{
  "status": "same_run_selected_mac_work_and_output_invariance_audited",
  "scope": "704x1280 TP1/SP4 eager diagnostic; 2 epochs, 10 chunks each; all recorded ranks summed once. Output40/40exact. DiT selection contains linear calls only, disjoint from flash attention core. VAE nominal dense convolution work may include halo/padded operations. Not full-model FLOPs, MFU or no-instrumentation timing. Excludes bias, normalization, activation, softmax, positional transformations, encoding and network. No full intermediate QKV/operator tensors for this metadata run.",
  "rows": [
    {
      "stage": "attention_cross",
      "calls": 16000,
      "nominal_mac_flops": 442918502400000
    },
    {
      "stage": "attention_self",
      "calls": 16000,
      "nominal_mac_flops": 41108373504000000
    },
    {
      "stage": "dit",
      "calls": 163200,
      "nominal_mac_flops": 34623521292288000
    },
    {
      "stage": "output",
      "calls": 20,
      "nominal_mac_flops": 0
    },
    {
      "stage": "vae",
      "calls": 7984,
      "nominal_mac_flops": 1635123619430400
    }
  ],
  "physical_mapping": [
    {
      "pid": 1893179,
      "rank": 0,
      "stage": "vae",
      "physical_gpu": 0,
      "uuid": "GPU-006eb78c-23cb-f37d-eb7c-0ccb578b8f11",
      "pci_bus_id": "00000000:08:00.0",
      "name": "NVIDIA L20X"
    },
    {
      "pid": 1896946,
      "rank": 0,
      "stage": "dit",
      "physical_gpu": 0,
      "uuid": "GPU-006eb78c-23cb-f37d-eb7c-0ccb578b8f11",
      "pci_bus_id": "00000000:08:00.0",
      "name": "NVIDIA L20X"
    },
    {
      "pid": 1893180,
      "rank": 1,
      "stage": "vae",
      "physical_gpu": 1,
      "uuid": "GPU-76ec3aee-7fec-b98c-562a-7697a4be04c7",
      "pci_bus_id": "00000000:7E:00.0",
      "name": "NVIDIA L20X"
    },
    {
      "pid": 1896947,
      "rank": 1,
      "stage": "dit",
      "physical_gpu": 1,
      "uuid": "GPU-76ec3aee-7fec-b98c-562a-7697a4be04c7",
      "pci_bus_id": "00000000:7E:00.0",
      "name": "NVIDIA L20X"
    },
    {
      "pid": 1893181,
      "rank": 2,
      "stage": "vae",
      "physical_gpu": 2,
      "uuid": "GPU-cbf66259-f4ab-0ede-1811-82037dde5924",
      "pci_bus_id": "00000000:A2:00.0",
      "name": "NVIDIA L20X"
    },
    {
      "pid": 1896948,
      "rank": 2,
      "stage": "dit",
      "physical_gpu": 2,
      "uuid": "GPU-cbf66259-f4ab-0ede-1811-82037dde5924",
      "pci_bus_id": "00000000:A2:00.0",
      "name": "NVIDIA L20X"
    },
    {
      "pid": 1893182,
      "rank": 3,
      "stage": "vae",
      "physical_gpu": 3,
      "uuid": "GPU-c2bead55-8621-7076-336f-43e4722b6a57",
      "pci_bus_id": "00000000:C6:00.0",
      "name": "NVIDIA L20X"
    },
    {
      "pid": 1896949,
      "rank": 3,
      "stage": "dit",
      "physical_gpu": 3,
      "uuid": "GPU-c2bead55-8621-7076-336f-43e4722b6a57",
      "pci_bus_id": "00000000:C6:00.0",
      "name": "NVIDIA L20X"
    }
  ],
  "sources": [
    {
      "path": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/operator_metadata_704_audit_v1.json",
      "sha256": "d7d7ace894dd819ee0ec0f353ceddd48c4ece5b783a83c9ca936b0b69e9680ca"
    },
    {
      "path": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/operator_metadata_704_attention_audit_v1.json",
      "sha256": "f39c5a1f1bf853f4b50faefd1df0fbaf4bf3bc0d746ee1e8ad59eab4611ba045"
    },
    {
      "path": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/operator_metadata_704_model_invariance_v1.json",
      "sha256": "041638ded13ced6ba24d953f0d1b4021fdaeece2be44c93fb0b893cc81f8c941"
    }
  ]
}