{
  "status": "TP4_NVME_SERVING_AND_BENCHMARKS_PASSED",
  "recorded": "2026-09-12",
  "base_image": "lmsysorg/sglang@sha256:3475d88ec3124867d9d6f7b3bd49afdcf8ef4d6a9f8454fab331d5adb5204de7",
  "base_index_digest": "sha256:e56358a68b06427362283c8c8a9d7d706448082ae53098ea1131b5d51aa1fd62",
  "base_platform": "linux/arm64",
  "sglang_commit": "e087e662ba1ac4ef7747537e2a9141085efd4561",
  "mia_url": "https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks",
  "mia_commit": "e59e6eb67479aa68f6fa700c600dc90a0729b5ec",
  "model_repo": "deepseek-ai/DeepSeek-V4.1-Flash",
  "model_revision": "df42c109f1defefcbfcedbe7d905718a12266e40",
  "model_config_sha256": "8be45ce0476004a3f529fd896115a4a2e800a129ad2d3ec05b16050f52e21879",
  "engram_total_bytes": 202758032400,
  "engram_average_bytes_per_rank": 50689508100.0,
  "base_declared_dependencies": {
    "flashinfer": "0.6.18",
    "cuda": "13.0.3",
    "nccl": "2.30.7"
  },
  "benchmark_source_sha256": "e0d6b2d25bd585d11fbdf39c2ddcdf7a4de8ab685af6bd42465e69f3ee6e80a8",
  "inspected_source_sha256": {
    "dspark_draft.py": "75ec519d908901cd589511bf98f280c67f7e6c378b3d17b41577e66601700f38",
    "server_args.py": "ebf54cd7a325b7e2bea506cc42185314e07ee08ead1ae9f9a556a403d29dde60",
    "dspark_worker_v2.py": "7361310ce8fb3643efa0ed6d98c3e31c291fb2e2a859d850d77bd6496106df4b",
    "metadata.py": "d5c4e3e75b0164bdc3d6a2b8e510814edd1d28ff615b56cefdabf5ea3bfaf56b",
    "engram.py": "e8aba7161f19f0320944dda1aa3608d014620ea808045e810de89bc32d615626",
    "fp8_utils.py": "c1b392059434e7b362b4f6afff16f2941d110840cb32632d68ff8459f2eafb09",
    "engram_gather.py": "86ab03bff1767085b5c1c3a22e5ac91bcb646d55a05a80bdcbebc4a761268306",
    "dspark_config.py": "73c3dc2986fa57eebcf61fdf47a817920a918f35e1414a724a0a0a3e44a7e577",
    "dspark_verify.py": "2ee03975f313ee6f5aacd92632d27daddf084e591758e588830563490f445db9",
    "fi0618__sparse_mla_sm120.py": "bddb1bd067554c3adef5179b46cd5214373218669d9d237344e10d38c3947698",
    "dspark_accept.py": "e233e4793b2e2a55f4aee846c24af93b21ffd5ff9ae3754c211490600b489c68",
    "environ.py": "27bfe2a3b90c76d8d233fc17d7c46906741c6102208fa3265f5a13b6faaac962"
  },
  "mia_base_note": "Mia uses a mutable dev-dsv41 tag and cites an older amd64 digest. This pins the ARM64 image resolved on 2026-09-11; it is not proof of the exact image used for Mia reported results.",
  "pinned_attention_source_sha256": "5d90df815f3161f3c43ec0126f9cb2eec85336e561917b111cab00e24d411132",
  "mia_attention_overlay_sha256": "fe7f78bead1f74efad00b5c1a5dfb36f71f55cffbb26518088e6f7ca99d971ae",
  "backports": [
    {
      "pr": 36655,
      "role": "Native H8/H16 decode with V4.1 fallback and page splitting retained"
    },
    {
      "pr": 39068,
      "role": "Five compression/index-processing files only; no TP4 WO-A or Q-RoPE changes"
    }
  ],
  "adapter_sha256": {
    "mxfp8_b12x.py": "2ec377aefa12682f34e2b4b000c0a131777294aeb5784bfd816b569d8f1f0276",
    "engram_backend.py": "40563d08621979227cec62a6feb14c226262a74044a1d00530f50866128fd175",
    "prefill_empty_cache.py": "24dadef526ec7dfc18527fd068d9b685a34cab1bf137a4591f0113e118d71b75",
    "sitecustomize.py": "8c3d022a828f31b5a0f1516d49e3ea4c285ee340c4443b2ce460dfa7781db277",
    "admission.py": "4f330f510983c79338a7eb3acdb9e700007a7186445c9473f51812de713f3a3b",
    "row_store.cpp": "9a9ed278c48bedd65a0e143535aa52391bd666bb67ca01a6bdb47aa4f8ecdb73"
  },
  "source_manifest": "patches/manifest.json",
  "expert_parallelism": {
    "global_tp": 4,
    "ep": 4,
    "moe_tp": 1,
    "target_local_experts": 96,
    "draft_local_experts": 32,
    "local_intermediate": 2304
  },
  "local_draft_context_fix": {
    "source": "python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py",
    "after_sha256": "8387704e671b1a67d116ccca9523092083eb0d299e654a5d6c7949f7245bf6df",
    "purpose": "Apply the existing speculative backend context through draft construction/capture/execution; record actual all-rank layout."
  },
  "description": "Measured TP4/EP4 NVMe Engram profile with four-rank RoCEnante; original checkpoint precision.",
  "sg17_reference_commit": "66aa6bd5085e63bbb011557075765e3191e807b2",
  "b12x_source_archive_sha256": "4922d0db9395fb8b537d7e216d509a5747867a45f20c18290e385786609dd05f",
  "runtime_configuration_sha256": "c1cc22b9aa0ecde171f81a1b1212e5e76be58a36b01725a9cf9aa4700faa2d5f",
  "base_native_head_image_id": "sha256:e7681b5276525821f9be286ed3bf0f51acb5239da27f69f60fdeb57e68c05581",
  "base_native_head_image_note": "Reuses the SG17 native-head image on each Spark; all eleven TP4 source overlays are mounted and verified. Runtime identity includes patches, adapters, frozen B12x sources and compiled row-store, not image ID alone.",
  "measured_runtime_commit": "25e7cb82f6524e3bbc46db6a9a54125e75e36d0e",
  "measurement_summary": "results/measurement-summary.json",
  "source_identity": "results/source-identity.json"
}