{
  "generated_at": "2026-09-10T05:44:54.844846+00:00",
  "status": "four_exact_input_hash_joins_checked",
  "rows": [
    {
      "case": "480_short",
      "input_sha256": "e86fd1d5515ffb6e8e8ba1c4a38e0b54b23bd7d8d4d931434a89f116a2186b96",
      "audit_sha256": "6267d155da94bb023f6a337e3b72c833d3d5ea3f077bbaef86c8c821dd810071",
      "backends": [
        {
          "backend": "vllm_fa2",
          "q_k_h_d": [
            4680,
            4680,
            10,
            128
          ],
          "qk_av_nominal_flops": 112140288000,
          "ncu_total_kernel_us": 332.48,
          "ncu_dram_bytes": 41279744.0,
          "nominal_flops_per_dram_byte": 2716.5935912780856,
          "nominal_tflops_under_ncu": 337.28431183830605,
          "main_kernel_sm_pct": 54.1837,
          "main_kernel_dram_pct": 2.580393,
          "main_kernel_l2_pct": 18.019752,
          "main_kernel_occupancy_pct": 10.632445,
          "clean_batch_mean_us": [
            334.333972632885,
            330.7521065076192,
            331.19584023952484
          ],
          "clean_timing_sources": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_480x832_tp1_sp4_v1/timings.jsonl"
          ],
          "clean_dispatch_traces": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_480x832_tp1_sp4_v1/qkv_rank0_6b4ac7dbc60c0c5a_vllm_fa2.json"
          ]
        },
        {
          "backend": "vllm_fa3",
          "q_k_h_d": [
            4680,
            4680,
            10,
            128
          ],
          "qk_av_nominal_flops": 112140288000,
          "ncu_total_kernel_us": 179.744,
          "ncu_dram_bytes": 41229824.0,
          "nominal_flops_per_dram_byte": 2719.882772237883,
          "nominal_tflops_under_ncu": 623.8889086701085,
          "main_kernel_sm_pct": 72.757391,
          "main_kernel_dram_pct": 4.846778,
          "main_kernel_l2_pct": 28.317732,
          "main_kernel_occupancy_pct": 18.720104,
          "clean_batch_mean_us": [
            173.89653384685516,
            187.16480056444803,
            185.86229339241982
          ],
          "clean_timing_sources": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_480x832_tp1_sp4_v1/timings.jsonl"
          ],
          "clean_dispatch_traces": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_480x832_tp1_sp4_v1/qkv_rank0_6b4ac7dbc60c0c5a_vllm_fa3.json"
          ]
        }
      ],
      "clean_mean_fa2_over_fa3": 1.8216106760294681,
      "ncu_fa2_over_fa3": 1.8497418550827844
    },
    {
      "case": "704_short",
      "input_sha256": "468353cb47a9d93b52d299d1edd226524978db730a64597becca38700ad6c050",
      "audit_sha256": "c27a41e71adf40bdccfc55a0ced42afc4f98d341365bc6f7eec8df6917d3a37b",
      "backends": [
        {
          "backend": "vllm_fa2",
          "q_k_h_d": [
            10560,
            10560,
            10,
            128
          ],
          "qk_av_nominal_flops": 570949632000,
          "ncu_total_kernel_us": 1594.7839999999999,
          "ncu_dram_bytes": 110585088.0,
          "nominal_flops_per_dram_byte": 5162.989353501261,
          "nominal_tflops_under_ncu": 358.01063466902104,
          "main_kernel_sm_pct": 55.544556,
          "main_kernel_dram_pct": 1.441039,
          "main_kernel_l2_pct": 22.594491,
          "main_kernel_occupancy_pct": 11.923779,
          "clean_batch_mean_us": [
            1617.91979153951,
            1628.97140900294,
            1629.786559343338
          ],
          "clean_timing_sources": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/timings.jsonl"
          ],
          "clean_dispatch_traces": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/qkv_rank0_b0c50d31de479cf8_vllm_fa2.json"
          ]
        },
        {
          "backend": "vllm_fa3",
          "q_k_h_d": [
            10560,
            10560,
            10,
            128
          ],
          "qk_av_nominal_flops": 570949632000,
          "ncu_total_kernel_us": 821.184,
          "ncu_dram_bytes": 102783232.0,
          "nominal_flops_per_dram_byte": 5554.890821102025,
          "nominal_tflops_under_ncu": 695.276128127192,
          "main_kernel_sm_pct": 75.16756,
          "main_kernel_dram_pct": 2.610327,
          "main_kernel_l2_pct": 23.927212,
          "main_kernel_occupancy_pct": 18.750561,
          "clean_batch_mean_us": [
            896.3633062442143,
            902.1849606434505,
            901.5195735295614
          ],
          "clean_timing_sources": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/timings.jsonl"
          ],
          "clean_dispatch_traces": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/qkv_rank0_b0c50d31de479cf8_vllm_fa3.json"
          ]
        }
      ],
      "clean_mean_fa2_over_fa3": 1.8061315670987816,
      "ncu_fa2_over_fa3": 1.9420543995012078
    },
    {
      "case": "704_long",
      "input_sha256": "8ab8d8e91a6d5757a64db4da0bf7d6b02a4ab0fb6c1bc2ca2c2524f9a1e2f113",
      "audit_sha256": "37098c366485c99cf8620310e633c72011438f54cf6babf115a5d98969c64cb7",
      "backends": [
        {
          "backend": "vllm_fa2",
          "q_k_h_d": [
            10560,
            63360,
            10,
            128
          ],
          "qk_av_nominal_flops": 3425697792000,
          "ncu_total_kernel_us": 9490.591999999999,
          "ncu_dram_bytes": 555490304.0,
          "nominal_flops_per_dram_byte": 6166.980354710206,
          "nominal_tflops_under_ncu": 360.95722922237104,
          "main_kernel_sm_pct": 55.830852,
          "main_kernel_dram_pct": 1.216341,
          "main_kernel_l2_pct": 24.511834,
          "main_kernel_occupancy_pct": 11.919707,
          "clean_batch_mean_us": [
            9629.564447402954,
            9627.407900492351,
            9623.962885538736
          ],
          "clean_timing_sources": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/timings.jsonl"
          ],
          "clean_dispatch_traces": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/qkv_rank0_7a044869db5ec3c0_vllm_fa2.json"
          ]
        },
        {
          "backend": "vllm_fa3",
          "q_k_h_d": [
            10560,
            63360,
            10,
            128
          ],
          "qk_av_nominal_flops": 3425697792000,
          "ncu_total_kernel_us": 4683.872,
          "ncu_dram_bytes": 569264640.0,
          "nominal_flops_per_dram_byte": 6017.759669738138,
          "nominal_tflops_under_ncu": 731.3815988139727,
          "main_kernel_sm_pct": 77.71574,
          "main_kernel_dram_pct": 2.527247,
          "main_kernel_l2_pct": 26.768563,
          "main_kernel_occupancy_pct": 18.748163,
          "clean_batch_mean_us": [
            5246.133125623067,
            5274.914247194925,
            5237.803192138672
          ],
          "clean_timing_sources": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/timings.jsonl"
          ],
          "clean_dispatch_traces": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/qkv_rank0_7a044869db5ec3c0_vllm_fa3.json"
          ]
        }
      ],
      "clean_mean_fa2_over_fa3": 1.8326803160160219,
      "ncu_fa2_over_fa3": 2.026227873007631
    },
    {
      "case": "704_cross",
      "input_sha256": "ee5e993223e93b3cc62c90d337f72eff3586309fffbc0dc2984b73efc41077c4",
      "audit_sha256": "a240d262d4e923f224e933b658b29686b466621076830d0028e0bca29f6ae0af",
      "backends": [
        {
          "backend": "vllm_fa2",
          "q_k_h_d": [
            10560,
            512,
            10,
            128
          ],
          "qk_av_nominal_flops": 27682406400,
          "ncu_total_kernel_us": 92.928,
          "ncu_dram_bytes": 39737600.0,
          "nominal_flops_per_dram_byte": 696.6300531486552,
          "nominal_tflops_under_ncu": 297.8909090909091,
          "main_kernel_sm_pct": 46.380532,
          "main_kernel_dram_pct": 8.888152,
          "main_kernel_l2_pct": 19.439977,
          "main_kernel_occupancy_pct": 11.87278,
          "clean_batch_mean_us": [
            95.68992001314957,
            95.87050673862298,
            95.90293320516744
          ],
          "clean_timing_sources": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/timings.jsonl"
          ],
          "clean_dispatch_traces": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/qkv_rank0_b37e08eb3a08f049_vllm_fa2.json"
          ]
        },
        {
          "backend": "vllm_fa3",
          "q_k_h_d": [
            10560,
            512,
            10,
            128
          ],
          "qk_av_nominal_flops": 27682406400,
          "ncu_total_kernel_us": 80.41600000000001,
          "ncu_dram_bytes": 39620864.0,
          "nominal_flops_per_dram_byte": 698.6825527075836,
          "nominal_tflops_under_ncu": 344.24003183446075,
          "main_kernel_sm_pct": 46.370683,
          "main_kernel_dram_pct": 10.610482,
          "main_kernel_l2_pct": 22.043751,
          "main_kernel_occupancy_pct": 18.452298,
          "clean_batch_mean_us": [
            80.02912017206351,
            80.39509368439515,
            80.34709346791108
          ],
          "clean_timing_sources": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/timings.jsonl"
          ],
          "clean_dispatch_traces": [
            "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/attention_replay_retry_704x1280_tp1_sp4_v1/qkv_rank0_b37e08eb3a08f049_vllm_fa3.json"
          ]
        }
      ],
      "clean_mean_fa2_over_fa3": 1.1939269805503285,
      "ncu_fa2_over_fa3": 1.1555909271786706
    }
  ],
  "scope": [
    "NCU and prior uninstrumented CUDA-event replay are different runs and allocations. Ratios shown separately; no profiler-overhead estimate.",
    "Nominal work includes QK and AV only (4QKHD), excludes softmax and auxiliary operations. Not executed instruction FLOPs or model MFU.",
    "DRAM low percent does not prove compute bound; SM aggregate is not tensor-core utilization. No controlled hardware bottleneck ablation.",
    "Three clean timing batches are within one process, not independent launches."
  ]
}
