{
  "status": "passed",
  "generated_at": "2026-09-10T02:08:30.020522+00:00",
  "allocated": "4",
  "input_sha256": "b972a90ab8c175e3761468ed98dc0c8304bd7ab3a3db99ebc2f9a89a7fb1b3d2",
  "reference_sha256": "1e6f38a5ef55eacc25edd64c2735b8f3a9e5f6b4bb4efa9ead8247a6e82f69c4",
  "context": {
    "phase": "commit",
    "step": null,
    "start_frame": 27,
    "role": "self",
    "layer": "transformer.blocks.0.self_attn"
  },
  "jobs": [
    {
      "version": 2,
      "output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_attention_fa2_v1",
      "oracle": {
        "max_abs": 0.010101945123846434,
        "relative_l2": 0.0016959539687613817,
        "allclose_atol_003_rtol_003": true,
        "full_finite": true,
        "query_samples": 32,
        "heads": 4,
        "all_valid_keys": 28080,
        "allocated_keys": 29640,
        "captured_output": {
          "exact": false,
          "allclose": true,
          "relative_l2": 0.0005383218211689153,
          "max_abs": 0.0625
        },
        "passed": true
      },
      "kernels": [
        {
          "name": "void flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 64, 4, 0, 0, bfloat16_t, Flash_kernel_traits<128, 128, 64, 4, bfloat16_t>>, 0, 0, 0, 0, 0, 1, 0, 0>(Flash_fwd_params)",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 1.907808,
              "unit": "ms"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 55.362668,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 2.102284,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 19.837116,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 10.613872,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 173.64096,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 19.356416,
              "unit": "Mbyte"
            }
          }
        }
      ],
      "elapsed_seconds": 22.124754905700684,
      "sha256": {
        "output_metadata.json": "4b1304aa377878d6b4939026faab2e02ca4f45eddd9e240fef68e905fac671ff",
        "ncu_attention_input.py": "9e58a354dc04dda2ab339e273e3df73d9f6328cae60cacd54543245176cfd565",
        "metrics.csv": "94dcee7744039f5489c4e08adff0080258362b2791e511079793dc1c214e71b1",
        "run.log": "597d2033e5732da7173d119b848fef20b8518f41451bc33c35c5b313023524ca",
        "output.pt": "9657a7e7a845c2aaffd87701c8194de2ed65ca33501fe9355d15e739b3e84f93",
        "attention.ncu-rep": "86ccb6c3d4fe5b809024101800721a537f1d1e9ae49e1733114f288afa11ad47",
        "identity.json": "9117d91e17902e9ba4648c75cd6bab84eb66ff39af8b6a020c2932cc73f027da",
        "manifest.json": "8d9b3a2b982d9e5201c9c3bd35525735e59bcf3a13102d176b5c0348c9f4e991",
        "guard_own_pid_history.json": "d42d4f622aa3fff22a0dcccf87a0b5424c5a6a9910834ba9de0e9018c9b12bff"
      }
    },
    {
      "version": 3,
      "output": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/ncu_capability_attention_fa3_v1",
      "oracle": {
        "max_abs": 0.010101945123846434,
        "relative_l2": 0.0016967686260984625,
        "allclose_atol_003_rtol_003": true,
        "full_finite": true,
        "query_samples": 32,
        "heads": 4,
        "all_valid_keys": 28080,
        "allocated_keys": 29640,
        "captured_output": {
          "exact": true,
          "allclose": true,
          "relative_l2": 0.0,
          "max_abs": 0.0
        },
        "passed": true
      },
      "kernels": [
        {
          "name": "void prepare_varlen_num_blocks_kernel<1, 0>(int, int, int, const int *, const int *, const int *, const int *, const int *, const int *, int, int, int, int, int, FastDivmod, FastDivmod, int *, int *, int *, int *, int *, bool, bool, bool, int)",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 2.752,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 0.005847,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 0.052288,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 0.390833,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 1.212442,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 0.006912,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 0.0,
              "unit": "Mbyte"
            }
          }
        },
        {
          "name": "void device_kernel<enable_sm90_or_later<FlashAttnFwdSm90<CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1>>, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128>>, 128, bfloat16_t, float, arch::Sm90, 0, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, bfloat16_t, 1>, CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160>>, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1>>, bfloat16_t, arch::Sm90, 256, 1, 1, 0, 0, 1>, VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, 0, 1, 1, 0, 0, 1>>>>(T1::Params)",
          "device": "0",
          "metrics": {
            "gpu__time_duration.sum": {
              "value": 918.976,
              "unit": "us"
            },
            "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 79.461747,
              "unit": "%"
            },
            "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 4.516504,
              "unit": "%"
            },
            "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
              "value": 22.351045,
              "unit": "%"
            },
            "sm__warps_active.avg.pct_of_peak_sustained_active": {
              "value": 18.747057,
              "unit": "%"
            },
            "dram__bytes_read.sum": {
              "value": 184.964352,
              "unit": "Mbyte"
            },
            "dram__bytes_write.sum": {
              "value": 14.760192,
              "unit": "Mbyte"
            }
          }
        }
      ],
      "elapsed_seconds": 23.789729833602905,
      "sha256": {
        "output_metadata.json": "fca6e84a0ef8c3c1587b6a9ab856da7c1001916c6df7995433d6f18a3b2f6e6e",
        "ncu_attention_input.py": "9e58a354dc04dda2ab339e273e3df73d9f6328cae60cacd54543245176cfd565",
        "metrics.csv": "4c60a5c660223beee793593e24a22056e72662278cfd5a19db6554c3676b341b",
        "run.log": "d10dadc98b00727ee3b783a1acf62b19a022e5b2f40f122e666aa5757756d479",
        "output.pt": "c1581c2a22db1956a4d404d1959d4c7bd713cde823dc59fc23cf10b7024a9474",
        "attention.ncu-rep": "1aaad44e2b246de1ab363a6edcaf573edf04c5942c24009d029672c09e8923d9",
        "identity.json": "131dad8e203d9b6a5b2367dc226c50d74e574e95b24d35d2c8b1914e0a33deb1",
        "manifest.json": "843b2a1f23a8c515fefdb88bc2885c72bc85499943fbb89aedea9c4c2df37cae",
        "guard_own_pid_history.json": "d29dba3520ae52d65f586500cab97f396a6272daedc5facb8671ed69e6f81710"
      }
    }
  ],
  "scope": "Same allocated GPU, one real input, FA2 then FA3, NCU replayed kernels. Sampled FP64 oracle uses unchanged .03/.03 allclose plus relative L2 <= .01. Not full-tensor FP64, clean timing, full-model speedup, or MFU."
}