{
  "status": "derived_from_audited_measurements",
  "generated_at": "2026-09-10T02:19:06.576397+00:00",
  "rows": [
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/dit_rank0_cd8cbc742f5c3f37.pt",
      "input_sha256": "824a77bcaead8af0dacf908827027966a81c2d964d919eb4200bac0771b4a369",
      "stage": "dit",
      "op": "aten.linear.default",
      "input_tensor": {
        "shape": [
          4680,
          144
        ],
        "stride": [
          144,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 1347840
      },
      "output_tensor": {
        "shape": [
          4680,
          5120
        ],
        "stride": [
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 47923200
      },
      "shape_parameters": {
        "M": 4680,
        "N": 5120,
        "K": 144
      },
      "mac_flops": 6900940800,
      "bias_additions": 23961600,
      "formula": "2*M*N*K (multiply and add each count one); bias additions listed separately",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 25.312,
      "dram_read_bytes": 2865920.0,
      "dram_write_bytes": 21970688.0,
      "dram_total_bytes": 24836608.0,
      "nominal_mac_tflops_per_s": 272.63514538558786,
      "nominal_mac_flops_per_dram_byte": 277.8535941783999,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_cd8cbc742f5c3f37_v1",
      "ncu_metrics_sha256": "48444dbcd0b825c247c4b7094c7b2e678413c213928495d57a944b3cf18dbc82",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/dit_rank0_9296b219534b05e0.pt",
      "input_sha256": "787b43c2c5ab957d3bd3c45047d646e6e407e506ea71b3249408d5d20bf00b0c",
      "stage": "dit",
      "op": "aten.linear.default",
      "input_tensor": {
        "shape": [
          1,
          4680,
          1536
        ],
        "stride": [
          7188480,
          1536,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 14376960
      },
      "output_tensor": {
        "shape": [
          1,
          4680,
          5120
        ],
        "stride": [
          23961600,
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 47923200
      },
      "shape_parameters": {
        "M": 4680,
        "N": 5120,
        "K": 1536
      },
      "mac_flops": 73610035200,
      "bias_additions": 23961600,
      "formula": "2*M*N*K (multiply and add each count one); bias additions listed separately",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 95.52,
      "dram_read_bytes": 46278400.0,
      "dram_write_bytes": 33586688.0,
      "dram_total_bytes": 79865088.0,
      "nominal_mac_tflops_per_s": 770.6243216080403,
      "nominal_mac_flops_per_dram_byte": 921.6797607485263,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_9296b219534b05e0_v1",
      "ncu_metrics_sha256": "14e3ba7326a14e47cdd8e4fb9bea081d2c716492736a50e51c69198ce0ac9403",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/dit_rank0_f56603026672560c.pt",
      "input_sha256": "4a5b492ee606c1a2c9fe4844a0c3125609fa599f4c027f56b4329c89d2879b7e",
      "stage": "dit",
      "op": "aten.linear.default",
      "input_tensor": {
        "shape": [
          1,
          4680,
          5120
        ],
        "stride": [
          23961600,
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 47923200
      },
      "output_tensor": {
        "shape": [
          1,
          4680,
          5120
        ],
        "stride": [
          23961600,
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 47923200
      },
      "shape_parameters": {
        "M": 4680,
        "N": 5120,
        "K": 5120
      },
      "mac_flops": 245366784000,
      "bias_additions": 23961600,
      "formula": "2*M*N*K (multiply and add each count one); bias additions listed separately",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 292.4479999999999,
      "dram_read_bytes": 314121984.0,
      "dram_write_bytes": 42865664.0,
      "dram_total_bytes": 356987648.0,
      "nominal_mac_tflops_per_s": 839.009957325747,
      "nominal_mac_flops_per_dram_byte": 687.3256970504481,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_dit_v1",
      "ncu_metrics_sha256": "e3b23b6928422a2d427cc67fa15c47e2fc6a548d66f2a4bf7b6d2626482cecaa",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/dit_rank0_10a11d3772070763.pt",
      "input_sha256": "ea1184eaad12555ac80bf5ee33de7aa2251926b289572cadadb4043884848587",
      "stage": "dit",
      "op": "aten.linear.default",
      "input_tensor": {
        "shape": [
          3,
          256
        ],
        "stride": [
          256,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 1536
      },
      "output_tensor": {
        "shape": [
          3,
          5120
        ],
        "stride": [
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 30720
      },
      "shape_parameters": {
        "M": 3,
        "N": 5120,
        "K": 256
      },
      "mac_flops": 7864320,
      "bias_additions": 15360,
      "formula": "2*M*N*K (multiply and add each count one); bias additions listed separately",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 4.672,
      "dram_read_bytes": 2649344.0,
      "dram_write_bytes": 0.0,
      "dram_total_bytes": 2649344.0,
      "nominal_mac_tflops_per_s": 1.683287671232877,
      "nominal_mac_flops_per_dram_byte": 2.9684027442264953,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_10a11d3772070763_v1",
      "ncu_metrics_sha256": "cb20c331c27111df939e9eed97b745c16f21a1653c79efb6a0be063ed79a11ed",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/dit_rank0_ca590b3a668a9ecf.pt",
      "input_sha256": "a6cb8d5375b991c9c5d1a1e5674fa86eb61d10648b2ccf71bb7ec40c56f2ac1e",
      "stage": "dit",
      "op": "aten.linear.default",
      "input_tensor": {
        "shape": [
          3,
          5120
        ],
        "stride": [
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 30720
      },
      "output_tensor": {
        "shape": [
          3,
          5120
        ],
        "stride": [
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 30720
      },
      "shape_parameters": {
        "M": 3,
        "N": 5120,
        "K": 5120
      },
      "mac_flops": 157286400,
      "bias_additions": 15360,
      "formula": "2*M*N*K (multiply and add each count one); bias additions listed separately",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 17.344,
      "dram_read_bytes": 52487168.0,
      "dram_write_bytes": 3208960.0,
      "dram_total_bytes": 55696128.0,
      "nominal_mac_tflops_per_s": 9.068634686346863,
      "nominal_mac_flops_per_dram_byte": 2.8240095972201154,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_ca590b3a668a9ecf_v1",
      "ncu_metrics_sha256": "ff0ed187a9a695411c47f2a3e2bc64fa92dfad31300a626d8085c077e414338a",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/dit_rank0_22805ed4eb786e66.pt",
      "input_sha256": "d6a2e382ecf0fdad711df2cd1bcb239a004321ae666d26aa04920718cf49bf81",
      "stage": "dit",
      "op": "aten.linear.default",
      "input_tensor": {
        "shape": [
          1,
          3,
          5120
        ],
        "stride": [
          15360,
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 30720
      },
      "output_tensor": {
        "shape": [
          1,
          3,
          30720
        ],
        "stride": [
          92160,
          30720,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 184320
      },
      "shape_parameters": {
        "M": 3,
        "N": 30720,
        "K": 5120
      },
      "mac_flops": 943718400,
      "bias_additions": 92160,
      "formula": "2*M*N*K (multiply and add each count one); bias additions listed separately",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 79.808,
      "dram_read_bytes": 314746880.0,
      "dram_write_bytes": 5089536.0,
      "dram_total_bytes": 319836416.0,
      "nominal_mac_tflops_per_s": 11.824859663191658,
      "nominal_mac_flops_per_dram_byte": 2.950628361218255,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_22805ed4eb786e66_v1",
      "ncu_metrics_sha256": "461281d7eeb06dfcbe6015480ea440edec821aa42f46c491f276a01086fb0ce4",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/dit_rank0_fcebe966800b94b1.pt",
      "input_sha256": "4f1e9988bf86b3f91f274f9b8c3bf33bf42b39945181ce6f09fbce8e0132cb6f",
      "stage": "dit",
      "op": "aten.linear.default",
      "input_tensor": {
        "shape": [
          1,
          1170,
          5120
        ],
        "stride": [
          5990400,
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 11980800
      },
      "output_tensor": {
        "shape": [
          1,
          1170,
          15360
        ],
        "stride": [
          17971200,
          15360,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 35942400
      },
      "shape_parameters": {
        "M": 1170,
        "N": 15360,
        "K": 5120
      },
      "mac_flops": 184025088000,
      "bias_additions": 17971200,
      "formula": "2*M*N*K (multiply and add each count one); bias additions listed separately",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 252.832,
      "dram_read_bytes": 262954496.0,
      "dram_write_bytes": 33932544.0,
      "dram_total_bytes": 296887040.0,
      "nominal_mac_tflops_per_s": 727.8552082014935,
      "nominal_mac_flops_per_dram_byte": 619.8488421724303,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_fcebe966800b94b1_v1",
      "ncu_metrics_sha256": "6ce786696e2bd2aa6bd27156d88c51f6bbad5cb8da821def44861f8877d32af8",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/dit_rank0_10a9e2fea034c6cf.pt",
      "input_sha256": "a64f7da6eb1fcc9779d40258d974eb529c2397aaba4df6e32363d7c5d4e73065",
      "stage": "dit",
      "op": "aten.linear.default",
      "input_tensor": {
        "shape": [
          1,
          1170,
          5120
        ],
        "stride": [
          5990400,
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 11980800
      },
      "output_tensor": {
        "shape": [
          1,
          1170,
          5120
        ],
        "stride": [
          5990400,
          5120,
          1
        ],
        "dtype": "torch.bfloat16",
        "logical_bytes": 11980800
      },
      "shape_parameters": {
        "M": 1170,
        "N": 5120,
        "K": 5120
      },
      "mac_flops": 61341696000,
      "bias_additions": 5990400,
      "formula": "2*M*N*K (multiply and add each count one); bias additions listed separately",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 87.104,
      "dram_read_bytes": 74980864.0,
      "dram_write_bytes": 9790720.0,
      "dram_total_bytes": 84771584.0,
      "nominal_mac_tflops_per_s": 704.2351212343865,
      "nominal_mac_flops_per_dram_byte": 723.6115347331483,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_dit_rank0_10a9e2fea034c6cf_v1",
      "ncu_metrics_sha256": "02cea0e01e66f689def70badde18b4711e2c75e3d183fa6c857fa909b05eebd9",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/vae_rank0_6638f430ac58a438.pt",
      "input_sha256": "32bcd9b4d8be26c876626649d4ff2e623424efe309222bcd7e80afc47310ded4",
      "stage": "vae",
      "op": "aten.conv3d.default",
      "input_tensor": {
        "shape": [
          1,
          16,
          3,
          60,
          104
        ],
        "stride": [
          299520,
          18720,
          6240,
          104,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 1198080
      },
      "output_tensor": {
        "shape": [
          1,
          16,
          3,
          60,
          104
        ],
        "stride": [
          299520,
          18720,
          6240,
          104,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 1198080
      },
      "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
      },
      "mac_flops": 9584640,
      "bias_additions": 299520,
      "formula": "2*output_elements*(Cin/groups)*Kd*Kh*Kw; padding already materialized in saved input, implicit padding zero; bias additions separate",
      "profiled_kernel_count": 2,
      "summed_profiled_kernel_us": 8.511999999999999,
      "dram_read_bytes": 2438400.0,
      "dram_write_bytes": 0.0,
      "dram_total_bytes": 2438400.0,
      "nominal_mac_tflops_per_s": 1.126015037593985,
      "nominal_mac_flops_per_dram_byte": 3.930708661417323,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_6638f430ac58a438_v1",
      "ncu_metrics_sha256": "bf58bc7e296e1e73240163db03528d9e83724ea72a4549ba0113bd6442d39bc7",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/vae_rank0_29ef146661136bf4.pt",
      "input_sha256": "2b9e80cee384305d79045df290612b04bcfe041bd67dc69a2bc096e45a6c5872",
      "stage": "vae",
      "op": "aten.conv3d.default",
      "input_tensor": {
        "shape": [
          1,
          16,
          3,
          62,
          28
        ],
        "stride": [
          83328,
          5208,
          1736,
          28,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 333312
      },
      "output_tensor": {
        "shape": [
          1,
          384,
          1,
          60,
          26
        ],
        "stride": [
          599040,
          1560,
          1560,
          26,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 2396160
      },
      "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
      },
      "mac_flops": 517570560,
      "bias_additions": 599040,
      "formula": "2*output_elements*(Cin/groups)*Kd*Kh*Kw; padding already materialized in saved input, implicit padding zero; bias additions separate",
      "profiled_kernel_count": 4,
      "summed_profiled_kernel_us": 26.944,
      "dram_read_bytes": 4486912.0,
      "dram_write_bytes": 0.0,
      "dram_total_bytes": 4486912.0,
      "nominal_mac_tflops_per_s": 19.20912114014252,
      "nominal_mac_flops_per_dram_byte": 115.35117247675015,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_29ef146661136bf4_v1",
      "ncu_metrics_sha256": "3aba0e06fff40239486c3f155fa2503d25526a0cfaa39b92bb9f8e2060fd3134",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/vae_rank0_551fa1ff3095f40c.pt",
      "input_sha256": "6309c12f24013edabd88edced7c2ce22672322d203206b12ba218c88696effb5",
      "stage": "vae",
      "op": "aten.conv3d.default",
      "input_tensor": {
        "shape": [
          1,
          384,
          3,
          62,
          28
        ],
        "stride": [
          1999872,
          5208,
          1736,
          28,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 7999488
      },
      "output_tensor": {
        "shape": [
          1,
          384,
          1,
          60,
          26
        ],
        "stride": [
          599040,
          1560,
          1560,
          26,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 2396160
      },
      "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
      },
      "mac_flops": 12421693440,
      "bias_additions": 599040,
      "formula": "2*output_elements*(Cin/groups)*Kd*Kh*Kw; padding already materialized in saved input, implicit padding zero; bias additions separate",
      "profiled_kernel_count": 5,
      "summed_profiled_kernel_us": 122.848,
      "dram_read_bytes": 52742144.0,
      "dram_write_bytes": 2068992.0,
      "dram_total_bytes": 54811136.0,
      "nominal_mac_tflops_per_s": 101.1143318572545,
      "nominal_mac_flops_per_dram_byte": 226.6271846655395,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_551fa1ff3095f40c_v1",
      "ncu_metrics_sha256": "12d48a0b8063b436147f604f8d269b1480e8e9947b99b78dfb9996f041d5da68",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/vae_rank0_c16d11a7624e1a8c.pt",
      "input_sha256": "1e43418295ca6422f27afcf0b9940aa818fbb12fcf1013c12dfcf7a60f86e7d7",
      "stage": "vae",
      "op": "aten.conv3d.default",
      "input_tensor": {
        "shape": [
          1,
          192,
          1,
          120,
          52
        ],
        "stride": [
          1198080,
          6240,
          1198080,
          52,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 4792320
      },
      "output_tensor": {
        "shape": [
          1,
          384,
          1,
          120,
          52
        ],
        "stride": [
          2396160,
          6240,
          6240,
          52,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 9584640
      },
      "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
      },
      "mac_flops": 920125440,
      "bias_additions": 2396160,
      "formula": "2*output_elements*(Cin/groups)*Kd*Kh*Kw; padding already materialized in saved input, implicit padding zero; bias additions separate",
      "profiled_kernel_count": 2,
      "summed_profiled_kernel_us": 21.152,
      "dram_read_bytes": 14717440.0,
      "dram_write_bytes": 3072.0,
      "dram_total_bytes": 14720512.0,
      "nominal_mac_tflops_per_s": 43.50063540090772,
      "nominal_mac_flops_per_dram_byte": 62.50634760530068,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_c16d11a7624e1a8c_v1",
      "ncu_metrics_sha256": "966631416bc3349cc1bd29b3a29b66c20b5642dd0523d5b553436567f63583c0",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/vae_rank0_081e405bd1e07dce.pt",
      "input_sha256": "3193a32ddc024e9d34eb5446f9909408dde38d963ebc5cb9654998ecf661a820",
      "stage": "vae",
      "op": "aten.conv3d.default",
      "input_tensor": {
        "shape": [
          1,
          192,
          3,
          122,
          54
        ],
        "stride": [
          3794688,
          19764,
          6588,
          54,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 15178752
      },
      "output_tensor": {
        "shape": [
          1,
          384,
          1,
          120,
          52
        ],
        "stride": [
          2396160,
          6240,
          6240,
          52,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 9584640
      },
      "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
      },
      "mac_flops": 24843386880,
      "bias_additions": 2396160,
      "formula": "2*output_elements*(Cin/groups)*Kd*Kh*Kw; padding already materialized in saved input, implicit padding zero; bias additions separate",
      "profiled_kernel_count": 6,
      "summed_profiled_kernel_us": 158.56,
      "dram_read_bytes": 67194880.0,
      "dram_write_bytes": 8898560.0,
      "dram_total_bytes": 76093440.0,
      "nominal_mac_tflops_per_s": 156.68129969727548,
      "nominal_mac_flops_per_dram_byte": 326.4852644327816,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_081e405bd1e07dce_v1",
      "ncu_metrics_sha256": "e65b3a2773150323d43a62874e0ab021c96847e6456ffed465200382f326d5a5",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/vae_rank0_8601bc3a88374d99.pt",
      "input_sha256": "c7f10fe8b8ca84229e785645c70c003456c3950c79cdd176694e5bc86c9247f4",
      "stage": "vae",
      "op": "aten.conv3d.default",
      "input_tensor": {
        "shape": [
          1,
          384,
          3,
          122,
          54
        ],
        "stride": [
          7589376,
          19764,
          6588,
          54,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 30357504
      },
      "output_tensor": {
        "shape": [
          1,
          384,
          1,
          120,
          52
        ],
        "stride": [
          2396160,
          6240,
          6240,
          52,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 9584640
      },
      "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
      },
      "mac_flops": 49686773760,
      "bias_additions": 2396160,
      "formula": "2*output_elements*(Cin/groups)*Kd*Kh*Kw; padding already materialized in saved input, implicit padding zero; bias additions separate",
      "profiled_kernel_count": 6,
      "summed_profiled_kernel_us": 287.104,
      "dram_read_bytes": 134126336.0,
      "dram_write_bytes": 26391808.0,
      "dram_total_bytes": 160518144.0,
      "nominal_mac_tflops_per_s": 173.06193490860457,
      "nominal_mac_flops_per_dram_byte": 309.53992191686444,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_8601bc3a88374d99_v1",
      "ncu_metrics_sha256": "f31995cc8246a57dfaac0d7653269ccfbd0f244a053a5a7a0ee54e5e41aacedb",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/vae_rank0_99475e040d65d084.pt",
      "input_sha256": "0cc0a21ce474602c53c3e5621e64598ed7be5a2c16603fc968441df41de71330",
      "stage": "vae",
      "op": "aten.conv3d.default",
      "input_tensor": {
        "shape": [
          1,
          192,
          3,
          242,
          106
        ],
        "stride": [
          14775552,
          76956,
          25652,
          106,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 59102208
      },
      "output_tensor": {
        "shape": [
          1,
          192,
          1,
          240,
          104
        ],
        "stride": [
          4792320,
          24960,
          24960,
          104,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 19169280
      },
      "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
      },
      "mac_flops": 49686773760,
      "bias_additions": 4792320,
      "formula": "2*output_elements*(Cin/groups)*Kd*Kh*Kw; padding already materialized in saved input, implicit padding zero; bias additions separate",
      "profiled_kernel_count": 5,
      "summed_profiled_kernel_us": 230.72,
      "dram_read_bytes": 165777408.0,
      "dram_write_bytes": 53530624.0,
      "dram_total_bytes": 219308032.0,
      "nominal_mac_tflops_per_s": 215.3552954230236,
      "nominal_mac_flops_per_dram_byte": 226.56157782675282,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_vae_rank0_99475e040d65d084_v1",
      "ncu_metrics_sha256": "8e04dd103a921257d405955a2acf864a485740280690469b34a8aabc8043b9a3",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/vae_rank0_0bd1bd0a5c44ff3b.pt",
      "input_sha256": "1e47965836d719f164296b5c8d52cac8c22a10d9ff8c5f4fdfa48498909f8856",
      "stage": "vae",
      "op": "aten.conv3d.default",
      "input_tensor": {
        "shape": [
          1,
          96,
          3,
          482,
          210
        ],
        "stride": [
          29151360,
          303660,
          101220,
          210,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 116605440
      },
      "output_tensor": {
        "shape": [
          1,
          96,
          1,
          480,
          208
        ],
        "stride": [
          9584640,
          99840,
          99840,
          208,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 38338560
      },
      "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
      },
      "mac_flops": 49686773760,
      "bias_additions": 9584640,
      "formula": "2*output_elements*(Cin/groups)*Kd*Kh*Kw; padding already materialized in saved input, implicit padding zero; bias additions separate",
      "profiled_kernel_count": 5,
      "summed_profiled_kernel_us": 466.624,
      "dram_read_bytes": 312005632.0,
      "dram_write_bytes": 161180160.0,
      "dram_total_bytes": 473185792.0,
      "nominal_mac_tflops_per_s": 106.48139349883417,
      "nominal_mac_flops_per_dram_byte": 105.00478797131761,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_vae_v1",
      "ncu_metrics_sha256": "9715aa5feeaf6b673c12be1eba462cb072e0fca573ef8dd4cb696fb4ef5e2a96",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/output_rank0_f18eb91117dda1c8.pt",
      "input_sha256": "897dd2bd3ad5fe7258ee5c3ae9cfd93da4e5e94644902267f0b3d8bd156f63b0",
      "stage": "output",
      "op": "aten.to.dtype",
      "input_tensor": {
        "shape": [
          1,
          3,
          9,
          480,
          832
        ],
        "stride": [
          10782720,
          3594240,
          399360,
          832,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 43130880
      },
      "output_tensor": {
        "shape": [
          1,
          3,
          9,
          480,
          832
        ],
        "stride": [
          10782720,
          3594240,
          399360,
          832,
          1
        ],
        "dtype": "torch.uint8",
        "logical_bytes": 10782720
      },
      "shape_parameters": {},
      "mac_flops": null,
      "bias_additions": 0,
      "formula": "dtype conversion: no useful floating-point FLOP rate assigned",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 29.824,
      "dram_read_bytes": 43166464.0,
      "dram_write_bytes": 5343232.0,
      "dram_total_bytes": 48509696.0,
      "nominal_mac_tflops_per_s": null,
      "nominal_mac_flops_per_dram_byte": null,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_batch_output_rank0_f18eb91117dda1c8_v1",
      "ncu_metrics_sha256": "21a33e9eac00633e4c8000dabf6158110f2d0b379b1006144b11779474998d09",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    },
    {
      "input": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/operator_capture_480x832_v2/operators/output_rank0_e753c57113db57aa.pt",
      "input_sha256": "a9aab61b63e7cab483e81582c9af41d89d525814ab14de753e8a7b16de6c4f1b",
      "stage": "output",
      "op": "aten.to.dtype",
      "input_tensor": {
        "shape": [
          1,
          3,
          12,
          480,
          832
        ],
        "stride": [
          14376960,
          4792320,
          399360,
          832,
          1
        ],
        "dtype": "torch.float32",
        "logical_bytes": 57507840
      },
      "output_tensor": {
        "shape": [
          1,
          3,
          12,
          480,
          832
        ],
        "stride": [
          14376960,
          4792320,
          399360,
          832,
          1
        ],
        "dtype": "torch.uint8",
        "logical_bytes": 14376960
      },
      "shape_parameters": {},
      "mac_flops": null,
      "bias_additions": 0,
      "formula": "dtype conversion: no useful floating-point FLOP rate assigned",
      "profiled_kernel_count": 1,
      "summed_profiled_kernel_us": 39.583999999999996,
      "dram_read_bytes": 57544448.0,
      "dram_write_bytes": 11594752.0,
      "dram_total_bytes": 69139200.0,
      "nominal_mac_tflops_per_s": null,
      "nominal_mac_flops_per_dram_byte": null,
      "ncu_output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_output_v1",
      "ncu_metrics_sha256": "ef8d7d820d04ff1bb9f5f0e42b670acbf7ea11499b7e808a05a41c42d6b8f866",
      "tensor_policy": {
        "matmul_precision": "highest",
        "matmul_allow_tf32": false,
        "cudnn_allow_tf32": true,
        "cudnn_benchmark": false,
        "cudnn_deterministic": false
      }
    }
  ],
  "scope": "18 saved operators only. FLOPs count nominal dense multiply-add work including explicitly materialized input padding; not executed instruction FLOPs. Rate divides by summed NCU kernel duration (including layout and bias kernels), not unprofiled operator latency. DRAM traffic sums each profiled kernel, reflects replay cache/writeback behavior, not required logical bytes. Bias additions separate; dtype conversion FLOPs undefined. No local bandwidth ceiling or full-model MFU claimed; no mapping from this isolated op to all integrated calls implied."
}