{
  "source_gpu_trace": "benchmarks\\gb10-post-fc2-warmed-step-stats_cuda_gpu_trace.csv",
  "source_kernel_exec_trace": "benchmarks\\gb10-post-fc2-warmed-step-stats_cuda_kern_exec_trace.csv",
  "gpu_span_ns": 20907172640,
  "gpu_operation_count": 2803,
  "kernel_count": 2744,
  "kernel_time_ns": 20895065376,
  "kernel_busy_percent_of_span": 99.94209038109325,
  "launch_gaps": {
    "positive_gap_count": 2743,
    "total_ns": 11518912,
    "average_ns": 4199.384615384615,
    "maximum_ns": 4797728
  },
  "cpu_gpu_overlap": {
    "scope": "CUDA kernel-launch API intervals intersected with GPU kernel intervals",
    "launch_api_union_ns": 13065552528,
    "launch_api_gpu_overlap_ns": 13062771136,
    "launch_api_overlap_percent": 99.97871202160002
  },
  "components": {
    "sage2": {
      "milliseconds": 13029.23856,
      "percent_of_kernel_time": 62.355576905566124,
      "launches": 250
    },
    "nvfp4_gemms": {
      "milliseconds": 4167.855808,
      "percent_of_kernel_time": 19.946603339117498,
      "launches": 200
    },
    "nvfp4_packing": {
      "milliseconds": 2026.853664,
      "percent_of_kernel_time": 9.700154689767265,
      "launches": 800
    },
    "norm_and_rope": {
      "milliseconds": 1084.3856,
      "percent_of_kernel_time": 5.1896731619970025,
      "launches": 202
    },
    "remaining_gate_add": {
      "milliseconds": 549.298592,
      "percent_of_kernel_time": 2.6288436150619683,
      "launches": 100
    },
    "other": {
      "milliseconds": 37.433152,
      "percent_of_kernel_time": 0.17914828849014078,
      "launches": 1192
    }
  },
  "classification_policy": {
    "sage2": "mainloop plus MeanScale/TransposePadPermute/QuantInt8 preparation",
    "nvfp4_gemms": "SM120 block-scaled CUTLASS GEMMs",
    "nvfp4_packing": "absmax, final-scale, zero-fill and NVFP4 quantization kernels",
    "norm_and_rope": "layer norm, mean reduction and fused RMS/RoPE kernels",
    "remaining_gate_add": "fused residual gate/add kernels",
    "other": "all unmatched kernels"
  }
}