{
  "generated_at": "2026-09-10T02:02:22.754833+00:00",
  "status": "passed",
  "jobs": [
    {
      "stage": "dit",
      "status": "passed",
      "output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_dit_v1",
      "elapsed_seconds": 14.154895782470703,
      "allocated": "0",
      "input_sha256": "4a5b492ee606c1a2c9fe4844a0c3125609fa599f4c027f56b4329c89d2879b7e",
      "correctness": {
        "op": "aten.linear.default",
        "stage": "dit",
        "shape": [
          1,
          4680,
          5120
        ],
        "exact": true,
        "max_abs": 0.0,
        "rmse": 0.0,
        "finite": true,
        "math_policy": {
          "matmul_precision": "highest",
          "matmul_allow_tf32": false,
          "cudnn_allow_tf32": true,
          "cudnn_benchmark": false,
          "cudnn_deterministic": false
        },
        "scope": "Original ATen op on saved real tensors; isolated diagnostic"
      },
      "kernel_count": 1,
      "kernels": [
        {
          "id": "0",
          "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 292.448,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 94.027733,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 25.371858,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 52.627935,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 14.649706,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 314.121984,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 42.865664,
              "unit": "Mbyte"
            },
            "dram__bytes.sum.per_second": {
              "value": 1.220688,
              "unit": "Tbyte/s"
            },
            "launch__registers_per_thread": {
              "value": 168.0,
              "unit": "register/thread"
            },
            "launch__shared_mem_per_block": {
              "value": 222.404,
              "unit": "Kbyte/block"
            }
          }
        }
      ],
      "sha256": {
        "replay_comparison.json": "369c62e668a96131ff716e2d7527937010ddb07ace3c1a1113f9beb7aca7da4c",
        "metrics.csv": "e3b23b6928422a2d427cc67fa15c47e2fc6a548d66f2a4bf7b6d2626482cecaa",
        "operator.ncu-rep": "11e4d03e257ff0eca8872072b023ba7829ac2b6b25569f0a020e641022958844",
        "run.log": "86ac88cc3a8cc99c3e583d48e0b6b129aeea977cf2fa3cd2c0975e477a2de364",
        "identity.json": "8ec3c6c72438622d51772a37ebce2f3fb62aea0a9363ed75d7629ca06656ba65",
        "replay_operator.py": "21e6ebbe13e53933591ad14feaa6ab4b189738f3647f2087ce8bf523b0e07450",
        "manifest.json": "31f9d06dd868b28f0f3b10250afb08f504aac34c692ff90e76b213ebde71d05f",
        "guard_own_pid_history.json": "021c81f7b2116cdc2aaf9598af44ee76ec6625e04b098aed3bdce6b9e2f47779"
      }
    },
    {
      "stage": "vae",
      "status": "passed",
      "output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_vae_v1",
      "elapsed_seconds": 16.918233633041382,
      "allocated": "7",
      "input_sha256": "1e47965836d719f164296b5c8d52cac8c22a10d9ff8c5f4fdfa48498909f8856",
      "correctness": {
        "op": "aten.conv3d.default",
        "stage": "vae",
        "shape": [
          1,
          96,
          1,
          480,
          208
        ],
        "exact": true,
        "max_abs": 0.0,
        "rmse": 0.0,
        "finite": true,
        "math_policy": {
          "matmul_precision": "highest",
          "matmul_allow_tf32": false,
          "cudnn_allow_tf32": true,
          "cudnn_benchmark": false,
          "cudnn_deterministic": false
        },
        "scope": "Original ATen op on saved real tensors; isolated diagnostic"
      },
      "kernel_count": 5,
      "kernels": [
        {
          "id": "0",
          "name": "void cudnn::nchwToNhwcKernel<float, float, float, 0, 1, 0>(cudnn::nchw2nhwc_params_t<T3>, const T1 *, T2 *)",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 70.336,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 37.540406,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 63.710272,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 80.815686,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 93.308252,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 116.620544,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 98.860032,
              "unit": "Mbyte"
            },
            "dram__bytes.sum.per_second": {
              "value": 3.063589,
              "unit": "Tbyte/s"
            },
            "launch__registers_per_thread": {
              "value": 32.0,
              "unit": "register/thread"
            },
            "launch__shared_mem_per_block": {
              "value": 5.248,
              "unit": "Kbyte/block"
            }
          }
        },
        {
          "id": "1",
          "name": "void cudnn::nchwToNhwcKernel<float, float, float, 0, 1, 0>(cudnn::nchw2nhwc_params_t<T3>, const T1 *, T2 *)",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 3.52,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 12.390404,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 5.953501,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 15.414774,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 24.41718,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 1.004288,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 0.0,
              "unit": "Mbyte"
            },
            "dram__bytes.sum.per_second": {
              "value": 0.285309,
              "unit": "Tbyte/s"
            },
            "launch__registers_per_thread": {
              "value": 32.0,
              "unit": "register/thread"
            },
            "launch__shared_mem_per_block": {
              "value": 5.248,
              "unit": "Kbyte/block"
            }
          }
        },
        {
          "id": "2",
          "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroupsize1x1x1_g1_execute_segment_k_off_kernel__5x_cudnn",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 345.6,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 29.68631,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 8.985101,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 60.273264,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 17.755377,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 117.670144,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 31.733248,
              "unit": "Mbyte"
            },
            "dram__bytes.sum.per_second": {
              "value": 0.432301,
              "unit": "Tbyte/s"
            },
            "launch__registers_per_thread": {
              "value": 168.0,
              "unit": "register/thread"
            },
            "launch__shared_mem_per_block": {
              "value": 232.448,
              "unit": "Kbyte/block"
            }
          }
        },
        {
          "id": "3",
          "name": "void cudnn::nhwcToNchwKernel<float, float, float, 1, 0, 0>(cudnn::nhwc2nchw_params_t<T3>, const T1 *, T2 *)",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 19.904,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 41.754029,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 58.781261,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 81.369093,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 89.587993,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 38.355968,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 17.823744,
              "unit": "Mbyte"
            },
            "dram__bytes.sum.per_second": {
              "value": 2.822534,
              "unit": "Tbyte/s"
            },
            "launch__registers_per_thread": {
              "value": 32.0,
              "unit": "register/thread"
            },
            "launch__shared_mem_per_block": {
              "value": 5.248,
              "unit": "Kbyte/block"
            }
          }
        },
        {
          "id": "4",
          "name": "void at::elementwise_kernel<128, 2, void at::gpu_kernel_impl_nocast<at::CUDAFunctor_add<float>>(at::TensorIteratorBase &, const T1 &)::[lambda(int) (instance 1)]>(int, T3)",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 27.264,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 59.158799,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 39.056992,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 60.717071,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 79.24734,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 38.354688,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 12.763136,
              "unit": "Mbyte"
            },
            "dram__bytes.sum.per_second": {
              "value": 1.87492,
              "unit": "Tbyte/s"
            },
            "launch__registers_per_thread": {
              "value": 18.0,
              "unit": "register/thread"
            },
            "launch__shared_mem_per_block": {
              "value": 1.024,
              "unit": "Kbyte/block"
            }
          }
        }
      ],
      "sha256": {
        "replay_comparison.json": "fbc0f1cc0275724c75a7685d0c3da84472740f5dee6ac7697f4160715261d45b",
        "metrics.csv": "9715aa5feeaf6b673c12be1eba462cb072e0fca573ef8dd4cb696fb4ef5e2a96",
        "operator.ncu-rep": "c9a1af7a6b873a50b0bf97afac127c0a95a95393722b41b707f51989c697a292",
        "run.log": "320072147bf1c95d44cd928524260654a333c7576eaa37dfeacd0a891ad9bc4a",
        "identity.json": "e5ea23ca715cea4c1ff59609f629ec7157e8de7ab3cc28bde2a26b0d351570a3",
        "replay_operator.py": "21e6ebbe13e53933591ad14feaa6ab4b189738f3647f2087ce8bf523b0e07450",
        "manifest.json": "d3bbc32df63026796e323fc0bdf4665c84eb041651437c3f366df91fc662f20e",
        "guard_own_pid_history.json": "3d693b44c8e29eed1ea64959d33f92802723acae456211428931452c84fb994d"
      }
    },
    {
      "stage": "output",
      "status": "passed",
      "output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_output_v1",
      "elapsed_seconds": 11.483896017074585,
      "allocated": "5",
      "input_sha256": "a9aab61b63e7cab483e81582c9af41d89d525814ab14de753e8a7b16de6c4f1b",
      "correctness": {
        "op": "aten.to.dtype",
        "stage": "output",
        "shape": [
          1,
          3,
          12,
          480,
          832
        ],
        "exact": true,
        "max_abs": 0.0,
        "rmse": 0.0,
        "finite": true,
        "math_policy": {
          "matmul_precision": "highest",
          "matmul_allow_tf32": false,
          "cudnn_allow_tf32": true,
          "cudnn_benchmark": false,
          "cudnn_deterministic": false
        },
        "scope": "Original ATen op on saved real tensors; isolated diagnostic"
      },
      "kernel_count": 1,
      "kernels": [
        {
          "id": "0",
          "name": "void unrolled_elementwise_kernel<direct_copy_kernel_cuda(TensorIteratorBase &)::[lambda() (instance 3)]::operator ()() lambda() (instance 1)]::operator ()() lambda(unsigned char) (instance 1)], array<char *, 2>, 4, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, LoadWithCast<1>, StoreWithCast<1>>(int, T1, T2, T4, T5, T6, T7)",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 39.584,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 55.888331,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 36.344303,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 44.802352,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 89.899207,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 57.544448,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 11.594752,
              "unit": "Mbyte"
            },
            "dram__bytes.sum.per_second": {
              "value": 1.746645,
              "unit": "Tbyte/s"
            },
            "launch__registers_per_thread": {
              "value": 32.0,
              "unit": "register/thread"
            },
            "launch__shared_mem_per_block": {
              "value": 1.024,
              "unit": "Kbyte/block"
            }
          }
        }
      ],
      "sha256": {
        "replay_comparison.json": "59ec25fab08d094e717140a7b0928a857ff389dbec068b03d6d3100ac8242bbb",
        "metrics.csv": "ef8d7d820d04ff1bb9f5f0e42b670acbf7ea11499b7e808a05a41c42d6b8f866",
        "operator.ncu-rep": "c82f979792422d72988256c4fd4707e899701d0a595c81f9689df3fbd3cb3c9f",
        "run.log": "eded896000896d7f5bd8b96428023fe17082d40a53629f054ead5b4381376ed5",
        "identity.json": "a9ba2df59f68800ced66deaaafc40607d26cbc1ecab3320c91dace7266f99fcd",
        "replay_operator.py": "21e6ebbe13e53933591ad14feaa6ab4b189738f3647f2087ce8bf523b0e07450",
        "manifest.json": "4bb8b7cfbf7552414ab9e7358a851a245455039059d9a25bf68a2ab932203d9d",
        "guard_own_pid_history.json": "0d2b4b506fea7c2607a01ec393f78d04c55e16056977ee3a8d65d9049fcc60c5"
      }
    }
  ],
  "scope": "Three isolated real-input operator replays, single device per job and distinct allocated devices across jobs. Exact reference output passed. NCU replay instrumentation and clocks/cache effects prevent treating these timings as clean integrated performance. No whole-model MFU or optimization acceptance. Guard history is a point-in-time check of each own PID."
}