{
  "generated_at": "2026-09-10T02:17:01.757714+00:00",
  "kernels": [
    {
      "kernel": "void cudnn::nchwToNhwcKernel<float, float, float, 0, 1, 0>(cudnn::nchw2nhwc_params_t<T3>, const T1 *, T2 *)",
      "metrics": {
        "gpu__time_duration.sum": {
          "value": 69.952,
          "unit": "us"
        },
        "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 37.664733,
          "unit": "%"
        },
        "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 64.014557,
          "unit": "%"
        },
        "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 81.220531,
          "unit": "%"
        },
        "sm__pipe_tensor_cycles_active.avg.pct_of_peak_sustained_elapsed": {
          "value": 0.0,
          "unit": "%"
        },
        "sm__warps_active.avg.pct_of_peak_sustained_active": {
          "value": 92.804217,
          "unit": "%"
        },
        "smsp__warps_eligible.avg.per_cycle_active": {
          "value": 0.610395,
          "unit": "warp"
        },
        "smsp__issue_active.avg.pct_of_peak_sustained_active": {
          "value": 32.410894,
          "unit": "%"
        },
        "launch__occupancy_limit_registers": {
          "value": 8.0,
          "unit": "block"
        },
        "launch__occupancy_limit_shared_mem": {
          "value": 19.0,
          "unit": "block"
        },
        "launch__occupancy_limit_warps": {
          "value": 8.0,
          "unit": "block"
        },
        "launch__waves_per_multiprocessor": {
          "value": 26.96,
          "unit": ""
        },
        "derived__local_spilling_requests": {
          "value": 0.0,
          "unit": ""
        },
        "sass__inst_executed_register_spilling": {
          "value": 0.0,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_long_scoreboard_per_issue_active.ratio": {
          "value": 27.541326,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_barrier_per_issue_active.ratio": {
          "value": 8.676464,
          "unit": "inst"
        }
      }
    },
    {
      "kernel": "void cudnn::nchwToNhwcKernel<float, float, float, 0, 1, 0>(cudnn::nchw2nhwc_params_t<T3>, const T1 *, T2 *)",
      "metrics": {
        "gpu__time_duration.sum": {
          "value": 3.52,
          "unit": "us"
        },
        "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 12.364265,
          "unit": "%"
        },
        "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 5.962006,
          "unit": "%"
        },
        "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 15.411113,
          "unit": "%"
        },
        "sm__pipe_tensor_cycles_active.avg.pct_of_peak_sustained_elapsed": {
          "value": 0.0,
          "unit": "%"
        },
        "sm__warps_active.avg.pct_of_peak_sustained_active": {
          "value": 24.673274,
          "unit": "%"
        },
        "smsp__warps_eligible.avg.per_cycle_active": {
          "value": 0.362063,
          "unit": "warp"
        },
        "smsp__issue_active.avg.pct_of_peak_sustained_active": {
          "value": 22.179784,
          "unit": "%"
        },
        "launch__occupancy_limit_registers": {
          "value": 8.0,
          "unit": "block"
        },
        "launch__occupancy_limit_shared_mem": {
          "value": 19.0,
          "unit": "block"
        },
        "launch__occupancy_limit_warps": {
          "value": 8.0,
          "unit": "block"
        },
        "launch__waves_per_multiprocessor": {
          "value": 0.27,
          "unit": ""
        },
        "derived__local_spilling_requests": {
          "value": 0.0,
          "unit": ""
        },
        "sass__inst_executed_register_spilling": {
          "value": 0.0,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_long_scoreboard_per_issue_active.ratio": {
          "value": 6.845452,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_barrier_per_issue_active.ratio": {
          "value": 0.563324,
          "unit": "inst"
        }
      }
    },
    {
      "kernel": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroupsize1x1x1_g1_execute_segment_k_off_kernel__5x_cudnn",
      "metrics": {
        "gpu__time_duration.sum": {
          "value": 343.968,
          "unit": "us"
        },
        "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 29.783099,
          "unit": "%"
        },
        "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 9.053182,
          "unit": "%"
        },
        "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 60.36853,
          "unit": "%"
        },
        "sm__pipe_tensor_cycles_active.avg.pct_of_peak_sustained_elapsed": {
          "value": 29.783099,
          "unit": "%"
        },
        "sm__warps_active.avg.pct_of_peak_sustained_active": {
          "value": 17.59726,
          "unit": "%"
        },
        "smsp__warps_eligible.avg.per_cycle_active": {
          "value": 0.1987,
          "unit": "warp"
        },
        "smsp__issue_active.avg.pct_of_peak_sustained_active": {
          "value": 18.475147,
          "unit": "%"
        },
        "launch__occupancy_limit_registers": {
          "value": 1.0,
          "unit": "block"
        },
        "launch__occupancy_limit_shared_mem": {
          "value": 1.0,
          "unit": "block"
        },
        "launch__occupancy_limit_warps": {
          "value": 5.0,
          "unit": "block"
        },
        "launch__waves_per_multiprocessor": {
          "value": 1.0,
          "unit": ""
        },
        "derived__local_spilling_requests": {
          "value": 0.0,
          "unit": ""
        },
        "sass__inst_executed_register_spilling": {
          "value": 0.0,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_long_scoreboard_per_issue_active.ratio": {
          "value": 8.368729,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_barrier_per_issue_active.ratio": {
          "value": 4.608407,
          "unit": "inst"
        }
      }
    },
    {
      "kernel": "void cudnn::nhwcToNchwKernel<float, float, float, 1, 0, 0>(cudnn::nhwc2nchw_params_t<T3>, const T1 *, T2 *)",
      "metrics": {
        "gpu__time_duration.sum": {
          "value": 19.968,
          "unit": "us"
        },
        "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 41.449384,
          "unit": "%"
        },
        "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 59.726043,
          "unit": "%"
        },
        "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 81.746478,
          "unit": "%"
        },
        "sm__pipe_tensor_cycles_active.avg.pct_of_peak_sustained_elapsed": {
          "value": 0.0,
          "unit": "%"
        },
        "sm__warps_active.avg.pct_of_peak_sustained_active": {
          "value": 88.769719,
          "unit": "%"
        },
        "smsp__warps_eligible.avg.per_cycle_active": {
          "value": 1.27777,
          "unit": "warp"
        },
        "smsp__issue_active.avg.pct_of_peak_sustained_active": {
          "value": 47.31937,
          "unit": "%"
        },
        "launch__occupancy_limit_registers": {
          "value": 8.0,
          "unit": "block"
        },
        "launch__occupancy_limit_shared_mem": {
          "value": 19.0,
          "unit": "block"
        },
        "launch__occupancy_limit_warps": {
          "value": 8.0,
          "unit": "block"
        },
        "launch__waves_per_multiprocessor": {
          "value": 8.86,
          "unit": ""
        },
        "derived__local_spilling_requests": {
          "value": 0.0,
          "unit": ""
        },
        "sass__inst_executed_register_spilling": {
          "value": 0.0,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_long_scoreboard_per_issue_active.ratio": {
          "value": 16.131762,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_barrier_per_issue_active.ratio": {
          "value": 4.391265,
          "unit": "inst"
        }
      }
    },
    {
      "kernel": "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)",
      "metrics": {
        "gpu__time_duration.sum": {
          "value": 27.616,
          "unit": "us"
        },
        "sm__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 58.140107,
          "unit": "%"
        },
        "gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 38.876093,
          "unit": "%"
        },
        "lts__throughput.avg.pct_of_peak_sustained_elapsed": {
          "value": 59.812679,
          "unit": "%"
        },
        "sm__pipe_tensor_cycles_active.avg.pct_of_peak_sustained_elapsed": {
          "value": 2.08986,
          "unit": "%"
        },
        "sm__warps_active.avg.pct_of_peak_sustained_active": {
          "value": 79.875376,
          "unit": "%"
        },
        "smsp__warps_eligible.avg.per_cycle_active": {
          "value": 1.542711,
          "unit": "warp"
        },
        "smsp__issue_active.avg.pct_of_peak_sustained_active": {
          "value": 64.835462,
          "unit": "%"
        },
        "launch__occupancy_limit_registers": {
          "value": 21.0,
          "unit": "block"
        },
        "launch__occupancy_limit_shared_mem": {
          "value": 32.0,
          "unit": "block"
        },
        "launch__occupancy_limit_warps": {
          "value": 16.0,
          "unit": "block"
        },
        "launch__waves_per_multiprocessor": {
          "value": 17.73,
          "unit": ""
        },
        "derived__local_spilling_requests": {
          "value": 0.0,
          "unit": ""
        },
        "sass__inst_executed_register_spilling": {
          "value": 0.0,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_long_scoreboard_per_issue_active.ratio": {
          "value": 13.23422,
          "unit": "inst"
        },
        "smsp__average_warps_issue_stalled_barrier_per_issue_active.ratio": {
          "value": 0.0,
          "unit": "inst"
        }
      }
    }
  ],
  "source_sha256": "a941df9d7f493f0e835a5bd5568ef2ba7ef956d1eda8f84287bcfa85bd03c611",
  "scope": "Single real-input VAE convolution, detailed profiler replay. Resource limit is a theoretical block-residency limit; warp wait metrics are not wall-time fractions. Low DRAM throughput does not disprove latency or locality constraints. No stall metric alone proves a causal bottleneck; use equal-boundary layout A/B as one counterexample.",
  "scheduler_note": "Guard briefly classified this PID as unknown user near successful completion, with no warning or signal and resolution process exited. Identity file proves UID2005 at startup. Teardown race is a possible explanation, not proven. Do not claim zero guard records."
}