{
  "status": "passed",
  "generated_at": "2026-09-10T02:54:50.446866+00:00",
  "rows": [
    {
      "input_sha256": "f665a3650c243b5eb40736ffa5b9bbc7082d55f2c02138cea5a754e662d6cfa6",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "recomputed_errors": {
        "fp64_sample": {
          "max_abs": 8.546650151765312e-05,
          "relative_l2": 0.0017510778660335417
        },
        "saved_output": {
          "max_abs": 0.000244140625,
          "relative_l2": 0.0007356730879964922
        }
      },
      "kernels": [
        "void pytorch_flash::flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<128, 128, 64, 4, cutlass::bfloat16_t> >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)"
      ],
      "output_sha256": "0011df15c47b3c6124c409ddcd9bf77997ed721b93414fedda9096c580c9b78d",
      "trace_sha256": "20791c141c44c154a89fc63bb98221b10c787407e34465b3986c0036617cf261"
    },
    {
      "input_sha256": "3616141a1f20908c94a715a3571503688dd093c39a364b756e7ac96ab3571840",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "recomputed_errors": {
        "fp64_sample": {
          "max_abs": 0.0006142374604229672,
          "relative_l2": 0.0020143029066039112
        },
        "saved_output": {
          "max_abs": 0.00048828125,
          "relative_l2": 0.0002644235921410514
        }
      },
      "kernels": [
        "void pytorch_flash::flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<128, 128, 64, 4, cutlass::bfloat16_t> >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)"
      ],
      "output_sha256": "f08d66ee2ff56fc3b55ae07aa86ec8f0b8a0f70678df2ffcd0ba7d45e8e1dab4",
      "trace_sha256": "3c074afee07b71c8892725302c4732aff6539fc6d9f2628533eca9782aede7c2"
    },
    {
      "input_sha256": "706a473a57c5ed3f0867bec0e79feba3c35c91a4a221518e46b6fd1073d48807",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "recomputed_errors": {
        "fp64_sample": {
          "max_abs": 0.0002558632665645988,
          "relative_l2": 0.0017170595682587538
        },
        "saved_output": {
          "max_abs": 0.00048828125,
          "relative_l2": 0.00028158001154948226
        }
      },
      "kernels": [
        "void pytorch_flash::flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<128, 128, 64, 4, cutlass::bfloat16_t> >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)"
      ],
      "output_sha256": "1eb367f154daa6c1efab87e9481b4c411615b2df308a829631c96e2562964512",
      "trace_sha256": "8f3f9e785068b2e437078b7e9a8152d61da0fe121ab2105bcc3f38b6ce52989d"
    },
    {
      "input_sha256": "cd0f1f60b8acee6834efea178c2f7c4a7e944a95f2df157498c454f9268aa3ac",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "recomputed_errors": {
        "fp64_sample": {
          "max_abs": 0.00018629015193590004,
          "relative_l2": 0.0018716524974285657
        },
        "saved_output": {
          "max_abs": 0.0001220703125,
          "relative_l2": 0.0007455699301395639
        }
      },
      "kernels": [
        "void pytorch_flash::flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<128, 128, 64, 4, cutlass::bfloat16_t> >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)"
      ],
      "output_sha256": "2d9e321596ff1fbbfb6884a5847264165a2edf2e3f85cbf4f13e676188885b30",
      "trace_sha256": "e818a43c66849e28e45cbdc4a429764e0497505b04c6ad8a3490f48ea8b46e06"
    },
    {
      "input_sha256": "1a9829e52a9c8d8b4d17353f833d4c0749bcacac2ec5ef35936baa3fc3efc969",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "recomputed_errors": {
        "fp64_sample": {
          "max_abs": 6.036895268362458e-05,
          "relative_l2": 0.0017721556104615425
        },
        "saved_output": {
          "max_abs": 0.000244140625,
          "relative_l2": 0.0008029395356618093
        }
      },
      "kernels": [
        "void pytorch_flash::flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<128, 128, 64, 4, cutlass::bfloat16_t> >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)"
      ],
      "output_sha256": "d214f5c8ac2473f36ea1b5e78f2c48bad2c56b8329f0d6befa8fed6443c64b4f",
      "trace_sha256": "96fe396bcfcc38b5fee78caafbf02fdbc50f277daecbc3f2a30556dd8a5055ae"
    },
    {
      "input_sha256": "05122a53f7b3cba34235e7557234a1eb08746b495c88c0dd6aa6283aa6d72796",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "recomputed_errors": {
        "fp64_sample": {
          "max_abs": 7.629575161268234e-05,
          "relative_l2": 0.0018663348506979109
        },
        "saved_output": {
          "max_abs": 0.0001220703125,
          "relative_l2": 0.0007221844784401737
        }
      },
      "kernels": [
        "void pytorch_flash::flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<128, 128, 64, 4, cutlass::bfloat16_t> >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)"
      ],
      "output_sha256": "4687ce269e90e5c39f050d0e9e7e1ef8a5835c7cb211e5a1f1eeb6f213afe795",
      "trace_sha256": "4c38196619e4c5ae5164251fe24f4ee7ca1ef2340393e0ec7e08a4a3a39d5aef"
    }
  ],
  "source_module_sha256": "cde23067a3c4a2b2ce4957c8ebc7dfcb71b092d745b7f7a0c725fd5848f02b79",
  "patch_sha256": "7560113e1030c1f1a7061c06507f2d6d5b54dcafcd14d9e1220867111e98bb54",
  "scope": "CPU re-audit of six actual SDPAImpl Flash-only saved outputs and actual PyTorch Flash kernel traces. Valid lengths and original thresholds retained. This is a local implementation probe, not model correctness or model timing."
}