{
  "schemaVersion": 1,
  "operationId": "20260823083139-1c0b23d8ce",
  "checkedAt": "2026-08-23T09:05:00Z",
  "status": "pinned-source-zero-runtime-hicache-audit",
  "revisions": {
    "sglangTag": "v0.5.18",
    "sglangCommit": "71de97b264b04dcd514cf904003028aefe9775c8",
    "glm52Fp8Revision": "ba978f7d347eaf65d22f1a86833408afdb953541"
  },
  "sourceReceipts": [
    {
      "label": "sglang-release-v0518",
      "url": "https://api.github.com/repos/sgl-project/sglang/releases/tags/v0.5.18",
      "bytes": 53374,
      "sha256": "2abaea076aa8553329dddd4201501ed94edde801dd93ca47669ef3a9de2ef8bf"
    },
    {
      "label": "sglang-glm52-cookbook",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/docs/cookbook/autoregressive/GLM/GLM-5.2.mdx",
      "bytes": 20574,
      "sha256": "ed999a20d87297acaf62a4f5c14fd0e3fec054077e71594edb630f54fcc92b11"
    },
    {
      "label": "sglang-hicache-best-practices",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/docs/docs/advanced_features/hicache_best_practices.mdx",
      "bytes": 8811,
      "sha256": "801690f88d2ec4c7c13e5de457c6ac51e8c6892e7b60df891fb454f6deb432d4"
    },
    {
      "label": "sglang-hicache-design",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/docs/docs/advanced_features/hicache_design.mdx",
      "bytes": 18982,
      "sha256": "7028577c9ea33607772d7a7add7832a82aab40cc4c2c2e23c507db7950ac83bc"
    },
    {
      "label": "sglang-server-args",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/python/sglang/srt/server_args.py",
      "bytes": 444424,
      "sha256": "5ed5badef8dca0ee3dd2b34a1c9c417e172ff1a947d9fe802f7983fb4a2be950"
    },
    {
      "label": "sglang-host-pool-base",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/python/sglang/srt/mem_cache/pool_host/base.py",
      "bytes": 13979,
      "sha256": "d5540a147704ce2724fb9dbd9b9e7eca052d572533b2c3922e603e68b58421d9"
    },
    {
      "label": "sglang-mla-host-pool",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/python/sglang/srt/mem_cache/pool_host/mla.py",
      "bytes": 26457,
      "sha256": "d2e1e26583cee8f19b3f5152ee86fb7bce6ba160490eaba9dcc1ae33068b49fc"
    },
    {
      "label": "sglang-host-pool-registry",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/python/sglang/srt/mem_cache/memory_pool_host.py",
      "bytes": 59309,
      "sha256": "65926ed8ab1e004cf49f975e687f1f5698a8d4c45b21c4e98acebdfb1f2d304e"
    },
    {
      "label": "sglang-device-memory-pool",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/python/sglang/srt/mem_cache/memory_pool.py",
      "bytes": 208853,
      "sha256": "41e87128dbdf9178dad65f695e7a98428ecfeebc1e3407fee57df5067f4205c3"
    },
    {
      "label": "sglang-hiradix-cache",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/python/sglang/srt/mem_cache/hiradix_cache.py",
      "bytes": 82244,
      "sha256": "1a88173716402ad9119ddf7dd07fb6cef663343827671dc84d5543581b88c903"
    },
    {
      "label": "sglang-hicache-storage",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/python/sglang/srt/mem_cache/hicache_storage.py",
      "bytes": 25362,
      "sha256": "de92d6c7228f3c93d889e1c054f867235055f74663b2808c5456e53ff040be86"
    },
    {
      "label": "sglang-dsa-host-pool-test",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/test/registered/unit/mem_cache/test_dsa_pool_host_unit.py",
      "bytes": 6366,
      "sha256": "e9051bf9673e9cbff2d33fd04c50730c2c6e9b80d1dd52bdd6bf2b34731aebed"
    },
    {
      "label": "sglang-hicache-variants-test",
      "url": "https://raw.githubusercontent.com/sgl-project/sglang/71de97b264b04dcd514cf904003028aefe9775c8/test/registered/hicache/test_hicache_variants.py",
      "bytes": 3777,
      "sha256": "3ff2d199e34d8e59f69ec0c9d94a1872855fe9f301c1502e1fffbcf366b8f192"
    },
    {
      "label": "glm52-fp8-config",
      "url": "https://huggingface.co/zai-org/GLM-5.2-FP8/resolve/ba978f7d347eaf65d22f1a86833408afdb953541/config.json",
      "bytes": 29464,
      "sha256": "22e49334abf8562fecf70ca3292ba3f5b33f5602fb2bf10b52dd64a66cfe65ff"
    },
    {
      "label": "glm52-fp8-model-card",
      "url": "https://huggingface.co/zai-org/GLM-5.2-FP8/resolve/ba978f7d347eaf65d22f1a86833408afdb953541/README.md",
      "bytes": 10909,
      "sha256": "de23c1b7cab43a99f0fedf4edf10de0d165882a2ecb2e2bad6c5796fcabf2e46"
    },
    {
      "label": "sglang-issue-30321",
      "url": "https://api.github.com/repos/sgl-project/sglang/issues/30321",
      "bytes": 12866,
      "sha256": "a5b1382ae281d352e1cc36602fd667bd3ad0d45f3005c5b6c81097891207cb9a"
    },
    {
      "label": "sglang-issue-30321-comments",
      "url": "https://api.github.com/repos/sgl-project/sglang/issues/30321/comments",
      "bytes": 50257,
      "sha256": "92ee298fe5853620910917241238c35b87829623392ae678e3bc10b9e2f5b6d0"
    },
    {
      "label": "sglang-pr-31443",
      "url": "https://api.github.com/repos/sgl-project/sglang/pulls/31443",
      "bytes": 34730,
      "sha256": "ba2ae2dc91e6dd5c1238fbfd29ee976a184c69b0790cd40eb60247d9c6017375"
    },
    {
      "label": "sglang-issue-31600",
      "url": "https://api.github.com/repos/sgl-project/sglang/issues/31600",
      "bytes": 9645,
      "sha256": "6ba2f83c2dcdd10df99507bcf97a2c615094c4c40e13675ee2dd720e2a02e961"
    },
    {
      "label": "sglang-issue-31600-comments",
      "url": "https://api.github.com/repos/sgl-project/sglang/issues/31600/comments",
      "bytes": 8703,
      "sha256": "666dfa0176fb54d2bb80d0d35ec8826617caa10bf4eab44c199d8f5f67e4a1e5"
    },
    {
      "label": "sglang-pr-30393",
      "url": "https://api.github.com/repos/sgl-project/sglang/pulls/30393",
      "bytes": 40632,
      "sha256": "62fb772a852d25c85e1262f1fdb5000f245cc991097d35eb89e953dd56579442"
    },
    {
      "label": "sglang-pr-32418",
      "url": "https://api.github.com/repos/sgl-project/sglang/pulls/32418",
      "bytes": 32216,
      "sha256": "8542ebde4c700b3953c3660ff560982d39ddb8b5b4e874b45fb43eb1200f80ad"
    },
    {
      "label": "sglang-pr-31427",
      "url": "https://api.github.com/repos/sgl-project/sglang/pulls/31427",
      "bytes": 34296,
      "sha256": "387240a08be2003aae815d754db08cc96e477afc71d0a589895b0f4609a367c7"
    },
    {
      "label": "sglang-pr-28896",
      "url": "https://api.github.com/repos/sgl-project/sglang/pulls/28896",
      "bytes": 40338,
      "sha256": "b8897b55ab6ac0710776ed592c80090450138fb476244d829a12552ae7719d7d"
    },
    {
      "label": "sglang-glm52-optimization-blog",
      "url": "https://lmsys.org/blog/2026-07-13-glm52-optimization/",
      "bytes": 70091,
      "sha256": "8bdc263748e08f0e3a30cb91326bc43158b398ce58e42fd506c2d769067c82db"
    },
    {
      "label": "zai-glm52-release",
      "url": "https://z.ai/blog/glm-5.2",
      "bytes": 598,
      "sha256": "63e94b9e11a2db64243ff18418011577f2a98ca7851ce9dcb0e97cfe34cc245a"
    },
    {
      "label": "zhipu-research",
      "url": "https://www.zhipuai.cn/zh/research",
      "bytes": 1236461,
      "sha256": "188772a5cb6b65eb65f02024d89d153b692ef3d3cb2c98ddd233352c0dd6ac80"
    }
  ],
  "model": {
    "artifact": "zai-org/GLM-5.2-FP8",
    "modelType": "glm_moe_dsa",
    "decoderLayers": 78,
    "mtpLayers": 1,
    "nativeContextTokens": 1048576,
    "kvLoraRank": 512,
    "qkRopeHeadDim": 64,
    "indexHeadDim": 128,
    "indexQuantBlockSize": 128,
    "fp8ScaleBytesPerBlock": 4,
    "bf16BytesPerElement": 2,
    "fp8BytesPerElement": 1
  },
  "geometry": {
    "bf16PrimaryBytesPerLayerToken": 1152,
    "fp8PrimaryBytesPerLayerToken": 656,
    "dsaIndexerBytesPerLayerToken": 132,
    "fp8PrimaryBreakdownBytes": {
      "quantizedNoRope": 512,
      "scales": 16,
      "bf16Rope": 128
    },
    "indexerBreakdownBytes": {
      "quantizedIndex": 128,
      "scale": 4
    }
  },
  "profiles": [
    {
      "id": "bf16-target-78",
      "primaryFormat": "bf16",
      "layers": 78,
      "primaryBytesPerLayerToken": 1152,
      "primaryBytesPerToken": 89856,
      "indexerBytesPerToken": 10296,
      "combinedBytesPerToken": 100152,
      "indexerOverheadPercentOfPrimary": 11.458333,
      "nativeContext": {
        "tokens": 1048576,
        "decimalGbPerRank": 105.016984,
        "gibPerRank": 97.804688,
        "decimalGbAcrossTp4": 420.067936,
        "decimalGbAcrossTp8": 840.135872
      },
      "fixedPools": [
        {
          "requestedPrimaryDecimalGb": 64,
          "rawTokenSlots": 712250,
          "pageAlignedTokenSlots": 712256,
          "primaryAllocatedDecimalGb": 64.000475,
          "dsaIndexerAllocatedDecimalGb": 7.333388,
          "combinedAllocatedDecimalGb": 71.333863,
          "combinedAllocatedGiB": 66.434837,
          "approximateFreeHostMemoryGateGiB": 76.434837,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 71.333863,
            "ranks4": 285.335452,
            "ranks8": 570.670904
          }
        },
        {
          "requestedPrimaryDecimalGb": 96,
          "rawTokenSlots": 1068376,
          "pageAlignedTokenSlots": 1068416,
          "primaryAllocatedDecimalGb": 96.003588,
          "dsaIndexerAllocatedDecimalGb": 11.000411,
          "combinedAllocatedDecimalGb": 107.003999,
          "combinedAllocatedGiB": 99.65524,
          "approximateFreeHostMemoryGateGiB": 109.65524,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 107.003999,
            "ranks4": 428.015996,
            "ranks8": 856.031992
          }
        },
        {
          "requestedPrimaryDecimalGb": 120,
          "rawTokenSlots": 1335470,
          "pageAlignedTokenSlots": 1335488,
          "primaryAllocatedDecimalGb": 120.00161,
          "dsaIndexerAllocatedDecimalGb": 13.750184,
          "combinedAllocatedDecimalGb": 133.751794,
          "combinedAllocatedGiB": 124.566065,
          "approximateFreeHostMemoryGateGiB": 134.566065,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 133.751794,
            "ranks4": 535.007176,
            "ranks8": 1070.014352
          }
        }
      ]
    },
    {
      "id": "bf16-target-plus-mtp-79",
      "primaryFormat": "bf16",
      "layers": 79,
      "primaryBytesPerLayerToken": 1152,
      "primaryBytesPerToken": 91008,
      "indexerBytesPerToken": 10428,
      "combinedBytesPerToken": 101436,
      "indexerOverheadPercentOfPrimary": 11.458333,
      "nativeContext": {
        "tokens": 1048576,
        "decimalGbPerRank": 106.363355,
        "gibPerRank": 99.058594,
        "decimalGbAcrossTp4": 425.45342,
        "decimalGbAcrossTp8": 850.90684
      },
      "fixedPools": [
        {
          "requestedPrimaryDecimalGb": 64,
          "rawTokenSlots": 703234,
          "pageAlignedTokenSlots": 703296,
          "primaryAllocatedDecimalGb": 64.005562,
          "dsaIndexerAllocatedDecimalGb": 7.333971,
          "combinedAllocatedDecimalGb": 71.339533,
          "combinedAllocatedGiB": 66.440118,
          "approximateFreeHostMemoryGateGiB": 76.440118,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 71.339533,
            "ranks4": 285.358132,
            "ranks8": 570.716264
          }
        },
        {
          "requestedPrimaryDecimalGb": 96,
          "rawTokenSlots": 1054852,
          "pageAlignedTokenSlots": 1054912,
          "primaryAllocatedDecimalGb": 96.005431,
          "dsaIndexerAllocatedDecimalGb": 11.000622,
          "combinedAllocatedDecimalGb": 107.006054,
          "combinedAllocatedGiB": 99.657153,
          "approximateFreeHostMemoryGateGiB": 109.657153,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 107.006054,
            "ranks4": 428.024216,
            "ranks8": 856.048432
          }
        },
        {
          "requestedPrimaryDecimalGb": 120,
          "rawTokenSlots": 1318565,
          "pageAlignedTokenSlots": 1318592,
          "primaryAllocatedDecimalGb": 120.002421,
          "dsaIndexerAllocatedDecimalGb": 13.750277,
          "combinedAllocatedDecimalGb": 133.752698,
          "combinedAllocatedGiB": 124.566907,
          "approximateFreeHostMemoryGateGiB": 134.566907,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 133.752698,
            "ranks4": 535.010792,
            "ranks8": 1070.021584
          }
        }
      ]
    },
    {
      "id": "fp8-target-78",
      "primaryFormat": "fp8",
      "layers": 78,
      "primaryBytesPerLayerToken": 656,
      "primaryBytesPerToken": 51168,
      "indexerBytesPerToken": 10296,
      "combinedBytesPerToken": 61464,
      "indexerOverheadPercentOfPrimary": 20.121951,
      "nativeContext": {
        "tokens": 1048576,
        "decimalGbPerRank": 64.449675,
        "gibPerRank": 60.023438,
        "decimalGbAcrossTp4": 257.7987,
        "decimalGbAcrossTp8": 515.5974
      },
      "fixedPools": [
        {
          "requestedPrimaryDecimalGb": 64,
          "rawTokenSlots": 1250781,
          "pageAlignedTokenSlots": 1250816,
          "primaryAllocatedDecimalGb": 64.001753,
          "dsaIndexerAllocatedDecimalGb": 12.878402,
          "combinedAllocatedDecimalGb": 76.880155,
          "combinedAllocatedGiB": 71.600224,
          "approximateFreeHostMemoryGateGiB": 81.600224,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 76.880155,
            "ranks4": 307.52062,
            "ranks8": 615.04124
          }
        },
        {
          "requestedPrimaryDecimalGb": 96,
          "rawTokenSlots": 1876172,
          "pageAlignedTokenSlots": 1876224,
          "primaryAllocatedDecimalGb": 96.00263,
          "dsaIndexerAllocatedDecimalGb": 19.317602,
          "combinedAllocatedDecimalGb": 115.320232,
          "combinedAllocatedGiB": 107.400335,
          "approximateFreeHostMemoryGateGiB": 117.400335,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 115.320232,
            "ranks4": 461.280928,
            "ranks8": 922.561856
          }
        },
        {
          "requestedPrimaryDecimalGb": 120,
          "rawTokenSlots": 2345215,
          "pageAlignedTokenSlots": 2345216,
          "primaryAllocatedDecimalGb": 120.000012,
          "dsaIndexerAllocatedDecimalGb": 24.146344,
          "combinedAllocatedDecimalGb": 144.146356,
          "combinedAllocatedGiB": 134.246756,
          "approximateFreeHostMemoryGateGiB": 144.246756,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 144.146356,
            "ranks4": 576.585424,
            "ranks8": 1153.170848
          }
        }
      ]
    },
    {
      "id": "fp8-target-plus-mtp-79",
      "primaryFormat": "fp8",
      "layers": 79,
      "primaryBytesPerLayerToken": 656,
      "primaryBytesPerToken": 51824,
      "indexerBytesPerToken": 10428,
      "combinedBytesPerToken": 62252,
      "indexerOverheadPercentOfPrimary": 20.121951,
      "nativeContext": {
        "tokens": 1048576,
        "decimalGbPerRank": 65.275953,
        "gibPerRank": 60.792969,
        "decimalGbAcrossTp4": 261.103812,
        "decimalGbAcrossTp8": 522.207624
      },
      "fixedPools": [
        {
          "requestedPrimaryDecimalGb": 64,
          "rawTokenSlots": 1234949,
          "pageAlignedTokenSlots": 1235008,
          "primaryAllocatedDecimalGb": 64.003055,
          "dsaIndexerAllocatedDecimalGb": 12.878663,
          "combinedAllocatedDecimalGb": 76.881718,
          "combinedAllocatedGiB": 71.60168,
          "approximateFreeHostMemoryGateGiB": 81.60168,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 76.881718,
            "ranks4": 307.526872,
            "ranks8": 615.053744
          }
        },
        {
          "requestedPrimaryDecimalGb": 96,
          "rawTokenSlots": 1852423,
          "pageAlignedTokenSlots": 1852480,
          "primaryAllocatedDecimalGb": 96.002924,
          "dsaIndexerAllocatedDecimalGb": 19.317661,
          "combinedAllocatedDecimalGb": 115.320585,
          "combinedAllocatedGiB": 107.400664,
          "approximateFreeHostMemoryGateGiB": 117.400664,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 115.320585,
            "ranks4": 461.28234,
            "ranks8": 922.56468
          }
        },
        {
          "requestedPrimaryDecimalGb": 120,
          "rawTokenSlots": 2315529,
          "pageAlignedTokenSlots": 2315584,
          "primaryAllocatedDecimalGb": 120.002825,
          "dsaIndexerAllocatedDecimalGb": 24.14691,
          "combinedAllocatedDecimalGb": 144.149735,
          "combinedAllocatedGiB": 134.249902,
          "approximateFreeHostMemoryGateGiB": 144.249902,
          "coResidentRankTotalsDecimalGb": {
            "ranks1": 144.149735,
            "ranks4": 576.59894,
            "ranks8": 1153.19788
          }
        }
      ]
    }
  ],
  "hicacheDefaults": {
    "pageSizeTokens": 64,
    "defaultRatio": 2,
    "defaultWritePolicy": "write_through",
    "defaultIoBackend": "kernel",
    "defaultMemoryLayout": "page_first",
    "defaultStoragePrefetchPolicy": "timeout",
    "storagePrefetchThresholdTokens": 256,
    "hostMemoryReserveGiB": 10,
    "prefetchTimeoutExamples": [
      {
        "tokens": 8192,
        "computedSeconds": 2.8
      },
      {
        "tokens": 65536,
        "computedSeconds": 8.4
      },
      {
        "tokens": 1048576,
        "computedSeconds": 30
      }
    ]
  },
  "compatibilityCases": [
    {
      "case": "L1 plus L2, radix cache enabled, MTP disabled",
      "sourceResult": "supported baseline shape",
      "pilotDecision": "start here"
    },
    {
      "case": "L1 plus L2 plus L3 with DCP size one",
      "sourceResult": "supported in principle with a configured storage backend",
      "pilotDecision": "add only after L2 parity"
    },
    {
      "case": "L3 with DCP size greater than one",
      "sourceResult": "unsupported",
      "pilotDecision": "reject"
    },
    {
      "case": "HiCache while radix cache is disabled",
      "sourceResult": "incompatible",
      "pilotDecision": "reject"
    },
    {
      "case": "optimistic prefill with L2 write_back",
      "sourceResult": "supported narrow combination",
      "pilotDecision": "separate experiment"
    },
    {
      "case": "optimistic prefill with L3 or write_through",
      "sourceResult": "unsupported combination",
      "pilotDecision": "reject"
    },
    {
      "case": "GLM-5.2 MTP with HiCache",
      "sourceResult": "packed and sidecar support merged, while related legacy and DSA restore work remains open",
      "pilotDecision": "enable only after target-only parity and acceptance-length gates"
    },
    {
      "case": "direct I/O with page_first",
      "sourceResult": "runtime resolves the layout to page_first_direct",
      "pilotDecision": "record the resolved layout"
    }
  ],
  "issueEvidence": [
    {
      "number": 30321,
      "stateAtCheck": "closed",
      "role": "historical GLM-5.2 MTP plus L2/L3 output-corruption canary",
      "resolutionEvidence": "PR 31443 merged usable-prefix clamping and the reporter said the L3 garbling stopped",
      "claimBoundary": "does not prove every v0.5.18 backend or topology is safe"
    },
    {
      "number": 31600,
      "stateAtCheck": "open",
      "role": "reported warm-hit TTFT improvement paired with worse end-to-end time and MTP acceptance",
      "reportedColdTtftSeconds": 31.5,
      "reportedStoreHitTtftSeconds": 4.2,
      "reportedColdTotalSeconds": 98.99,
      "reportedStoreHitTotalSeconds": 152.62,
      "claimBoundary": "user-reported SGLang v0.5.14 measurements are a gate-design input, not a transferable benchmark"
    }
  ],
  "runtimeBoundary": {
    "localGpuRuns": 0,
    "modelCalls": 0,
    "upstreamTestsExecutedLocally": 0,
    "sourceTestPresenceProvesExecution": false,
    "calculatedMemoryIsMeasuredRss": false,
    "issueMeasurementsArePublisherBenchmarks": false
  },
  "serp": {
    "requestId": "serpapi-4919986fe54f468a8d1225de11ad98cb",
    "query": "\"GLM-5.2\" HiCache",
    "organicResults": 9,
    "value": "high-value",
    "decisionImpact": "Confirmed a distinct GLM-5.2 HiCache intent, surfaced official SGLang guidance and two GLM-specific failure reports, and supported a sizing-plus-rollout audit rather than another prompt-caching page."
  },
  "overlapAudit": {
    "prepublicationSitemapCanonicals": 85,
    "registeredScopedPages": 78,
    "exactHiCachePages": 0,
    "nearestPage": "/guides/glm-5-2-prompt-caching/",
    "nearestPageIntent": "hosted API prompt-cache billing and behavior",
    "newPageIntent": "self-hosted SGLang L2/L3 KV offload memory and rollout",
    "distinctIntent": true
  },
  "decision": {
    "classification": "l2-first-fail-closed-pilot",
    "firstPilot": "Pinned SGLang v0.5.18, GLM-5.2, radix cache enabled, L2 only, write_back, MTP off, one fixed corpus, and a reversible HiCache flag.",
    "addL3When": "L2 answer parity, restored-prefix parity, TTFT, end-to-end latency, decode throughput, cache accounting, memory headroom, and eviction stress all pass.",
    "addMtpWhen": "Target-only L2/L3 is stable and restored hits preserve output parity, acceptance length, acceptance rate, and end-to-end latency.",
    "rejectWhen": "DCP greater than one with L3, disabled radix cache, unpinned runtime, missing per-rank host-memory budget, or a rollout judged only by warm-hit TTFT."
  },
  "invariants": {
    "twentySevenSourceReceipts": true,
    "allSourceHashesValid": true,
    "glm52DsaGeometry": true,
    "bf16PrimaryIs1152BytesPerLayerToken": true,
    "fp8PrimaryIs656BytesPerLayerToken": true,
    "indexerIs132BytesPerLayerToken": true,
    "targetFp8CombinedIs61464BytesPerToken": true,
    "targetBf16CombinedIs100152BytesPerToken": true,
    "everyFixedPoolIncludesIndexerAndReserve": true,
    "nativeContextCalculatedForEveryProfile": true,
    "prefetchTimeoutClampedAtThirtySeconds": true,
    "sourceTestsAreNotExecutions": true,
    "zeroLocalRuntime": true,
    "distinctFromApiPromptCaching": true
  }
}