{ "schemaVersion": 1, "kind": "pp-doclayout-v3-webgpu-fp16-experiment", "generatedAt": "2026-07-13T07:08:17Z", "input": { "path": "tmp/pdfs/2307.01197.pdf", "pageCount": 15 }, "runtime": { "name": "in-app-browser-webgpu", "modelSource": "dev", "pipelineWorkers": 1, "layoutBatchSize": 1, "probeCount": 1 }, "artifacts": { "fp32": { "id": "pp-doclayout-v3-jaxjs", "dtype": "float32", "weightsSha256": "13492ac1f96d5340caa0f75e8c894a3b7d8e95d7c2e485fd5b671bee9a3df6f3", "weightsSizeBytes": 130847572 }, "fp16": { "id": "pp-doclayout-v3-fp16-jaxjs", "dtype": "float16", "weightsSha256": "da8613530ea128d8e85d00be530fe02a29830e59308f6dc6c67e12fdb03e6a8f", "weightsSizeBytes": 65459939, "sizeDeltaBytes": -65387633, "sizeDeltaPercent": -49.9723701407 } }, "selection": { "profile": "medium", "canonicalModelSetSha256": "6a1599f083822d6e010a48afdb6230f31acf7ad62ca3231173212cd480fcb6e7", "resolvedLayoutModel": "pp-doclayout-v3-fp16-jaxjs", "hostedArtifactRevision": "78ad71f021ad480ca365c3465aea6d8ed13a4f11", "hostedReleaseDescriptorRevision": "f18021b5c38a47c8fa76fab9fa294a74862aec84", "fp32FallbackPathExists": false, "contractValidatedBeforeRun": true, "networkResourceTimingAvailable": false }, "baseline": { "model": "pp-doclayout-v3-jaxjs", "fullDocumentMs": 89237, "page5CheckpointMs": 16676, "page10CheckpointMs": 46628, "completedPages": 15, "failedPages": 0, "detectedTables": 14, "pagesPerSecond": 0.1680917108 }, "candidate": { "model": "pp-doclayout-v3-fp16-jaxjs", "fullDocumentMs": 74845, "page5CheckpointMs": 21067, "page10CheckpointMs": 50260, "completedPages": 15, "failedPages": 0, "detectedTables": 14, "pagesPerSecond": 0.2004141893, "deviceLossOrTimeout": false }, "deltas": { "fullDocumentMs": -14392, "fullDocumentPercent": -16.1278393491, "pagesPerSecondPercent": 19.2290734184, "page5CheckpointMs": 4391, "page5CheckpointPercent": 26.3312544975, "page10CheckpointMs": 3632, "page10CheckpointPercent": 7.7893111435 }, "quality": { "completedPagesExact": true, "failedPagesExact": true, "tableCountExact": true, "frozenFp16LayoutOracle": "pass", "frozenOracleEvidence": "precision-experiment.json" }, "safety": { "systemMemoryFreePercentBefore": 53, "systemMemoryFreePercentAfter": 51, "memoryPressureNormalAfter": true, "browserWorkerRssMeasured": false, "gpuMemoryMeasured": false, "liveProbeCountAfter": 0 }, "measurementLimitations": [ "Full-document and checkpoint values are conservative first-observed polling times from one directional run, not a variance estimate.", "Page 5 and page 10 were observed at coarse intervals and are upper bounds; both diagnostics regressed relative to the saved FP32 checkpoints.", "The table ONNX worker emitted repeated VerifyOutputSizes warnings without page failures; measured wall time includes that logging overhead.", "Browser worker network timing and attributable RSS/GPU allocation were unavailable, so no transfer-time or memory gain is claimed." ], "decision": { "status": "accept-canonical-medium-default", "canonicalProfile": "medium", "canonicalLayoutModel": "pp-doclayout-v3-fp16-jaxjs", "cliWasmRisk": "known-warm-memory-failure", "hostedHfAvailability": "published-immutable", "productionPublicationAllowed": true, "reason": "One bounded WebGPU run preserved completion and table-count gates while reducing conservative full-document wall time by 16.13%. The canonical medium profile selects FP16 by explicit product decision and is now published through immutable artifact and release-descriptor commits; the prior WASM memory rejection still applies." } }