Publish accepted Thor LPK 0.5.16 and performance report

.gitattributes

@@ -36,3 +36,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 cloud.lazycat.aipod.glm53f-v0.5.0.lpk filter=lfs diff=lfs merge=lfs -text
 cloud.lazycat.aipod.glm53f-v0.5.1.lpk filter=lfs diff=lfs merge=lfs -text
 cloud.lazycat.aipod.glm53f-v0.5.2.lpk filter=lfs diff=lfs merge=lfs -text
+cloud.lazycat.aipod.glm53f-v0.5.16.lpk filter=lfs diff=lfs merge=lfs -text

PERFORMANCE-0.5.16.md

@@ -0,0 +1,63 @@
+# LPK 0.5.16 默认 8 路安装后复测
+
+0.5.16 小包通过 LPK / Model Host 实际安装并部署到 192.168.1.124 和 192.168.1.104。默认最大并发改为 8 路，单请求仍为 1M 上下文，共享 KV 池为 2,676,736 tokens（4 路时为 2,684,928 tokens）。
+
+推理镜像和所有运行源码、原生模块与上次验收相同；仅 LPK 控制器的默认并发及对应 64 行共享验证窗口改变。安装前基线使用此前在同一双机、同一推理镜像、临时 8 路配置下采集的三轮原始成绩。
+
+## 相同 8 路配置的三轮中位数
+
+未修改的 sparkDash 1.8.9 采集器；每路 400 输出 tokens、32-token 解码预热、temperature=0、top_p=1、关闭思考。冷 Prefill 使用独立 nonce，并由累计缓存计数确认未命中。
+
+| 指标 | 安装前 tok/s | 安装后 tok/s | 吞吐变化 | 安装前 TTFT ms | 安装后 TTFT ms |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| C1 decode aggregate | 49.31 | 49.63 | +0.65% | 261.93 | 261.86 |
+| C8 decode aggregate | 106.12 | 105.50 | -0.58% | 648.76 | 651.14 |
+| 8K cold prefill | 791.84 | 820.75 | +3.65% | 10372.04 | 10010.35 |
+| 32K cold prefill | 865.77 | 905.86 | +4.63% | 37878.35 | 36204.30 |
+
+预先声明的工程门禁均通过：各项吞吐下降不超过 3%，冷 Prefill TTFT 增长不超过 3%，短提示 TTFT 增长不超过 10%。完整三轮样本和门禁计算保留在发行证据中；这不等于统计等价性证明。
+
+8 并发各流生成速度的中位数为 13.81 tok/s；表中 8 并发值为合计吞吐。
+
+## 完整 1M 与 SSD 历史恢复
+
+安装后在默认 8 路配置下，新随机请求实际计算 1,048,448 个输入 tokens，缓存命中 0，Prefill 667.91 tok/s，TTFT 1570.87 秒；前、中、后三处检索与完整 token 计数通过。
+
+此前默认 4 路、同长度的单次结果为 661.92 tok/s、TTFT 1588.17 秒。并发上限与随机输入均不同，仅供参考，不用于上述 3% 门禁。
+
+安装前保留的三个 64K 历史在新版 LPK 停止/启动后从 SSD 恢复，逐条验证磁盘加载计数增加：
+
+| 历史 | 缓存命中 | TTFT 秒 |
+| --- | ---: | ---: |
+| 64K-1 | 100.000% | 1.124 |
+| 64K-2 | 100.000% | 0.365 |
+| 64K-3 | 100.000% | 0.366 |
+
+新 1M 历史在实际 LPK 重启后恢复：缓存命中 100.000%，TTFT 9.562 秒，输出与冷请求逐字一致，SSD 加载计数增加。
+
+并发从 4 改为 8 会改变 TensorFold 的 SSD 兼容标识。两机旧命名空间各有 39 份快照；原件全部保留，缺失的 38 份/节点经保留原件的头部标识迁移后，新服务索引有 45 项。三条 64K 和一条旧 1M 历史已跨配置验证 100% 缓存命中、逐字输出一致和实际读盘；其余文件完成双 rank 结构/启动扫描，未逐条重放。旧 1M 续问 TTFT 为 9.56 秒。
+
+## 包、代码与来源
+
+LPK 大小 6,652,416 bytes，SHA256 `5721567582001061f2fece5ac0157e42f83d7867558989b9ccddb1844cf1f82d`。
+
+最终推理镜像 ID `sha256:5149447c3091bee8b7f0816e26f2e1f8fd72968197544ba0fda65a7caa5103c6`。两端源文件和原生模块再次读回校验；模型文件的路径、大小、mtime 与 inode 在安装前后完全一致。
+
+LPK 与运行归档仅通过用户 manateelazycat 的 Hugging Face / ModelScope 账户发布。小包复用下列已经公开且固定 revision 的运行归档依赖链：
+
+| 归档 | bytes | HF 固定 revision | MS 固定 revision |
+| --- | ---: | --- | --- |
+| glm-5.3-flash-dual-t5000-runtime-0.5.14.tar.gz | 3,173,705,404 | `fb520c24f72d189d8d43136c8291a101fce21b3c` | `f47dfedb7662453cdbaf197626d0b6c14bb19eb9` |
+| glm-5.3-flash-dual-t5000-runtime-0.5.15-source.tar.gz | 1,686,212 | `547f480bd081b253c15f7959a54f9175d5eb38fe` | `6702209a8869792e7ccc2a3107013e255576102e` |
+| glm-5.3-flash-dual-t5000-runtime-0.5.15-native.tar.gz | 5,926,786 | `aba9a81a3cd0e18a60ec04b33713414048b0bdd4` | `e7de9add79531bbaed198f3cedb123a745c649a2` |
+| glm-5.3-flash-dual-t5000-runtime-0.5.15-incremental.tar.gz | 132,134 | `00d273025d5562b26ac98dde962c2be6fbc01984` | `6dd1c279757c83cd075382df3de73d908129f547` |
+
+运行归档曾在空隔离 Docker 中按依赖顺序实际导入并逐文件校验；本次 LPK 仅改变部署配置，没有重传 3.2 GB 基础归档。模型权重、KV 历史和编译缓存均不进入公开包。
+
+0.5.15 默认 4 路安装后首轮 8K Prefill 曾超出 3% 门禁；失败样本、缓存诊断与复测数据保存在 `dist/mia-release-0.5.15/`。本版的验收基线是此前真实 8 路测量，不把 4 路与 8 路吞吐混作安装回归。
+
+0.5.16 初次安装验证暴露了 SSD 快照兼容标识改变，旧会话未自动命中；已迁移并保留旧快照，失败样本在 `installed-verification-first-ssd-failed/`。第二次验证的功能脚本仍限制 4 路，修正脚本后完整重跑；旧日志在 `installed-verification-second-harness-failed/`。
+
+Thor 上游补丁仍覆盖固定的 107 项中适用的 99 项；新增 24 项，其余此前已覆盖，8 项硬件不适用保持排除。见 [Thor 补丁报告](BENCHMARK-THOR-ALL-PATCHES-20261014.md)。
+
+全部安装收据、三轮样本、1M 请求/SSE、SSD 恢复和双节点镜像审计保存在 `dist/mia-release-0.5.16/`；同名 JSON 报告绑定关键证据 SHA256。模型来源证明沿用原完整验收，本轮未重新下载或重算全部 178 GB 权重的哈希。

SHA256SUMS-0.5.16

@@ -0,0 +1,3 @@
+72262db8bffde23a73cc76b46d47dd612143ef64ab2bbb377a4c50d8e3313ecb  PERFORMANCE-0.5.16.md
+5721567582001061f2fece5ac0157e42f83d7867558989b9ccddb1844cf1f82d  cloud.lazycat.aipod.glm53f-v0.5.16.lpk
+3d11868aad5abe776f2961ad59cb5b47b17bbbda3b9e3baa3d22979b8929028d  release-0.5.16.json

cloud.lazycat.aipod.glm53f-v0.5.16.lpk

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5721567582001061f2fece5ac0157e42f83d7867558989b9ccddb1844cf1f82d
+size 6652416

release-0.5.16.json

@@ -0,0 +1,701 @@
+{
+  "passed": true,
+  "version": "0.5.16",
+  "validationScope": "existing-thor-tp2-upgrade",
+  "hosts": [
+    "192.168.1.124",
+    "192.168.1.104"
+  ],
+  "defaultConcurrency": 8,
+  "package": {
+    "name": "cloud.lazycat.aipod.glm53f-v0.5.16.lpk",
+    "size": 6652416,
+    "sha256": "5721567582001061f2fece5ac0157e42f83d7867558989b9ccddb1844cf1f82d"
+  },
+  "imageId": "sha256:5149447c3091bee8b7f0816e26f2e1f8fd72968197544ba0fda65a7caa5103c6",
+  "runtimeArchiveImportOrder": [
+    {
+      "name": "glm-5.3-flash-dual-t5000-runtime-0.5.14.tar.gz",
+      "size": 3173705404,
+      "sha256": "38e8275d106031d9d53c6c6194f156d331ad68231f317de106bf10d88b4f4821",
+      "imageId": "sha256:81825a22506ab353fc71aa06ca4061a3a5a72f10ffac559a159b0acc6cda8072",
+      "sources": {
+        "huggingface": {
+          "repository": "manateelazycat/GLM-5.3-Flash-TensorFold-Runtime",
+          "revision": "fb520c24f72d189d8d43136c8291a101fce21b3c"
+        },
+        "modelscope": {
+          "repository": "manateelazycat/GLM-5.3-Flash-TensorFold-Runtime",
+          "revision": "f47dfedb7662453cdbaf197626d0b6c14bb19eb9"
+        }
+      }
+    },
+    {
+      "name": "glm-5.3-flash-dual-t5000-runtime-0.5.15-source.tar.gz",
+      "size": 1686212,
+      "sha256": "c0f42dbcdf49e6157baba49ee83ff2b70e56cb4d1212da0998050429f5992d0c",
+      "imageId": "sha256:012953fc13b47d3de38087fe157a4c91194347d896eb68a3f68f38ea1b78496b",
+      "sources": {
+        "huggingface": {
+          "repository": "manateelazycat/GLM-5.3-Flash-TensorFold-Runtime",
+          "revision": "547f480bd081b253c15f7959a54f9175d5eb38fe"
+        },
+        "modelscope": {
+          "repository": "manateelazycat/GLM-5.3-Flash-TensorFold-Runtime",
+          "revision": "6702209a8869792e7ccc2a3107013e255576102e"
+        }
+      }
+    },
+    {
+      "name": "glm-5.3-flash-dual-t5000-runtime-0.5.15-native.tar.gz",
+      "size": 5926786,
+      "sha256": "03f911d06ef7f7c9bc60c3c87171c8e14f4b9dbc972669377ae6945a5ee8b084",
+      "imageId": "sha256:1f35a90d0b10391c9ea4b510c02d3bb702f2e9ee86a91b319a0953b62352d820",
+      "sources": {
+        "huggingface": {
+          "repository": "manateelazycat/GLM-5.3-Flash-TensorFold-Runtime",
+          "revision": "aba9a81a3cd0e18a60ec04b33713414048b0bdd4"
+        },
+        "modelscope": {
+          "repository": "manateelazycat/GLM-5.3-Flash-TensorFold-Runtime",
+          "revision": "e7de9add79531bbaed198f3cedb123a745c649a2"
+        }
+      }
+    },
+    {
+      "name": "glm-5.3-flash-dual-t5000-runtime-0.5.15-incremental.tar.gz",
+      "size": 132134,
+      "sha256": "1d3bf12806daad88c7a95663e82d6cbfb80dc2c37c4a72863fe5c789905ff69f",
+      "imageId": "sha256:5149447c3091bee8b7f0816e26f2e1f8fd72968197544ba0fda65a7caa5103c6",
+      "sources": {
+        "huggingface": {
+          "repository": "manateelazycat/GLM-5.3-Flash-TensorFold-Runtime",
+          "revision": "00d273025d5562b26ac98dde962c2be6fbc01984"
+        },
+        "modelscope": {
+          "repository": "manateelazycat/GLM-5.3-Flash-TensorFold-Runtime",
+          "revision": "6dd1c279757c83cd075382df3de73d908129f547"
+        }
+      }
+    }
+  ],
+  "protocol": {
+    "label": "thor-r4-c1-c8-20261014",
+    "base": "http://192.168.1.124:19170/",
+    "model": "glm-5.3-flash",
+    "repetitions": 3,
+    "stages": [
+      "decode",
+      "prefill"
+    ],
+    "contextSizes": [
+      8192,
+      32768
+    ],
+    "decodeConcurrencies": [
+      1,
+      8
+    ],
+    "sparkdashRevision": "ead41f78ebcf3758bd5f93bcbf3b49d70a83041a",
+    "protocol": "Unmodified sparkDash 1.8.9 collectors; 400 output tokens; temperature 0; top_p 1; thinking off; 32-token decode warmup; client first-to-last-token aggregate",
+    "sources": [
+      {
+        "path": "server/collectors/DecodeBench.js",
+        "sha256": "65fc472092b6c577f05efe7498c938d5522a784a5762480f71d17abe47a3f8ea"
+      },
+      {
+        "path": "server/collectors/PrefillBench.js",
+        "sha256": "6b14c4eb3fb510806024c4a6ccb01d5c4e011014ebc40919a306a1519f1d67d4"
+      },
+      {
+        "path": "server/collectors/LlmStreaming.js",
+        "sha256": "1e4d06b85e2742822bc87c1fa2e86c2432c160e43c4b5cf0fec1fd5786be845e"
+      },
+      {
+        "path": "src/shared/llmPrompts.js",
+        "sha256": "f930fcfa13a041fcfe1ff6cfbf882f2c946293dc1ac29aba9f3768fd1ad327aa"
+      }
+    ],
+    "startedUtc": "2026-10-14T01:12:44.720Z"
+  },
+  "comparisons": [
+    {
+      "metric": "C1 decode aggregate",
+      "beforeTokensPerSecond": 49.31,
+      "afterTokensPerSecond": 49.63,
+      "throughputChangePercent": 0.648955587102007,
+      "beforeTtftMs": 261.93,
+      "afterTtftMs": 261.86,
+      "ttftChangePercent": -0.026724697438240685
+    },
+    {
+      "metric": "C8 decode aggregate",
+      "beforeTokensPerSecond": 106.12,
+      "afterTokensPerSecond": 105.5,
+      "throughputChangePercent": -0.5842442517904334,
+      "beforeTtftMs": 648.76,
+      "afterTtftMs": 651.14,
+      "ttftChangePercent": 0.3668536901165309
+    },
+    {
+      "metric": "8K cold prefill",
+      "beforeTokensPerSecond": 791.84,
+      "afterTokensPerSecond": 820.75,
+      "throughputChangePercent": 3.6509900990098876,
+      "beforeTtftMs": 10372.04,
+      "afterTtftMs": 10010.35,
+      "ttftChangePercent": -3.487163566665774
+    },
+    {
+      "metric": "32K cold prefill",
+      "beforeTokensPerSecond": 865.77,
+      "afterTokensPerSecond": 905.86,
+      "throughputChangePercent": 4.630560079466828,
+      "beforeTtftMs": 37878.35,
+      "afterTtftMs": 36204.3,
+      "ttftChangePercent": -4.419543089918109
+    }
+  ],
+  "longContext": {
+    "inputTokens": 1048448,
+    "cachedTokens": 0,
+    "installedEightPrefillTokensPerSecond": 667.910777219144,
+    "installedEightTtftSeconds": 1570.8711028569378,
+    "oldFourPrefillTokensPerSecond": 661.9214300383794,
+    "oldFourTtftSeconds": 1588.1729277479462,
+    "checks": {
+      "http": true,
+      "sse": true,
+      "quality": true,
+      "usage": true,
+      "oneRequest": true,
+      "healthyAndIdle": true,
+      "fullPromptComputed": true,
+      "completionCounter": true,
+      "drafter": true
+    },
+    "scope": "One fresh random full 1M cold request after installing the default-eight LPK. Old four-stream value is context only; it is not a like-for-like performance gate."
+  },
+  "ssd": {
+    "passed": true,
+    "retainedHistories": {
+      "path": "installed-verification/ssd-retained-histories/summary.json",
+      "sha256": "d2a5a149aca0e75aeb1b383b52ed5072ef733162eda57d2ba0ce320d70d9eeb0",
+      "cachedPercent": [
+        100.0,
+        100.0,
+        100.0
+      ],
+      "ttftSeconds": [
+        1.1237383079715073,
+        0.3654726690147072,
+        0.3659856639569625
+      ]
+    },
+    "oneMillionAfterLPKRestart": {
+      "path": "installed-one-million-ssd/receipt.json",
+      "sha256": "54f51703fb763f9fba8c087885242306e16177ad7d04f1873aee1e51af0500a7",
+      "cachedPercent": 100.0,
+      "ttftSeconds": 9.561538559035398
+    },
+    "scope": "Three 64K histories and the new 1M history restored from SSD after actual LPK restarts; exact 1M output and disk load counter checked."
+  },
+  "ssdMigration": {
+    "passed": true,
+    "sourceConcurrency": 4,
+    "targetConcurrency": 8,
+    "oldSnapshotsPerNode": 39,
+    "migratedOldCopiesPerNode": 38,
+    "targetSpillEntries": 45,
+    "oldNamespacePreserved": true,
+    "representative64KExactParity": true,
+    "representative1MExactParity": true,
+    "representative1MTtftSeconds": 9.556760090054013,
+    "scope": "All old 4-way files have a paired 8-way namespace file. Three 64K and one 1M exact histories verified; remaining files structurally scanned on both ranks but not individually replayed. No model or runtime image change.",
+    "evidence": {
+      "migration-fixtures/receipt.json": "641d06c5f1885441d639c89078906fa412f0944fb5f9840885ca12651b1b0979",
+      "ssd-cross-config-acceptance.json": "a2d1ca702b1885d00deba6aaf801a19e31a70b3f31d95eaa767c4da42a795206",
+      "migration-old-million/receipt.json": "e94540ffc3f826270a8e8d757b4eadfae102b357843598b16bea33114a0eaeaf",
+      "old-million-cross-config-replay/receipt.json": "3633ad065882293ca7d47d3d088aeeb4bc75236a9ee6fe54b1fc873c1bdaacf4",
+      "migration-all-remaining/receipt.json": "88f9e0dd9f1e1e52cc7989d52458bb10e223bc581962dea1c8f255636efa059c"
+    },
+    "completedUtc": "2026-10-14T07:57:35.779629+00:00"
+  },
+  "standardPerformanceGate": {
+    "passed": true,
+    "criteria": {
+      "version": "0.5.16",
+      "declaredUtc": "2026-10-14T06:25:25.325796+00:00",
+      "repetitions": 3,
+      "maximumThroughputRegressionFraction": 0.03,
+      "maximumColdTtftRegressionFraction": 0.03,
+      "maximumShortTtftRegressionFraction": 0.1,
+      "ordinaryBaseline": "validation/thor-all-patches-20261013/performance-c1-c8-20261014/standard/summary.json",
+      "eightStreamBaseline": "validation/thor-all-patches-20261013/performance-c1-c8-20261014/standard/summary.json",
+      "interpretation": "Same protocol medians; 3% existing engineering throughput limit, 3% cold TTFT limit, 10% short-prompt TTFT jitter limit. Preserve all samples. This is not proof of statistical equivalence."
+    },
+    "comparisons": [
+      {
+        "kind": "decode",
+        "concurrency": 1,
+        "before": 49.31,
+        "after": 49.63,
+        "changePercent": 0.648955587102007,
+        "beforeTtftMs": 261.93,
+        "afterTtftMs": 261.86,
+        "ttftChangePercent": -0.026724697438240685,
+        "throughputPassed": true,
+        "ttftPassed": true
+      },
+      {
+        "kind": "decode",
+        "concurrency": 8,
+        "before": 106.12,
+        "after": 105.5,
+        "changePercent": -0.5842442517904334,
+        "beforeTtftMs": 648.76,
+        "afterTtftMs": 651.14,
+        "ttftChangePercent": 0.3668536901165309,
+        "throughputPassed": true,
+        "ttftPassed": true
+      },
+      {
+        "kind": "prefill",
+        "targetTokens": 8192,
+        "before": 791.84,
+        "after": 820.75,
+        "changePercent": 3.6509900990098876,
+        "beforeTtftMs": 10372.04,
+        "afterTtftMs": 10010.35,
+        "ttftChangePercent": -3.487163566665774,
+        "throughputPassed": true,
+        "ttftPassed": true
+      },
+      {
+        "kind": "prefill",
+        "targetTokens": 32768,
+        "before": 865.77,
+        "after": 905.86,
+        "changePercent": 4.630560079466828,
+        "beforeTtftMs": 37878.35,
+        "afterTtftMs": 36204.3,
+        "ttftChangePercent": -4.419543089918109,
+        "throughputPassed": true,
+        "ttftPassed": true
+      }
+    ]
+  },
+  "eightStreamReceipt": {
+    "passed": true,
+    "version": "0.5.16",
+    "imageId": "sha256:5149447c3091bee8b7f0816e26f2e1f8fd72968197544ba0fda65a7caa5103c6",
+    "packageSHA256": "5721567582001061f2fece5ac0157e42f83d7867558989b9ccddb1844cf1f82d",
+    "deploymentId": "e8b4efd55cc2964444d4b701",
+    "installedLPKStopStartPassed": true,
+    "noRuntimeCodeOverlays": true,
+    "allSourceHashesMatch": true,
+    "allFourRailsRXBytes": [
+      {
+        "mgbe0_0": 61004296924,
+        "mgbe1_0": 35483480818,
+        "mgbe2_0": 35480588888,
+        "mgbe3_0": 35481057692
+      },
+      {
+        "mgbe0_0": 60938164120,
+        "mgbe1_0": 35483183310,
+        "mgbe2_0": 35478587622,
+        "mgbe3_0": 35479995036
+      }
+    ],
+    "upstreamRecipePort": {
+      "upstreamRepository": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks-TensorFold",
+      "upstreamCommit": "549fdc562c477d477244d2df0590d6c14bda4889",
+      "recipeVersion": "1.10",
+      "performancePatches": [
+        85,
+        86
+      ],
+      "cacheAndReliabilityPatches": [
+        54,
+        60,
+        63,
+        71,
+        74,
+        77,
+        78,
+        79,
+        92,
+        96,
+        97
+      ],
+      "adaptation97": "Preserve TP2 slot_bytes(text, rows=0), no Spark display reservation; default cache share zero and existing eight kept entries.",
+      "defaults": {
+        "TF_GLM_MULTI_LONE": "0",
+        "TENSORFOLD_NUCLEUS_UNION": "1",
+        "TF_GLM_EFFORT_TAIL": "0"
+      },
+      "evidence": "dist/tensorfold-066-v110-thor-20261008/comparison.json",
+      "files": [
+        {
+          "path": "glm53f_mia/families/glm5_next/cuda/sparse.py",
+          "beforeSHA256": "183e59caee59339cd285a11a7e97a8d385389b1bc0cb43d059debe11b11e6a89",
+          "sha256": "77ab4f88e854fd7c8ef6869131fa7e4e9488d9b9def63085aa772938e246c696"
+        },
+        {
+          "path": "glm53f_mia/vision/glm.py",
+          "beforeSHA256": "e373c78ba414a8aa8ce006fa806944307c738ec0fe9ded67a8917a7230f4cd19",
+          "sha256": "43b7fb6d6179262ff76efedb7a1eebd483c44096273d43d2ead7bf99e8df0eb6"
+        },
+        {
+          "path": "glm53f_mia/families/glm5_next/prompts.py",
+          "beforeSHA256": "02592186c6b9065f73499078b0d3bfd6174ee3967d6d246bb62631ec115fc77d",
+          "sha256": "ed81dc186eea5986dec5766a8f7a49d2afe091ea3e1d9842303b51fd461f31a7"
+        },
+        {
+          "path": "glm53f_mia/families/glm5_next/cuda/multi.py",
+          "beforeSHA256": "ea06e44089ae0d704d4d1f15d8b6b3e4bb525ea3baad5c07d6d582363bc20dc3",
+          "sha256": "582233ff8d1a6e99ce311c785a5866305f74ef779b619c7b51e7ec3ee9d7b0c3"
+        },
+        {
+          "path": "glm53f_mia/families/glm5_next/cuda/engine.py",
+          "beforeSHA256": "e981e1d6f5146e801455455c1b922d7205b9ef1fbd92e4c8cd9b210bb5c53303",
+          "sha256": "22460cdac8ad9e5f30131902d11f7e730ed8a87f2ef0683b0ea5358111612f33"
+        },
+        {
+          "path": "glm53f_mia/families/glm5_next/cuda/dflash2.py",
+          "beforeSHA256": "734b229a3400ca1a08097c5421080468e5ff239290a396ca3434cd6792e219a6",
+          "sha256": "8f353657488ef5829971ffab91ee8b4643c2eedad8970847e8a1c90cc04f02b5"
+        },
+        {
+          "path": "glm53f_mia/families/glm5_next/cuda/dflash2_multi.py",
+          "beforeSHA256": "af6cd761f088b8385ca27e08712546d374243dc6fe0dafa0ba79a741fbd660ec",
+          "sha256": "70e67c81e59816678d9a73e6c599426b1c098654cd852a036125e766cd7ca1f8"
+        },
+        {
+          "path": "glm53f_mia/families/glm5_next/cuda/app.py",
+          "beforeSHA256": "278d4b68c06d382247dbba761f9c57daf73d773479e8a22b76462cac011d158d",
+          "sha256": "111e374156bd6b7b8a4f7c646f8f80c32b31cae080fc9c38debefba190ceaee8"
+        }
+      ],
+      "tensorfoldVersion": "0.6.6",
+      "tensorfoldCommit": "cb2ebf0540f42604e2759b2ddef497861e928248"
+    },
+    "performance": [
+      {
+        "kind": "decode",
+        "concurrency": 1,
+        "before": 49.31,
+        "after": 49.63,
+        "changePercent": 0.648955587102007,
+        "beforeTtftMs": 261.93,
+        "afterTtftMs": 261.86,
+        "ttftChangePercent": -0.026724697438240685,
+        "throughputPassed": true,
+        "ttftPassed": true
+      },
+      {
+        "kind": "decode",
+        "concurrency": 8,
+        "before": 106.12,
+        "after": 105.5,
+        "changePercent": -0.5842442517904334,
+        "beforeTtftMs": 648.76,
+        "afterTtftMs": 651.14,
+        "ttftChangePercent": 0.3668536901165309,
+        "throughputPassed": true,
+        "ttftPassed": true
+      },
+      {
+        "kind": "prefill",
+        "targetTokens": 8192,
+        "before": 791.84,
+        "after": 820.75,
+        "changePercent": 3.6509900990098876,
+        "beforeTtftMs": 10372.04,
+        "afterTtftMs": 10010.35,
+        "ttftChangePercent": -3.487163566665774,
+        "throughputPassed": true,
+        "ttftPassed": true
+      },
+      {
+        "kind": "prefill",
+        "targetTokens": 32768,
+        "before": 865.77,
+        "after": 905.86,
+        "changePercent": 4.630560079466828,
+        "beforeTtftMs": 37878.35,
+        "afterTtftMs": 36204.3,
+        "ttftChangePercent": -4.419543089918109,
+        "throughputPassed": true,
+        "ttftPassed": true
+      }
+    ],
+    "benchmark": {
+      "passed": true,
+      "label": "installed-0.5.16-four-rail",
+      "checkedUtc": "2026-10-14T08:25:48.603Z",
+      "provenance": {
+        "label": "installed-0.5.16-four-rail",
+        "base": "http://192.168.1.124:19170/",
+        "model": "glm-5.3-flash",
+        "repetitions": 3,
+        "stages": [
+          "decode",
+          "prefill"
+        ],
+        "contextSizes": [
+          8192,
+          32768
+        ],
+        "decodeConcurrencies": [
+          1,
+          8
+        ],
+        "sparkdashRevision": "ead41f78ebcf3758bd5f93bcbf3b49d70a83041a",
+        "protocol": "Unmodified sparkDash 1.8.9 collectors; 400 output tokens; temperature 0; top_p 1; thinking off; 32-token decode warmup; client first-to-last-token aggregate",
+        "sources": [
+          {
+            "path": "server/collectors/DecodeBench.js",
+            "sha256": "65fc472092b6c577f05efe7498c938d5522a784a5762480f71d17abe47a3f8ea"
+          },
+          {
+            "path": "server/collectors/PrefillBench.js",
+            "sha256": "6b14c4eb3fb510806024c4a6ccb01d5c4e011014ebc40919a306a1519f1d67d4"
+          },
+          {
+            "path": "server/collectors/LlmStreaming.js",
+            "sha256": "1e4d06b85e2742822bc87c1fa2e86c2432c160e43c4b5cf0fec1fd5786be845e"
+          },
+          {
+            "path": "src/shared/llmPrompts.js",
+            "sha256": "f930fcfa13a041fcfe1ff6cfbf882f2c946293dc1ac29aba9f3768fd1ad327aa"
+          }
+        ],
+        "startedUtc": "2026-10-14T08:21:24.289Z"
+      },
+      "decode": [
+        {
+          "concurrency": 1,
+          "medianAggregateDecodeTps": 49.63,
+          "medianPerStreamDecodeTps": 49.63,
+          "medianTtftMs": 261.86
+        },
+        {
+          "concurrency": 8,
+          "medianAggregateDecodeTps": 105.5,
+          "medianPerStreamDecodeTps": 13.81,
+          "medianTtftMs": 651.14
+        }
+      ],
+      "prefill": [
+        {
+          "targetTokens": 8192,
+          "medianPrefillTps": 820.75,
+          "medianTtftMs": 10010.35,
+          "allColdVerifiedViaCumulativeCounter": true
+        },
+        {
+          "targetTokens": 32768,
+          "medianPrefillTps": 905.86,
+          "medianTtftMs": 36204.3,
+          "allColdVerifiedViaCumulativeCounter": true
+        }
+      ]
+    },
+    "decodeRegressionGatePassed": true,
+    "maximumDecodeRegressionFraction": 0.03,
+    "maximumContextEvidence": {
+      "scope": "Installed LPK after LPK stop/start; full prompt, zero cached tokens",
+      "path": "/home/andy/lzc-aipod-glm53f/dist/mia-release-0.5.16/installed-verification/needle1m-guided.json",
+      "sha256": "948f6f62590e125922b77332ae4cb2358a21efb15c2b81096737e2d958e5bb8b"
+    },
+    "health": {
+      "ok": true,
+      "backend": "tensorfold",
+      "busy": false,
+      "requests_running": 0,
+      "requests_total": 132,
+      "prompt_tokens_total": 1406457,
+      "completion_tokens_total": 21253,
+      "prefill_seconds_total": 1803.8343,
+      "decode_seconds_total": 1120.1853,
+      "cached_tokens_total": 196736,
+      "rounds_total": 9662,
+      "drafted_total": 16563,
+      "accepted_total": 11459,
+      "streams": {
+        "decoding": 0,
+        "prefilling": 0,
+        "max": 8,
+        "filling": 0,
+        "paused": 0
+      },
+      "pool_tokens": 2676736,
+      "pool_free_tokens": 1509376,
+      "kept_prompts": 8,
+      "kept_bytes": 726949120,
+      "kept_bytes_cap": 4294967296,
+      "kept_mix": {
+        "own": 8,
+        "shared": 0,
+        "chats": 8,
+        "rows": 1164160,
+        "by_chat": [
+          [
+            128,
+            1,
+            512
+          ],
+          [
+            129,
+            1,
+            8192
+          ],
+          [
+            130,
+            1,
+            32768
+          ],
+          [
+            131,
+            1,
+            1048448
+          ],
+          [
+            104,
+            1,
+            32768
+          ],
+          [
+            115,
+            1,
+            512
+          ],
+          [
+            116,
+            1,
+            8192
+          ],
+          [
+            117,
+            1,
+            32768
+          ]
+        ]
+      },
+      "multi_prefill": {
+        "chunks": 9,
+        "pieces": 51,
+        "rows": 1772
+      },
+      "spill": {
+        "highwater": 0.7,
+        "writers": 2,
+        "direct": true,
+        "entries": 52,
+        "pending": 0,
+        "stored_gib": 24.98,
+        "stored_bytes": 26826768384,
+        "written": 7,
+        "skipped": 0,
+        "full": 0,
+        "no_space": 0,
+        "deferred": 2,
+        "behind": 7,
+        "unwritten_evicted": 3,
+        "write_failed": 0,
+        "loaded": 3,
+        "load_failed": 0,
+        "crc_failed": 0,
+        "load_unfit": 0,
+        "pruned": 0,
+        "save_s": 0.24,
+        "write_s": 1.05,
+        "copy_wait_s": 0.0,
+        "copy_waits": 0,
+        "behind_s": 0.46,
+        "written_gib": 1.52,
+        "written_bytes": 1630575360,
+        "loaded_gib": 1.4,
+        "loaded_bytes": 1505613568,
+        "load_s": 0.57,
+        "loaded_tokens": 196608,
+        "restore_ms_p50": 189.2,
+        "restore_ms_p95": 199.5,
+        "disk_free_bytes": 1445691621376,
+        "unwritten_short": 3
+      },
+      "iteration_s": 0.0,
+      "context_length": 1048576
+    }
+  },
+  "evidence": {
+    "installed": {
+      "path": "lpk-installed/receipt.json",
+      "sha256": "03142eab57dd7999d3fd4050b962edd9ddac2001436d926d4131796f464d9e8a"
+    },
+    "deployment": {
+      "path": "lpk-deployment/receipt.json",
+      "sha256": "dd953203d0f807e8f2056a08f0ad18e6d1df3ac26ad6f2963e8aa0982751fc23"
+    },
+    "comparison": {
+      "path": "installed-verification/standard-performance-gate.json",
+      "sha256": "9a8267f2ea22500862d69e39903f52aeb3a7c07eec4a9859ff7f57db6d36b5e7"
+    },
+    "environments": {
+      "path": "final-runtime-audit.json",
+      "sha256": "7399a1ebc36ba4d02b09f9f3166a02b6f007babb77e62ebe47b4fdb9440a31f8"
+    },
+    "hfModels": {
+      "path": "../mia-release-0.5.0/hf-source-verification/receipt.json",
+      "sha256": "00434034a96abd33b0cdec63bf3deb967f83ead0d487827863be485cfa2ff645"
+    },
+    "msModels": {
+      "path": "../mia-release-0.5.0/models/host-models-downloaded.json",
+      "sha256": "e15469dae0d9285470e3c88aefedc2d5c266e1341947709050a2d1f355a875e4"
+    },
+    "runtimePublication": {
+      "path": "publication/publication.json",
+      "sha256": "bc6c1b6453f8471a9f7f717862ecd3aa0bb95fb0d3c6b11ca5717de9ee9f8f17"
+    },
+    "runtimeColdImport": {
+      "path": "../mia-release-0.5.15/layered-cold-import.json",
+      "sha256": "593d1e5c6d9bcc3a42c32df52e4ae7dfbdde52e8d4c0361caac48f678c00eadb"
+    },
+    "ssd": {
+      "path": "installed-ssd-acceptance.json",
+      "sha256": "71a8e8c0387b5178e220e002f5742a559c2e7d6606ee76a31e1b480b3d6f904b"
+    },
+    "ssdMigration": {
+      "path": "ssd-migration-acceptance.json",
+      "sha256": "064df5b082fdc6153498d8d3e04f3daa36ad03a346276957c51c7922d9ace12b"
+    },
+    "eight": {
+      "path": "installed-verification/receipt.json",
+      "sha256": "ca15b83776a73ae1d71f874f2ba7ea510888149e2ad24a0b862da63e64cc55f9"
+    },
+    "modelIdentity": {
+      "path": "model-identity-comparison.json",
+      "sha256": "bd4a944ea9a78cf30ba771e3d18eed79fcb61e518736cef30dc21dca395cf365"
+    },
+    "fullInstalledVerification": {
+      "path": "installed-verification/receipt.json",
+      "sha256": "ca15b83776a73ae1d71f874f2ba7ea510888149e2ad24a0b862da63e64cc55f9"
+    },
+    "reuse": {
+      "path": "reused-evidence.json",
+      "sha256": "4b1455fce0776b3dd172c8e31f283658ab8b6e4830a97ed590d5830dde2fdd86"
+    },
+    "additional": {
+      "path": "additional-installed-verification-complete.json",
+      "sha256": "c8d1cdcd7c1ca73b5ed1824969c64419eb37e40e156b2b68371a6ca814c4363a"
+    },
+    "acceptance": {
+      "path": "installed-final-acceptance/acceptance.json",
+      "sha256": "7bacab7fb47b79bc224a8f20040adab371b3a8b58fa034941ad37873622a4174"
+    }
+  },
+  "measuredUtc": "2026-10-14T08:57:04.625696+00:00",
+  "retainedModelSourceProofScope": "Weights were unchanged on both Thor nodes; original full model-source receipts retained. No new full 178 GB hash claim."
+}