{
  "scope": "Same saved input, audited isolated interfaces. Distinct dispatch does not establish the cause of timing differences. Three within-process groups, not independent-launch significance.",
  "rows": [
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_034f783e4e41e1c0.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12187722663084666,
      "varlen_ms": 0.12889802659551303,
      "native_over_varlen_time": 0.9455321376897576,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_3019e6cfd6e84b2b.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 4.14625013589859,
      "varlen_ms": 4.1462031141916915,
      "native_over_varlen_time": 1.0000113409077183,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_53c9bd15f7acb7c0.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11889023996889592,
      "varlen_ms": 0.12857141335805258,
      "native_over_varlen_time": 0.9247019758412703,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_7c21104bed6fe513.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12042431977887948,
      "varlen_ms": 0.1294914132853349,
      "native_over_varlen_time": 0.9299791910798283,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_825f8b3d6e43b9fd.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12312213328977425,
      "varlen_ms": 0.12921663999557495,
      "native_over_varlen_time": 0.9528349699697392,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_8b7d4bb8fe8219a6.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12132501343886058,
      "varlen_ms": 0.12951669365167617,
      "native_over_varlen_time": 0.9367519353540139,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_a2b4332ea3af9d3b.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.7015460254748662,
      "varlen_ms": 0.7336599463224411,
      "native_over_varlen_time": 0.9562277850814266,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_a43a6e1d65520b5b.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 2.773186244169871,
      "varlen_ms": 2.7717079456647236,
      "native_over_varlen_time": 1.0005333529124016,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_b9e7999e0b6e56f3.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1287707734356324,
      "varlen_ms": 0.12860223973790805,
      "native_over_varlen_time": 1.0013105035967322,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_cccc73c58e3e0a6e.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.6859724791844686,
      "varlen_ms": 0.723376426299413,
      "native_over_varlen_time": 0.9482925545330633,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_eb19931361a68319.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 4.125397028923035,
      "varlen_ms": 4.136651938756307,
      "native_over_varlen_time": 0.9972792224243416,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp1_sp1_v1/captures/qkv_rank0_f38973a7f296b1fc.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 2.770487360159556,
      "varlen_ms": 2.757614399592082,
      "native_over_varlen_time": 1.0046681510545414,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_0c1b0b9bae2b8b48.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12500234675904115,
      "varlen_ms": 0.07314112027486165,
      "native_over_varlen_time": 1.7090570432786223,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_0e33ce888d278925.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.3441518916686376,
      "varlen_ms": 0.36670069257418314,
      "native_over_varlen_time": 0.9385089764972725,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_23b4681596dc64ac.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.3762392544746398,
      "varlen_ms": 1.363146344423294,
      "native_over_varlen_time": 1.0096049188738314,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_325560d6739aa74a.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1146253868440787,
      "varlen_ms": 0.07272234675784905,
      "native_over_varlen_time": 1.5762058288047063,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_37c9b2a3872f7438.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 2.038186988433202,
      "varlen_ms": 2.066770348151525,
      "native_over_varlen_time": 0.9861700358997857,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_5f58cf168d29ac00.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.338245120048523,
      "varlen_ms": 0.3616702934106191,
      "native_over_varlen_time": 0.9352305849031936,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_60db4e364fe9a358.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1151480533182621,
      "varlen_ms": 0.0726522666712602,
      "native_over_varlen_time": 1.5849203141766863,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_89e6d37b3d6f79ec.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11838815987110138,
      "varlen_ms": 0.07280533323685329,
      "native_over_varlen_time": 1.626091861786501,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_994f8af95ec0756f.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11624981331328552,
      "varlen_ms": 0.0734181334078312,
      "native_over_varlen_time": 1.5833937464403807,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_b29cb74704615250.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11583146661520004,
      "varlen_ms": 0.07326389340062936,
      "native_over_varlen_time": 1.5810170772906946,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_b3c02cfce9305c38.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.3413798391819,
      "varlen_ms": 1.3667474122842154,
      "native_over_varlen_time": 0.9814394577415596,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_480x832_tp1_sp2_v1/captures/qkv_rank0_c345c054e4ffb254.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 2.0687572276592254,
      "varlen_ms": 2.071820477247238,
      "native_over_varlen_time": 0.9985214695859737,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_13f5e4136c72875f.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.9871550919612249,
      "varlen_ms": 1.0039885846773784,
      "native_over_varlen_time": 0.9832333823580646,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_2419ddf76b22ab88.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10883381322026253,
      "varlen_ms": 0.0664642133563757,
      "native_over_varlen_time": 1.6374798966882296,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_5d233cbdc47812ff.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.16793258721629778,
      "varlen_ms": 0.17683818687995276,
      "native_over_varlen_time": 0.9496398384264109,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_6b4ac7dbc60c0c5a.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.17163765360911687,
      "varlen_ms": 0.1858622933924198,
      "native_over_varlen_time": 0.92346678003553,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_6bef1f209a2b6a6e.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1144806394726038,
      "varlen_ms": 0.06130080003291369,
      "native_over_varlen_time": 1.8675227633430027,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_6c601440672149ef.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11027957305312157,
      "varlen_ms": 0.06215882666409016,
      "native_over_varlen_time": 1.774157894727947,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_7baca46400faa6d9.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11105397333701451,
      "varlen_ms": 0.061795093193650244,
      "native_over_varlen_time": 1.7971325488416914,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_8ec3a76708f5e241.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11055807987848917,
      "varlen_ms": 0.06250271997104088,
      "native_over_varlen_time": 1.7688522984233896,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_a7ce4999e4f2ba20.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10886144027113914,
      "varlen_ms": 0.06251637314756711,
      "native_over_varlen_time": 1.7413268683097878,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_c6a5fcb1e5f12d0d.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.6779844270149866,
      "varlen_ms": 0.6801240543524424,
      "native_over_varlen_time": 0.9968540631319195,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_cf2e5e4d322f9347.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.9954527986049652,
      "varlen_ms": 1.0302844790617625,
      "native_over_varlen_time": 0.966192172002322,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/qkv_rank0_df3a5d6865cf4937.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.6672954668601354,
      "varlen_ms": 0.6754716779788336,
      "native_over_varlen_time": 0.9878955530109519,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp1_sp4_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_0c1b0b9bae2b8b48.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11991839987536271,
      "varlen_ms": 0.0734310399244229,
      "native_over_varlen_time": 1.6330750592499546,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_2e0b4394096e26d5.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.3645372811953227,
      "varlen_ms": 1.3757801612218221,
      "native_over_varlen_time": 0.9918279966935163,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_325560d6739aa74a.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11382688018182914,
      "varlen_ms": 0.07304693341255188,
      "native_over_varlen_time": 1.5582704826082394,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_60db4e364fe9a358.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11382186683515708,
      "varlen_ms": 0.07257024005055428,
      "native_over_varlen_time": 1.568437237576531,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_745464bfa49527f7.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 2.0705624544620513,
      "varlen_ms": 2.069998505115509,
      "native_over_varlen_time": 1.0002724394945932,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_7c5bdfbbb4c3e603.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.34783050686120986,
      "varlen_ms": 0.36884202698866525,
      "native_over_varlen_time": 0.943033823181703,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_89e6d37b3d6f79ec.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11410293338199456,
      "varlen_ms": 0.07260896007219951,
      "native_over_varlen_time": 1.5714718027711052,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_994f8af95ec0756f.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11776469337443511,
      "varlen_ms": 0.07306720013419787,
      "native_over_varlen_time": 1.6117312988337338,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_b29cb74704615250.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11590752022961776,
      "varlen_ms": 0.07280309334397315,
      "native_over_varlen_time": 1.592068618320775,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_e8b549dc16c13a1e.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 2.057416640520096,
      "varlen_ms": 2.063999364376068,
      "native_over_varlen_time": 0.9968106948240453,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_fb84ccff3c768374.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.33750986665487287,
      "varlen_ms": 0.3634714670976003,
      "native_over_varlen_time": 0.9285732091984096,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp1_v1/captures/qkv_rank0_fe264be86166ca15.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.3687993613878886,
      "varlen_ms": 1.3579630931218465,
      "native_over_varlen_time": 1.0079797958581704,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_13f5e4136c72875f.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.0151226669549942,
      "varlen_ms": 1.0288913071155548,
      "native_over_varlen_time": 0.9866179837798803,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_2419ddf76b22ab88.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10965600006282329,
      "varlen_ms": 0.060343040078878404,
      "native_over_varlen_time": 1.8172104010584258,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_5d233cbdc47812ff.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.1694986661275228,
      "varlen_ms": 0.18349130680163703,
      "native_over_varlen_time": 0.9237422147238781,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_6b4ac7dbc60c0c5a.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.1749981865286827,
      "varlen_ms": 0.18862197359402974,
      "native_over_varlen_time": 0.9277720044713907,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_6bef1f209a2b6a6e.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11002581348021825,
      "varlen_ms": 0.05992597340295712,
      "native_over_varlen_time": 1.8360288074150648,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_6c601440672149ef.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10822325331469376,
      "varlen_ms": 0.06040725339204073,
      "native_over_varlen_time": 1.7915605699257513,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_7baca46400faa6d9.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10873066671192647,
      "varlen_ms": 0.059257493341962494,
      "native_over_varlen_time": 1.8348846800600342,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_8ec3a76708f5e241.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10743231996893883,
      "varlen_ms": 0.060073600026468436,
      "native_over_varlen_time": 1.7883449622064291,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_a7ce4999e4f2ba20.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10944373354315758,
      "varlen_ms": 0.05992010666678349,
      "native_over_varlen_time": 1.8264943043538897,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_c6a5fcb1e5f12d0d.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.6897617067893346,
      "varlen_ms": 0.6921579736471176,
      "native_over_varlen_time": 0.9965379769517692,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_cf2e5e4d322f9347.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.0237431488434474,
      "varlen_ms": 1.0456963183482488,
      "native_over_varlen_time": 0.9790061711802925,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp2_sp2_v1/captures/qkv_rank0_df3a5d6865cf4937.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.6823570150136947,
      "varlen_ms": 0.6898210148016611,
      "native_over_varlen_time": 0.9891798022562237,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp2_sp2_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_2419ddf76b22ab88.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11035093322396278,
      "varlen_ms": 0.06385600009312233,
      "native_over_varlen_time": 1.7281216027160498,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_3c2cc055c01fd729.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.0042486411333085,
      "varlen_ms": 1.025128747622172,
      "native_over_varlen_time": 0.979631722808188,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_59302977417c4056.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.1736502399543921,
      "varlen_ms": 0.18202730630834899,
      "native_over_varlen_time": 0.9539790676253465,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_6bef1f209a2b6a6e.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10927936000128587,
      "varlen_ms": 0.0619287467127045,
      "native_over_varlen_time": 1.764598281122126,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_6c601440672149ef.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11114218672116598,
      "varlen_ms": 0.06335040009270111,
      "native_over_varlen_time": 1.75440386419866,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_7baca46400faa6d9.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.10930591995517412,
      "varlen_ms": 0.06229568000882864,
      "native_over_varlen_time": 1.7546308177338,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_8e3efe1e7972bcf2.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.16513066704074542,
      "varlen_ms": 0.1816379728416602,
      "native_over_varlen_time": 0.9091197421846106,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_8ec3a76708f5e241.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.11133994673689207,
      "varlen_ms": 0.0630199467142423,
      "native_over_varlen_time": 1.7667413659004192,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_97529d7d3b67351f.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.6693049595753352,
      "varlen_ms": 0.6715546679496766,
      "native_over_varlen_time": 0.9966499996475194,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_9bc0adfc73179efd.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.0006252815326055,
      "varlen_ms": 1.019128535191218,
      "native_over_varlen_time": 0.9818440431999672,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_a7ce4999e4f2ba20.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1509335463742415,
      "varlen_ms": 0.0627017600958546,
      "native_over_varlen_time": 2.4071660212329538,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "480x832_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_480x832_tp4_sp1_v1/captures/qkv_rank0_bf28d4d0099672a9.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.6760780807336172,
      "varlen_ms": 0.6790842670202255,
      "native_over_varlen_time": 0.9955731763602782,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_480x832_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_003f18b0d1c40f62.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.18776767988999685,
      "varlen_ms": 0.2647998934984207,
      "native_over_varlen_time": 0.7090927319089602,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_146bc622ce0024e2.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.18752384006977083,
      "varlen_ms": 0.266061332821846,
      "native_over_varlen_time": 0.7048143301429536,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_15833641ab5dc2f0.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 3.3910914103190106,
      "varlen_ms": 3.433674867947896,
      "native_over_varlen_time": 0.9875982848503256,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_227f18fd0026d1aa.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1871487998465697,
      "varlen_ms": 0.2653903989990552,
      "native_over_varlen_time": 0.7051830079476084,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_4624946207ce8f34.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 3.4813514661788942,
      "varlen_ms": 3.546550501982371,
      "native_over_varlen_time": 0.9816162110853819,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_6c1a0557389da998.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 20.206677010854087,
      "varlen_ms": 20.032400023142497,
      "native_over_varlen_time": 1.0086997557711634,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_6e8918fca4760b4e.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.188549013286829,
      "varlen_ms": 0.26605045328537624,
      "native_over_varlen_time": 0.7086964557229446,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_71ccfd9a9e128870.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 13.45603263537089,
      "varlen_ms": 13.434210577011108,
      "native_over_varlen_time": 1.0016243647689373,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_7cb64f86f262cbae.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 13.5761643409729,
      "varlen_ms": 13.530414950052897,
      "native_over_varlen_time": 1.0033812260073978,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_847f3e9ceb4b37a4.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.18822314699490866,
      "varlen_ms": 0.2654146131873131,
      "native_over_varlen_time": 0.7091664800764851,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_883584e23546203b.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 20.039921480814616,
      "varlen_ms": 20.037266171773275,
      "native_over_varlen_time": 1.0001325185291534,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_dit_only_704x1280_tp1_sp1_v2/captures/qkv_rank0_ef7fffa69ac75786.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1891430400311947,
      "varlen_ms": 0.26510207980871203,
      "native_over_varlen_time": 0.7134724864009871,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_021f565f50432429.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12725717278818288,
      "varlen_ms": 0.14241983994841576,
      "native_over_varlen_time": 0.893535429012385,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_03f37ea3bc112c48.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 6.928581328392029,
      "varlen_ms": 6.884411938985189,
      "native_over_varlen_time": 1.0064158550938411,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_2da4c70927569963.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.6880451174577078,
      "varlen_ms": 1.7203141311804453,
      "native_over_varlen_time": 0.9812423712984354,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_3682f84de105b397.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 10.287673508326213,
      "varlen_ms": 10.262020371754964,
      "native_over_varlen_time": 1.0024998134521206,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_468f321174193448.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 6.86066645304362,
      "varlen_ms": 6.853254612286886,
      "native_over_varlen_time": 1.001081506696606,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_61c120430b1af4e0.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12527754647036393,
      "varlen_ms": 0.14249194661776224,
      "native_over_varlen_time": 0.8791903643959872,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_c0f6ec677589277b.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1260874665280183,
      "varlen_ms": 0.14299178640047708,
      "native_over_varlen_time": 0.8817811826959427,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_c334b2b13ebc634a.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 10.346855341593425,
      "varlen_ms": 10.299845457077026,
      "native_over_varlen_time": 1.004564134938947,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_cbcd7690956bb027.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.7614090653260548,
      "varlen_ms": 1.765483413139979,
      "native_over_varlen_time": 0.9976922197152349,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_dfcab6f0a5779c76.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12259776016076405,
      "varlen_ms": 0.1421386665602525,
      "native_over_varlen_time": 0.8625222335879655,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_f5a94ff018466c49.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12324864027400811,
      "varlen_ms": 0.1418315732975801,
      "native_over_varlen_time": 0.8689788698558487,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp1_sp2_v1/captures/qkv_rank0_ff91739f1c7bab2a.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12637834635873635,
      "varlen_ms": 0.14192298675576845,
      "native_over_varlen_time": 0.8904712988898517,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_1ef9f10991fa086e.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12820618701477846,
      "varlen_ms": 0.08081055983901024,
      "native_over_varlen_time": 1.5865028935598167,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_6be6af8ca00bf828.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12592618681490422,
      "varlen_ms": 0.08117194664975007,
      "native_over_varlen_time": 1.5513510764779959,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_6e53ae344202af4b.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 3.4943771688143412,
      "varlen_ms": 3.516112316449483,
      "native_over_varlen_time": 0.9938184148630697,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_7a044869db5ec3c0.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 5.221868057250976,
      "varlen_ms": 5.246133125623067,
      "native_over_varlen_time": 0.9953746754435995,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_8d9062a68f7c0930.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.8616356253623962,
      "varlen_ms": 0.8817827206850052,
      "native_over_varlen_time": 0.9771518596928755,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_b0c50d31de479cf8.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.8846096009016037,
      "varlen_ms": 0.9015195735295614,
      "native_over_varlen_time": 0.981242811443624,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_b37e08eb3a08f049.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12548309328655402,
      "varlen_ms": 0.08034709346791108,
      "native_over_varlen_time": 1.5617626957055926,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_be6bbd23e5a64d4b.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12656277348597844,
      "varlen_ms": 0.08139509332676728,
      "native_over_varlen_time": 1.5549189553463842,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_c0afb7cc208e55f9.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 5.240111893018087,
      "varlen_ms": 5.275665820439657,
      "native_over_varlen_time": 0.9932607696105726,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_d12872a3fe20c4e4.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12633279994130134,
      "varlen_ms": 0.0809340801090002,
      "native_over_varlen_time": 1.5609345256183693,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_db45c3c5819dc1f4.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 3.5063396374384563,
      "varlen_ms": 3.533585387865702,
      "native_over_varlen_time": 0.9922894885968208,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp1_sp4",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_retry_704x1280_tp1_sp4_v1/captures/qkv_rank0_fcd1494e1f8885be.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1248234668125709,
      "varlen_ms": 0.08090720017751057,
      "native_over_varlen_time": 1.542798002386783,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp1_sp4_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_021f565f50432429.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.13284063977499802,
      "varlen_ms": 0.14160202672084174,
      "native_over_varlen_time": 0.9381266839978487,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_260748fc9c594def.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.6796830908457439,
      "varlen_ms": 1.7114493874708812,
      "native_over_varlen_time": 0.9814389506007651,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_61c120430b1af4e0.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12253002673387528,
      "varlen_ms": 0.14248928010463716,
      "native_over_varlen_time": 0.85992452656014,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_af3cea46fecfaf39.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 6.74043893178304,
      "varlen_ms": 6.745012811024984,
      "native_over_varlen_time": 0.9993218872417162,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_b5faac666beaef04.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 10.089247048695881,
      "varlen_ms": 10.101725021998087,
      "native_over_varlen_time": 0.9987647680693117,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_bb7d90c82a1e72ca.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 6.783341323534648,
      "varlen_ms": 6.7615478340784705,
      "native_over_varlen_time": 1.0032231509694183,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_c02ae993eecc25ee.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 1.7068827748298645,
      "varlen_ms": 1.7262891705830892,
      "native_over_varlen_time": 0.9887583169240007,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_c0f6ec677589277b.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12312277344365914,
      "varlen_ms": 0.14204736029108364,
      "native_over_varlen_time": 0.8667726960314911,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_c3e120abc9b34d7b.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 10.1260444800059,
      "varlen_ms": 10.123783798217774,
      "native_over_varlen_time": 1.0002233040366315,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_dfcab6f0a5779c76.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12221130671600501,
      "varlen_ms": 0.1424833065768083,
      "native_over_varlen_time": 0.8577236846347661,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_f5a94ff018466c49.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12442752023537954,
      "varlen_ms": 0.14222015981872876,
      "native_over_varlen_time": 0.8748936887286064,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp1_v1/captures/qkv_rank0_ff91739f1c7bab2a.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12378442679842314,
      "varlen_ms": 0.1426755196849505,
      "native_over_varlen_time": 0.8675940138277273,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_1ef9f10991fa086e.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12229823991656304,
      "varlen_ms": 0.08137845339874426,
      "native_over_varlen_time": 1.5028331801455717,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_6be6af8ca00bf828.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.13287893325090408,
      "varlen_ms": 0.08161482647061348,
      "native_over_varlen_time": 1.6281224747656473,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_6e53ae344202af4b.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 3.5638381838798523,
      "varlen_ms": 3.563025493621826,
      "native_over_varlen_time": 1.0002280899363423,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_7a044869db5ec3c0.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 5.335509451230367,
      "varlen_ms": 5.3451931762695315,
      "native_over_varlen_time": 0.9981883302025161,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_8d9062a68f7c0930.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.8730627202987671,
      "varlen_ms": 0.8971169084310532,
      "native_over_varlen_time": 0.9731872313338137,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_b0c50d31de479cf8.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.8995394146442414,
      "varlen_ms": 0.9164026641845703,
      "native_over_varlen_time": 0.9815984280716445,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_b37e08eb3a08f049.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12265642662843068,
      "varlen_ms": 0.08069472009936969,
      "native_over_varlen_time": 1.5200056023168331,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_be6bbd23e5a64d4b.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12084629322091739,
      "varlen_ms": 0.08226112008094788,
      "native_over_varlen_time": 1.4690572302200642,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_c0afb7cc208e55f9.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 5.345107639630635,
      "varlen_ms": 5.379622505505879,
      "native_over_varlen_time": 0.9935841472445475,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_d12872a3fe20c4e4.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12621503981451193,
      "varlen_ms": 0.08116416017214458,
      "native_over_varlen_time": 1.5550587789834454,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_db45c3c5819dc1f4.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 3.580957446893056,
      "varlen_ms": 3.5928801012039187,
      "native_over_varlen_time": 0.996681588593266,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp2_sp2",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp2_sp2_v1/captures/qkv_rank0_fcd1494e1f8885be.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12138101321955522,
      "varlen_ms": 0.0811755733191967,
      "native_over_varlen_time": 1.49528987916431,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp2_sp2_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_1ef9f10991fa086e.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12671733332177004,
      "varlen_ms": 0.08057013337810834,
      "native_over_varlen_time": 1.5727581426121895,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_5608acc7409c5075.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.8584346669912338,
      "varlen_ms": 0.8806887443860372,
      "native_over_varlen_time": 0.9747310527849229,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_597583b95ac01e65.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 0.8837016546726226,
      "varlen_ms": 0.8933813357353211,
      "native_over_varlen_time": 0.9891651183257244,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_5ff40347734fb848.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 5.208139723141988,
      "varlen_ms": 5.271096954345703,
      "native_over_varlen_time": 0.9880561424407475,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_6be6af8ca00bf828.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1246984534462293,
      "varlen_ms": 0.08048394689957301,
      "native_over_varlen_time": 1.5493580800880289,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_909ee6942527a295.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 3.511685225168864,
      "varlen_ms": 3.520047365029653,
      "native_over_varlen_time": 0.9976244240506921,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_aff3e85d5a3411c1.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 3.4977959537506105,
      "varlen_ms": 3.506738775571187,
      "native_over_varlen_time": 0.9974498180808692,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_b37e08eb3a08f049.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12439552033940951,
      "varlen_ms": 0.08086815988024076,
      "native_over_varlen_time": 1.538250907695059,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_be6bbd23e5a64d4b.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.1257201066116492,
      "varlen_ms": 0.0812822399288416,
      "native_over_varlen_time": 1.5467106556328993,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_c1119485a5572ce6.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 27,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "native_ms": 5.238194672266642,
      "varlen_ms": 5.285219944318135,
      "native_over_varlen_time": 0.9911024947784723,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_d12872a3fe20c4e4.pt",
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 9,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12443103996415933,
      "varlen_ms": 0.08056277344624202,
      "native_over_varlen_time": 1.544522794354762,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    },
    {
      "configuration": "704x1280_tp4_sp1",
      "input_file": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_evening_704x1280_tp4_sp1_v1/captures/qkv_rank0_fcd1494e1f8885be.pt",
      "context": {
        "phase": "commit",
        "step": null,
        "start_frame": 27,
        "role": "cross",
        "layer": "transformer.blocks.0.cross_attn"
      },
      "native_ms": 0.12475338677565256,
      "varlen_ms": 0.08072554647922515,
      "native_over_varlen_time": 1.5454015762873534,
      "both_sm90_fa3": true,
      "identical_main_kernel_names": false,
      "native_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<176>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<176> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, false, false, false, false>, flash::StaticPersistentTileScheduler<false> > >::Params)"
      ],
      "varlen_main_kernels": [
        "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > > >(flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<128>, cute::C<160>, cute::C<128> >, 128, cutlass::bfloat16_t, float, cutlass::arch::Sm90, false, false, false, true, false, false, false, true, true, true, false, false, cutlass::bfloat16_t, 1>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<128>, cute::C<128>, cute::C<160> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false, false, 1>, flash::VarlenDynamicPersistentTileScheduler<128, 160, 256, 128, false, true, true, false, false, true> > >::Params)"
      ],
      "native_has_varlen_preparation": false,
      "varlen_has_preparation": true,
      "audit": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/analysis/attention_evening_704x1280_tp4_sp1_audit.json"
    }
  ]
}
