha256": "1e43418295ca6422f27afcf0b9940aa818fbb12fcf1013c12dfcf7a60f86e7d7",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_c16d11a7624e1a8c_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     3,
     122,
     54
    ],
    "weight": [
     384,
     192,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     1,
     120,
     52
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 24843386880.0,
   "kernel_us": 158.56,
   "dram_bytes": 76093440.0,
   "tflops": 156.68129969727548,
   "intensity": 326.4852644327816,
   "input_sha256": "3193a32ddc024e9d34eb5446f9909408dde38d963ebc5cb9654998ecf661a820",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_081e405bd1e07dce_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     3,
     122,
     54
    ],
    "weight": [
     384,
     384,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     1,
     120,
     52
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 49686773760.0,
   "kernel_us": 287.104,
   "dram_bytes": 160518144.0,
   "tflops": 173.06193490860457,
   "intensity": 309.53992191686444,
   "input_sha256": "c7f10fe8b8ca84229e785645c70c003456c3950c79cdd176694e5bc86c9247f4",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_8601bc3a88374d99_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     3,
     242,
     106
    ],
    "weight": [
     192,
     192,
     3,
     3,
     3
    ],
    "output": [
     1,
     192,
     1,
     240,
     104
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 49686773760.0,
   "kernel_us": 230.72,
   "dram_bytes": 219308032.0,
   "tflops": 215.3552954230236,
   "intensity": 226.56157782675282,
   "input_sha256": "0cc0a21ce474602c53c3e5621e64598ed7be5a2c16603fc968441df41de71330",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_99475e040d65d084_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     96,
     3,
     482,
     210
    ],
    "weight": [
     96,
     96,
     3,
     3,
     3
    ],
    "output": [
     1,
     96,
     1,
     480,
     208
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 49686773760.0,
   "kernel_us": 466.624,
   "dram_bytes": 473185792.0,
   "tflops": 106.48139349883417,
   "intensity": 105.00478797131761,
   "input_sha256": "1e47965836d719f164296b5c8d52cac8c22a10d9ff8c5f4fdfa48498909f8856",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_vae_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "output",
   "op": "aten.to.dtype",
   "dtype": "torch.float32",
   "shape_parameters": {},
   "flops": null,
   "kernel_us": 29.824,
   "dram_bytes": 48509696.0,
   "tflops": null,
   "intensity": null,
   "input_sha256": "897dd2bd3ad5fe7258ee5c3ae9cfd93da4e5e94644902267f0b3d8bd156f63b0",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_output_rank0_f18eb91117dda1c8_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "no_nominal_mac_flops_or_dram_bytes (convert/copy: bandwidth-only, N/A for MFU)",
   "family": "convert/other"
  },
  {
   "source": "operator_work_traffic.json",
   "resolution": "480x832",
   "stage": "output",
   "op": "aten.to.dtype",
   "dtype": "torch.float32",
   "shape_parameters": {},
   "flops": null,
   "kernel_us": 39.583999999999996,
   "dram_bytes": 69139200.0,
   "tflops": null,
   "intensity": null,
   "input_sha256": "a9aab61b63e7cab483e81582c9af41d89d525814ab14de753e8a7b16de6c4f1b",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_output_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "no_nominal_mac_flops_or_dram_bytes (convert/copy: bandwidth-only, N/A for MFU)",
   "family": "convert/other"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 1170,
    "N": 13824,
    "K": 5120
   },
   "flops": 165622579200.0,
   "kernel_us": 215.52,
   "dram_bytes": 229586944.0,
   "tflops": 768.4789309576837,
   "intensity": 721.3937182769417,
   "input_sha256": "b8c205d5140f4a6d1aabaa7d0946d8eb3aa5bdc6007bff3e8efd26fcf1740c87",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_80ac9ebe75667de4_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 1170,
    "N": 5120,
    "K": 13824
   },
   "flops": 165622579200.0,
   "kernel_us": 220.60799999999998,
   "dram_bytes": 216986624.0,
   "tflops": 750.755091383812,
   "intensity": 763.2847414594552,
   "input_sha256": "d89811d1f0f1ab9382656380f7ff354f19e521d815cd71b6e39b7d2a72def4ec",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_ef1fe591ff44862b_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 1170,
    "N": 5120,
    "K": 5120
   },
   "flops": 61341696000.0,
   "kernel_us": 85.696,
   "dram_bytes": 84834304.0,
   "tflops": 715.8058252427185,
   "intensity": 723.0765516741907,
   "input_sha256": "7de91ad08752fccf4380732c881403df8972bc20e43941ce68d7ee08b8d5d769",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_6d01e6f8a09a63ac_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     6,
     242,
     106
    ],
    "weight": [
     192,
     192,
     3,
     3,
     3
    ],
    "output": [
     1,
     192,
     4,
     240,
     104
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 198747095040.0,
   "kernel_us": 872.4159999999999,
   "dram_bytes": 798215168.0,
   "tflops": 227.81229945347175,
   "intensity": 248.9893740530874,
   "input_sha256": "7ea1828a641121c7e50c8fec13aa00d78a4f4b0f419741b8b6e74c3fd8f2b955",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_41fea95495c28b56_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     96,
     6,
     482,
     210
    ],
    "weight": [
     96,
     96,
     3,
     3,
     3
    ],
    "output": [
     1,
     96,
     4,
     480,
     208
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 198747095040.0,
   "kernel_us": 1672.16,
   "dram_bytes": 1630753024.0,
   "tflops": 118.85650598028897,
   "intensity": 121.87442985848482,
   "input_sha256": "3da5719db48ceca956fb363d6dbfa423964bbde96b9ce8b8dd9399261d9405da",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_74aff8f080d200c8_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     4,
     122,
     54
    ],
    "weight": [
     384,
     384,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     2,
     120,
     52
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 99373547520.0,
   "kernel_us": 486.23999999999995,
   "dram_bytes": 250029568.0,
   "tflops": 204.37139585389932,
   "intensity": 397.44718320674775,
   "input_sha256": "538cab189b8721f3989e60d331a5e8f3a0b3ea6e3f542b4f5be45a3f1807f370",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_7fc47a1ee259ec1d_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     4,
     122,
     54
    ],
    "weight": [
     384,
     192,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     2,
     120,
     52
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 49686773760.0,
   "kernel_us": 253.02399999999997,
   "dram_bytes": 123802112.0,
   "tflops": 196.3717819653472,
   "intensity": 401.340275681242,
   "input_sha256": "6fa08b3e20b8b3a0f9f34750ff5afbc4a19f18369bd70ccbf0ed0d1071ca5f08",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_e49e1cfe69d26b07_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     4,
     120,
     52
    ],
    "weight": [
     768,
     384,
     3,
     1,
     1
    ],
    "output": [
     1,
     768,
     2,
     120,
     52
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 22083010560.0,
   "kernel_us": 168.224,
   "dram_bytes": 259840512.0,
   "tflops": 131.27146281148944,
   "intensity": 84.986788203373,
   "input_sha256": "3547fbf5eb18cf7da7abff35f3a785e4dd1cfc51a243c23698a162ff25086677",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_c87dc674a51efcf2_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     96,
     6,
     482,
     210
    ],
    "weight": [
     3,
     96,
     3,
     3,
     3
    ],
    "output": [
     1,
     3,
     4,
     480,
     208
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 6210846720.0,
   "kernel_us": 704.128,
   "dram_bytes": 988795904.0,
   "tflops": 8.82062170514452,
   "intensity": 6.281222135806906,
   "input_sha256": "3558a52a268a76189298b60a739c4bc3a01099db809b69c10633a6c964ffd111",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_27b8a5265a151590_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 4680,
    "N": 64,
    "K": 5120
   },
   "flops": 3067084800.0,
   "kernel_us": 17.6,
   "dram_bytes": 51621120.0,
   "tflops": 174.26618181818182,
   "intensity": 59.41530908279402,
   "input_sha256": "34457f152c46e24a2e317b09c8ddc97428430b2e02f99cc2ddeb74c8c157ed9e",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_301783745952cfe0_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     3,
     60,
     26
    ],
    "weight": [
     768,
     384,
     3,
     1,
     1
    ],
    "output": [
     1,
     768,
     1,
     60,
     26
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 2760376320.0,
   "kernel_us": 51.711999999999996,
   "dram_bytes": 31157504.0,
   "tflops": 53.379801980198025,
   "intensity": 88.59426993895275,
   "input_sha256": "03221b37d95f7c65a91055eae771f796fbf8e0d7758ce6840701ea0c066358ea",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_d35bb120bedeb7d0_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     2,
     120,
     52
    ],
    "weight": [
     384,
     192,
     1,
     1,
     1
    ],
    "output": [
     1,
     384,
     2,
     120,
     52
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 1840250880.0,
   "kernel_us": 43.936,
   "dram_bytes": 41000704.0,
   "tflops": 41.88480699198834,
   "intensity": 44.883397124107915,
   "input_sha256": "5da1ee3cce399169645cbf03383f0c7594f9d9b878d0bc38e2e663bb1e06edd8",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_6a92f0b3792a465a_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_work_traffic_v1.json",
   "resolution": "480x832",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     96,
     3,
     482,
     210
    ],
    "weight": [
     3,
     96,
     3,
     3,
     3
    ],
    "output": [
     1,
     3,
     1,
     480,
     208
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 1552711680.0,
   "kernel_us": 284.0,
   "dram_bytes": 340329472.0,
   "tflops": 5.467294647887323,
   "intensity": 4.562377953561424,
   "input_sha256": "480f421d034402c9b8b81d025dce79d2061ec150ba4509115415f8dd5d3b37b5",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing_ddb198f9c3730e1d_v1",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 2640,
    "N": 5120,
    "K": 5120
   },
   "flops": 138412032000.0,
   "kernel_us": 185.344,
   "dram_bytes": 185954560.0,
   "tflops": 746.7845303867404,
   "intensity": 744.3325509199667,
   "input_sha256": "8e60010c4ab22e3ad90412ff00e301e28227cf54db05320e456d91009e843787",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_255d8b6427b0b5a7_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 2640,
    "N": 15360,
    "K": 5120
   },
   "flops": 415236096000.0,
   "kernel_us": 518.112,
   "dram_bytes": 527754752.0,
   "tflops": 801.4408004446916,
   "intensity": 786.7974554969047,
   "input_sha256": "5486c3201c4910ee4de3cbdc2b72d468ede3117a03f1e731537cf900e60ecc9d",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_db47f7eb1f64a850_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 2640,
    "N": 5120,
    "K": 13824
   },
   "flops": 373712486400.0,
   "kernel_us": 472.896,
   "dram_bytes": 458745344.0,
   "tflops": 790.263580998782,
   "intensity": 814.6403909878156,
   "input_sha256": "75bf41dc2a7a283f82bdcd0635998b9bc8c93a30b0ebb2b209ce44fd3b9e360c",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_1b9e93116d235408_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 2640,
    "N": 13824,
    "K": 5120
   },
   "flops": 373712486400.0,
   "kernel_us": 451.03999999999996,
   "dram_bytes": 453526016.0,
   "tflops": 828.5573040085136,
   "intensity": 824.0155431347956,
   "input_sha256": "c08dc8cdb79c43328ef6849ee7e60049c20bbb243f50823bd9a80415eb91dc94",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_9ba227caa2b6dfb2_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 2640,
    "N": 5120,
    "K": 5120
   },
   "flops": 138412032000.0,
   "kernel_us": 183.52,
   "dram_bytes": 184763904.0,
   "tflops": 754.2068003487359,
   "intensity": 749.1291805568256,
   "input_sha256": "984423718cf2b4425d4e6c3c3cc45ee71db6a1b3d36906b61da17257f41a950c",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_697ad043e70ad841_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     96,
     6,
     706,
     322
    ],
    "weight": [
     96,
     96,
     3,
     3,
     3
    ],
    "output": [
     1,
     96,
     4,
     704,
     320
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 448454983680.0,
   "kernel_us": 3677.9519999999998,
   "dram_bytes": 3759372480.0,
   "tflops": 121.93062434746294,
   "intensity": 119.28985118282294,
   "input_sha256": "361110b0df425093397551c589c4c8f8e119f195c1b7c8ce7fe463e598338f6e",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_0551e0aa26bb7201_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     6,
     354,
     162
    ],
    "weight": [
     192,
     192,
     3,
     3,
     3
    ],
    "output": [
     1,
     192,
     4,
     352,
     160
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 448454983680.0,
   "kernel_us": 2130.8480000000004,
   "dram_bytes": 1939374848.0,
   "tflops": 210.4584577032242,
   "intensity": 231.2368772558197,
   "input_sha256": "836c7ee7ca56334d7997004c7d80dc3a7742322bc7befd6ad4a0f29a4ec8ad0f",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_1b34e3d0553bbca4_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 10560,
    "N": 5120,
    "K": 5120
   },
   "flops": 553648128000.0,
   "kernel_us": 650.3359999999999,
   "dram_bytes": 678032640.0,
   "tflops": 851.3262805688138,
   "intensity": 816.5508492334528,
   "input_sha256": "70fed7084bfb383625de9935f9cc6ad7e398ccdf921a0e71276f5cf1bdbdcb0d",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_606cff6747219bac_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     4,
     178,
     82
    ],
    "weight": [
     384,
     384,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     2,
     176,
     80
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 224227491840.0,
   "kernel_us": 1129.5360000000003,
   "dram_bytes": 625349376.0,
   "tflops": 198.5129219785823,
   "intensity": 358.5635493462138,
   "input_sha256": "ae178ca2bbf227553643572af23755eb3ed51289f32398e11a61f88fb81ea3a1",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_ac3438e4ac0a9485_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     3,
     90,
     42
    ],
    "weight": [
     384,
     384,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     1,
     88,
     40
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 28028436480.0,
   "kernel_us": 180.32,
   "dram_bytes": 83944192.0,
   "tflops": 155.43720319432123,
   "intensity": 333.89369546853226,
   "input_sha256": "fd2ac936cea7b4d5d60612a6045f5c612e27536de2424aa111e4bc689fc1e795",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_a50d349fe3a8545f_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 10560,
    "N": 5120,
    "K": 1536
   },
   "flops": 166094438400.0,
   "kernel_us": 208.928,
   "dram_bytes": 169742848.0,
   "tflops": 794.9841017001072,
   "intensity": 978.5062543548227,
   "input_sha256": "94f48ff6a98bf87d9ac8185c5db13ca77c238c34547a0c86dd5cac7b7d5b326f",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_4cfc56167ab9dd35_v1",
   "scheduler_note": [
    {
     "gpu_id": 0,
     "pid": 2019660,
     "user": "unknown",
     "kind": "foreign",
     "holder": "zhoutaichang",
     "holder_account": "zhoutaichang",
     "command": "[No data]",
     "memory_mb": 2266,
     "first_seen": "2026-09-10T05:20:53.219014752Z",
     "last_seen": "2026-09-10T05:20:54.100960726Z",
     "sightings": 3,
     "warn_count": 0,
     "last_warned": null,
     "last_signal": null,
     "end_time": "2026-09-10T05:20:55.549497427Z",
     "resolution": "process exited"
    }
   ],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     4,
     178,
     82
    ],
    "weight": [
     384,
     192,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     2,
     176,
     80
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 112113745920.0,
   "kernel_us": 603.9679999999998,
   "dram_bytes": 325180416.0,
   "tflops": 185.62861926459684,
   "intensity": 344.7739790086252,
   "input_sha256": "146fb336db70349e18bee5b6332d0ab63940d20d95e88e53d065f0d87a4673e8",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_053cda5cb0c0dcf6_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     4,
     176,
     80
    ],
    "weight": [
     768,
     384,
     3,
     1,
     1
    ],
    "output": [
     1,
     768,
     2,
     176,
     80
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 49828331520.0,
   "kernel_us": 299.13599999999997,
   "dram_bytes": 662414848.0,
   "tflops": 166.57417201540437,
   "intensity": 75.22224429365448,
   "input_sha256": "05ade70ee900de950adaf23171001399a4d6cc88017bf23770bc74ef28c01d3b",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_4273bf5ee693569e_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 10560,
    "N": 5120,
    "K": 144
   },
   "flops": 15571353600.0,
   "kernel_us": 53.472,
   "dram_bytes": 86605056.0,
   "tflops": 291.2057450628366,
   "intensity": 179.79728111947645,
   "input_sha256": "31fe6110c4fdd22ab61ff0c539192126b4bc8da9a7355f9916c23de05f620d74",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_3d9ee58b5b1e2e00_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     96,
     3,
     706,
     322
    ],
    "weight": [
     96,
     96,
     3,
     3,
     3
    ],
    "output": [
     1,
     96,
     1,
     704,
     320
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 112113745920.0,
   "kernel_us": 1006.016,
   "dram_bytes": 1154930176.0,
   "tflops": 111.44330300909726,
   "intensity": 97.07404676904035,
   "input_sha256": "85c0ba2d39aa4aa6ffbabaa237b9fa5d7721eefd855682c8cf482c661a5943d9",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_1ae7dbe8166a0b6b_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     3,
     354,
     162
    ],
    "weight": [
     192,
     192,
     3,
     3,
     3
    ],
    "output": [
     1,
     192,
     1,
     352,
     160
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 112113745920.0,
   "kernel_us": 562.496,
   "dram_bytes": 554916608.0,
   "tflops": 199.31474342928658,
   "intensity": 202.0371066637818,
   "input_sha256": "a9625d3e84c044a16224205940b2918538dcbebc7b9e36a51af82b718a774552",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_996e506f74aef428_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     3,
     178,
     82
    ],
    "weight": [
     384,
     384,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     1,
     176,
     80
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 112113745920.0,
   "kernel_us": 585.4079999999999,
   "dram_bytes": 387858688.0,
   "tflops": 191.5138602820597,
   "intensity": 289.05823019748885,
   "input_sha256": "9b2cc0033aa13a80bc2dd24c544a6b1097e94f10f8c1f3748cbca111210d5fa1",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_e70c8df9d1c257cc_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     96,
     6,
     706,
     322
    ],
    "weight": [
     3,
     96,
     3,
     3,
     3
    ],
    "output": [
     1,
     3,
     4,
     704,
     320
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 14014218240.0,
   "kernel_us": 1515.04,
   "dram_bytes": 2271778896.0,
   "tflops": 9.250064843172456,
   "intensity": 6.168830190594393,
   "input_sha256": "c09536053b980c5e962fadc017011cff0cd6c94409f5c2270984dfec83c493a3",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_0e33e5013ffbea1c_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "dit",
   "op": "aten.linear.default",
   "dtype": "torch.bfloat16",
   "shape_parameters": {
    "M": 10560,
    "N": 64,
    "K": 5120
   },
   "flops": 6920601600.0,
   "kernel_us": 31.135999999999996,
   "dram_bytes": 111464960.0,
   "tflops": 222.27009249743065,
   "intensity": 62.08768746698514,
   "input_sha256": "26ebef86e4633eac87d2d256ccdab318b493bdf4c53c9c855e742e4fe4f4f537",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_5a8bc19ef82213ac_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "GEMM (bf16)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     384,
     3,
     88,
     40
    ],
    "weight": [
     768,
     384,
     3,
     1,
     1
    ],
    "output": [
     1,
     768,
     1,
     88,
     40
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 6228541440.0,
   "kernel_us": 74.752,
   "dram_bytes": 66484224.0,
   "tflops": 83.32273972602741,
   "intensity": 93.68450235652897,
   "input_sha256": "9dd2254d46cabc12367da79ca54c5220c508380e8b7fd1b985050405f5f81ee3",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_fd5ad4e08e7ff9d8_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     2,
     176,
     80
    ],
    "weight": [
     384,
     192,
     1,
     1,
     1
    ],
    "output": [
     1,
     384,
     2,
     176,
     80
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 4152360960.0,
   "kernel_us": 80.96,
   "dram_bytes": 128120832.0,
   "tflops": 51.28904347826087,
   "intensity": 32.4097252193929,
   "input_sha256": "bccd4b59806daff52b45f9c87ba32579441de0ed0adb3dd99356b7ebbcc1f4c7",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_9a2a65226b52659e_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     3,
     178,
     82
    ],
    "weight": [
     384,
     192,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     1,
     176,
     80
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 56056872960.0,
   "kernel_us": 306.40000000000003,
   "dram_bytes": 195641344.0,
   "tflops": 182.95324073107048,
   "intensity": 286.52876643497194,
   "input_sha256": "d56deba2cf2a7667416ca96682c1bcff892969b687097e1b3ea3921ac1a1d6ac",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_740d67537eaca4e1_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     16,
     3,
     90,
     42
    ],
    "weight": [
     384,
     16,
     3,
     3,
     3
    ],
    "output": [
     1,
     384,
     1,
     88,
     40
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 1167851520.0,
   "kernel_us": 42.367999999999995,
   "dram_bytes": 8282368.0,
   "tflops": 27.564471299093658,
   "intensity": 141.00454362810248,
   "input_sha256": "f8fb1466dc2a1a6e997b5eff656e0d892480f377234df948eb6b06009d0b88b2",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_9e24c4a75ab83dea_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     96,
     3,
     706,
     322
    ],
    "weight": [
     3,
     96,
     3,
     3,
     3
    ],
    "output": [
     1,
     3,
     1,
     704,
     320
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 3503554560.0,
   "kernel_us": 502.5599999999999,
   "dram_bytes": 794126080.0,
   "tflops": 6.971415472779371,
   "intensity": 4.411836669562597,
   "input_sha256": "a2673addbb1775eb79e779d7635c4fbcf9dd34b0d4953bdf81e740bec14aa1ef",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_aa00674f1a72afcd_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     192,
     1,
     176,
     80
    ],
    "weight": [
     384,
     192,
     1,
     1,
     1
    ],
    "output": [
     1,
     384,
     1,
     176,
     80
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 2076180480.0,
   "kernel_us": 36.704,
   "dram_bytes": 36730624.0,
   "tflops": 56.56551002615519,
   "intensity": 56.524508813136414,
   "input_sha256": "85c1d12d0656fbd621625fe39cf33eef4823fe7d3338e672b380cbbe7e08c3bc",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_2ea2cf3e2f86821c_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "vae",
   "op": "aten.conv3d.default",
   "dtype": "torch.float32",
   "shape_parameters": {
    "input": [
     1,
     16,
     3,
     88,
     160
    ],
    "weight": [
     16,
     16,
     1,
     1,
     1
    ],
    "output": [
     1,
     16,
     3,
     88,
     160
    ],
    "stride": [
     1,
     1,
     1
    ],
    "padding": [
     0,
     0,
     0
    ],
    "dilation": [
     1,
     1,
     1
    ],
    "groups": 1
   },
   "flops": 21626880.0,
   "kernel_us": 10.495999999999999,
   "dram_bytes": 5448448.0,
   "tflops": 2.0604878048780493,
   "intensity": 3.9693652210684585,
   "input_sha256": "28dd946d4a53d1263bf5d552d2aacaa22bc7162bea85950320f6d0e02d5750e7",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_b611cd0d7be4ea16_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "conv3d (fp32/tf32)"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "output",
   "op": "aten.to.dtype",
   "dtype": "torch.float32",
   "shape_parameters": {},
   "flops": null,
   "kernel_us": 64.63999999999999,
   "dram_bytes": 116176384.0,
   "tflops": null,
   "intensity": null,
   "input_sha256": "80d26723a59d26f80a9dfde0de9bf8e28bb650baf110fa77f33200b5bb9e4198",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_59ef4c5b540ee17f_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "no_nominal_mac_flops_or_dram_bytes (convert/copy: bandwidth-only, N/A for MFU)",
   "family": "convert/other"
  },
  {
   "source": "missing_operator_704_work_traffic_v1.json",
   "resolution": "704x1280",
   "stage": "output",
   "op": "aten.to.dtype",
   "dtype": "torch.float32",
   "shape_parameters": {},
   "flops": null,
   "kernel_us": 84.224,
   "dram_bytes": 157341184.0,
   "tflops": null,
   "intensity": null,
   "input_sha256": "ff3ca616760a5d842212457937cd00ab117a7e2049cc892ea9847400317b91f6",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_missing704_99c188aeb0ac98c3_v1",
   "scheduler_note": [],
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "no_nominal_mac_flops_or_dram_bytes (convert/copy: bandwidth-only, N/A for MFU)",
   "family": "convert/other"
  },
  {
   "source": "attention_hardware_480_short_audit_v1.json",
   "resolution": "480x832",
   "stage": "attention",
   "op": "flash_attn_v2",
   "dtype": "torch.bfloat16",
   "shape_parameters": "{\"case\": \"480_short\", \"role\": \"self\", \"valid_keys\": 4680}",
   "flops": 112140288000.0,
   "kernel_us": 332.48,
   "dram_bytes": 41279744.0,
   "tflops": 337.28431183830605,
   "intensity": 2716.5935912780856,
   "input_sha256": "e86fd1d5515ffb6e8e8ba1c4a38e0b54b23bd7d8d4d931434a89f116a2186b96",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_hardware_extension_v1/480_short/fa2",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "attention (bf16)"
  },
  {
   "source": "attention_hardware_480_short_audit_v1.json",
   "resolution": "480x832",
   "stage": "attention",
   "op": "flash_attn_v3",
   "dtype": "torch.bfloat16",
   "shape_parameters": "{\"case\": \"480_short\", \"role\": \"self\", \"valid_keys\": 4680}",
   "flops": 112140288000.0,
   "kernel_us": 179.744,
   "dram_bytes": 41229824.0,
   "tflops": 623.8889086701085,
   "intensity": 2719.882772237883,
   "input_sha256": "e86fd1d5515ffb6e8e8ba1c4a38e0b54b23bd7d8d4d931434a89f116a2186b96",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_hardware_extension_v1/480_short/fa3",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "attention (bf16)"
  },
  {
   "source": "attention_hardware_704_short_audit_v1.json",
   "resolution": "704x1280",
   "stage": "attention",
   "op": "flash_attn_v2",
   "dtype": "torch.bfloat16",
   "shape_parameters": "{\"case\": \"704_short\", \"role\": \"self\", \"valid_keys\": 10560}",
   "flops": 570949632000.0,
   "kernel_us": 1.594784,
   "dram_bytes": 110585088.0,
   "tflops": 358010.634669021,
   "intensity": 5162.989353501261,
   "input_sha256": "468353cb47a9d93b52d299d1edd226524978db730a64597becca38700ad6c050",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_hardware_extension_v1/704_short/fa2",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "attention (bf16)"
  },
  {
   "source": "attention_hardware_704_short_audit_v1.json",
   "resolution": "704x1280",
   "stage": "attention",
   "op": "flash_attn_v3",
   "dtype": "torch.bfloat16",
   "shape_parameters": "{\"case\": \"704_short\", \"role\": \"self\", \"valid_keys\": 10560}",
   "flops": 570949632000.0,
   "kernel_us": 821.184,
   "dram_bytes": 102783231.99999999,
   "tflops": 695.276128127192,
   "intensity": 5554.890821102026,
   "input_sha256": "468353cb47a9d93b52d299d1edd226524978db730a64597becca38700ad6c050",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_hardware_extension_v1/704_short/fa3",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "attention (bf16)"
  },
  {
   "source": "attention_hardware_704_long_audit_v1.json",
   "resolution": "704x1280",
   "stage": "attention",
   "op": "flash_attn_v2",
   "dtype": "torch.bfloat16",
   "shape_parameters": "{\"case\": \"704_long\", \"role\": \"self\", \"valid_keys\": 63360}",
   "flops": 3425697792000.0,
   "kernel_us": 9.490592,
   "dram_bytes": 555490304.0,
   "tflops": 360957.22922237095,
   "intensity": 6166.980354710206,
   "input_sha256": "8ab8d8e91a6d5757a64db4da0bf7d6b02a4ab0fb6c1bc2ca2c2524f9a1e2f113",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_hardware_extension_v1/704_long/fa2",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "attention (bf16)"
  },
  {
   "source": "attention_hardware_704_long_audit_v1.json",
   "resolution": "704x1280",
   "stage": "attention",
   "op": "flash_attn_v3",
   "dtype": "torch.bfloat16",
   "shape_parameters": "{\"case\": \"704_long\", \"role\": \"self\", \"valid_keys\": 63360}",
   "flops": 3425697792000.0,
   "kernel_us": 4.683872,
   "dram_bytes": 569264640.0,
   "tflops": 731381.5988139727,
   "intensity": 6017.759669738138,
   "input_sha256": "8ab8d8e91a6d5757a64db4da0bf7d6b02a4ab0fb6c1bc2ca2c2524f9a1e2f113",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_hardware_extension_v1/704_long/fa3",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "attention (bf16)"
  },
  {
   "source": "attention_hardware_704_cross_audit_v1.json",
   "resolution": "704x1280",
   "stage": "attention",
   "op": "flash_attn_v2",
   "dtype": "torch.bfloat16",
   "shape_parameters": "{\"case\": \"704_cross\", \"role\": \"cross\", \"valid_keys\": 512}",
   "flops": 27682406400.0,
   "kernel_us": 92.928,
   "dram_bytes": 39737600.0,
   "tflops": 297.89090909090913,
   "intensity": 696.6300531486552,
   "input_sha256": "ee5e993223e93b3cc62c90d337f72eff3586309fffbc0dc2984b73efc41077c4",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_hardware_extension_v1/704_cross/fa2",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "attention (bf16)"
  },
  {
   "source": "attention_hardware_704_cross_audit_v1.json",
   "resolution": "704x1280",
   "stage": "attention",
   "op": "flash_attn_v3",
   "dtype": "torch.bfloat16",
   "shape_parameters": "{\"case\": \"704_cross\", \"role\": \"cross\", \"valid_keys\": 512}",
   "flops": 27682406400.0,
   "kernel_us": 80.41600000000001,
   "dram_bytes": 39620864.00000001,
   "tflops": 344.24003183446075,
   "intensity": 698.6825527075835,
   "input_sha256": "ee5e993223e93b3cc62c90d337f72eff3586309fffbc0dc2984b73efc41077c4",
   "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_hardware_extension_v1/704_cross/fa3",
   "scheduler_note": null,
   "l2_bytes": null,
   "l2_status": "not_collected",
   "roofline_status": "plotted",
   "family": "attention (bf16)"
  }
 ],
 "families": {
  "GEMM (bf16)": 21,
  "conv3d (fp32/tf32)": 34,
  "convert/other": 4,
  "attention (bf16)": 8
 },
 "l2_layer": "collected: analysis/ncu_l2_stall_audit_v1.json (lts__t_bytes, l1tex__t_bytes, tensor pipe, warp stalls) for 59 operators + 6 attention inputs; right panel",
 "scope": "Isolated replays on assigned GPUs (identities in each ncu_output/identity.json); nominal MAC FLOPs include materialized padding; time is summed profiled kernel duration; not integrated runtime, not model MFU. Points above the same-GPU empirical line are on other GPUs or measure different kernels; spec ceilings are documentation values, not measured on this node."
}