{ "schemaVersion": 1, "operationId": "20260826042447-42217f9b8b", "checkedAt": "2026-08-26T04:52:00Z", "status": "pinned-source-zero-gpu-glm52-vllm-h200-runtime-oom-audit", "revisions": { "vllmLatestStableTag": "v0.27.1", "vllmLatestStableCommit": "6e448d0ea9bf3d88d898b65449ca6dc2aec170ac", "vllmMainCommit": "044b05220ee9eb42ae3b90b984d0c7e9eb442f0e", "flashMlaPinnedCommit": "a8f794d1251cbfd88a5011445dd5582289c727e4", "glm52Fp8Revision": "ba978f7d347eaf65d22f1a86833408afdb953541", "chunkingPr": 49357, "flashMlaFixPr": 19, "integrationPr": 53755 }, "sourceReceipts": [ { "id": "zai-release", "url": "https://z.ai/blog/glm-5.2", "purpose": "fixed first-party release check and model context", "httpStatus": 200, "bytes": 598, "sha256": "63e94b9e11a2db64243ff18418011577f2a98ca7851ce9dcb0e97cfe34cc245a" }, { "id": "zhipu-research", "url": "https://www.zhipuai.cn/zh/research", "purpose": "fixed first-party research-index check", "httpStatus": 200, "bytes": 1236461, "sha256": "188772a5cb6b65eb65f02024d89d153b692ef3d3cb2c98ddd233352c0dd6ac80" }, { "id": "hf-model-card", "url": "https://huggingface.co/zai-org/GLM-5.2-FP8/raw/ba978f7d347eaf65d22f1a86833408afdb953541/README.md", "purpose": "pinned first-party checkpoint documentation", "httpStatus": 200, "bytes": 10909, "sha256": "de23c1b7cab43a99f0fedf4edf10de0d165882a2ecb2e2bad6c5796fcabf2e46" }, { "id": "hf-config", "url": "https://huggingface.co/zai-org/GLM-5.2-FP8/raw/ba978f7d347eaf65d22f1a86833408afdb953541/config.json", "purpose": "pin GLM-5.2 architecture, query heads, index geometry, and context", "httpStatus": 200, "bytes": 29464, "sha256": "22e49334abf8562fecf70ca3292ba3f5b33f5602fb2bf10b52dd64a66cfe65ff" }, { "id": "vllm-release-0271", "url": "https://api.github.com/repos/vllm-project/vllm/releases/tags/v0.27.1", "purpose": "pin the latest stable release at audit time", "httpStatus": 200, "bytes": 13964, "sha256": "0c6224baeffb5d951d0af0452d4ffc33a1a4dbb478c74c3c43bd143e495636e3" }, { "id": "vllm-tag-flashmla-sparse", "url": "https://raw.githubusercontent.com/vllm-project/vllm/6e448d0ea9bf3d88d898b65449ca6dc2aec170ac/vllm/v1/attention/backends/mla/flashmla_sparse.py", "purpose": "audit the v0.27.1 mixed FP8 path, head padding, and unchunked full-token call", "httpStatus": 200, "bytes": 34530, "sha256": "0544c431bbf311d7283fee7f34cd7ad8a2df7961380646f2c12db2e861608ee4" }, { "id": "vllm-tag-flashmla-pin", "url": "https://raw.githubusercontent.com/vllm-project/vllm/6e448d0ea9bf3d88d898b65449ca6dc2aec170ac/cmake/external_projects/flashmla.cmake", "purpose": "verify the FlashMLA source revision bundled by v0.27.1", "httpStatus": 200, "bytes": 8021, "sha256": "3cd4062759f9710f2f1ead18cc726605780b98cb3f584101e1fe9f5ab60e35cc" }, { "id": "vllm-tag-scheduler", "url": "https://raw.githubusercontent.com/vllm-project/vllm/6e448d0ea9bf3d88d898b65449ca6dc2aec170ac/vllm/config/scheduler.py", "purpose": "pin max-num-batched-tokens and speculative scheduling semantics", "httpStatus": 200, "bytes": 12302, "sha256": "3cf5d41a5d662ab0eada6f3128e2538abc0a250d6aef82acab7df7dd84e498bc" }, { "id": "vllm-tag-cache", "url": "https://raw.githubusercontent.com/vllm-project/vllm/6e448d0ea9bf3d88d898b65449ca6dc2aec170ac/vllm/config/cache.py", "purpose": "pin GPU memory utilization and KV cache configuration semantics", "httpStatus": 200, "bytes": 14663, "sha256": "0b603decec7487138e97b9cafbad9684ec22c8052f52a227236d1151a03aa34f" }, { "id": "flashmla-pinned-source", "url": "https://raw.githubusercontent.com/vllm-project/FlashMLA/a8f794d1251cbfd88a5011445dd5582289c727e4/csrc/api/sparse_decode.h", "purpose": "reproduce the unconditional FP32 split-KV accumulator shapes", "httpStatus": 200, "bytes": 18739, "sha256": "9333a66b2a2d13583a72c4617ff76a5f361b1528da8d97b40d7ab26237a741e9" }, { "id": "vllm-issue-53413", "url": "https://api.github.com/repos/vllm-project/vllm/issues/53413", "purpose": "capture the GLM-5.2 H200 delayed-OOM report and exact reproduction profile", "httpStatus": 200, "bytes": 8756, "sha256": "656a83676c2fe103c24597b48083b22dbf9a0998fc60e793882e9de25c0bc105" }, { "id": "vllm-issue-53413-comments", "url": "https://api.github.com/repos/vllm-project/vllm/issues/53413/comments?per_page=100", "purpose": "capture maintainer workaround guidance and separately attributed operator evidence", "httpStatus": 200, "bytes": 5958, "sha256": "a210c7ca8265d8dc46a4575a07b4f255480d1d443b1043270dbf7b106a0cbb6e" }, { "id": "vllm-pr-49357", "url": "https://api.github.com/repos/vllm-project/vllm/pulls/49357", "purpose": "audit the open mixed-batch chunking workaround and its throughput boundary", "httpStatus": 200, "bytes": 19486, "sha256": "7358a4cb1fe3746bb774d12ae5d43e7f9e50167e8e394ebcfc3789941347f159" }, { "id": "vllm-pr-53755", "url": "https://api.github.com/repos/vllm-project/vllm/pulls/53755", "purpose": "audit the open FlashMLA integration fix and reported allocation test", "httpStatus": 200, "bytes": 21866, "sha256": "68626c42217640a8320494dd74d7bfda38c501d5ae57a489e1d11fdd72844aa8" }, { "id": "flashmla-pr-19", "url": "https://api.github.com/repos/vllm-project/FlashMLA/pulls/19", "purpose": "audit the open upstream no-split workspace fix", "httpStatus": 200, "bytes": 16907, "sha256": "b2fcf55b6b19b91ebf066dc8686e00c7b760476ee343f50b3bb9bce6df038445" }, { "id": "flashmla-pr-19-files", "url": "https://api.github.com/repos/vllm-project/FlashMLA/pulls/19/files?per_page=100", "purpose": "verify the proposed conditional allocation and regression test diff", "httpStatus": 200, "bytes": 8505, "sha256": "c30ed90955652c42c31e6f15c011b29e4454b5e0af34dc0b68685008c46258db" }, { "id": "vllm-glm52-recipe", "url": "https://raw.githubusercontent.com/vllm-project/recipes/f65701647c2a123ba529dfebb52163c4947ffac9/models/zai-org/GLM-5.2.yaml", "purpose": "pin official GLM-5.2 serving guidance and model profile", "httpStatus": 200, "bytes": 15247, "sha256": "517662028453e8a0b62f53a8b507d6b8ac502aaecbd5452b396c37ea2fbae4a7" }, { "id": "vllm-ascend-glm52", "url": "https://raw.githubusercontent.com/vllm-project/vllm-ascend/b7de842253ad8da059c31ff7e5e9357974d7929a/docs/source/tutorials/models/GLM5.2.md", "purpose": "verify the documented MTP decode-budget relationship in a disaggregated GLM-5.2 profile", "httpStatus": 200, "bytes": 71812, "sha256": "c9ab1c9c885fb107d7a0e64d8553460bde3b3f2c9bd867eeeec6d84f1af7cb04" }, { "id": "llmd-h200-guide", "url": "https://raw.githubusercontent.com/llm-d/llm-d/0a3dd07b4bde6406388c31355fccf8a1dc051780/guides/agentic-serving/glm-5-2-h200.md", "purpose": "verify that an official ecosystem H200 route uses prefill-decode disaggregation", "httpStatus": 200, "bytes": 13000, "sha256": "854d68148d80b12b068f08fbb277dc642ba8dac805f96714413df59ad9004a66" }, { "id": "vllm-main-flashmla-pin", "url": "https://raw.githubusercontent.com/vllm-project/vllm/044b05220ee9eb42ae3b90b984d0c7e9eb442f0e/cmake/external_projects/flashmla.cmake", "purpose": "verify audited main still pins the affected FlashMLA revision", "httpStatus": 200, "bytes": 8021, "sha256": "3cd4062759f9710f2f1ead18cc726605780b98cb3f584101e1fe9f5ab60e35cc" }, { "id": "vllm-main-flashmla-sparse", "url": "https://raw.githubusercontent.com/vllm-project/vllm/044b05220ee9eb42ae3b90b984d0c7e9eb442f0e/vllm/v1/attention/backends/mla/flashmla_sparse.py", "purpose": "verify audited main retains the full-token mixed FP8 call", "httpStatus": 200, "bytes": 39072, "sha256": "d1a4db2384f99cb748a00afcb46572a0425c9a43c2b2da9c5c032c6f3bd271ae" } ], "model": { "artifact": "zai-org/GLM-5.2-FP8", "architecture": "GlmMoeDsaForCausalLM", "modelType": "glm_moe_dsa", "decoderLayers": 78, "globalQueryHeads": 64, "indexTopK": 2048, "indexHeadDim": 128, "indexHeads": 32, "kvLoraRank": 512, "valueHeadDim": 256, "maxPositionEmbeddings": 1048576 }, "affectedProfile": { "hardware": "8x NVIDIA H200", "computeCapability": "SM90", "checkpoint": "zai-org/GLM-5.2-FP8", "vllmReportedVersion": "0.24.0", "tensorParallelSize": 8, "expertParallelEnabled": true, "kvCacheDtype": "fp8", "maxModelLength": 262144, "maxNumSequences": 32, "maxNumBatchedTokens": 32768, "mtpSpeculativeTokens": 5, "reportedRuntimeBeforeFailureHours": 61, "reportedFreeGiBBeforeFailedAllocation": 6.4, "reportedFailedAllocationGiB": 8, "reportedKvCacheUsagePercentRange": [ 2, 5.4 ], "boundary": "Issue-reporter evidence; GLM52.ai did not reproduce this GPU run." }, "versionBoundary": [ { "name": "vLLM v0.27.1", "commit": "6e448d0ea9bf3d88d898b65449ca6dc2aec170ac", "releasedAt": "2026-08-11T10:47:49Z", "flashMlaCommit": "a8f794d1251cbfd88a5011445dd5582289c727e4", "noSplitWorkspaceFixPresent": false, "decision": "affected source shape remains present" }, { "name": "audited vLLM main", "commit": "044b05220ee9eb42ae3b90b984d0c7e9eb442f0e", "checkedAt": "2026-08-26T04:52:00Z", "flashMlaCommit": "a8f794d1251cbfd88a5011445dd5582289c727e4", "noSplitWorkspaceFixPresent": false, "decision": "affected FlashMLA pin and full-token mixed path remain present" }, { "name": "open vLLM chunking PR 49357", "state": "open", "strategy": "bound mixed batches by chunking with an acknowledged throughput tradeoff", "decision": "source-evaluation candidate, not a production recommendation" }, { "name": "open FlashMLA PR 19 plus vLLM PR 53755", "state": "open", "strategy": "skip unused FP32 split-KV accumulators when the scheduler has one SM partition", "reportedTestHardware": "one H100 80GB", "reportedPeakBeforeMiB": 42.3467, "reportedPeakAfterMiB": 8.2822, "decision": "wait for merge, pinned packaging, and H200 workload validation" } ], "workspaceFormula": { "totalSplits": 2, "accumulatorBytesPerToken": 262144, "accumulatorMiBPerToken": 0.25, "lseBytesPerToken": 512, "outputBytesPerToken": 65536, "dominantAccumulatorFormula": "(batch + SM partitions) x scheduled tokens x padded query heads x kernel value dimension x FP32 bytes", "boundary": "No-split shape matching the reported allocation and open upstream fix. Real allocations depend on the resolved kernel path, scheduler metadata, concurrency, graph pools, allocator state, and other workspaces." }, "workspaceRows": [ { "maxNumBatchedTokens": 4096, "dominantFp32AccumulatorGiB": 1, "lseAccumulatorMiB": 2, "bf16OutputGiB": 0.25, "dominantAccumulatorPlusOutputGiB": 1.25 }, { "maxNumBatchedTokens": 8192, "dominantFp32AccumulatorGiB": 2, "lseAccumulatorMiB": 4, "bf16OutputGiB": 0.5, "dominantAccumulatorPlusOutputGiB": 2.5 }, { "maxNumBatchedTokens": 16384, "dominantFp32AccumulatorGiB": 4, "lseAccumulatorMiB": 8, "bf16OutputGiB": 1, "dominantAccumulatorPlusOutputGiB": 5 }, { "maxNumBatchedTokens": 32768, "dominantFp32AccumulatorGiB": 8, "lseAccumulatorMiB": 16, "bf16OutputGiB": 2, "dominantAccumulatorPlusOutputGiB": 10 }, { "maxNumBatchedTokens": 65536, "dominantFp32AccumulatorGiB": 16, "lseAccumulatorMiB": 32, "bf16OutputGiB": 4, "dominantAccumulatorPlusOutputGiB": 20 } ], "reportedFailureReconciliation": { "calculatedDominantFp32AccumulatorGiB": 8, "reportedFailedAllocationGiB": 8, "reportedFreeGiBBeforeFailedAllocation": 6.4, "reportedScratchDeficitGiB": 1.6, "mtpTokensPerSequencePerDecodeStep": 6, "decodeOnlyScheduledTokens": 192, "mixedBatchHeadroomTokens": 32576, "interpretation": "The pinned no-split accumulator shape reproduces the reported 8 GiB request. Decode-only MTP occupancy is small relative to the 32K cap; mixed prefill is required to approach the reported budget." }, "mitigationGates": [ "confirm the exact vLLM commit or immutable image digest and resolved FlashMLA source revision", "confirm GLM-5.2-FP8, FP8 sparse MLA, mixed prefill plus decode, and per-rank head padding select the affected path", "lower max-num-batched-tokens as the direct temporary control and test the resulting throughput and latency", "measure minimum free HBM and peak allocated and reserved bytes during a production-shaped mixed burst, not only startup KV cache usage", "disable MTP for the first control, then add five speculative tokens only after the plain mixed-batch arm passes", "consider prefill-decode disaggregation only when the extra hardware and transfer path are justified", "do not deploy an open PR or switch cache dtype without output-parity, failure, and rollback evidence", "promote only a pinned build containing the merged fix after soak, burst, cancellation, restart, and rollback tests" ], "runtimeBoundary": { "localGpuRuns": 0, "modelCalls": 0, "upstreamTestsExecutedLocally": 0, "issueReportsTreatedAsLocalBenchmarks": false, "openPullRequestsRecommendedForProduction": false, "generatedIllustrationIsBenchmark": false }, "aiHot": { "fingerprint": "f1-b94ea55430da42a9", "batchArtifact": "tmp/aihot-seo-operations/batches/20260826042447-42217f9b8b.json", "batchSha256": "402a026f0b5294927c9be9ac1650ecbdf61ad04fd0e4cfbfb7bb560a258e596f", "classifications": [ { "id": "cmt96t9tm0augrolyx2c3vi84", "permalink": "https://aihot.virxact.com/items/cmt96t9tm0augrolyx2c3vi84", "classification": "duplicate-adjacent-intent", "reason": "The LangChain and Airbyte ingestion item is adjacent to existing LangChain, LlamaIndex, and ColBERT retrieval pages but provides no direct new GLM-5.2 runtime decision." }, { "id": "cmt90wf1p06j9roly5e34fmu3", "permalink": "https://aihot.virxact.com/items/cmt90wf1p06j9roly5e34fmu3", "classification": "weak-glm-link", "reason": "OpenWorker may accept local open models, but the discovery item did not verify a GLM-5.2 adapter, protocol, or reproducible task." }, { "id": "cmt8z2eko055crolytaitdxv8", "permalink": "https://aihot.virxact.com/items/cmt8z2eko055crolytaitdxv8", "classification": "weak-glm-link", "reason": "Claude memory is a different vendor product and has no direct GLM-5.2 technical decision." }, { "id": "cmt8uphwp3qmhro73czv8pbpv", "permalink": "https://aihot.virxact.com/items/cmt8uphwp3qmhro73czv8pbpv", "classification": "weak-glm-link", "reason": "WeatherNext cyclone prediction has no direct GLM-5.2 training, inference, or application boundary." }, { "id": "cmt8vaqtw3r29ro73d67v03bc", "permalink": "https://aihot.virxact.com/items/cmt8vaqtw3r29ro73d67v03bc", "classification": "weak-glm-link", "reason": "The compute-supply forecast is broad commentary without a specific GLM-5.2 reader task or reproducible evidence block." }, { "id": "cmt924nw007a4rolyhmqbswmh", "permalink": "https://aihot.virxact.com/items/cmt924nw007a4rolyhmqbswmh", "classification": "duplicate-intent", "reason": "Existing GLM-5.2 OpenRouter routing and API-provider pages already cover model and provider selection; another broad selector page would compete with them." }, { "id": "cmt8zzh0w05nirolyozmvq7i0", "permalink": "https://aihot.virxact.com/items/cmt8zzh0w05nirolyozmvq7i0", "classification": "weak-glm-link", "reason": "A video-generation API is not a compatible use of the text-only GLM-5.2 checkpoint." } ] }, "searchSupply": { "queries": [ "GLM-5.2 vLLM OOM H200 sparse_decode_fwd", "GLM-5.2 max-num-batched-tokens OOM" ], "method": "public web search used only for result-supply discovery", "serpApiRequests": 0, "serpApiBoundary": "No paid request was made because the site-specific 250-request monthly ceiling was already exceeded in the shared ledger.", "finding": "The dedicated result was the upstream issue itself; adjacent results were generic deployment guides or different OOM paths. No result reconciled the latest stable pin, audited main pin, workspace arithmetic, temporary cap, and two open fixes." }, "overlapAudit": { "prepublicationCanonicalCount": 90, "primaryIntent": "audit and bound the delayed GLM-5.2 FP8 FlashMLA sparse-decode workspace OOM on H200 under vLLM", "readerJob": "identify the affected path, calculate the batch-token workspace, choose a temporary cap or disaggregation gate, and wait for a merged pinned fix before production", "closestPages": [ { "url": "/guides/glm-5-2-vllm-decode-context-parallelism/", "difference": "DCP version and correctness gates; it does not size the no-split FlashMLA accumulator or repair a delayed runtime OOM." }, { "url": "/guides/glm-5-2-mtp-speculative-decoding/", "difference": "MTP configuration and acceptance benchmarking; it does not audit the mixed FP8 sparse workspace failure." }, { "url": "/guides/run-glm-5-2-locally/", "difference": "Broad capacity and route selection; it does not diagnose an H200 service that starts successfully and later dies." }, { "url": "/guides/glm-5-2-hicache-kv-offload/", "difference": "SGLang host and storage cache tiers; the reported vLLM failure is not KV-cache pressure." } ], "distinctIntent": true }, "decision": { "classification": "affected-stable-and-main-use-bounded-temporary-control-wait-for-merged-fix", "stable": "Treat vLLM v0.27.1 as affected for the audited GLM-5.2 FP8 mixed sparse path because it pins the unfixed FlashMLA source.", "currentMain": "Treat audited main 044b0522 as affected because it retains the same FlashMLA pin and full-token mixed path.", "temporaryControl": "Reduce max-num-batched-tokens, disable MTP for the first control, and require a mixed prefill/decode burst plus soak test with measured HBM headroom. The 8192 value is a canary candidate, not a universal safe default.", "alternative": "Use prefill-decode disaggregation only when its extra GPUs, transfer path, and failure surface are justified and separately validated.", "fixBoundary": "Do not deploy either open PR as a production fix. Wait for merge and pinned packaging, then repeat correctness, allocation, service, failure, soak, and rollback tests." }, "invariants": { "twentyOneSourceReceipts": true, "allSourcesReturned200": true, "allSourceHashesValid": true, "glm52ArchitecturePinned": true, "stablePinsAffectedFlashMla": true, "auditedMainPinsAffectedFlashMla": true, "bothFixPathsRemainOpen": true, "reportedEightGiBReproduced": true, "reportedPhysicalHeadroomInsufficient": true, "mtpDecodeOnlyBelowMixedTokenCap": true, "noLocalGpuOrModelRun": true, "openPrNotProductionRecommendation": true, "distinctIntent": true } }