{
  "scope": "H0 local source mapping for H1\u2013H5. Source evidence plus one actual QKV metadata example; complete compiled dispatch/transport attribution and external PR status remain unverified.",
  "rows": [
    {
      "experiment_id": "H1",
      "reference_url": "https://github.com/vllm-project/vllm-omni/pull/6340",
      "inspected_sha": "b9cdea7b6da42d51cd5772548c616655fec21668",
      "local_symbol": {
        "path": "/data/zhoutaichang/feature/vllm-omni-uspbench/vllm_omni/diffusion/models/lingbot_world/transformer.py",
        "symbol": "LingBot self attention forward: fused QKV projection and 5D QKV exchange; _LingBotRMSNorm.forward",
        "sha256": "ffb0dc9fdc429b4871cbca9e0a4a332663f0469c79c2d05057d9220a9b3e04b1",
        "archived": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/h0_lingbot_source_mapping_v2/transformer.py"
      },
      "preconditions": [
        "Preserve head/sequence ownership",
        "Measure incremental packing/unpacking savings and actual transport",
        "Retain exact output and graph lifecycle semantics"
      ],
      "already_enabled": "Fused QKV 5D exchange already in pinned LingBot self path when Ulysses>1; observed qkv_exchange metadata confirms before/after5D shapes.",
      "hypothesis": "Potential layout/materialization reduction if compatible and not already enabled",
      "rejection_reason": "Do not claim combining three Q/K/V exchanges as a new optimization. Current inspected implementation uses dist.all_to_all_single and layout materializations; native SymmMem is not evidenced by this path.",
      "status": "existing_fused_exchange_incremental_layout_optimization_pending",
      "additional_evidence": {
        "path": "/data/zhoutaichang/feature/vllm-omni-uspbench/vllm_omni/diffusion/distributed/comm.py",
        "symbol": "all_to_all_5D: reshape/transpose/contiguous, dist.all_to_all_single",
        "sha256": "daac23a33e8110c3f42f56cf6d02492542e3625361c39f9be113ee4d5bedbf5c",
        "archived": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/h0_lingbot_source_mapping_v2/comm.py"
      }
    },
    {
      "experiment_id": "H3",
      "reference_url": "https://github.com/vllm-project/vllm-omni/pull/6607",
      "inspected_sha": "b9cdea7b6da42d51cd5772548c616655fec21668",
      "local_symbol": {
        "path": "/data/zhoutaichang/feature/vllm-omni-uspbench/vllm_omni/diffusion/models/minimax_h3/ops/vae/dispatch.py",
        "symbol": "H3VAEOperatorSet / resolve_h3_vae_operators",
        "sha256": "7c7df96a99d6e3bb63fdfe902c142bd78c9dbcb614d4b9b05601ce296a535f83"
      },
      "preconditions": [
        "Matching qk_norm_rope or scaled_residual semantics and shape",
        "Preserve LingBot FP32 Wan decoder semantics"
      ],
      "already_enabled": false,
      "hypothesis": "Inspect equivalent operations and data movement before adapting model-specific fusion",
      "rejection_reason": "Frozen LingBot worker loads AutoencoderKLWan FP32 and installs width spatial sharding; H3 operator set is not a drop-in decoder replacement",
      "additional_evidence": {
        "path": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/async_vae_usp4_spatial4_20260908/measured/D_r0/async_vae_spatial_worker.py",
        "symbol": "main: AutoencoderKLWan / install_wan_spatial_shard_decode",
        "sha256": "c0d863bdd20ae55596048b327f612cd70d33568e6e7077781b6de3d4bb4556e8"
      },
      "status": "direct_replacement_incompatible_fusion_opportunities_unverified"
    },
    {
      "experiment_id": "H4",
      "reference_url": "https://vllm.ai/blog/2026-09-01-minimax-h3-production-serving",
      "inspected_sha": "b9cdea7b6da42d51cd5772548c616655fec21668",
      "local_symbol": {
        "path": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/async_vae_usp4_spatial4_20260908/measured/D_r0/async_vae_spatial_worker.py",
        "symbol": "pixels round/clamp/to(uint8)/cpu",
        "sha256": "c0d863bdd20ae55596048b327f612cd70d33568e6e7077781b6de3d4bb4556e8"
      },
      "preconditions": [
        "Byte-accurate output and rounding/channel semantics",
        "Timing includes conversion and D2H"
      ],
      "already_enabled": "uint8 conversion before CPU transfer present in frozen worker",
      "hypothesis": "Only incremental fusion/staging/reuse savings can be claimed",
      "rejection_reason": "Cannot claim existing uint8 payload reduction as a new optimization",
      "status": "existing_source_path_no_new_ab"
    },
    {
      "experiment_id": "H5",
      "reference_url": "https://github.com/vllm-project/vllm-omni/issues/6872",
      "inspected_sha": "b9cdea7b6da42d51cd5772548c616655fec21668",
      "local_symbol": {
        "path": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/async_vae_usp4_spatial4_20260908/measured/D_r0/async_vae_spatial_lingbot.py",
        "symbol": "mode D submit/collect_decode sequential path",
        "sha256": "3f527395e30fa5db12eb4bc42e234906cbed5a2cf38ca646ac840d9ed8ba28b4"
      },
      "preconditions": [
        "Chunk ownership, bounded queues, event dependencies",
        "Backpressure/cancel/flush correctness",
        "Measured resource contention"
      ],
      "already_enabled": "Frozen Case D is sequential; alternative mode E source presence does not prove current use",
      "hypothesis": "Independent overlap experiments after bottleneck attribution",
      "rejection_reason": null,
      "status": "not_executed"
    },
    {
      "experiment_id": "H2",
      "reference_url": "https://vllm.ai/blog/2026-09-01-minimax-h3-production-serving",
      "inspected_sha": "b9cdea7b6da42d51cd5772548c616655fec21668",
      "local_symbol": {
        "path": "/data/zhoutaichang/feature/vllm-omni-uspbench/vllm_omni/diffusion/models/lingbot_world/transformer.py",
        "symbol": "LingBot self attention forward: fused QKV projection and 5D QKV exchange; _LingBotRMSNorm.forward",
        "sha256": "ffb0dc9fdc429b4871cbca9e0a4a332663f0469c79c2d05057d9220a9b3e04b1",
        "archived": "/data/zhoutaichang/feature/Lingbot_world_realtime/runs/analysis_report_20260908/raw/h0_lingbot_source_mapping_v2/transformer.py"
      },
      "preconditions": [
        "Preserve FP32 RMS accumulation and TP global sum-of-squares reduction",
        "Preserve rotary layout, modulation and output dtype",
        "Compare actual compiled dispatch before claiming further fusion"
      ],
      "already_enabled": "QKVParallelLinear packed projection exists; source RMSNorm/RoPE/modulation expressions may be fused by compile, exact compiled operator mapping still missing",
      "hypothesis": "Reduce incremental norm/rotary/materialization overhead only where trace and counters justify it",
      "rejection_reason": "Cannot replace TP global RMS statistics with independent local/head normalization or claim packed QKV as new",
      "status": "source_mapped_runtime_fusion_attribution_pending"
    }
  ],
  "observed_qkv_example": {
    "file": "Lingbot_world_realtime/runs/analysis_report_20260908/raw/capture_cross_480x832_tp1_sp4_v1/captures/metadata_rank0_pid1956321.jsonl",
    "line": 4,
    "record": {
      "kind": "qkv_exchange",
      "rank": 0,
      "context": {
        "phase": "denoise",
        "step": 0,
        "start_frame": 0,
        "role": "self",
        "layer": "transformer.blocks.0.self_attn"
      },
      "before": {
        "shape": [
          1,
          1170,
          3,
          40,
          128
        ],
        "stride": [
          17971200,
          15360,
          5120,
          128,
          1
        ],
        "dtype": "torch.bfloat16",
        "device": "cuda:0",
        "bytes": 35942400
      },
      "after": {
        "shape": [
          1,
          4680,
          3,
          10,
          128
        ],
        "stride": [
          17971200,
          3840,
          1280,
          128,
          1
        ],
        "dtype": "torch.bfloat16",
        "device": "cuda:0",
        "bytes": 35942400
      }
    },
    "limitation": "Tensor bytes describe logical buffer size, not measured NCCL wire traffic or latency."
  }
}