System requirementsF1.1Requirements, SLOs & baselinesWorkloads & requirementsFoundationsno BaselineF1.1Requirements, SLOs & baselinesWorkloads & requirementsFoundationsno Proxy objectiveF1.3Objectives, proxies & decision policiesWorkloads & requirementsFoundationsno Service-level objectiveF1.1Requirements, SLOs & baselinesWorkloads & requirementsFoundationsno Workload characterizationF1.2Workload characterizationWorkloads & requirementsFoundationsno Execution placementF1.2Workload characterizationWorkloads & requirementsFoundationsno Data qualityP1.3Data validation & feature engineeringData engineeringLifecycleno Label qualityP1.1Data collection & labelingData engineeringLifecycleno Feature engineeringP1.3Data validation & feature engineeringData engineeringLifecycleno Missing dataP1.3Data validation & feature engineeringData engineeringLifecycleno Data leakageP1.6Leakage & training-serving skewData engineeringLifecycleno Point-in-time correctnessP1.6Leakage & training-serving skewData engineeringLifecycleno Data lineageP1.5Data contracts, versioning & lineageData engineeringLifecycleno Training-serving skewP1.6Leakage & training-serving skewData engineeringLifecycleno Dataset deduplicationP1.2Data selection & curationData engineeringLifecycleno Data selectionP1.2Data selection & curationData engineeringLifecycleno TokenizationF2.2Architecture building blocksModels & architecturesFoundationsno Distribution shiftP5.3Drift, feedback loops & retrainingMLOps & operationsLifecycleno Linear modelsF2.1Model familiesModels & architecturesFoundationsyes Tree ensemblesF2.1Model familiesModels & architecturesFoundationsyes Kernel methodsF2.1Model familiesModels & architecturesFoundationsyes Nearest-neighbor modelsF2.1Model familiesModels & architecturesFoundationsyes Neural networksF2.1Model familiesModels & architecturesFoundationsyes Inductive biasF2.1Model familiesModels & architecturesFoundationsyes EmbeddingsF2.2Architecture building blocksModels & architecturesFoundationsno AttentionF2.2Architecture building blocksModels & architecturesFoundationsno Position encodingF2.2Architecture building blocksModels & architecturesFoundationsno Training lossP2.2Optimization & training dynamicsTraining & adaptationLifecycleyes Automatic differentiationS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Optimization algorithmP2.2Optimization & training dynamicsTraining & adaptationLifecycleyes RegularizationP2.2Optimization & training dynamicsTraining & adaptationLifecycleyes Training batch sizeP2.2Optimization & training dynamicsTraining & adaptationLifecycleno Training stateP2.3Training memory & stateTraining & adaptationLifecycleno Activation recomputationP2.3Training memory & stateTraining & adaptationLifecycleno Empirical scaling lawsF3.4Scaling laws & compute-optimal allocationQuantitative system modelsFoundationsno Hyperparameter searchP2.4Hyperparameter search & experiment managementTraining & adaptationLifecycleno Experiment-budget allocationP2.4Hyperparameter search & experiment managementTraining & adaptationLifecycleno Transfer learningP2.5Adaptation: fine-tuning, prompting & transferTraining & adaptationLifecycleno Parameter-efficient adaptationP2.5Adaptation: fine-tuning, prompting & transferTraining & adaptationLifecycleno Tensor layoutS2.2Tensor operators & memory layoutFrameworks, compilers & runtimesStackno Matrix multiplicationS2.2Tensor operators & memory layoutFrameworks, compilers & runtimesStackno Numerical precisionS2.7Numerical formats & precisionFrameworks, compilers & runtimesStackno Numerical stabilityS2.7Numerical formats & precisionFrameworks, compilers & runtimesStackno Mixed-precision trainingS2.7Numerical formats & precisionFrameworks, compilers & runtimesStackno QuantizationQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno SparsityQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Knowledge distillationQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Graph compilationS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Kernel fusionS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Host dispatch overheadS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Memory capacityS1.2Memory hierarchy & bandwidthHardware & acceleratorsStackno Memory bandwidthS1.2Memory hierarchy & bandwidthHardware & acceleratorsStackno Compute throughputS1.1Accelerator architectureHardware & acceleratorsStackno Arithmetic intensityF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Roofline boundF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Data localityF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno TilingS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno ProfilingQ1.1Profiling & utilizationPerformance, efficiency & benchmarkingQualitiesno Resource utilizationQ1.1Profiling & utilizationPerformance, efficiency & benchmarkingQualitiesno Serial fractionF3.2Amdahl's law & scaling efficiencyQuantitative system modelsFoundationsno Collective communicationS3.3Collective communicationParallelism & distributed executionStackno Data parallelismS3.1Parallelism strategiesParallelism & distributed executionStackno Tensor parallelismS3.1Parallelism strategiesParallelism & distributed executionStackno Pipeline parallelismS3.1Parallelism strategiesParallelism & distributed executionStackno Training-state shardingS3.1Parallelism strategiesParallelism & distributed executionStackno Communication overlapS3.3Collective communicationParallelism & distributed executionStackno Network topologyS4.2Network fabricsDatacenter infrastructure & fleetStackno StragglersS3.3Collective communicationParallelism & distributed executionStackno Recovery checkpointQ2.2Checkpointing & elastic recoveryReliability & fault toleranceQualitiesno Consistent checkpoint cutQ2.2Checkpointing & elastic recoveryReliability & fault toleranceQualitiesno Replicated-log consensusS3.4Coordination & consensusParallelism & distributed executionStackyes FencingS3.4Coordination & consensusParallelism & distributed executionStackyes Federated learningS3.5Federated & decentralized learningParallelism & distributed executionStackno Arrival rateF3.3Queueing & capacity modelsQuantitative system modelsFoundationsno Service capacityF3.3Queueing & capacity modelsQuantitative system modelsFoundationsno Queueing delayF3.3Queueing & capacity modelsQuantitative system modelsFoundationsno Tail latencyP4.3Latency, throughput & goodputDeployment & servingLifecycleno Request batchingP4.4Batching, admission & autoscalingDeployment & servingLifecycleno Admission controlP4.4Batching, admission & autoscalingDeployment & servingLifecycleno AutoscalingP4.4Batching, admission & autoscalingDeployment & servingLifecycleno KV cacheP4.5LLM & generative inferenceDeployment & servingLifecycleno Prefill and decodeP4.5LLM & generative inferenceDeployment & servingLifecycleno Paged KV allocationP4.5LLM & generative inferenceDeployment & servingLifecycleno Prefix cachingP4.5LLM & generative inferenceDeployment & servingLifecycleno Speculative decodingP4.5LLM & generative inferenceDeployment & servingLifecycleno SLO goodputP4.3Latency, throughput & goodputDeployment & servingLifecycleno GeneralizationP3.2Metrics, calibration & uncertaintyModel evaluation & testingLifecycleyes Evaluation splitP3.1Evaluation design & experimentsModel evaluation & testingLifecycleno Model-selection optimismP3.1Evaluation design & experimentsModel evaluation & testingLifecycleno Evaluation metricsP3.2Metrics, calibration & uncertaintyModel evaluation & testingLifecycleno Probability calibrationP3.2Metrics, calibration & uncertaintyModel evaluation & testingLifecycleno Statistical uncertaintyP3.2Metrics, calibration & uncertaintyModel evaluation & testingLifecycleno Benchmark protocolQ1.2System benchmarkingPerformance, efficiency & benchmarkingQualitiesno Online experimentationP3.1Evaluation design & experimentsModel evaluation & testingLifecycleno Sliced evaluationP3.3Sliced & robustness evaluationModel evaluation & testingLifecycleno Selective predictionP3.2Metrics, calibration & uncertaintyModel evaluation & testingLifecycleno Artifact versioningP5.1Versioning, rollout & rollbackMLOps & operationsLifecycleno Production monitoringP5.2Monitoring & observabilityMLOps & operationsLifecycleno Staged rolloutP5.1Versioning, rollout & rollbackMLOps & operationsLifecycleno RollbackP5.1Versioning, rollout & rollbackMLOps & operationsLifecycleno Deployment feedback loopP5.3Drift, feedback loops & retrainingMLOps & operationsLifecycleno RetrainingP5.3Drift, feedback loops & retrainingMLOps & operationsLifecycleno Incident responseP5.4Incident responseMLOps & operationsLifecycleno System observabilityP5.2Monitoring & observabilityMLOps & operationsLifecycleno Threat modelingQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Data poisoningQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Adversarial evasionQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Privacy requirementsQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Differential privacyQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Fairness criteriaQ5.1Fairness & biasResponsible AIQualitiesno AuthorizationQ3.4Access control & isolationSecurity & privacyQualitiesno Execution isolationQ3.4Access control & isolationSecurity & privacyQualitiesno Human oversightQ5.3Human oversight & contestabilityResponsible AIQualitiesno Operating energyQ6.1Energy & carbon accountingSustainability & costQualitiesno Lifecycle carbonQ6.1Energy & carbon accountingSustainability & costQualitiesno Cost per useful outcomeQ6.3Cost, TCO & unit economicsSustainability & costQualitiesno Capacity planningS4.5Capacity planning & multi-tenancyDatacenter infrastructure & fleetStackno Agent trajectoryT1.3Agent harness & control loopAgentic systemsSystem typesno Agent planningT1.3Agent harness & control loopAgentic systemsSystem typesno Context windowF2.4Long-context & efficient sequence modelsModels & architecturesFoundationsno RetrievalP4.8Retrieval & vector searchDeployment & servingLifecycleno Approximate nearest-neighbor searchP4.8Retrieval & vector searchDeployment & servingLifecycleno HallucinationP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Tool contractT1.2Tool use & action interfacesAgentic systemsSystem typesno IdempotencyQ2.1Failure models, retries & degradationReliability & fault toleranceQualitiesno Compensating actionQ2.1Failure models, retries & degradationReliability & fault toleranceQualitiesno Prompt injectionQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Trajectory evaluationP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Reward hackingP2.6Preference & RL post-trainingTraining & adaptationLifecycleno Physical dynamicsT2.3Motion planning & controlEmbodied & physical systemsSystem typesyes Sensor calibrationT2.1Sensing, perception & fusionEmbodied & physical systemsSystem typesno State estimationT2.2State estimation, localization & mappingEmbodied & physical systemsSystem typesyes State observabilityT2.2State estimation, localization & mappingEmbodied & physical systemsSystem typesyes Observation ageT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Feedback delayT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Worst-case execution timeT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Feedback controlT2.3Motion planning & controlEmbodied & physical systemsSystem typesyes Trajectory feasibilityT2.3Motion planning & controlEmbodied & physical systemsSystem typesno Contact and complianceT2.3Motion planning & controlEmbodied & physical systemsSystem typesyes Simulation-to-real transferT2.5Robot learning & sim-to-realEmbodied & physical systemsSystem typesno Runtime assuranceQ4.2Runtime assurance & safety filtersSafety & assuranceQualitiesno Stopping envelopeQ4.2Runtime assurance & safety filtersSafety & assuranceQualitiesno Operational design domainQ4.1Hazard analysis & operating domainsSafety & assuranceQualitiesno Safety caseQ4.4Safety cases, standards & certificationSafety & assuranceQualitiesno Structured output constraintP4.6Decoding, sampling & structured outputDeployment & servingLifecycleno Data acquisition and samplingP1.1Data collection & labelingData engineeringLifecycleno Data contractP1.5Data contracts, versioning & lineageData engineeringLifecycleno Delayed outcomesP5.3Drift, feedback loops & retrainingMLOps & operationsLifecycleno Decision policyF1.3Objectives, proxies & decision policiesWorkloads & requirementsFoundationsno Robustness evaluationP3.3Sliced & robustness evaluationModel evaluation & testingLifecycleno Supervised learningP2.1Learning paradigmsTraining & adaptationLifecycleyes Self-supervised learningP2.1Learning paradigmsTraining & adaptationLifecycleyes Unsupervised learningP2.1Learning paradigmsTraining & adaptationLifecycleyes Reinforcement learningP2.1Learning paradigmsTraining & adaptationLifecycleyes Imitation learningP2.1Learning paradigmsTraining & adaptationLifecycleyes Model artifactF2.5Model artifacts & formatsModels & architecturesFoundationsno Failure modelQ2.1Failure models, retries & degradationReliability & fault toleranceQualitiesno BackpressureP4.4Batching, admission & autoscalingDeployment & servingLifecycleno Privacy leakageQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Execution timeQ1.1Profiling & utilizationPerformance, efficiency & benchmarkingQualitiesno Irreversible effectsT1.2Tool use & action interfacesAgentic systemsSystem typesno Dataset storage layoutP1.4Data pipelines & storage formatsData engineeringLifecycleno Data ingestionP1.4Data pipelines & storage formatsData engineeringLifecycleno Schema evolutionP1.4Data pipelines & storage formatsData engineeringLifecycleno Test oracleP3.5ML testing & debuggingModel evaluation & testingLifecycleno Integration testingP3.5ML testing & debuggingModel evaluation & testingLifecycleno Failure injectionQ2.4Fault injection & resilience testingReliability & fault toleranceQualitiesno Dependency environmentP5.5ML pipelines & reproducibilityMLOps & operationsLifecycleno Workflow orchestrationP5.5ML pipelines & reproducibilityMLOps & operationsLifecycleno Experiment reproducibilityP5.5ML pipelines & reproducibilityMLOps & operationsLifecycleno Accelerator execution architectureS1.1Accelerator architectureHardware & acceleratorsStackno Memory hierarchyS1.2Memory hierarchy & bandwidthHardware & acceleratorsStackno Host-device interconnectS1.3Host-device & scale-up interconnectHardware & acceleratorsStackno Power delivery and integrityS1.4Chip power & thermal limitsHardware & acceleratorsStackno Thermal behavior and throttlingS1.4Chip power & thermal limitsHardware & acceleratorsStackno Service-level indicatorF1.1Requirements, SLOs & baselinesWorkloads & requirementsFoundationsno Error budgetF1.1Requirements, SLOs & baselinesWorkloads & requirementsFoundationsno Latency budgetF1.1Requirements, SLOs & baselinesWorkloads & requirementsFoundationsno Operation and byte accountingF1.2Workload characterizationWorkloads & requirementsFoundationsno Memory footprintF1.2Workload characterizationWorkloads & requirementsFoundationsno Workload traceF1.2Workload characterizationWorkloads & requirementsFoundationsno Cost-sensitive decisionF1.3Objectives, proxies & decision policiesWorkloads & requirementsFoundationsno Pareto frontierF1.3Objectives, proxies & decision policiesWorkloads & requirementsFoundationsno TransformerF2.1Model familiesModels & architecturesFoundationsno Deep recommendation modelF2.1Model familiesModels & architecturesFoundationsno Diffusion modelF2.1Model familiesModels & architecturesFoundationsno ConvolutionF2.2Architecture building blocksModels & architecturesFoundationsno Recurrent computationF2.2Architecture building blocksModels & architecturesFoundationsno Normalization layersF2.2Architecture building blocksModels & architecturesFoundationsno Mixture of expertsF2.3Sparse & conditional computationModels & architecturesFoundationsno Expert routingF2.3Sparse & conditional computationModels & architecturesFoundationsno Expert capacityF2.3Sparse & conditional computationModels & architecturesFoundationsno Load-balancing lossF2.3Sparse & conditional computationModels & architecturesFoundationsno Active parametersF2.3Sparse & conditional computationModels & architecturesFoundationsno Early exitF2.3Sparse & conditional computationModels & architecturesFoundationsno Sparse and sliding-window attentionF2.4Long-context & efficient sequence modelsModels & architecturesFoundationsno Linear attentionF2.4Long-context & efficient sequence modelsModels & architecturesFoundationsno State-space sequence modelF2.4Long-context & efficient sequence modelsModels & architecturesFoundationsno Grouped-query attentionF2.4Long-context & efficient sequence modelsModels & architecturesFoundationsno Context-length extensionF2.4Long-context & efficient sequence modelsModels & architecturesFoundationsno Weight serialization formatF2.5Model artifacts & formatsModels & architecturesFoundationsno Model exchange formatF2.5Model artifacts & formatsModels & architecturesFoundationsno Model exportF2.5Model artifacts & formatsModels & architecturesFoundationsno Model signatureF2.5Model artifacts & formatsModels & architecturesFoundationsno Ridge pointF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Compute-bound vs. memory-boundF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Data-movement energyF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Amdahl's lawF3.2Amdahl's law & scaling efficiencyQuantitative system modelsFoundationsno Strong scalingF3.2Amdahl's law & scaling efficiencyQuantitative system modelsFoundationsno Weak scalingF3.2Amdahl's law & scaling efficiencyQuantitative system modelsFoundationsno Parallel scaling efficiencyF3.2Amdahl's law & scaling efficiencyQuantitative system modelsFoundationsno Little's lawF3.3Queueing & capacity modelsQuantitative system modelsFoundationsno Latency-throughput curveF3.3Queueing & capacity modelsQuantitative system modelsFoundationsno Arrival burstinessF3.3Queueing & capacity modelsQuantitative system modelsFoundationsno Compute-optimal allocationF3.4Scaling laws & compute-optimal allocationQuantitative system modelsFoundationsno Training compute estimateF3.4Scaling laws & compute-optimal allocationQuantitative system modelsFoundationsno Systolic arrayS1.1Accelerator architectureHardware & acceleratorsStackno Matrix engineS1.1Accelerator architectureHardware & acceleratorsStackno SIMT executionS1.1Accelerator architectureHardware & acceleratorsStackno Accelerator dataflowS1.1Accelerator architectureHardware & acceleratorsStackno Domain-specific architectureS1.1Accelerator architectureHardware & acceleratorsStackno High-bandwidth memoryS1.2Memory hierarchy & bandwidthHardware & acceleratorsStackno On-chip SRAM and scratchpadsS1.2Memory hierarchy & bandwidthHardware & acceleratorsStackno Memory wallS1.2Memory hierarchy & bandwidthHardware & acceleratorsStackno Scale-up interconnectS1.3Host-device & scale-up interconnectHardware & acceleratorsStackno Cache-coherent interconnectS1.3Host-device & scale-up interconnectHardware & acceleratorsStackno Peer-to-peer device transferS1.3Host-device & scale-up interconnectHardware & acceleratorsStackno Thermal design powerS1.4Chip power & thermal limitsHardware & acceleratorsStackno Dynamic voltage and frequency scalingS1.4Chip power & thermal limitsHardware & acceleratorsStackno Dark siliconS1.4Chip power & thermal limitsHardware & acceleratorsStackno Power cappingS1.4Chip power & thermal limitsHardware & acceleratorsStackno Microcontroller-class deviceS1.5Edge & embedded hardwareHardware & acceleratorsStackno Heterogeneous mobile SoCS1.5Edge & embedded hardwareHardware & acceleratorsStackno On-device neural acceleratorS1.5Edge & embedded hardwareHardware & acceleratorsStackno Energy per inferenceQ6.1Energy & carbon accountingSustainability & costQualitiesno On-device memory budgetS1.5Edge & embedded hardwareHardware & acceleratorsStackno TensorS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Computation graphS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Eager vs. graph executionS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Tracing and JIT captureS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno IO-aware attention kernelS2.2Tensor operators & memory layoutFrameworks, compilers & runtimesStackno Convolution loweringS2.2Tensor operators & memory layoutFrameworks, compilers & runtimesStackno Vendor kernel libraryS2.2Tensor operators & memory layoutFrameworks, compilers & runtimesStackno ML compiler IRS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Kernel autotuningS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Kernel DSLS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Static memory planningS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Device streamsS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Caching memory allocatorS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Captured graph replayS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Loss scalingS2.7Numerical formats & precisionFrameworks, compilers & runtimesStackno Block-scaled number formatsS2.7Numerical formats & precisionFrameworks, compilers & runtimesStackno Accumulation precisionS2.7Numerical formats & precisionFrameworks, compilers & runtimesStackno Expert parallelismS3.1Parallelism strategiesParallelism & distributed executionStackno Sequence and context parallelismS3.1Parallelism strategiesParallelism & distributed executionStackno Hybrid parallelismS3.1Parallelism strategiesParallelism & distributed executionStackno Pipeline bubbleS3.1Parallelism strategiesParallelism & distributed executionStackno Ring all-reduceS3.3Collective communicationParallelism & distributed executionStackno Tree and hierarchical all-reduceS3.3Collective communicationParallelism & distributed executionStackno All-to-all exchangeS3.3Collective communicationParallelism & distributed executionStackno Alpha-beta communication modelS3.3Collective communicationParallelism & distributed executionStackno Gradient compressionS3.3Collective communicationParallelism & distributed executionStackno LeaseS3.4Coordination & consensusParallelism & distributed executionStackyes Leader electionS3.4Coordination & consensusParallelism & distributed executionStackyes Bounded stalenessS3.4Coordination & consensusParallelism & distributed executionStackno Federated averagingS3.5Federated & decentralized learningParallelism & distributed executionStackno Statistical heterogeneityS3.5Federated & decentralized learningParallelism & distributed executionStackno Client selectionS3.5Federated & decentralized learningParallelism & distributed executionStackno Cross-device vs. cross-silo federationS3.5Federated & decentralized learningParallelism & distributed executionStackno Power usage effectivenessS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno Liquid coolingS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno Rack power densityS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno Power oversubscriptionS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno Warehouse-scale computerS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno Remote direct memory accessS4.2Network fabricsDatacenter infrastructure & fleetStackno Lossless network fabricS4.2Network fabricsDatacenter infrastructure & fleetStackno Datacenter congestion controlS4.2Network fabricsDatacenter infrastructure & fleetStackno Fat-tree (Clos) topologyS4.2Network fabricsDatacenter infrastructure & fleetStackno Rail-optimized topologyS4.2Network fabricsDatacenter infrastructure & fleetStackno Bisection bandwidthS4.2Network fabricsDatacenter infrastructure & fleetStackno In-network reductionS4.2Network fabricsDatacenter infrastructure & fleetStackno Parallel file systemS4.3Storage & data deliveryDatacenter infrastructure & fleetStackno Object storageS4.3Storage & data deliveryDatacenter infrastructure & fleetStackno Training input pipelineS2.6Data loading & input pipelinesFrameworks, compilers & runtimesStackno Checkpoint I/OS4.3Storage & data deliveryDatacenter infrastructure & fleetStackno Storage tiering and cachingS4.3Storage & data deliveryDatacenter infrastructure & fleetStackno Gang schedulingS4.4Cluster scheduling & orchestrationDatacenter infrastructure & fleetStackno Job preemptionS4.4Cluster scheduling & orchestrationDatacenter infrastructure & fleetStackno Dominant resource fairnessS4.4Cluster scheduling & orchestrationDatacenter infrastructure & fleetStackno Elastic trainingS4.4Cluster scheduling & orchestrationDatacenter infrastructure & fleetStackno Topology-aware placementS4.4Cluster scheduling & orchestrationDatacenter infrastructure & fleetStackno Bin packingS4.4Cluster scheduling & orchestrationDatacenter infrastructure & fleetStackno Multi-tenancyS4.5Capacity planning & multi-tenancyDatacenter infrastructure & fleetStackno Resource quotaS4.5Capacity planning & multi-tenancyDatacenter infrastructure & fleetStackno Accelerator sharing and partitioningS4.5Capacity planning & multi-tenancyDatacenter infrastructure & fleetStackno Spot and preemptible capacityS4.5Capacity planning & multi-tenancyDatacenter infrastructure & fleetStackno Performance interferenceS4.5Capacity planning & multi-tenancyDatacenter infrastructure & fleetStackno Weak supervisionP1.1Data collection & labelingData engineeringLifecycleno Active learningP1.1Data collection & labelingData engineeringLifecycleno Inter-annotator agreementP1.1Data collection & labelingData engineeringLifecycleno Quality filteringP1.2Data selection & curationData engineeringLifecycleno Data mixtureP1.2Data selection & curationData engineeringLifecycleno Synthetic dataP1.2Data selection & curationData engineeringLifecycleno Data augmentationP1.2Data selection & curationData engineeringLifecycleno Data validationP1.3Data validation & feature engineeringData engineeringLifecycleno Feature storeP1.3Data validation & feature engineeringData engineeringLifecycleno Stream processingP1.4Data pipelines & storage formatsData engineeringLifecycleno Feature backfillP1.4Data pipelines & storage formatsData engineeringLifecycleno Dataset versioningP1.5Data contracts, versioning & lineageData engineeringLifecycleno Data dependency debtP1.5Data contracts, versioning & lineageData engineeringLifecycleno Feature freshnessP1.6Leakage & training-serving skewData engineeringLifecycleno Learning-rate scheduleP2.2Optimization & training dynamicsTraining & adaptationLifecycleno Critical batch sizeP2.2Optimization & training dynamicsTraining & adaptationLifecycleno Training instabilityP2.2Optimization & training dynamicsTraining & adaptationLifecycleno Gradient accumulationP2.3Training memory & stateTraining & adaptationLifecycleno Training-state offloadingP2.3Training memory & stateTraining & adaptationLifecycleno Experiment trackingP2.4Hyperparameter search & experiment managementTraining & adaptationLifecycleno Hyperparameter transferP2.4Hyperparameter search & experiment managementTraining & adaptationLifecycleno Supervised fine-tuningP2.5Adaptation: fine-tuning, prompting & transferTraining & adaptationLifecycleno Low-rank adaptationP2.5Adaptation: fine-tuning, prompting & transferTraining & adaptationLifecycleno RL from human feedbackP2.6Preference & RL post-trainingTraining & adaptationLifecycleno Reward modelP2.6Preference & RL post-trainingTraining & adaptationLifecycleno Direct preference optimizationP2.6Preference & RL post-trainingTraining & adaptationLifecycleno RL with verifiable rewardsP2.6Preference & RL post-trainingTraining & adaptationLifecycleno RL rollout generationP2.6Preference & RL post-trainingTraining & adaptationLifecycleno Policy weight synchronizationP2.6Preference & RL post-trainingTraining & adaptationLifecycleno Policy stalenessP2.6Preference & RL post-trainingTraining & adaptationLifecycleno Online learningP2.7Continual & online learningTraining & adaptationLifecycleno Incremental trainingP2.7Continual & online learningTraining & adaptationLifecycleno Catastrophic forgettingP2.7Continual & online learningTraining & adaptationLifecycleno Experience replayP2.7Continual & online learningTraining & adaptationLifecycleno On-device trainingP2.7Continual & online learningTraining & adaptationLifecycleno Cross-validationP3.1Evaluation design & experimentsModel evaluation & testingLifecycleyes Sequential testingP3.1Evaluation design & experimentsModel evaluation & testingLifecycleno Experiment interferenceP3.1Evaluation design & experimentsModel evaluation & testingLifecycleno Conformal predictionP3.2Metrics, calibration & uncertaintyModel evaluation & testingLifecycleyes Slice discoveryP3.3Sliced & robustness evaluationModel evaluation & testingLifecycleno Model-based gradingP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno pass@kP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Pairwise preference evaluationP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Human evaluationP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Benchmark contaminationP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Time to trainQ1.2System benchmarkingPerformance, efficiency & benchmarkingQualitiesno Inference load scenarioQ1.2System benchmarkingPerformance, efficiency & benchmarkingQualitiesno Metamorphic testingP3.5ML testing & debuggingModel evaluation & testingLifecycleno Behavioral testingP3.5ML testing & debuggingModel evaluation & testingLifecycleno Production-readiness assessmentP3.5ML testing & debuggingModel evaluation & testingLifecycleno Online inferenceP4.1Serving architecturesDeployment & servingLifecycleno Batch inferenceP4.1Serving architecturesDeployment & servingLifecycleno Model serverP4.1Serving architecturesDeployment & servingLifecycleno Multi-model servingP4.1Serving architecturesDeployment & servingLifecycleno Model cascadeP4.2Compound AI systems & model routingDeployment & servingLifecycleno Model routingP4.2Compound AI systems & model routingDeployment & servingLifecycleno Time to first tokenP4.3Latency, throughput & goodputDeployment & servingLifecycleno Time per output tokenP4.3Latency, throughput & goodputDeployment & servingLifecycleno Hedged requestsP4.3Latency, throughput & goodputDeployment & servingLifecycleno Continuous batchingP4.4Batching, admission & autoscalingDeployment & servingLifecycleno Request load balancingP4.4Batching, admission & autoscalingDeployment & servingLifecycleno Model cold startP4.4Batching, admission & autoscalingDeployment & servingLifecycleno Chunked prefillP4.5LLM & generative inferenceDeployment & servingLifecycleno Prefill-decode disaggregationP4.5LLM & generative inferenceDeployment & servingLifecycleno Decoding and sampling controlsP4.6Decoding, sampling & structured outputDeployment & servingLifecycleno Constrained decodingP4.6Decoding, sampling & structured outputDeployment & servingLifecycleno Output validationP4.6Decoding, sampling & structured outputDeployment & servingLifecycleno Reasoning budgetP4.7Test-time computeDeployment & servingLifecycleno Best-of-n samplingP4.7Test-time computeDeployment & servingLifecycleno Self-consistency votingP4.7Test-time computeDeployment & servingLifecycleno VerifierP4.7Test-time computeDeployment & servingLifecycleno Process reward modelP4.7Test-time computeDeployment & servingLifecycleno Inference-time searchP4.7Test-time computeDeployment & servingLifecycleno Test-time scalingP4.7Test-time computeDeployment & servingLifecycleno Retrieval-augmented generationP4.8Retrieval & vector searchDeployment & servingLifecycleno Dense retrievalP4.8Retrieval & vector searchDeployment & servingLifecycleno Lexical retrievalP4.8Retrieval & vector searchDeployment & servingLifecycleyes RerankingP4.8Retrieval & vector searchDeployment & servingLifecycleno Model conversionP4.9On-device & edge deploymentDeployment & servingLifecycleno On-device inference runtimeP4.9On-device & edge deploymentDeployment & servingLifecycleno Over-the-air model updateP4.9On-device & edge deploymentDeployment & servingLifecycleno Device-cloud partitioningP4.9On-device & edge deploymentDeployment & servingLifecycleno Device heterogeneityP4.9On-device & edge deploymentDeployment & servingLifecycleno Model registryP5.1Versioning, rollout & rollbackMLOps & operationsLifecycleno Shadow deploymentP5.1Versioning, rollout & rollbackMLOps & operationsLifecycleno CI/CD for MLP5.1Versioning, rollout & rollbackMLOps & operationsLifecycleno Release gateP5.1Versioning, rollout & rollbackMLOps & operationsLifecycleno Drift detectionP5.2Monitoring & observabilityMLOps & operationsLifecycleno Prediction loggingP5.2Monitoring & observabilityMLOps & operationsLifecycleno AlertingP5.2Monitoring & observabilityMLOps & operationsLifecycleno Concept driftP5.3Drift, feedback loops & retrainingMLOps & operationsLifecycleno Continuous trainingP5.3Drift, feedback loops & retrainingMLOps & operationsLifecycleno Blameless postmortemP5.4Incident responseMLOps & operationsLifecycleyes ML failure attributionP5.4Incident responseMLOps & operationsLifecycleno ML configuration managementP5.5ML pipelines & reproducibilityMLOps & operationsLifecycleno Infrastructure as codeP5.5ML pipelines & reproducibilityMLOps & operationsLifecycleyes ML platformP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno ML technical debtP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno EntanglementP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno Undeclared consumersP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno MLOps maturityP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno ML team rolesP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno Model deprecationP5.7Model retirementMLOps & operationsLifecycleno Artifact archivalP5.7Model retirementMLOps & operationsLifecycleno Data retentionP5.7Model retirementMLOps & operationsLifecycleno Model FLOPs utilizationQ1.1Profiling & utilizationPerformance, efficiency & benchmarkingQualitiesno Bottleneck analysisQ1.1Profiling & utilizationPerformance, efficiency & benchmarkingQualitiesno Training goodputQ1.1Profiling & utilizationPerformance, efficiency & benchmarkingQualitiesno PruningQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Structured pruningQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Low-rank factorizationQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Quantization-aware trainingQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Neural architecture searchQ1.4Hardware-aware model designPerformance, efficiency & benchmarkingQualitiesno Hardware-aware NASQ1.4Hardware-aware model designPerformance, efficiency & benchmarkingQualitiesno Depthwise separable convolutionF2.2Architecture building blocksModels & architecturesFoundationsno Accuracy-efficiency frontierQ1.4Hardware-aware model designPerformance, efficiency & benchmarkingQualitiesno Model-hardware co-designQ1.4Hardware-aware model designPerformance, efficiency & benchmarkingQualitiesno Retry with backoffQ2.1Failure models, retries & degradationReliability & fault toleranceQualitiesyes Circuit breakerQ2.1Failure models, retries & degradationReliability & fault toleranceQualitiesyes Replication and failoverQ2.1Failure models, retries & degradationReliability & fault toleranceQualitiesno Graceful degradationQ2.1Failure models, retries & degradationReliability & fault toleranceQualitiesno Optimal checkpoint intervalQ2.2Checkpointing & elastic recoveryReliability & fault toleranceQualitiesno Asynchronous checkpointingQ2.2Checkpointing & elastic recoveryReliability & fault toleranceQualitiesno Silent data corruptionQ2.3Hardware faults & silent data corruptionReliability & fault toleranceQualitiesno Mean time between failuresQ2.3Hardware faults & silent data corruptionReliability & fault toleranceQualitiesno AvailabilityQ2.3Hardware faults & silent data corruptionReliability & fault toleranceQualitiesno Gray failureQ2.3Hardware faults & silent data corruptionReliability & fault toleranceQualitiesno Failure domainQ2.3Hardware faults & silent data corruptionReliability & fault toleranceQualitiesno Algorithm-based fault toleranceQ2.3Hardware faults & silent data corruptionReliability & fault toleranceQualitiesno Chaos engineeringQ2.4Fault injection & resilience testingReliability & fault toleranceQualitiesyes Game dayQ2.4Fault injection & resilience testingReliability & fault toleranceQualitiesyes Hardware fault injectionQ2.4Fault injection & resilience testingReliability & fault toleranceQualitiesno Blast radiusQ2.4Fault injection & resilience testingReliability & fault toleranceQualitiesno Backdoor attackQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno JailbreakQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Adversarial trainingQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Model extractionQ3.2Model & supply-chain securitySecurity & privacyQualitiesno Model watermarkingQ3.2Model & supply-chain securitySecurity & privacyQualitiesno Model artifact signingQ3.2Model & supply-chain securitySecurity & privacyQualitiesno Unsafe model deserializationQ3.2Model & supply-chain securitySecurity & privacyQualitiesno ML supply-chain compromiseQ3.2Model & supply-chain securitySecurity & privacyQualitiesno Membership inferenceQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Model inversionQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno DP-SGDQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Machine unlearningQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Secure aggregationQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Least privilegeQ3.4Access control & isolationSecurity & privacyQualitiesno Confidential computingQ3.4Access control & isolationSecurity & privacyQualitiesno Remote attestationQ3.4Access control & isolationSecurity & privacyQualitiesno Side-channel attackQ3.4Access control & isolationSecurity & privacyQualitiesno Energy efficiencyQ6.1Energy & carbon accountingSustainability & costQualitiesno Grid carbon intensityQ6.1Energy & carbon accountingSustainability & costQualitiesno Operational carbonQ6.1Energy & carbon accountingSustainability & costQualitiesno Emission scopesQ6.1Energy & carbon accountingSustainability & costQualitiesno Carbon-aware schedulingQ6.1Energy & carbon accountingSustainability & costQualitiesno Embodied carbonQ6.2Embodied carbon & lifecycle impactSustainability & costQualitiesno Life cycle assessmentQ6.2Embodied carbon & lifecycle impactSustainability & costQualitiesno Water footprintQ6.2Embodied carbon & lifecycle impactSustainability & costQualitiesno Hardware service lifetimeQ6.2Embodied carbon & lifecycle impactSustainability & costQualitiesno Electronic wasteQ6.2Embodied carbon & lifecycle impactSustainability & costQualitiesno Total cost of ownershipQ6.3Cost, TCO & unit economicsSustainability & costQualitiesno Cost per tokenQ6.3Cost, TCO & unit economicsSustainability & costQualitiesno Cost attributionQ6.3Cost, TCO & unit economicsSustainability & costQualitiesno Hazard analysisQ4.1Hazard analysis & operating domainsSafety & assuranceQualitiesno System-theoretic process analysisQ4.1Hazard analysis & operating domainsSafety & assuranceQualitiesno Failure modes and effects analysisQ4.1Hazard analysis & operating domainsSafety & assuranceQualitiesno Functional insufficiency (SOTIF)Q4.1Hazard analysis & operating domainsSafety & assuranceQualitiesno Safety filterQ4.2Runtime assurance & safety filtersSafety & assuranceQualitiesno Control barrier functionQ4.2Runtime assurance & safety filtersSafety & assuranceQualitiesno Simplex architectureQ4.2Runtime assurance & safety filtersSafety & assuranceQualitiesno Minimal-risk conditionQ4.2Runtime assurance & safety filtersSafety & assuranceQualitiesno GuardrailsQ4.2Runtime assurance & safety filtersSafety & assuranceQualitiesno Scenario-based testingQ4.3Safety verification & validationSafety & assuranceQualitiesno Simulation-based falsificationQ4.3Safety verification & validationSafety & assuranceQualitiesno Formal verificationQ4.3Safety verification & validationSafety & assuranceQualitiesno Neural network verificationQ4.3Safety verification & validationSafety & assuranceQualitiesno Rare-event safety validationQ4.3Safety verification & validationSafety & assuranceQualitiesno X-in-the-loop testingQ4.3Safety verification & validationSafety & assuranceQualitiesno Goal Structuring NotationQ4.4Safety cases, standards & certificationSafety & assuranceQualitiesyes Functional safetyQ4.4Safety cases, standards & certificationSafety & assuranceQualitiesno Safety integrity levelQ4.4Safety cases, standards & certificationSafety & assuranceQualitiesyes Safety performance indicatorQ4.4Safety cases, standards & certificationSafety & assuranceQualitiesno Group fairness metricsQ5.1Fairness & biasResponsible AIQualitiesno Individual fairnessQ5.1Fairness & biasResponsible AIQualitiesno Fairness criteria incompatibilityQ5.1Fairness & biasResponsible AIQualitiesno Bias mitigationQ5.1Fairness & biasResponsible AIQualitiesno Feature attributionQ5.2Explainability & documentationResponsible AIQualitiesno Counterfactual explanationQ5.2Explainability & documentationResponsible AIQualitiesno Inherently interpretable modelQ5.2Explainability & documentationResponsible AIQualitiesno Model cardQ5.2Explainability & documentationResponsible AIQualitiesno Datasheet for datasetsQ5.2Explainability & documentationResponsible AIQualitiesno Contestability and recourseQ5.3Human oversight & contestabilityResponsible AIQualitiesno Automation biasQ5.3Human oversight & contestabilityResponsible AIQualitiesno Authority handoverQ5.3Human oversight & contestabilityResponsible AIQualitiesno AI risk managementQ5.4AI governance, audit & regulationResponsible AIQualitiesno Algorithmic impact assessmentQ5.4AI governance, audit & regulationResponsible AIQualitiesno Algorithmic auditQ5.4AI governance, audit & regulationResponsible AIQualitiesno Risk-tiered AI regulationQ5.4AI governance, audit & regulationResponsible AIQualitiesno Frontier safety frameworkQ5.4AI governance, audit & regulationResponsible AIQualitiesno Context compactionT1.1Context engineering & agent memoryAgentic systemsSystem typesno Long-context degradationF2.4Long-context & efficient sequence modelsModels & architecturesFoundationsno Long-term agent memoryT1.1Context engineering & agent memoryAgentic systemsSystem typesno Function callingT1.2Tool use & action interfacesAgentic systemsSystem typesno Tool discovery protocolT1.2Tool use & action interfacesAgentic systemsSystem typesno Computer useT1.2Tool use & action interfacesAgentic systemsSystem typesno Agent harnessT1.3Agent harness & control loopAgentic systemsSystem typesno Agent execution budgetT1.3Agent harness & control loopAgentic systemsSystem typesno Action approval gateT1.3Agent harness & control loopAgentic systemsSystem typesno Durable executionT1.4Durable execution & agent stateAgentic systemsSystem typesno Record and replayT1.4Durable execution & agent stateAgentic systemsSystem typesno Suspend and resumeT1.4Durable execution & agent stateAgentic systemsSystem typesno Multi-agent orchestrationT1.5Multi-agent orchestrationAgentic systemsSystem typesno Shared agent stateT1.5Multi-agent orchestrationAgentic systemsSystem typesno Multi-agent coordination failureT1.5Multi-agent orchestrationAgentic systemsSystem typesno Correlated agent errorsT1.5Multi-agent orchestrationAgentic systemsSystem typesno Robot perceptionT2.1Sensing, perception & fusionEmbodied & physical systemsSystem typesyes Sensor fusionT2.1Sensing, perception & fusionEmbodied & physical systemsSystem typesno Sensor time synchronizationT2.1Sensing, perception & fusionEmbodied & physical systemsSystem typesno Recursive state filteringT2.2State estimation, localization & mappingEmbodied & physical systemsSystem typesyes Simultaneous localization and mappingT2.2State estimation, localization & mappingEmbodied & physical systemsSystem typesyes Learned world modelT2.2State estimation, localization & mappingEmbodied & physical systemsSystem typesno Motion planningT2.3Motion planning & controlEmbodied & physical systemsSystem typesyes Model predictive controlT2.3Motion planning & controlEmbodied & physical systemsSystem typesyes Real-time schedulingT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Mixed-criticality partitioningT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Vision-language-action modelT2.5Robot learning & sim-to-realEmbodied & physical systemsSystem typesno Teleoperation dataT2.5Robot learning & sim-to-realEmbodied & physical systemsSystem typesno Action chunkingT2.5Robot learning & sim-to-realEmbodied & physical systemsSystem typesno Domain randomizationT2.5Robot learning & sim-to-realEmbodied & physical systemsSystem typesno Thread hierarchyS2.3Kernel & GPU programming modelFrameworks, compilers & runtimesStackno Memory coalescing and bank conflictsS2.3Kernel & GPU programming modelFrameworks, compilers & runtimesStackno Occupancy and latency hidingS2.3Kernel & GPU programming modelFrameworks, compilers & runtimesStackno Asynchronous data movementS2.3Kernel & GPU programming modelFrameworks, compilers & runtimesStackno Matrix-engine programmingS2.3Kernel & GPU programming modelFrameworks, compilers & runtimesStackno Embedding-table shardingS3.1Parallelism strategiesParallelism & distributed executionStackno Embedding cache and tieringS1.2Memory hierarchy & bandwidthHardware & acceleratorsStackno Parameter serverS3.1Parallelism strategiesParallelism & distributed executionStackno Multi-stage retrieval and rankingP4.1Serving architecturesDeployment & servingLifecycleno Multimodal modelF2.1Model familiesModels & architecturesFoundationsno Modality encoderF2.2Architecture building blocksModels & architecturesFoundationsno Multimodal encoder disaggregationP4.5LLM & generative inferenceDeployment & servingLifecycleno Automatic parallelizationS3.2Parallelization planning & placementParallelism & distributed executionStackno Sharding annotation and propagationS3.2Parallelization planning & placementParallelism & distributed executionStackno Device meshS3.2Parallelization planning & placementParallelism & distributed executionStackno Composable function transformationsS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Distributed task and actor runtimeS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Single- vs. multi-controller executionS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Analytical performance modelF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Performance simulationF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Simulator fidelityF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Post-training quantizationQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Weight-only quantizationQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Activation outliersQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno KV-cache quantizationQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno N:M structured sparsityQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno KV-cache offloading and tieringP4.5LLM & generative inferenceDeployment & servingLifecycleno KV-cache transferP4.5LLM & generative inferenceDeployment & servingLifecycleno Compound AI systemP4.2Compound AI systems & model routingDeployment & servingLifecycleno LM program optimizationP4.2Compound AI systems & model routingDeployment & servingLifecycleno Multi-adapter servingP4.1Serving architecturesDeployment & servingLifecycleno Denoising-step reductionP4.5LLM & generative inferenceDeployment & servingLifecycleno Cross-step feature cachingP4.5LLM & generative inferenceDeployment & servingLifecycleno Graph-based vector indexP4.8Retrieval & vector searchDeployment & servingLifecycleno Partitioned and quantized vector indexP4.8Retrieval & vector searchDeployment & servingLifecycleno Filtered vector searchP4.8Retrieval & vector searchDeployment & servingLifecycleno Vector index freshnessP4.8Retrieval & vector searchDeployment & servingLifecycleno Numerical nondeterminismS2.7Numerical formats & precisionFrameworks, compilers & runtimesStackno Batch invarianceS2.7Numerical formats & precisionFrameworks, compilers & runtimesStackno Checkpoint reshardingQ2.2Checkpointing & elastic recoveryReliability & fault toleranceQualitiesno Distributed batch data processingP1.4Data pipelines & storage formatsData engineeringLifecycleno Distributed tracingP5.2Monitoring & observabilityMLOps & operationsLifecycleno Model and agent call tracingP5.2Monitoring & observabilityMLOps & operationsLifecycleno Evaluation harnessP3.1Evaluation design & experimentsModel evaluation & testingLifecycleno Energy measurement methodologyQ6.1Energy & carbon accountingSustainability & costQualitiesno Idle vs. dynamic powerQ6.1Energy & carbon accountingSustainability & costQualitiesno Homomorphic-encryption inferenceQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Secure multi-party computation inferenceQ3.3Privacy leakage & differential privacySecurity & privacyQualitiesno Red teamingQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Dangerous-capability evaluationP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Compute-in-memoryS1.1Accelerator architectureHardware & acceleratorsStackno Emerging non-volatile memoryS1.2Memory hierarchy & bandwidthHardware & acceleratorsStackno Embedded ML DSLS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Differentiable programmingS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Persistent kernelS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Tensor graph superoptimizationS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Large-model training frameworkS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Cloud service modelsS4.5Capacity planning & multi-tenancyDatacenter infrastructure & fleetStackno Data warehouse and lakehouseP1.4Data pipelines & storage formatsData engineeringLifecycleno Synchronous training power swingsS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno UPS and battery energy storageS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno Grid interconnectionS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno On-site power generationS4.1Datacenter power, cooling & facilitiesDatacenter infrastructure & fleetStackno Data cleaningP1.3Data validation & feature engineeringData engineeringLifecycleno Class imbalanceP1.1Data collection & labelingData engineeringLifecycleno Class rebalancingP1.1Data collection & labelingData engineeringLifecycleno Feature hashingP1.3Data validation & feature engineeringData engineeringLifecycleno Feature crossP1.3Data validation & feature engineeringData engineeringLifecycleno Data governanceP1.5Data contracts, versioning & lineageData engineeringLifecycleno Consent and purpose limitationP1.5Data contracts, versioning & lineageData engineeringLifecycleno Exploratory data analysisP1.3Data validation & feature engineeringData engineeringLifecycleno Model and prediction visualizationP3.5ML testing & debuggingModel evaluation & testingLifecycleno Model selectionP2.4Hyperparameter search & experiment managementTraining & adaptationLifecycleno PretrainingP2.1Learning paradigmsTraining & adaptationLifecycleno Causal language modelingP2.1Learning paradigmsTraining & adaptationLifecycleno Generative adversarial networkF2.1Model familiesModels & architecturesFoundationsno Mode collapseP2.2Optimization & training dynamicsTraining & adaptationLifecycleno PerplexityP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Reference-match scoringP3.4Evaluating generative & agentic systemsModel evaluation & testingLifecycleno Error analysisP3.5ML testing & debuggingModel evaluation & testingLifecycleno Training diagnosticsP3.5ML testing & debuggingModel evaluation & testingLifecycleno Request correlation keysP4.1Serving architecturesDeployment & servingLifecycleno Stateless model servingP4.1Serving architecturesDeployment & servingLifecycleno Model gatewayP4.2Compound AI systems & model routingDeployment & servingLifecycleno Semantic cacheP4.2Compound AI systems & model routingDeployment & servingLifecycleno Glue codeP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno Pipeline jungleP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno Build-versus-buy decisionP5.6ML platforms, teams & technical debtMLOps & operationsLifecycleno User feedback collectionP5.3Drift, feedback loops & retrainingMLOps & operationsLifecycleno Human-AI interaction designQ5.3Human oversight & contestabilityResponsible AIQualitiesno Datacenter demand responseQ6.1Energy & carbon accountingSustainability & costQualitiesno Prompt engineeringP2.5Adaptation: fine-tuning, prompting & transferTraining & adaptationLifecycleno In-context learningP2.5Adaptation: fine-tuning, prompting & transferTraining & adaptationLifecycleno Streaming speech recognitionP4.1Serving architecturesDeployment & servingLifecycleno Voice interaction pipelineP4.1Serving architecturesDeployment & servingLifecycleno KV-cache evictionP4.5LLM & generative inferenceDeployment & servingLifecycleno Dynamic sparse attentionF2.4Long-context & efficient sequence modelsModels & architecturesFoundationsno CPU-GPU hybrid inferenceP4.5LLM & generative inferenceDeployment & servingLifecycleno Heterogeneous-accelerator training and servingS3.2Parallelization planning & placementParallelism & distributed executionStackno Graph neural networkF2.1Model familiesModels & architecturesFoundationsno Neighbor samplingS2.6Data loading & input pipelinesFrameworks, compilers & runtimesStackno Dynamic re-parallelizationS3.2Parallelization planning & placementParallelism & distributed executionStackno Contextual activation sparsityQ1.3Model compressionPerformance, efficiency & benchmarkingQualitiesno Non-prefix KV reuseP4.5LLM & generative inferenceDeployment & servingLifecycleno ML-aware cluster schedulingS4.4Cluster scheduling & orchestrationDatacenter infrastructure & fleetStackno Adaptive routingS4.2Network fabricsDatacenter infrastructure & fleetStackno Incast congestionS4.2Network fabricsDatacenter infrastructure & fleetStackno Context selectionT1.1Context engineering & agent memoryAgentic systemsSystem typesno Context stalenessT1.1Context engineering & agent memoryAgentic systemsSystem typesno Iterative multi-hop retrievalP4.8Retrieval & vector searchDeployment & servingLifecycleno Capability-based authorizationQ3.4Access control & isolationSecurity & privacyQualitiesno Ephemeral sandboxQ3.4Access control & isolationSecurity & privacyQualitiesno Network egress controlQ3.4Access control & isolationSecurity & privacyQualitiesno Information-flow controlQ3.4Access control & isolationSecurity & privacyQualitiesno Compounding error in sequential decisionsP2.1Learning paradigmsTraining & adaptationLifecycleno Interactive imitation learningP2.1Learning paradigmsTraining & adaptationLifecycleno Loss masking and token normalizationP2.5Adaptation: fine-tuning, prompting & transferTraining & adaptationLifecycleno Credit assignmentP2.6Preference & RL post-trainingTraining & adaptationLifecycleno Actuator limitsT2.3Motion planning & controlEmbodied & physical systemsSystem typesno Watchdog timerT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Closed-loop policy evaluationQ4.3Safety verification & validationSafety & assuranceQualitiesno Event data recorderQ4.4Safety cases, standards & certificationSafety & assuranceQualitiesno Decision accountability recordQ5.4AI governance, audit & regulationResponsible AIQualitiesno Agent interruption and cancellationT1.3Agent harness & control loopAgentic systemsSystem typesno Agent loop detectionT1.3Agent harness & control loopAgentic systemsSystem typesno Task and motion planningT2.3Motion planning & controlEmbodied & physical systemsSystem typesno Multi-object trackingT2.2State estimation, localization & mappingEmbodied & physical systemsSystem typesno Safety change-impact analysisQ4.4Safety cases, standards & certificationSafety & assuranceQualitiesno Sparse tensor kernelsS2.2Tensor operators & memory layoutFrameworks, compilers & runtimesStackno Local SGDS3.3Collective communicationParallelism & distributed executionStackno Expert replication and placementS3.2Parallelization planning & placementParallelism & distributed executionStackno Memory-efficient optimizerP2.3Training memory & stateTraining & adaptationLifecycleno Operator execution traceF3.1Roofline, performance models & simulationQuantitative system modelsFoundationsno Offline reinforcement learningP2.1Learning paradigmsTraining & adaptationLifecycleno Byzantine-robust aggregationQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Cluster managerS4.4Cluster scheduling & orchestrationDatacenter infrastructure & fleetStackno Rebound effectQ6.1Energy & carbon accountingSustainability & costQualitiesno Data cascadeP1.3Data validation & feature engineeringData engineeringLifecycleno Data valuationP1.2Data selection & curationData engineeringLifecycleno Diffusion policyT2.5Robot learning & sim-to-realEmbodied & physical systemsSystem typesno Sensor spoofingQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Paired comparison and noninferiority testingP3.1Evaluation design & experimentsModel evaluation & testingLifecycleno CPU execution architectureS1.1Accelerator architectureHardware & acceleratorsStackno CPU threading and NUMA placementS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Dynamic shapes and shape specializationS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Rule-based graph rewritingS2.4ML compilers & kernel optimizationFrameworks, compilers & runtimesStackno Sparse and ragged tensor representationS2.2Tensor operators & memory layoutFrameworks, compilers & runtimesStackno Execution providers and operator fallbackS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Large language modelF2.1Model familiesModels & architecturesFoundationsno Inference compute and memory estimateF3.4Scaling laws & compute-optimal allocationQuantitative system modelsFoundationsno Synchronous vs. asynchronous trainingS3.1Parallelism strategiesParallelism & distributed executionStackno Input prefetchingS2.6Data loading & input pipelinesFrameworks, compilers & runtimesStackno Event-time windowing and watermarksP1.4Data pipelines & storage formatsData engineeringLifecycleno Token streamingP4.5LLM & generative inferenceDeployment & servingLifecycleno Serverless inferenceP4.1Serving architecturesDeployment & servingLifecycleno LLM request scheduling and preemptionP4.4Batching, admission & autoscalingDeployment & servingLifecycleno ContainerizationP5.5ML pipelines & reproducibilityMLOps & operationsLifecycleno Distributed job hang diagnosisQ2.3Hardware faults & silent data corruptionReliability & fault toleranceQualitiesno Benchmark representativenessQ1.2System benchmarkingPerformance, efficiency & benchmarkingQualitiesno Data and model licensingP1.5Data contracts, versioning & lineageData engineeringLifecycleno Content provenanceQ5.4AI governance, audit & regulationResponsible AIQualitiesno Fleet data engineP5.3Drift, feedback loops & retrainingMLOps & operationsLifecycleno Robotics middleware and dataflowT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Parallel physics simulationT2.5Robot learning & sim-to-realEmbodied & physical systemsSystem typesno Agent credentials and delegated authorityT1.2Tool use & action interfacesAgentic systemsSystem typesno Graph partitioning for GNN trainingS3.1Parallelism strategiesParallelism & distributed executionStackno Host-side data loadingS2.6Data loading & input pipelinesFrameworks, compilers & runtimesStackno Streaming inferenceP4.1Serving architecturesDeployment & servingLifecycleno Inference API and payload serializationP4.1Serving architecturesDeployment & servingLifecycleno Request throughputP4.3Latency, throughput & goodputDeployment & servingLifecycleno Latency percentilesP4.3Latency, throughput & goodputDeployment & servingLifecycleno Cache-aware request routingP4.4Batching, admission & autoscalingDeployment & servingLifecycleno Distributed job launch and process groupsS3.3Collective communicationParallelism & distributed executionStackno Agentic workloadT1.3Agent harness & control loopAgentic systemsSystem typesno Performance regression testingQ1.2System benchmarkingPerformance, efficiency & benchmarkingQualitiesno Performance measurement variabilityQ1.2System benchmarkingPerformance, efficiency & benchmarkingQualitiesno PII detection and de-identificationP1.5Data contracts, versioning & lineageData engineeringLifecycleno Gradient clippingP2.2Optimization & training dynamicsTraining & adaptationLifecycleyes Weight initializationP2.2Optimization & training dynamicsTraining & adaptationLifecycleyes Processing delivery semanticsP1.4Data pipelines & storage formatsData engineeringLifecycleno Online and offline feature consistencyP1.6Leakage & training-serving skewData engineeringLifecycleno Out-of-distribution detectionP3.2Metrics, calibration & uncertaintyModel evaluation & testingLifecycleno Model and feature compatibilityP5.1Versioning, rollout & rollbackMLOps & operationsLifecycleno Workload identity and secrets managementQ3.4Access control & isolationSecurity & privacyQualitiesno Data deletion propagationP5.7Model retirementMLOps & operationsLifecycleno Runtime memory fragmentationS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Perception latency budgetT2.1Sensing, perception & fusionEmbodied & physical systemsSystem typesno Onboard compute partitioningT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Health checks and readiness probesQ2.1Failure models, retries & degradationReliability & fault toleranceQualitiesno Python GIL and host-language concurrencyS2.1ML frameworks & programming modelsFrameworks, compilers & runtimesStackno Flash write enduranceQ2.3Hardware faults & silent data corruptionReliability & fault toleranceQualitiesno Energy-harvesting intermittent computingS1.5Edge & embedded hardwareHardware & acceleratorsStackno Memory-mapped model weightsS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno OS memory pressure and pagingS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Managed-runtime interop and garbage collectionS2.5Runtime execution & dispatchFrameworks, compilers & runtimesStackno Zero-copy sensor input via DMAS2.6Data loading & input pipelinesFrameworks, compilers & runtimesStackno Active sensor interferenceT2.1Sensing, perception & fusionEmbodied & physical systemsSystem typesno Network propagation delayS4.2Network fabricsDatacenter infrastructure & fleetStackno Consistency-availability trade-offS3.4Coordination & consensusParallelism & distributed executionStackno Decentralized gossip averagingS3.5Federated & decentralized learningParallelism & distributed executionStackno Interrupt latencyT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Encoder-only, decoder-only, and encoder-decoder TransformersF2.1Model familiesModels & architecturesFoundationsno Data transfer costQ6.3Cost, TCO & unit economicsSustainability & costQualitiesno Duty cycling and sleep statesS1.4Chip power & thermal limitsHardware & acceleratorsStackno Priority inversionT2.4Real-time & mixed-criticality computingEmbodied & physical systemsSystem typesno Certified robustness by randomized smoothingQ3.1Threat models & adversarial attacksSecurity & privacyQualitiesno Sensor interconnect bandwidthS1.5Edge & embedded hardwareHardware & acceleratorsStackno Finite buffers and overflowF3.3Queueing & capacity modelsQuantitative system modelsFoundationsno Asynchronous on-device inferenceP4.9On-device & edge deploymentDeployment & servingLifecycleno determinessets acceptance thresholds forThe operational outcome, evaluation population, and threshold must be declared. determinessets targets forService targets are part of the stated deployment requirements. part-ofis the reference point withinThe baseline must be measured using the same protocol. measuresstands in for, and can diverge from,Improving a measurable proxy need not improve the intended outcome. determinesdetermines resource demand forSizing must include input and request distributions, not just a representative average. determineschanges attainableCommunication paths, locality, and available service resources depend on placement. limitsbounds achievableThe effect depends on task, coverage, errors, and the learning procedure. determinesdetermines the reliability of the supervision inFor supervised learning; label errors change the optimization signal. determinesforces imputation and indicator decisions inTreatment must preserve the meaning and availability of inputs. requiresmust preserveFor predictions whose inputs become available over time. increasescan invalidateWhen evaluation uses information unavailable at the intended prediction boundary. mitigatespreventsOnly when source availability and temporal joins are modeled correctly. enablesrecords the dependencies behindProvenance must identify versions and transformations sufficient for reconstruction. reducescan degrade deployedA deployed pipeline may compute different inputs from those used in development. part-ofis a curation stage ofRemoving duplicates changes sample weights and must preserve required coverage. determinesshapes the data supportingSelection can reduce cost while changing coverage and distribution. determinessets the sequence length processed byFor a fixed text, changing vocabulary and segmentation can change sequence length. determinesdefines the lookup identities ofEmbedding rows must correspond to the model's tokenizer and vocabulary. reducescan degradeCalibration measured on one distribution need not hold after a shift. mitigatesmitigates, when triggered by,A response requires diagnosis and evidence; not every detected shift warrants retraining. usesoperates onA linear score can use engineered or learned input features. usescan useApproximation trades exact recovery for search cost in compatible applications. usesis trained usingFor gradient-based learning with supported differentiable objectives and derivative rules. determineschanges data needed forArchitecture and task alignment affect which patterns can be learned efficiently. requiresmust be resident withinResidency depends on table dimensions, format, and placement or sharding. usesmakes order-sensitiveAttention without another position-sensitive mechanism is permutation equivariant. usesreuses history throughFor compatible autoregressive attention architectures. requiresrequires an objective fromOptimizing the training objective alone does not certify task success. enablesproduces the gradients consumed byOptimizer state and update rules are separate from differentiation. determinesshapesEffects depend on the data, model, objective, and optimization regime. determinessets gradient noise and step count forChanging the batch can alter gradient noise and statistical efficiency. limitslimits the per-deviceActivation and workspace demand depend on execution and accumulation choices. increasesadds optimizer variables toDifferent update rules retain different auxiliary state. limitslimits the per-deviceCount resident parameters, gradients, optimizer state, and other retained data separately. determinesinformsA fitted relationship has a validity domain and does not guarantee extrapolated gains. causescan induceSelecting among noisy validation scores can produce an optimistic winner. implementsimplementsParameter-efficient adaptation is one option, not all transfer learning. reducesshrinks trainable and optimizer state inThe method may select existing parameters or introduce new parameterizations; frozen parameters and activations can still remain resident. determinessets the data-movement cost ofViews and physical copies have different costs; kernels support particular layouts. determinessets storage width withinDifferent state components can use different formats. determinesaffectsRange, rounding, accumulation, and intermediate representation all matter. requiresrequires managingLoss scaling or wider accumulation is useful only for the relevant numerical regime. reduceslowers the numerical precision ofThe encoding, scale, and rounding rule determine the approximation. requiresrequires re-evaluation ofQuality and calibration must be evaluated for the resulting deployed system. producesproduces a smaller studentThe student architecture and task evidence determine the actual trade-off. enablescan applyOnly when semantics and execution dependencies permit combination. reducescan reduceFewer launches can reduce host work, depending on the original execution boundary. determinessets the sloped ceiling ofThe bandwidth must correspond to the memory boundary used for arithmetic intensity. determinessets the flat ceiling ofThe ceiling must match the precision and operation mix of the workload. determinesselects the active bound inThe compute and bandwidth ceilings describe the declared operating model. increasesimprovesTile size, storage capacity, and execution support constrain useful reuse. increasescan increaseWhen reuse reduces transfers across the measured memory boundary. measuresestimates realizedInstrumentation and measurement boundaries affect interpretation. limitslimits end-to-end gains from addedAccelerating the changed portion leaves unchanged work as a residual bound. usesusesSynchronous replicas commonly aggregate gradients; other learning semantics differ. requiresrequires per-layerPartitioned operations require data exchange according to their decomposition. limitsstallDependencies and pipeline bubbles expose unequal stage or worker times. reducesreduces per-worker copies ofReplication, communication, and reconstruction depend on the sharding stage. reduceshides some exposed cost ofOnly independent work overlaps; contention or dependencies can expose residual communication. limitsconstrains achievable rates ofAvailable paths and contention determine achievable exchange rates. part-ofis persisted, in selected form, inResumption semantics determine which state must be saved. requiresrequires, for distributed jobs, aThe saved worker states and dependencies must form a recoverable state. enablesmaintains the epoch authority used byResource-enforced epochs remain necessary; a consensus log alone does not fence effects. requiresmust addressKeeping raw records local alone does not establish privacy of updates or outputs. limitssets the demand placed onThe request mix and service-time distribution determine demand. increasesincreases, near saturation,With fixed service conditions, waits typically rise near saturation; this is not a universal scalar model. increasescontributes toService, batching, network, and other delays also contribute. increasescan increaseWorkload, memory, and kernel efficiency bound the gain. increasescan increaseWaiting to form a batch can add delay; the net effect depends on queueing and service changes. requiresadmits work against a knownLimits must be expressed in resources relevant to the workload. determinesadjusts provisionedProvisioning delay, state transfer, and warm-up affect when capacity becomes useful. limitslimits the residentSize depends on architecture, sequence length, format, and concurrency. implementsreduces allocation waste inBlock metadata, fragmentation, and compatible sharing determine the resulting capacity. usesreuses compatibleToken prefix, model state, position conventions, and access boundaries must match. limitsbounds the decode phase ofThe dominant resource depends on architecture, batch, context length, and implementation. increasescan raise effective decode throughput ofAcceptance rate, proposer cost, verification, and workload determine the net gain. determinesdefines the counting rule forThe counting rule must specify latency, quality, failures, and the workload population. measuressupports unbiased estimates ofThe partition and sampling unit must match the intended deployment question. increasesinflates reportedIndependent final evaluation is needed after selection on noisy scores. enablessupportsAbstention thresholds still require risk and coverage evaluation under the deployment distribution. limitsbounds the precision of conclusions fromDependence, sample selection, and repeated comparison matter. measuresspecifies how to measureMeasurement needs a declared load model, warm-up, boundary, and quantile method. measuresestimates causal effects onCausal interpretation depends on assignment, interference, observation, and stopping assumptions. measurestests specifiedAn aggregate score cannot establish performance for every subgroup. requiresrequires pinnedEvidence and compatibility must refer to the exact candidate being deployed. detectsobserves signs ofSignals require diagnosis and do not establish harmfulness by themselves. usescan useA rollout can use rollback when acceptance criteria fail; state compatibility and external effects may limit recovery. reducescan degradeDeployed decisions can alter which labels and examples are subsequently observed. requiresproduces a newAn updated learned artifact must be reevaluated and versioned. enablessupportsUseful diagnosis requires recorded context and trustworthy instrumentation. determinesscopes the attack surfaces considered forThe adversary's ability to modify training inputs or signals must be specified. determinesscopes the attack surfaces considered forThe adversary's knowledge, perturbations, and target behavior must be specified. implementsprovides a formal guarantee forThe guarantee bounds output-distribution change under a declared neighboring relation and accounting parameters; it does not establish every privacy requirement. requiresrequires actionableRecorded signals must be interpretable and available in time to support intervention. increasescontributes operational emissions toConversion depends on operational accounting boundaries and energy-source emissions; embodied emissions are separate. requiresrequires an acceptance criterion fromCounting attempted work as successful outcomes understates the unit cost. requiresis sized againstProvisioning decisions must account for demand variation and resource constraints. limitsaccumulates demand forOnly selected observations and records fit in a model invocation. determinesrevises the actions withinExecution observations and changing constraints can invalidate a prospective plan. limitsselects what entersRelevance, authorization, provenance, and token budget constrain what may be supplied. usescan useApproximation and index freshness affect retrieval quality and cost. determinesdeclares the authority and effects checked byTyped arguments alone do not grant authority to perform an action. enablesenables safe retries withinStable operation identity and effect reconciliation must match the external tool semantics. usescan use, to remedy earlier steps,A remedy may not restore the original state or undo all consequences. mitigatescan cause violations of intendedUntrusted content must not silently acquire authority through model interpretation. measuresmeasures accepted outcomes ofEvaluate external effects and constraint compliance, not only final text. causescan inviteA reward is not itself evidence that the intended task was performed. limitsconstrainsPhysical states, inputs, disturbances, and constraints determine admissible motion. enablesprovides observations forCalibration, geometry, timing, and model assumptions affect the inference. limitslimitsA low reported covariance does not by itself establish that the state is distinguishable. determinesinflates the uncertainty ofState can evolve between capture and use; propagation depends on the plant and disturbances. limitslimits the stability margin ofDelay can reduce stability margins; the effect depends on sampling and dynamics. limitscontributes to boundingExecution-time bounds alone do not include all queueing, scheduling, sensing, and transport delays. determineschangesContact modes, stiffness, damping, and impact alter physical response. requiresrequires assessing simulator mismatch inSimulation evidence must be evaluated against the actual deployment conditions. limitsconstrains executed actions toA numerical filter supplies only the guarantees justified by its model, timing, and implementation assumptions. increasesshrinksDelay adds motion before braking; braking authority and uncertainty also matter. limitsbounds the claims inEvidence must match the intended operating conditions and configuration. determinesstates the premises relied on byAn argument is not certification of a general numerical filter. implementsenforces the access boundaries ofIsolation can restrict resource access, but it does not choose the authorization policy or prove external-effect correctness. determinesdetermines the observed population forSelection and missing outcomes can change the population to which an estimate applies. determinesspecifies the validity criteria forA schema check alone cannot establish semantic correctness or task fitness. limitslimits timely conclusions fromOutcome maturation and selective observation require explicit evaluation windows. usesuses outputs assessed byA probability threshold needs task costs and deployment-distribution evidence as well as calibration. requiresuses targets assessed byThe target semantics and observation process must match the intended task. determinesconstructs targets for its ownTargets are derived from the data under a specified objective and corruption or prediction scheme. implementscan produce learnedLearned structure or representations require separate downstream evaluation. usesoptimizes behavior against a reward acting asThe reward and interaction model may differ from the desired real-world outcome. causesinduces compounding state-distribution shiftPolicy actions change the states visited relative to demonstration data. requiresmust be resident withinResidency can include parameters, tree structure, support vectors, or stored examples, depending on the model. determinesdetermines the contents ofFeature transforms and output rules are part of the execution contract. determinesdetermines the contents ofTree structures, thresholds, leaf values, and aggregation rules require storage and compatible execution. determinesdetermines the size ofSupport vectors or an explicit approximation affect the stored state and evaluation cost. determinesdetermines the contents ofStored vectors, metadata, and optional indexes contribute capacity and search cost. limitsbounds the guarantees ofFault classes, membership, quorum availability, and timing assumptions bound the agreement claim. determinesdetermines the recovery requirements forRecovery semantics must cover the faults and state dependencies actually assumed. limitslimits offeredThe signal must reach producers in time and avoid shifting overload into unbounded buffers. limitscan violateRisk depends on the information, recipient, purpose, consent, and declared protections. limitsimplies a lower bound onFor fixed arithmetic work, a throughput ceiling implies a lower execution-time bound. Additional overheads can lengthen actual runtime. increasescontributes toExecution contributes to each request latency; component quantiles cannot generally be added to obtain an end-to-end quantile. limitslimits feasibleModel state, request state, intermediates, and allocator overhead must fit for the selected execution strategy. measuresmeasureThe metric, sampling protocol, and target distribution determine what the estimate means. determinesgate advancement duringRelease criteria and populations must be specified before interpreting candidate results. reducesreduces traffic demand onSmaller representations can reduce transferred bytes; dequantization and execution support determine total cost. measurestracks deployedObserved task metrics can be delayed, biased, or unavailable; system health alone does not establish task quality. increasesreplicates, multiplying the total footprint of,Conventional replicated data parallelism keeps model and update state at each worker; ZeRO and hybrid strategies shard selected components. usespartitionsFor tensor decompositions whose operand and reduction partitions require compatible communication and layout. measuresmeasures observedMeasurement boundaries, instrumentation, and workload determine the observation; a sample is not a universal time bound. measuressamples, but cannot by itself bound,A justified worst-case bound requires assumptions and analysis covering the relevant executions and hardware states, beyond observed samples. limitsconstrains acceptableThe protected populations, outcome criterion, operating conditions, and trade-offs must be stated; one criterion does not establish all fairness goals. determinesselects attack cases forTests exercise a declared adversary and interface; passing selected attacks does not establish universal robustness. limitsbounds planned recovery inA design's recovery guarantees are limited to its declared faults and assumptions; operations may also encounter unmodeled incidents. usesuses estimated state fromFor controllers using estimated state, uncertainty and observation timing matter; other output-feedback designs need not reconstruct full state. determinesselects inputs that shapeThe closed-loop response depends on the plant, disturbances, constraints, sensing, and controller implementation. usesrelies on qualifiedState-dependent checks need valid observation timing and error bounds; the guarantee remains conditional on those premises. determinesscopes the relevant conditions forThe intended domain identifies relevant stresses and coupled corners, but finite selected tests do not exhaust continuous operating conditions. limitsbounds sustainableOnly for matching work units, counting boundary, workload, operating state, and horizon. A measured throughput is not a universal capacity ceiling. increasesadds monetary cost toConvert energy through the applicable tariff or cost model; do not add it again when already included in cloud or facility charges. limitscan prevent full reversal byRestoring software or configuration cannot necessarily undo physical, informational, financial, or human consequences already caused. limitscan remain afterA remedy may address an obligation or consequence without restoring the original world state. mitigatescan gate prospectiveOnly when the mechanism mediates the relevant pre-actuation path with sufficient timing and authority. Detection can occur after an effect, and fallback can itself be irreversible. causescan createDisclosed information may be copied or retained outside the system's control, so subsequent deletion cannot guarantee full retraction. limitscan constrainPeople need usable evidence, sufficient time, and effective authority; a nominal approval step alone does not establish meaningful control. determinesdetermines the contents ofExecution requires compatible parameters, operators, and configuration. Training state can include additional data not needed for inference. measurescan assess support forFor tasks requiring factual or source-grounded assertions, evaluate support in addition to fluency or schema validity; not every task has the same truth criterion. determinessets the I/O required byEffects depend on access pattern, projection/filtering, encoding, storage medium, and caching; a layout does not guarantee higher throughput. requiresrequires compatibility rules inContracts must identify producers/consumers, field meaning, allowed changes, defaults, and migration expectations; format-supported changes need not preserve application semantics. requirescan requireWhen offered record arrival exceeds sustainable processing or downstream capacity, the chosen ingestion system needs an explicit overload policy; not all sources permit slowing producers. requiresrequires event-time-preservingWhen downstream features or evaluation depend on event-time correctness, ingestion/availability timestamps and late-data policies must be retained; arrival order alone is insufficient. requiresrequires expectations from aThe oracle may be partial, probabilistic, or human-assisted and only covers declared properties and tested cases. detectscan exposeTests compare relevant preprocessing and model interfaces across representative paths; untested distributions or deployment differences can remain. measurestests compatibility withinTest environments must represent the target packages, binaries, resources, and configuration; passing one environment does not validate all others. determinesselects candidate scenarios forInjected faults instantiate selected assumptions; realistic timing, combinations, observation criteria, and containment must be declared. measuresexercisesA controlled scenario can test detection and response paths; success does not demonstrate preparedness for every incident. requiresrequires observable acceptance criteria fromCriteria may evaluate latency, error rates, recovery behavior, or invariants; injecting a fault without assessing outcomes is not sufficient evidence. requiresrequires a pinnedPackage identity alone is insufficient: hardware, data, seeds, nondeterministic operations, external services, and acceptance criteria may also matter. enablesidentifies the exact inputs forReproduction requires resolvable versions and the execution procedure/environment; version labels alone do not reproduce results. producescan record execution evidence asOnly when actual input/output artifacts, parameters, steps, and execution identities are captured; orchestration alone does not establish complete provenance. requiresrequires retry-safe steps throughWhen steps can be retried after ambiguous completion, repeated external effects need explicit handling; idempotency is one option rather than a universal orchestration guarantee. requiresrequiresThe benchmark protocol must state tolerances, repetitions, environment, and workload; reproducibility does not establish external validity. enablesprovides storage levels forThe algorithm, placement, and implementation determine which levels actually serve an access; effective rates depend on contention and access pattern. increasescan add transfer delay toWhen the measured execution boundary includes traffic crossing the host-device path, transport, topology, contention, reuse, and overlap determine its contribution; fully hidden transfers need not increase end-to-end time. determinesconstrains supported mappings ofOperations, data types, shapes, and implementation must match the execution resources; peak throughput is not achieved throughput. limitsconstrains feasible operating points forSupply limits, voltage integrity, and platform power management can constrain sustained execution; power is distinct from energy. increasescan increaseWhen thermal controls reduce frequency or service rate or suspend operation, a cold unloaded timing measurement need not characterize sustained or shared operation. increasesadds cooling overhead toThe system boundary and cooling overhead must be stated; carbon conclusions additionally require energy mix and lifecycle assumptions. part-ofis the measured quantity inThe indicator's population, window, and aggregation must match the objective's. determinessets the size ofThe objective is stated as a fraction over a defined window. determinespacesTeams agree to slow or halt changes when the remaining budget is exhausted. limitsallocates permittedStage latencies compose along a request path whose end-to-end target is fixed. measuressupplies the operation and byte counts forOperation and byte counts are taken at the same memory boundary. limitscapsAll resident state is counted at its peak, including runtime buffers. usesgrounds demand estimates inThe trace is representative of the arrival, size, and mix distribution being provisioned for. part-ofestimates the operations and bytes ofModel shapes and input sizes are known. determinessets thresholds inError costs and class priors can be stated for the deployment context. limitsselect an operating point onTwo or more objectives are measured on the same workload and cannot all be optimized at once. usesis built fromThe architecture uses attention layers interleaved with feed-forward blocks. usesis dominated by lookups inCategorical features are encoded through large embedding tables. increasesmultiplies by sampling step countGeneration runs the denoising network once per sampling step. increasestypically raisesFilters are reused across many spatial positions within one kernel. increasesadds sequential dependence toEach step depends on the previous hidden state, preventing parallelism across time steps. increasesimprovesActivation statistics would otherwise drift with depth or batch composition. reducesreduces relative to total parametersEach token is routed to only a small subset of experts. requiresstill requires storing all experts withinEvery expert must be resident somewhere in the system even if inactive for a token. determinesdetermines token traffic inExperts are placed on different devices from the tokens routed to them. limitsbounds assignments made byA fixed per-expert buffer is allocated per batch. determinesregularizes toward balanceThe auxiliary loss weight is large enough to counter router collapse without dominating the task loss. determinesdetermines per-token compute inCompute is counted for the experts actually activated rather than all stored parameters. reducesreduces average, and makes input-dependent,Exit criteria differ across inputs, so latency varies with input difficulty. increasesscales the size ofAutoregressive decoding retains keys and values for every prior position. determinesraises the pairwise score cost ofFull attention compares every query with every key. enablesmakes long contexts affordable by linearizing attention cost overPositions attend only within a window or fixed sparse pattern, possibly at a quality cost on long-range tasks. reducesreplaces with a fixed-size stateThe attention kernel is approximated by a feature map that permits a recurrent formulation. specializesis a parallel-trainable form ofThe state update is linear (or selectively gated) so training can use scans or convolutions. reducesreduces the size ofFewer key-value heads are stored than query heads. usesrescales or interpolatesThe model uses a position encoding, such as rotary embeddings, that can be rescaled or interpolated. part-ofpersists the parameters ofThe artifact's weights are persisted for later loading. enablesprovides input toThe target runtime or compiler supports the exchanged operator set. producescan emitThe program's operators and control flow are expressible in the format. measureschecks callers againstCallers and the artifact agree on input and output names, types, and shapes. determinesseparatesThroughput and bandwidth ceilings are measured for the same device and memory level. determinesclassifies prefill as compute-bound and decode as memory-bound inPrefill processes many tokens per weight read while decode processes few. increasesoften dominatesOperands are fetched from off-chip memory rather than reused locally. reducesreducesReused data is served from nearer, lower-energy storage levels. usesparameterizesThe unaccelerated fraction is measured on the same workload as the speedup. limitsbounds speedup underProblem size is fixed while resources grow. reducesreduceWorkers synchronize so the slowest determines step time. reducescan reduceCommunication time is exposed rather than overlapped with computation. usesrelates system occupancy toThe system is stable and averages are taken over a long window. measuressizes concurrent state held inEach in-flight request holds cache state for its duration. determineslocates the maximum load meetingLatency is measured at the percentile the objective specifies. increasesincreasesMean utilization is held fixed while inter-arrival variability grows. requiresis derived fromThe fitted law is used within the range of model and data sizes it was fit on. usesusesDense models are trained where the six-operations-per-parameter-per-token approximation holds. trades-offcan trade off against lifecycleLarge inference volume can favor smaller models trained beyond the compute-optimal point. specializesis a special case ofThe model is a dense transformer and attention cost is small relative to parameter compute. implementsexecutesOperand shapes are tiled to the array dimensions. requiresreaches peak throughput only at supportedOperations use input and accumulation formats the unit implements. determinesloses achieved throughput under divergent control flow inThreads in a group take different control-flow paths. determinesdeterminesThe accelerator's local storage cannot hold all operands at once. increasesraises, on its target operations,The target operations dominate the workload the hardware runs. increasesprovides highData resides in device memory rather than host memory. limitsbounds tile size forTiles must fit in on-chip storage for reuse to occur. determinesshifts workloads toward the memory-bound side ofPeak arithmetic throughput grows faster than memory bandwidth across hardware generations. limitsbounds the efficient group size ofTensor-parallel collectives occur every layer and need the highest-bandwidth links. increasescan extendPooled or attached memory is acceptable at its higher latency. usesbypasses host staging onDevices, drivers, and topology support direct transfers. limitssets the design envelope forSustained workloads, rather than brief bursts, determine heat removal needs. reducescan reduce, at some cost in speed,Workload phases tolerate lower clocks, such as memory-bound phases. enablesmotivatesPower, not transistor count, limits how much of a chip can be active. reducescan lowerThe cap binds below the power the workload would otherwise draw. determinesimposesWeights and activations must fit in on-chip flash and SRAM. requiresoften requiresThe full-precision model exceeds flash or SRAM limits. requiresrequires supported lowThe accelerator implements only specific integer or reduced-precision formats. requiresrequires operator-levelDifferent blocks support different operators, precisions, and power states. specializesis a per-input case ofThe boundary is one inference on one device, including memory and idle overhead. part-ofis metadata carried byThe framework records strides or format metadata with each tensor. usestraversesDerivatives are composed in reverse over recorded operations. determinesincurs per-operationEach operation is launched separately from the host program. enablesenablesTraced shapes and control flow remain valid for later inputs or guards trigger recompilation. requiresdepends on the capacity ofTile sizes are chosen to fit on-chip memory. usesmaps convolution ontoThe im2col-style transformation is used, at the cost of duplicated input data. implementsimplementsA lowering algorithm is chosen per layer shape. reducescan limit opportunities forLibrary calls are opaque to the compiler and cannot be fused with neighbors. enablesenablesThe IR exposes producer-consumer structure across operators. determinesselects parameters forThe schedule space is large and performance is hard to predict analytically. usesis written in terms ofKernels are written in terms of blocks that the compiler maps to threads and memory. reducesreduces peakTensor lifetimes are known at compile time. enablesenablesCommunication and computation are issued on independent streams with correct event dependencies. reducesreducesThe launch sequence and tensor addresses are static across iterations. part-ofis a component ofGradients would otherwise underflow in the reduced-precision format. increasesextends the dynamic range available at lowValues within each block have similar magnitude. determinesaffectsLong reductions accumulate rounding error in low precision. requiresrequiresTokens are dispatched to and gathered from experts on other devices. enablesdistributes the experts ofExperts do not fit, or run too slowly, on one device. increasesextends the feasibleActivation or attention memory for one sequence exceeds one device. determinesshapes the group mapping ofLink bandwidth differs between levels of the hardware hierarchy. producesincursMicrobatches per step are few relative to pipeline stages. implementsis an algorithm forParticipants can be arranged in a ring with comparable link bandwidth. measurespredicts the cost ofLink latency and bandwidth are roughly uniform and contention is ignored. alternative-tois an alternative toMessages are small or participants many, so latency terms dominate. limitsbounds throughput ofTraffic crosses the network bisection rather than staying within local groups. usesreduces communication volume ofConvergence is preserved, for example through error feedback, at the chosen compression ratio. requiresneeds epoch fencing to be safe, as inClocks are bounded in drift so expired holders can be rejected safely. usescommonly relies onReplicas must agree on a single leader under partial failure. mitigatesreduces waiting onConvergence tolerates updates computed on stale parameters within the bound. implementsis a baseline algorithm forClients can run local optimization steps between communication rounds. limitsslows convergence ofClient data distributions differ and clients take many local steps. increasescan amplify bias fromParticipation correlates with client characteristics such as connectivity or device class. determinesdetermines constraints onDevice availability and reliability differ between cross-device and cross-silo settings. determinesscales IT energy to facility energy inThe accounting boundary includes facility overheads. enablesraises the feasibleAir cooling cannot remove the heat generated at the target density. limitsconstrains the physical extent ofShort copper links require accelerators to be packed into few racks. requiresrelies onSimultaneous peaks across equipment are rare but must be bounded when they occur. reduceslowers per-message overhead ofTransfers bypass host network stacks and remote CPUs. enablesprovides drop-free transport forThe RDMA transport performs poorly under packet loss, as with RoCE. mitigatescan reduce network-inducedSynchronized collective traffic or incast creates switch queue buildup. increasescan provide fullSwitch tiers are not oversubscribed. enableskeeps collective traffic within one switch hop forParallel groups are placed so peers share the same rail. implementsoffloads reduction steps ofSwitches support aggregation for the data types and operations used. measuressummarizes the capacity ofTraffic is uniform enough that the worst-case cut is relevant. enablessupplies aggregate read bandwidth forMany workers read training data concurrently from shared storage. part-ofcommonly forms the capacity tier inHot data is cached closer to compute. usesreads through the nearer tiers ofWorking sets fit in nearer tiers or are prefetched. determinessets the cost of takingTraining state is large relative to available storage bandwidth. reducescan reduce clusterResources sit idle while waiting for a full allocation, unless backfilled. requiresdepends onPreempted work is resumed rather than restarted from scratch. implementsallocates shares among tenants inJobs demand several resource types in different proportions. mitigatesreduces lost work fromThe job can reshard state and continue with fewer workers. reducesreduces the cost ofCommunicating workers are placed on shared fast links or switches. specializesrefinesPlacement decisions are made at node and device granularity. trades-offtrades off againstDense packing fragments the contiguous allocations that locality requires. requiresrequiresTenants must not observe or disrupt one another's work. part-ofpartitions shared capacity withinShared capacity is divided among tenants with stated limits. increasescan raiseIndividual workloads cannot saturate a full accelerator. increasescan inflateCo-located workloads contend for unpartitioned shared resources. causesexposes workloads toThe provider reclaims capacity on short notice. reducescan lowerInterruption overhead is smaller than the price discount. determinesproduces higher-volume, noisierHolds when labeling sources are noisy and their errors must be modeled or denoised before training. determinesbiases sampling inThe selected examples depend on the current model, so the labeled set is not a random sample of the deployment population. measuresmeasures reliability ofHolds when multiple independent annotations of the same items are collected. implementsimplementsThe quality score must correlate with downstream value; filtering on a narrow proxy can remove useful diversity. determinesshifts across domainsHolds under a fixed training budget, where raising one domain's share reduces exposure to others. causescan introduceHolds when the generator's output distribution differs from the deployment distribution or repeats the generator's errors. increasescan improveThe transformations must preserve labels and reflect invariances present at deployment. implementsenforcesThe contract's schema and validity rules must be machine-checkable. mitigatesreducesHolds when training and serving read features from the same definitions; it does not prevent skew from logic outside the store. enablessupports point-in-time joins forRequires timestamped feature values and as-of joins when building training sets. limitsbounds achievableProcessing latency and windowing determine how recent a served feature value can be. requiresmust preserveBackfilled values must be computed only from data available at each historical timestamp. requiresrequiresHolds whenever training or evaluation data can change after a run. increasesamplifiesHolds when unstable or unused inputs remain in a model whose behavior depends jointly on all inputs. causesstale values can causeHolds when serving-time feature age differs from the age assumed when training examples were built. mitigatesmitigatesWarmup and decay reduce early and late divergence for many optimizers but do not remove instabilities from data or numerics. limitslimits useful scaling ofHolds when adding replicas increases the global batch beyond the critical batch size. mitigatesenables restarts that recover fromHolds when loss spikes are handled by rolling back to an earlier checkpoint and changing data order or hyperparameters. enablesenables large effective batch sizes within limited memory forHolds for losses that sum over examples; per-batch statistics such as batch normalization still see micro-batches. reducesenables state larger than accelerator memory forBenefit depends on host-device bandwidth and how well transfers overlap with compute. enablesrecords evidence forTracking must capture code, data, configuration, and environment versions, not only metrics. reducesreduces large-scale spend ofHolds only under a parameterization in which optimal hyperparameters remain stable as width or depth grows. specializesis an instance ofAlways; LoRA restricts updates to low-rank added matrices. causescan causeHolds when fine-tuning data are narrow relative to pretraining and no replay or regularization is used. part-ofsupplies the training signal inHolds for RLHF pipelines that learn a reward model from human comparisons. causesis exposed toHolds when the policy optimizes a learned reward beyond the region where it tracks human preference; a reference-model penalty reduces but does not remove it. requiresrequiresRewards are only as reliable as the programmatic checker; exploitable checkers produce reward hacking. limitsbounds the throughput ofHolds when rollout generation, not policy update, dominates step time, so serving optimizations raise training throughput. determinesfrequency determinesHolds when rollout workers continue generating between synchronizations. causescan induceHolds when the update rule assumes on-policy data and lag is not corrected by importance weighting or clipping. determinesshortens the cycle ofHolds when the model's own predictions influence the data it next learns from. specializesis a warm-start, lower-cost kind ofHolds when the model architecture and feature set are unchanged; schema or architecture changes require training from scratch. mitigatesmitigatesEffectiveness depends on buffer size and how representative stored examples are of earlier tasks. limitsconstrainsHolds on devices whose memory cannot hold full activations and optimizer state, forcing sparse or partial updates. reducesreduces variance of estimates underHolds for i.i.d. data; temporal or grouped data need splits that respect dependence. enablesenables valid early stopping ofHolds when the analysis uses a sequential procedure; repeated peeking at fixed-horizon tests inflates false positives. limitscan bias estimates fromHolds when units share resources, markets, or networks, so treatment of one unit affects others. enablesprovides coverage-calibrated sets forCoverage holds marginally under exchangeability between calibration and test data, not under distribution shift. enablesproposes slices forDiscovered slices require human review for coherence and multiple-comparison effects. requiresmust be calibrated againstHolds whenever model-graded scores are used as evidence; judges show position, length, and self-preference biases. limitsupper-bounds accuracy ofBest-of-n reaches pass@n only with a perfect selector; real verifiers lose part of the gap. limitsrequires reporting ofHolds because rankings from finite pairwise votes carry sampling error that can reorder close models. specializesis a form ofHolds when evaluation items or near-duplicates appear in training corpora; detection needs overlap checks against training data. mitigatesreducesHolds when deduplication is run against evaluation sets, not only within the training corpus. part-ofis the headline metric of a trainingHolds for training benchmarks that fix model, dataset, and quality target under run rules. limitssets the latency-percentile constraint onHolds for server-style scenarios that report throughput only at a stated latency percentile. implementssubstitutes output relations for a missingHolds when a valid metamorphic relation between inputs and outputs is known. determinessupplies criteria forHolds when rubric items are converted into automated or reviewed checks. reducesincreases the staleness (age at use) ofHolds when precomputed predictions are served after their inputs have changed. measuresis the judging metric ofHolds for user-facing request paths with latency targets. useshostsHolds for servers that queue concurrent requests to batch them on shared accelerators. increasesincreases exposure to load-timeHolds when more models than fit in memory share resources, so infrequently used models must be loaded on demand. reducescan reduceHolds when most requests are resolved by cheap models and the escalation criterion is reliable. requiresrelies onEach stage must abstain and defer when confidence is low; poorly calibrated scores misroute requests. determinessets the total allocated byHolds when the end-to-end SLO is decomposed into per-stage targets that sum within it. determinesprefill phase dominatesHolds when queueing delay is small relative to prompt processing time. limitsboundsHolds in decode at small batch, where each step reads weights and KV cache with low arithmetic intensity. reducesreducesHolds when slow responses come from replica-specific variability and the extra load from duplicates is small. specializesspecializesApplies to iterative generation, where requests finish at different steps. increasesincreasesHolds when KV memory admits enough concurrent sequences; interference between prefill and decode can raise latency. determinesdetermines hit rate ofHolds when cached prefixes are replica-local, so cache-aware routing raises reuse but can unbalance load. limitsbounds how quickly it can react:Holds when new replicas must load large weights before serving, so scaling must anticipate demand. reducesreduces decode stalls inHolds when long prefills would otherwise stall decode iterations of co-batched requests; it can raise TTFT for the long prompt. determinesprovisions separate resources for each phase ofPays off when interference between phases costs more than transferring KV state between pools. increasescan increaseHolds when TTFT and TPOT targets would otherwise force over-provisioning of a colocated deployment. implementsenforcesHolds when the constraint compiles to an automaton aligned with the tokenizer's vocabulary. alternative-tois the post-hoc alternative toValidation catches semantic rule violations decoding cannot express, at the cost of retries. requiresrequires sampling diversity fromVoting gains require sampled paths to differ; greedy decoding yields identical candidates. increaseslarger budgets increaseHolds because generated reasoning length varies widely across requests and adds decode steps. limitsprecision bounds the gains ofHolds as n grows, since selection pressure increasingly exploits verifier errors. specializesis a step-level form ofRequires step-level labels or estimates of step correctness. usesuses for pruningHolds for searches that score partial solutions before completion. specializesextends to inference computeHolds within task families where extra samples or reasoning steps improve accuracy; gains saturate and vary by difficulty. determinessets the quality gained per unit ofHolds whenever additional per-query compute is billed or consumes shared serving capacity. mitigatescan reduceHolds when retrieved passages are relevant and the model grounds its claims in them; irrelevant retrievals can mislead. requiresdepends onGeneration quality is bounded by retrieval recall for the facts the answer needs. usesuses at scaleHolds when the collection is too large for exact similarity search within latency limits. alternative-tois a complementary alternative toLexical matching handles rare terms and exact identifiers that embeddings may miss; hybrid systems combine both scores. part-ofis the precision stage ofHolds when first-stage recall is adequate and reranker cost fits the latency budget. usesoften appliesHolds for integer-only targets; converted models must be re-evaluated because numerics change. limitsconstrainsHolds on microcontrollers and mobile devices where weights, activations, and runtime must fit in limited RAM and flash. usesapplies staged exposure fromHolds for device fleets where updates are staged by cohort and devices may be offline. specializesis a form ofHolds when part of the computation can run on either side of a network link. increasesmultiplies targets and effort forHolds when devices differ in supported operators, precision, or accelerators. implementsstores and governsHolds when every deployable model is registered with its lineage and stage. part-ofis the zero-exposure stage beforeHolds when candidate outputs can be compared without user exposure; shadowing cannot measure effects of users reacting to the new model. implementsautomatesHolds when gate criteria are machine-checkable; some require human review. determinescontrols advancement ofHolds when each rollout stage has predeclared pass criteria. detectsdetectsHolds for shifts visible in monitored statistics; detected drift need not degrade quality, and some harmful shifts are not detected. causescan increase exposure toHolds when logs retain personal inputs or outputs without redaction, access control, or retention limits. part-ofinitiatesHolds when alerts are actionable; excessive alerting degrades response through fatigue. mitigatesmitigates, given fresh labels,Holds because input monitoring alone cannot reveal changed input-target relations; updated outcomes are needed. implementsautomatesHolds when data validation and release gates are automated so retrained models can be promoted without manual steps. part-ofcloses the loop onHolds when corrective actions are tracked to completion. requiresdepends onDistinguishing data, model, and infrastructure causes requires versioned, correlated signals from each layer. reducesreduces configurationHolds when configuration is reviewed, tested, and versioned like code. determinesdeclares and reproducesHolds when declarations pin versions and are applied without manual drift. reducescan amortizeHolds when many teams adopt shared components; a poorly maintained platform becomes debt itself. causescan create hiddenHolds when unknown consumers act on model outputs in ways that change future training data. increasesis a source ofHolds for models where inputs interact, so isolated changes have non-local effects. part-ofis a defining capability of higher levels ofHolds for maturity models that define levels by degree of pipeline automation. determinesassigns ownership ofHolds in organizations large enough to separate platform engineering from model development. limitscan block or breakHolds when consumers cannot be identified and migrated before the model is removed. enablespreservesHolds when archives include data references, environments, and configuration as well as weights. implementsimplements requirements ofHolds when retention limits apply to raw data, logs, and derived artifacts such as features and models. trades-offtrades off againstHolds when audit or reproduction needs conflict with deletion obligations. specializesspecializesHolds when utilization is defined as model-required FLOPs per second over aggregate peak FLOP rate, excluding recomputed work. limitsbounds attainableFor a given kernel mix, attainable MFU cannot exceed the roofline bound implied by each kernel's arithmetic intensity. enablessupplies evidence forWhen the profile covers the same workload, configuration, and execution boundary as the bottleneck claim. usesusesWhen the candidate bottleneck is a kernel's compute or memory-bandwidth bound rather than I/O, scheduling, or communication. producesproducesWhen pruning removes individual weights, leaving an irregular zero pattern whose speedup depends on sparse-execution support. specializesspecializesWhen the pruned unit is a whole channel, filter, head, or layer rather than an individual weight. alternative-tois a dense-hardware-friendly alternative to irregularCoarse granularity yields speedups on dense hardware but usually loses more quality than unstructured sparsity at equal parameter reduction. reducesreduces the cost ofWhen weight matrices are well approximated at a rank much smaller than their dimensions and the smaller products still run efficiently. specializesmitigates quality loss fromWhen the training-time quantization simulation matches the deployed encoding and rounding rules. specializesis a structured form ofWhen architectural choices are treated as a structured search space evaluated under an experiment budget. specializesspecializesWhen the search objective includes a cost measured or predicted on the target platform. reducesoptimizes measuredWhen latency on the target device is measured or predicted directly, since FLOP counts can rank architectures differently from real latency. reduceslowersFewer operations per output come with less data reuse, so memory-bound depthwise layers may not speed up in proportion to the FLOP reduction. requiresis defined only for statedA frontier is defined only for a stated quality metric, cost measure, and target platform. determinescan shift outwardWhen model structure and hardware or kernels are designed together rather than one being fixed while the other is tuned. determinesjointly specifiesWhen the accelerator's dataflow, precision support, or memory hierarchy is a design variable alongside the model. requiresrequiresRetrying an operation with external effects is safe only when the operation is idempotent or deduplicated. limitsboundsWhen the breaker's error threshold trips before retries saturate the failing dependency. increasesraisesWhen replica failures are independent and failover completes faster than the outage it masks. limitslimits the protection ofWhen replicas share a failure domain, a single cause can take out all of them together. determinesdeterminesUnder approximately exponential failure times, where the optimal interval scales with the square root of checkpoint cost times MTBF. increasesraisesFor a repairable component in steady state, availability is approximately MTBF divided by MTBF plus mean time to repair. determinesmaximizes, when well chosen,Under a stated MTBF and checkpoint write cost, the interval trades checkpoint overhead against expected lost work. reducesshortensWhen the blocking snapshot to host memory costs much less than the durable write, lowering the effective checkpoint cost. usesresumes fromWhen the job restarts on a changed worker set and state must be resharded from the latest consistent checkpoint. increasescan raiseWhen replacing failed nodes takes longer than continuing at reduced width and the optimizer tolerates the batch-size change. detectsdetectsFor linear-algebra operations whose input checksums predict output checksums, within the floating-point tolerance used for comparison. causescan manifest asIn synchronous distributed jobs, where one degraded but live node slows every step. usesusesWhen injected faults are framed as experiments against a stated steady-state hypothesis. limitsboundsWhen experiments run against production traffic and must limit the users and components they can affect. limitsboundsWhen isolation between failure domains holds, so a fault does not propagate across shared dependencies. measuresrehearsesWhen the exercise uses the same detection, paging, and runbook paths as real incidents. measuresestimates model sensitivity toWhen the injected fault model, location, and rate approximate the hardware's actual failure modes. specializesis a targeted form ofWhen poisoned training samples pair a trigger pattern with an attacker-chosen output while clean accuracy is preserved. causescan implant aWhen pretrained weights or fine-tuning data come from parties outside the defender's trust boundary. specializesspecializesWhen the adversarial input is a user prompt aimed at bypassing safety training rather than at a task prediction; it differs from prompt injection, where untrusted third-party content crosses an authority boundary. mitigatesreduces success ofOnly within the perturbation set and threat model used to generate training adversaries. trades-offtrades off against cleanRobust training typically costs clean accuracy and multiplies training compute by the cost of generating adversaries. producesproduces a functional copy ofWhen the attacker has sufficient query access and output detail, such as confidence scores, relative to model complexity. enablesenables transfer attacks forWhen adversarial inputs crafted against the extracted surrogate transfer to the deployed model. detectsprovides ownership evidence againstWhen the watermark survives the adversary's fine-tuning, pruning, or distillation of the copied model. measuresverifies integrity ofWhen consumers check signatures against trusted identities before loading and the signing key is not compromised. usesbinds provenance fromWhen the signature covers provenance metadata such as training data and build inputs, not only the weight bytes. enablesenablesWhen model files in executable serialization formats are loaded from untrusted sources without a data-only format such as safetensors. causespropagates intoWhen dependencies are resolved from mutable or unverified sources at build or load time. measuresmeasuresWhen attack advantage over a baseline is used as an empirical estimate of record-level leakage. limitsbounds the advantage ofAt the unit defined by the neighboring-dataset relation; the bound weakens as epsilon and delta grow. causescausesWhen outputs, confidence scores, or shared gradients carry enough information about sensitive attributes or inputs. implementsimplementsWhen clipping norm, noise multiplier, sampling rate, and step count are tracked by a privacy accountant. reducesincreases demand onPer-example gradient computation and clipping add compute and memory relative to standard minibatch SGD. alternative-tois a cheaper, approximate alternative to fullExact unlearning equals retraining without the removed records; approximate methods trade the strength of the guarantee for lower cost. requiresrequiresTo locate every model, shard, and derived artifact that was trained on the records to be removed. usescan use, to hide individual client updates,When enough clients participate per round that the aggregate does not reveal any single client's update. increaseshinders detection of, increasing exposure toBecause the server sees only aggregated updates, it cannot inspect individual client updates for anomalies. limitsconstrains grants made byWhen permissions are scoped to the task and expire or are revoked when the task ends. mitigateslimits the consequences ofWhen a model-driven component processes untrusted content and can act only through narrowly scoped credentials and tools. specializesis a hardware-attested form ofWhen the trusted execution environment's hardware and firmware are trusted and the host operator is not. enablesestablishes trust inWhen the verifier checks attestation evidence against reference values before releasing keys, data, or model weights. limitscan bypassWhen tenants share accelerators, caches, memory, or interconnect whose timing or contention is observable across the boundary. part-ofis a component ofLifecycle carbon sums operational and embodied components under one declared accounting boundary. part-ofis a component ofLifecycle carbon sums operational and embodied components under one declared accounting boundary. increasesscalesOperational carbon is energy consumed times the carbon intensity of supply at the time and place of consumption. usesexploits variation inWhen workloads are deferrable or relocatable and intensity varies materially across hours or regions. trades-offtrades off againstWhen shifting work in time or place delays completion or adds latency beyond a stated objective. reduceslowers, for a fixed amount of work,When useful work and energy are measured over the same system boundary and workload. determinesclassifies as value-chain (Scope 3)For an operator buying hardware, manufacturing emissions fall in Scope 3 while purchased electricity falls in Scope 2. measuresestimatesWhen the assessment's goal, scope, and functional unit are stated and manufacturer data are available or proxied. increasesincreasesThrough on-site evaporative cooling and off-site water used in electricity generation, both of which scale with energy use. trades-offcan conflict withWhen low-carbon hours or regions coincide with high cooling water demand, such as hot sunny afternoons. reducesamortizes, lowering per-unit-workWhen embodied emissions are attributed per year or per unit of work over the asset's service life. trades-offtrades off againstKeeping older hardware saves embodied emissions and capital but forgoes newer generations' work per joule. reduceslowers the annual rate ofFor a fleet of fixed size, shorter refresh cycles retire more equipment per year. part-ofsupplies the numerator ofWhen the TCO window and the count of accepted outcomes cover the same period and system boundary. determinessets the amortization period inCapital cost enters TCO through amortization over the assumed service life. measuresapproximatesOnly when tokens per accepted outcome are stable; longer reasoning, retries, or context resends make cheap tokens an expensive outcome. reducesreducesWhen larger or continuous batches raise tokens per accelerator-second without violating latency objectives. requiresdepends onWhen routers or cascade gates escalate on predicted quality or confidence, which must be calibrated to the request distribution. enablesenablesWhen shared spend is allocated to the workloads that produce the counted outcomes by a declared allocation rule. requiresargues only over hazards fromA safety case is complete only relative to the hazards identified for its stated operating domain. specializesis a method ofSTPA addresses hazards from unsafe interactions and control flaws that component-failure methods can miss. specializesis a method ofFMEA is bottom-up over component failures and does not by itself capture hazards without a failed component. limitsscopes the triggering conditions forSOTIF analysis applies to known and unknown triggering conditions inside the declared operating domain. causescan exposeHolds when deployment conditions fall outside those a learned component's training data covered. enablesprovides the certificate forGuarantees depend on an accurate dynamics model, bounded disturbance, and a feasible filtering problem at each tick. implementsimplementsThe filter must run independently of, and with authority over, the component it constrains. implementsimplementsSwitching must occur while the verified controller can still recover the state. usesfalls back towardHolds when the safety controller's goal is to reach a defined minimal-risk state rather than continue the mission. limitsbounds reachability ofA stop-type fallback is feasible only while the state remains inside its stopping envelope. detectspartially detectsClassifier guardrails are probabilistic and can be evaded, so they do not replace authority boundaries. specializesis a content-level form ofApplies when the safe set is a content or action policy rather than a physical state region. usessearches the parameter space ofFound counterexamples are only as realistic as the scenario model and simulator. usesdraws statistical evidence fromBounds hold for the sampled operating distribution and configuration, not for unmodeled conditions. specializesspecializesExact methods scale poorly with network size; bound-propagation methods trade completeness for scale. measurescertifies absence, within a bounded region, ofHolds only for the specified perturbation set, typically a norm ball around given inputs. measuresprogressively testsEach rung adds real software, timing, or hardware but still relies on simulated or recorded environments. measurescan certifyCertification covers the barrier condition under the modeled dynamics, not the fidelity of the model. implementsexpressesNotation makes the argument structure reviewable; it does not make the argument sound. measuresmonitors validity ofEach indicator must trace to a specific claim and have a threshold that triggers review. requiresis collected throughHolds when field telemetry captures the events the indicator counts. determinesassignsIntegrity levels derive from severity, exposure, and controllability of analyzed hazards. measuresoperationalizesA metric measures one criterion on a chosen population and grouping; other criteria remain unmeasured. limitsforces a choice amongHolds when base rates differ across groups and the predictor is imperfect. trades-offcan conflict withHolds when similarity under the task metric does not align with group membership. trades-offtrades off against aggregateReducing a disparity often lowers aggregate accuracy or shifts error to other groups or metrics. increasesadds toHolds when explanations are computed on the request path; perturbation-based methods need many extra model evaluations. enablessupportsUseful for recourse only when the suggested changes are feasible and actionable for the person. alternative-toremoves the need for post-hocHolds in high-stakes settings where an interpretable model reaches comparable accuracy. usesreports results ofA model card is informative only for the conditions and groups actually evaluated. usessummarizes, in human-readable form, provenance fromA datasheet summarizes provenance; machine-checkable lineage is still needed for reproducibility. reducesunderminesHolds when automation is usually right, error cues are weak, and reviewers face time pressure. limitsbounds, through human reaction time,Takeover is safe only if human reaction and re-orientation time fits within the remaining safety margin. part-ofextends to affected peopleRequires a channel, decision records, and authority to reverse outcomes. part-ofis part ofMapping risks to affected parties precedes measuring and managing them. usesdraws evidence fromAudits also need access to data, code, and logs; documentation alone is insufficient evidence. requiresmandates for high-risk usesApplies to systems classified high-risk under the governing regulation. requiresrequires conformity assessment byApplies to high-risk categories; obligations differ by role (provider, deployer). specializesspecializes for frontier modelsEffectiveness depends on whether capability evaluations elicit the capabilities they test for. reducesinvalidates reuse inRewriting earlier context breaks prefix reuse, so compaction pays off only when saved tokens exceed recompute. limitslimits useful size ofHolds even when the context fits the nominal window; strength varies by model and task. usesis read throughMemory recall quality is bounded by the retrieval contract and index freshness. enablescan corruptHolds when untrusted content can be written into memory and later retrieved as trusted context. requiresis validated againstThe harness must validate calls against the contract; schema conformance does not imply authority. enablesstandardizes tool discovery forDiscovering a tool through a protocol does not grant authority to use it. increasesraises exposure toGUI actions often lack typed preconditions, idempotency keys, or undo, unlike designed tool APIs. increaseswidens the surface forHolds when on-screen content from untrusted sources is read as part of the agent's context. producesproduces and recordsThe harness, not the model, owns continuation and the authoritative trajectory record. determinescaps per-trajectoryBudgets cap spend per trajectory; cost per accepted task also depends on success rate. implementsimplementsEffective only if approvals bind to the exact action and reviewers can judge it; see automation bias. mitigatesguards againstHolds when the gate is placed before dispatch of actions classified irreversible. requiresrelies onResuming after a crash between dispatch and record can re-issue a step, which is safe only for idempotent or deduplicated effects. usesreuses the log ofReplay is exact only if every nondeterministic input, including model samples and tool results, was recorded. enablesenables long waits forResumption must revalidate state and authority that may have changed during the wait. requiresrequiresHolds when a resumed worker could race a stale one that still believes it owns the trajectory. causescan introduceEach delegation and handoff adds a point where specification or context can be lost. requiresmust be justified against a single-agentComparisons should match total compute or token budget. usesusesHolds when leases can expire while a slow agent still attempts writes. limitslimits the benefit ofHolds when agents share a base model, prompts, or retrieved evidence. requiresrequiresTimestamp error translates to spatial error proportional to relative motion. requiresrequires extrinsic calibration fromFused estimates inherit calibration errors between sensor frames. increasesadds inference latency toInference time adds directly to the age of the state used by planning and control. implementsimplementsKalman-family filters assume approximately Gaussian, well-modeled noise; violations make reported covariance overconfident. specializesspecializesDrift accumulates without loop closure or external references. usescan use as its prediction model aPlans exploit model errors, so performance depends on accuracy in the regions the optimizer visits. requiresmust satisfyGeometric planners need a later step to enforce dynamic and actuation limits. requiresrequires a solver WCET within its control period, i.e. aHolds when each optimization must finish within the control period; warm starts and horizon length trade quality for time. specializesis a form ofRe-solving from each new state estimate closes the loop. requirestakes as inputSchedulability results are valid only if WCET bounds hold on the deployed platform. enablesprotects the timing ofHolds when the safety monitor shares hardware with learned components that can overrun or contend for memory. requiresrequires temporal and spatialShared caches, memory controllers, and buses can break isolation that address-space separation provides. increasesincreasesHolds when billion-parameter inference on robot compute runs slower than the control loop. usesuses, to decouple control rate from inference rate,Chunks executed open-loop delay reaction to new observations until the next inference. enablessupplies demonstrations forPolicy quality is bounded by demonstration coverage, consistency, and action-label timing. enablesimprovesHolds when the randomized ranges cover the real parameters without making the task unlearnable. usesis executed throughWarps of a thread block are scheduled together on one streaming multiprocessor. determinesdetermines achievedUncoalesced or bank-conflicted accesses waste transactions, so effective bandwidth falls below peak. implementsapplies within a device the idea ofLoads into on-chip memory proceed while matrix units compute on the previous tile. enablesis required to reach peakPeak matrix throughput is available only to instructions with supported shapes, layouts, and precisions. implementsmaps output tiles onto thread blocks inEach thread block typically computes one output tile staged through shared memory. enablesmakes trainableEmbedding tables exceed one device's memory and must be partitioned across devices. requiresrequires all-to-allLookups for a batch touch rows held on other devices, so activations are exchanged. increasesextends effectiveLookup skew must concentrate accesses on a small hot set of rows. usescommonly operates underWorkers may read parameters that lag the latest updates by a bounded number of steps. limitsbounds the per-stage cost ofEach stage receives a share of the end-to-end budget and a candidate count it can afford to score. usesuses for candidate generationThe first stage retrieves candidates by approximate similarity before learned ranking. usesis composed withNon-text inputs are encoded into tokens the shared backbone attends over. reducesreducesLarge media inputs would otherwise block token generation for requests sharing the batch. producesderivesThe compiler's cost model must reflect the target device mesh's compute and link bandwidths for the chosen plan to be competitive. implementsimplementsHolds when users supply seed annotations and the compiler propagates layouts through operators with known sharding rules. producesinsertsCollectives are inserted wherever an operator's input layouts differ from what its sharded implementation requires. requiresis defined over the axes ofTensor dimensions are sharded over named mesh axes, so mesh shape bounds which layouts are expressible. implementsexposes as a transformHolds when programs are pure functions so differentiation composes with vectorization and compilation. requiresrelies onTransformations are applied to a traced intermediate representation, so Python side effects and data-dependent control flow are not captured. enablesorchestratesApplies when rollout workers, trainers, and reward models run as separate stateful actors exchanging data through the runtime. determinesis a design axis ofA single controller simplifies heterogeneous pipelines but must keep per-step dispatch off the critical path to scale. usescomposesPer-operator roofline bounds are summed or overlapped with communication and queueing terms under stated overlap assumptions. usescomposesCommunication terms use per-collective latency and bandwidth parameters measured or specified for the target fabric. measuresbounds trust inPredictions are credible only within the configurations where validation error against measurement has been reported. alternative-totrades speed for detail againstSimulation captures contention and scheduling effects closed forms omit, at much higher evaluation cost per configuration. specializesappliesHolds when the calibration set is representative of deployment inputs; otherwise chosen ranges misfit and accuracy drops. specializesis usually a form ofHolds for GPTQ- and AWQ-class methods that set weight codes from calibration data without retraining. limitscan degrade the accuracy ofApplies to per-tensor activation quantization of large transformers; per-channel handling or scale migration mitigates it. reducesshrinksCapacity gains translate into throughput only when serving is KV-memory-bound and attention kernels read the quantized format directly. requiresrequires support inSpeedup materializes only on matrix engines with sparse-operand instructions for the specific N:M ratio. specializesspecializesConstrains sparsity to a fixed per-group pattern so hardware can index nonzeros with small metadata. enablesincreases the capacity and hit rate ofPays off when reload time from host or SSD is shorter than recomputing the prefix's prefill. requiresrequiresDisaggregation is profitable only when KV transfer between pools is faster than the decode-latency gain it buys. usescomposesHolds when the program chooses among models per step, so routing decisions are optimized against end-to-end quality and cost. specializesis a kind ofRetrieval-augmented generation is the two-component case of a retriever feeding one model call. determinestunes the prompts and weights ofRequires a declared program structure and an end-to-end metric evaluable on a training set of inputs. usesserves many instances ofHolds when adapters share one frozen base model so base weights are loaded once and only low-rank factors differ per request. specializesspecializesApplies when hosted variants differ only by adapters, letting requests for different variants share one batch. implementsimplementsBest recall-latency trade-off when the full graph and vectors fit in memory. implementsimplementsPreferred when memory per vector is the binding constraint, accepting lower recall or a reranking pass on exact vectors. alternative-totrades memory against recall versusCompared at equal query latency on the same corpus and recall target. specializesis a predicate-constrained form ofRecall or cost degrades most when the predicate is selective and uncorrelated with vector proximity. limitsupdate and delete cost bounds attainableDeletes and in-place updates degrade graph connectivity unless repaired or periodically rebuilt. reduceslimitsBitwise reproducibility requires fixing kernels, reduction order, parallel layout, and hardware, not only random seeds. determinessets the magnitude ofOrder-dependent rounding differences grow as accumulation precision falls and reduction length grows. mitigatesremoves the batch-composition source ofAchieved by fixing reduction strategies independent of batch size, typically at some kernel throughput cost. enablesaligns inference numerics with training forMatters when rollout log-probabilities from the serving engine are reused in on-policy training objectives. requiresrequiresHolds whenever a restart changes world size or parallel layout relative to the saved checkpoint. usesremaps between layouts ofRequires checkpoints that record each shard's global tensor coordinates rather than rank-local layouts. enablesexecutes at corpus scaleApplies when the corpus exceeds one machine, so near-duplicate detection is partitioned and shuffled across workers. enablesexecutesBackfills over long histories run as bounded batch jobs whose logic must match the serving feature definition. part-ofprovides request-level evidence withinRequires trace context to be propagated across every service and asynchronous hop the request traverses. specializesspecializesAdds model-specific span attributes such as token counts, model version, and cost under a shared semantic convention. measuresinstrumentsEach loop iteration, model call, and tool call is emitted as a child span of the task's root span. implementsoperationalizesComparability holds only when prompt templates, decoding settings, and scorer versions are pinned with the results. usesrecords results intoHolds when each run logs checkpoint, suite version, and configuration alongside scores. measuresdetermines the reported value ofReported energy is comparable only across measurements with the same boundary, sensor class, and sampling rate. determinessets attribution rules forMatters when hardware is shared or partly idle, so idle power must be charged by an explicit allocation rule. alternative-totrades compute overhead for hardware trust versusHomomorphic encryption avoids trusting server hardware but costs orders of magnitude more compute than enclave execution. alternative-totrades communication for computation versusHybrid protocols use homomorphic encryption for linear layers and secret sharing or garbled circuits for nonlinear ones. mitigatesprevents input disclosure inProtects inputs during computation but not leakage from the revealed outputs themselves. detectsdiscoversCoverage depends on attacker diversity and budget, so finding no failures does not establish their absence. requiresmust be scoped byThe adversary capabilities, interfaces, and harms probed must be declared for findings to be interpretable. part-ofsupplies the threshold tests that trigger safeguards inHolds when the framework ties stated capability thresholds to required mitigations before further training or deployment. usesuses elicitation fromCapability estimates are lower bounds unless elicitation, including expert red teaming, is strong. specializesspecializesThe objectives are model quality and a stated cost measure on a stated platform. measuresis the per-input reciprocal ofBoth are measured within the same declared system boundary. determinesdetermine feasibleLatency, privacy, connectivity, and power requirements rule out some device, edge, or datacenter placements. requirescompares candidates againstA candidate advances only if it beats or matches the incumbent baseline on the gated metrics under the same protocol. part-ofserves as the control arm ofThe incumbent system or model is served to the control group during the experiment. requiresis only meaningful for a statedBenchmark results transfer only to deployments whose input, request, and model distributions match the benchmark workload. usesrecords an empirical sample ofThe trace window covers the arrival, size, and mix variation that matters for the decision being evaluated. causescan invitePolicies optimized hard against a learned reward model can exploit its errors, scoring higher while true quality falls. producescomputesMonitoring records the events and counts over the population and window the indicator definition specifies. measurescommonly measures, as a percentile,Latency indicators are computed as a high percentile over a stated request population and window. measurescommonly measuresAvailability indicators count successful requests or uptime over a defined population and window. determinescan freeze releases throughTeams agree that an exhausted budget blocks or slows promotion of new versions. determinessets burn-rate thresholds forAlerts fire when the budget is being consumed faster than the window allows. limitscaps the intermediate tokens allowed byInteractive requests must answer within a fixed end-to-end latency target, so extra reasoning tokens consume the same budget. reducesreduces the peakRetained activations dominate training memory and spare compute is available to recompute them. increasesadds per-request state toAutoregressive serving retains keys and values for every in-flight token; the term grows with batch size and context length. measuresreveals theInter-arrival times are recorded at fine enough resolution to show bursts rather than only averages. usesreplaysTrace-driven simulation is used to compare scheduling, capacity, or serving policies on realistic demand. requiresrequires calibrated probabilities fromExpected-cost thresholds are computed from predicted probabilities, which must match observed frequencies on the deployment distribution. determinessets the abstention threshold inDeferring to a human or fallback has a known cost that can be compared with the expected cost of each error type. determinessets how much text fits inFor a fixed window, a tokenizer that produces more tokens per word admits less source text. alternative-tois an alternative toBoth are standard choices for tabular supervised prediction; trees capture interactions linear scores need engineered features for. specializesis, when sparse, a kind ofHolds when features are meaningful and few enough that coefficients can be read directly. usessplits onTrees operate on tabular features without requiring scaling; feature engineering still determines what interactions are available. increaseshas a quadratic trainingExact training forms or stores an n-by-n kernel matrix, so memory grows with the square of the training set size unless approximated. determinesencodes its inductive bias through the choice of kernel inThe kernel function fixes which similarity structure the model can exploit. usessearches overNeighbors are found by distance in a learned or engineered vector space whose quality determines result quality. usesspends most compute inDense and attention layers dominate; small or irregular layers can be bound by other operations. implementsencodes translation equivariance and locality as anWeights are shared across positions, so the model assumes local patterns recur across the input. limitsbounds the usable length ofWithout extension, sequences longer than the positions seen in training degrade quality. usesstabilizes residual blocks withEach block applies layer or RMS normalization around attention and feed-forward sublayers. usesinjects order withAttention alone is permutation equivariant, so the architecture adds absolute, relative, or rotary positions. specializesis a kind ofThe architecture is a differentiable composition trained by gradient descent. usesruns in its ranking stages aCandidate generation narrows the set before costly deep ranking models score it within the latency budget. specializesis a factorized kind ofA standard convolution is split into depthwise and pointwise steps. reducescuts parameter storage inParameter memory is a significant share of the deployed footprint, as on mobile and embedded targets. alternative-tois an alternative toBoth mix information across sequence positions; recurrence keeps fixed state but serializes time, attention parallelizes but grows cost with length. mitigatesmitigatesDeep networks whose activation scales would otherwise drift with depth or batch composition. part-ofis the gating component ofThe layer routes each token to a small subset of experts. part-ofis a buffer-size parameter ofExperts process fixed-size per-batch buffers so computation shapes stay static. increasesincreases padded token volume inA higher capacity factor allocates larger per-expert buffers that are exchanged whether or not they are full. mitigatesreduces hot-expert imbalance behindUnder expert parallelism, an overloaded expert's device finishes last and holds back the step. part-ofis an auxiliary objective in training ofRouting is learned and can otherwise collapse onto a few experts. determinessets the parameter count used inFor conditionally computed models, the six-operations-per-parameter-per-token estimate uses active rather than total parameters. requiresrelies on confidence quality assessed byExit decisions threshold an intermediate confidence; overconfident exits return wrong answers early. reducesreduces averageMany inputs are easy enough to exit at an early classifier and the exit overhead is small. specializesis a restricted kind ofEach position attends only within a window or fixed sparse pattern. reducescaps the size ofDuring decoding only the most recent window of keys and values must be retained. specializesis a kernelized variant ofSoftmax similarity is replaced by a feature-map inner product, changing model quality. alternative-tois an alternative toBoth reduce attention cost growth for long sequences, trading different quality losses. alternative-tois an alternative toBoth serve as the sequence-mixing layer; state-space layers trade exact long-range lookup for linear cost and fixed state. reduceseliminates the growingInference carries a constant-size state per layer instead of keys and values for every prior token. specializesis a kind ofMultiple query heads share each key-value head. reducesreducesDecode is bound by reading the KV cache, so fewer key-value heads cut bytes read per token. increasesextends the usable length ofThe model is fine-tuned briefly on longer sequences after rescaling positions; quality at the new length must be checked. determinesdetermines exposure toExecutable formats such as pickle run code on load; data-only formats such as safetensors do not. determinessets weight-loading time inMemory-mappable, zero-copy formats load faster than formats requiring full deserialization. usescommonly passes through aTraining and target runtimes differ and both support the exchange format's operator set. producesproduces a deployableOperators, shapes, and weights are fixed so the artifact runs without the original program. usescaptures the graph withThe program is traced or captured on example inputs; untraced control flow is lost. part-ofis the interface declared inThe artifact records input and output names, types, and shapes. requiresvalidates requests againstThe server decodes and checks requests using the artifact's declared inputs and outputs. specializesis a kind ofThe model consumes images and language and emits actions through a shared backbone. usescommonly uses as its shared backbone aEncoded modality tokens are processed together with text tokens by a transformer. usesschedules separately theImage or audio encoding is heavy enough to stall token generation if co-scheduled. mitigatesmitigatesOlder material is summarized so relevant information stays short and near the end of the window. mitigatesmitigates position effects behindReranked evidence is placed where models use it best, near the start or end of the context. increasesretains forward activations, increasingReverse-mode differentiation stores intermediates from the forward pass for use in the backward pass. determinesdetermines whether accesses achieveStrides decide whether adjacent threads touch adjacent addresses. reducesreduces activation and gradient bytes inActivations and gradients are stored in 16-bit formats while a master copy of weights stays in higher precision. usesgains throughput fromThe matrix engine supports the reduced input formats with wider accumulation. usespasses between operations aEdges carry tensors whose shapes and types the graph records. usescarries an element type drawn fromEach tensor records a numerical format that kernels must support. usestakes as input aThe compiler needs a captured graph rather than one operation at a time. part-ofis an alternative toBoth are ways to run framework programs; eager keeps host flexibility while capture enables whole-graph optimization. determinesforecloses cross-operationEach operation is launched as the host issues it, so no compiler sees neighboring operations. implementsexactly computesThe result matches standard softmax attention up to floating-point reordering. usesis built onQuery, key, and value blocks are staged through on-chip memory with an online softmax. increasescan increase, through duplicated patches,im2col-style lowering copies overlapping input patches into a larger matrix. implementsprovides tunedVendor libraries supply GEMM and convolution kernels selected per shape and precision. alternative-tois an alternative toBoth supply operator kernels; libraries cover common shapes, DSLs let engineers write fused or novel kernels. useslowers throughGraph, tensor, and loop-level representations carry the program between passes. usespropagates layouts throughSharding annotations are attached to operations in the compiler IR. usesmeasures candidates withCandidate schedules are compiled and timed on the target, or a cost model is trained from such measurements. part-ofis a search stage withinThe compiler exposes a schedule space whose performance is hard to predict analytically. useslets the compiler managePrograms are written over blocks while the compiler allocates shared memory and registers. enablesenables customEngineers can write a single kernel that combines operations no library provides fused. part-ofis a pass withinStatic shapes let the compiler compute tensor lifetimes ahead of execution. alternative-tois an alternative toBoth manage device buffers; static planning needs known lifetimes, caching allocation handles dynamic shapes. reduceshides, through asynchronous enqueue, exposedThe host can run ahead of the device, so launch work overlaps device execution until queues drain. usesis captured from and replayed onWork enqueued on a stream is recorded as a graph and replayed as one launch. reducesavoids synchronizing allocation calls inFreed blocks are reused instead of calling the device allocator, which can synchronize the host. reducesreduces per-step launch overhead inDecode steps launch many small kernels with static shapes, so host overhead is a large share of step time. increaseskeeps small gradients representable, improvingReduced-precision gradient formats would otherwise underflow; overflow is detected and the step skipped. mitigateslimits the range damage fromA shared scale per small block confines an outlier's effect to its own block. usescan encode values inHardware or kernels support the block-scaled element and scale formats. determinesfixes the availableThe matrix engine implements specific accumulation formats for each input format. usesexposes as a transformPure array functions are traced and handed to an XLA-style compiler. determinessets where per-step dispatch cost falls inOne controller issuing work to many devices must keep dispatch asynchronous and off the critical path. usesshares within a thread blockThreads of a block cooperate through shared memory sized per block. determinesdetermines, through thread-to-address mapping,Adjacent threads of a warp should access adjacent addresses. limitsboundsShared memory and registers per block limit how many warps can be resident. determinessets, through block size and per-thread resources,Block shape and register or shared-memory use determine resident warps per multiprocessor. usesstages operands intoBulk copies land in shared memory while compute proceeds on the previous buffer. increasesraises matrix-engineThe kernel would otherwise stall its matrix units waiting on loads. usesissues instructions toFragments are loaded in the shapes and layouts the unit accepts. requiresrequires specific operandMatrix instructions accept only particular fragment layouts, so data may need rearrangement. enablesenables bitwise-reproducible serving withinKernels use fixed reduction strategies independent of batch size and co-batched requests. specializesis a kind ofProcessing elements pass operands to neighbors under a fixed schedule. specializesis a kind ofThread groups issue the same instruction and divergent paths serialize. part-ofis a functional unit withinThe accelerator pairs matrix engines with vector or scalar units. part-ofis the device-memory level ofThe accelerator uses stacked DRAM as off-chip device memory. part-ofis the fastest managed level ofThe platform exposes caches, shared memory, or scratchpads near compute. limitsbounds the throughput ofOffloaded optimizer state, parameters, or activations cross the host link every step. limitsbounds reload rate inOffloaded KV blocks return to device memory over the host link on reuse. limitsboundsFacility and rack feeds, conversion stages, and breakers limit sustained draw per rack. causescan cause, through insufficient voltage margin,Droop or marginal supply under heavy load leaves circuits below the voltage they need for correct timing. reducesreducesOperands are reused across processing elements instead of being refetched from register files or memory. determinesdeterminesLocal storage cannot hold every operand, so the schedule chooses which data is reused in place. usesoperates under a chosenThe array keeps one operand stationary, such as weights, while others stream. increasesraisesTarget operations dominate the workload and are supported by the specialized hardware. determinessets deviceStack count and stack height fix the capacity attached to each accelerator. mitigatesnarrows, without closing,Stacking and wide interfaces raise bandwidth, though compute still grows faster across generations. reducesdepressesMemory-bound phases such as decode cannot feed the arithmetic units at peak rate. part-offorms the fastest tier ofThe cluster topology is hierarchical, with a node or rack domain joined by a slower scale-out fabric. specializesis a kind ofThe link adds hardware coherence and load-store memory semantics over a host-device physical layer. enablescan provide a pooled tier forAttached or pooled memory is cheaper than device memory and its higher latency can be overlapped with compute. usesmoves state withDirect device or NIC transfers avoid staging KV blocks through host memory. enablescan read and write device memory throughThe NIC and driver support direct access to accelerator memory. determinessets, with accelerators per rack,Racks are provisioned for the sustained power of the chips they hold. determinessets the fraction ofTransistor counts grow while the power envelope does not. usesenforces caps throughThe device lowers clocks and voltage when measured power exceeds the cap. usesthrottles throughThermal controls lower frequency and voltage when temperature limits are approached. usesruns models through a lightweightNo operating system or virtual memory is available, so the runtime is statically linked and memory-planned. part-ofis one block ofThe SoC exposes the NPU alongside CPU, GPU, and DSP under shared memory and power. usesdelegates operators acrossThe runtime partitions supported operators onto the NPU, GPU, or DSP and falls back to CPU. reducesreducesThe model's operators and precisions are supported, so inference avoids CPU fallback. usessearches withinThe search treats flash and SRAM limits as hard constraints on candidate models. limitsboundsTraining must fit weights, activations, and update state in on-chip memory. usesserves hot rows from anLookups are skewed so a small set of rows receives most accesses. specializesis a kind ofRows are placed across accelerator memory, host memory, and SSD by access frequency. reducesdivides per-device parameterEach stage holds only its layers' parameters and their state. part-ofis one dimension ofPipeline stages are combined with data and tensor parallelism across nodes. usesgathers and scatters state throughSharded parameters are all-gathered before use and gradients reduce-scattered after. reducesdivides per-workerParameters, gradients, or optimizer state are partitioned instead of replicated. specializesis a kind ofThe fabric is built from tiers of commodity switches. specializesis a kind ofSame-index accelerators across nodes share a leaf switch. part-ofis one dimension ofMoE layers place experts across devices alongside data and tensor parallelism. part-ofis one dimension ofLong-sequence training adds a sequence axis to the parallel mapping. usesexchanges keys and values throughBlocks of keys and values circulate among devices while attention is computed. reducesshrinks, through more microbatches,More microbatches per optimizer step fill the pipeline relative to the number of stages. reduceslowersDevices idle during fill and drain while their peak FLOPs still count in the denominator. implementsis an algorithm forParticipants are arranged in trees or hierarchical groups. specializesis a kind ofEvery participant sends distinct data to every other. measurespredicts the cost ofLink latency and bandwidth are roughly uniform and contention is ignored. reducescan reduceTraining is communication-bound and compression preserves convergence at the chosen ratio. usescan transmit only top-magnitude gradients, usingResiduals are accumulated locally so dropped values are sent later. usesholds leadership through aThe leader renews a time-bounded grant so a failed leader's authority lapses. enablesenables timedA standby can take over once the failed holder's lease has expired. requiresrequiresA deposed leader may still act; resources must reject its stale epoch. usescan combine updates throughThe server must not see individual client updates. part-ofis the per-round selection step ofEach round samples a subset of available clients. specializesis a cross-client form ofClients' data-generating distributions differ from each other and from the population. determinesdetermines exposure toCross-device federations span varied hardware and runtimes; cross-silo ones are few and uniform. determinesdetermines theCross-device clients drop out routinely, while cross-silo participants are expected to stay available. part-ofis the model-parallel part ofDense layers are data-parallel while embedding tables are partitioned. alternative-tois an alternative toBoth aggregate gradients for data-parallel training; servers suit sparse or asynchronous updates, all-reduce suits dense synchronous ones. usescommonly trains sparse tables through aUpdates touch only the rows looked up in each batch. usesranks plans withThe search evaluates candidate partitionings with a cost model of compute and communication. determinesdetermines the axis alignment ofThe busiest parallel axis is mapped onto the fastest links. usesassigns parallel dimensions to the axes ofEach parallelism dimension corresponds to a named mesh axis. measuresmeasures facility overhead ofIT and total facility energy are metered over the same period. reducescan reduceCooling energy is a large share of facility overhead and liquid loops need less fan and chiller work. mitigatesmitigates throttling inAir cooling cannot hold chips below their thermal limits at the target power. reducesreducesMore servers share provisioned power capacity because their peaks rarely coincide. usesprovisions power throughFleet-wide power measurements show aggregate peaks below summed nameplate. determinesis designed to minimizeFacility, hardware, and software choices are evaluated as one cost model. requiresrequiresPriority flow control alone spreads congestion and pauses; end-to-end rate control keeps queues short. increasescan inflatePause frames cause head-of-line blocking that delays unrelated flows under congestion. mitigatesmitigates networkJobs share switch queues on a common fabric. usesrelies for cross-rail traffic onTraffic between different rails is forwarded through the node's scale-up domain. reducesreduces link traffic and latency ofSwitches support aggregation for the data types and reduction operations used. alternative-tois an alternative toBoth perform all-reduce; switch aggregation shortens paths while rings need only end-host support. alternative-tois an alternative toBoth serve shared training data and checkpoints; file systems offer lower latency, object stores cheaper scalable capacity. usescommonly writes to aSharded checkpoints are written concurrently by many ranks. determinesfavors large sequential shards inPer-request latency is high, so many small reads are slow. implementshides the training stall fromA fast snapshot to host memory lets persistence proceed in the background. determinessets the write-cost input toThe optimal interval scales with the square root of checkpoint cost times MTBF. requiresrequiresSynchronous replicas cannot start or step until every worker holds its allocation. increasescan increase jobLarge jobs wait until enough resources are free at once. increasesraisesJobs can share nodes without violating resource or isolation constraints. increasescan increaseDense co-location puts jobs on shared caches, memory bandwidth, and links. usescan share devices throughTenant workloads are too small to saturate whole accelerators. part-ofis a policy withinTenants share a cluster under stated limits and guarantees. alternative-tois an alternative toBoth divide shared capacity among tenants; quotas are static limits, DRF adapts shares to demand. mitigatesmitigates, with hardware partitions,Partitions give each workload dedicated memory and compute slices; time-slicing does not. enablesmakes usableTraining continues while preempted workers leave and replacements join. part-ofis the corner ofBoth ceilings refer to the same device and memory level. increasesmoves to higher intensity thePeak compute grows faster than bandwidth across hardware generations. measuresmeasuresProblem size is fixed while resources grow. measuresmeasuresWork per resource is fixed while resources and problem size grow. alternative-tois an alternative toBoth describe how to use added resources; one grows the problem, the other fixes it. limitsbounds useful data-parallelGrowing the global batch with worker count stops reducing optimization steps beyond the critical batch size. usessizes concurrency withArrival rate and time in system are measured for a stable operating point. determinesmoves the operating point alongLarger batches raise throughput while adding batch-formation and execution delay. determinesbends upward theUtilization approaches service capacity, so waiting time grows sharply. determinesraises headroom required byCapacity is provisioned for bursts at the target latency percentile, not mean load. measuresbounds trust inSimulator predictions are validated against measurement on reference workloads. requiresrequires measurements fromValidation compares predictions with profiled or benchmarked execution. limitsboundsThe unparallelized fraction is fixed while workers are added to a fixed-size problem. determinessets the read and decode cost ofRecord grouping, compression, and columnar layout decide how many bytes each batch must fetch and decode. limitsbounds sustained all-unitPower and thermal limits prevent every unit from running at full frequency at once. reduceslowers network cost withinIdentical commodity switches replace large high-radix core switches at equal bisection bandwidth. detectsreveals undeclared and underusedLineage must be captured at the feature or column level and compared against what the model actually consumes. mitigatesmakes explicit and versionsHolds when every model input is consumed through a declared, owned contract rather than an ad hoc signal. detectsexposes subset regressions caused byApplies when a change to one input or hyperparameter is evaluated on the same fixed slices before and after. enablespins the settings required forHolds when every run records the exact reviewed configuration version it executed. enablesmakes configuration changes revertible throughApplies when configuration is versioned and deployed through the same pipeline as code. part-ofis a shared component ofTypical when an organization centralizes feature definitions for many teams. part-ofis a shared component ofTypical when promotion and deployment are standardized across teams. enablesprovisions repeatable environments forHolds when platform compute, storage, and services are declared in versioned files. determinesassigns on-call ownership inApplies when data, model, and platform owners are named for each production model. increasesadd hidden coupling toHolds whenever model outputs are consumed without access control or registration. mitigatesprevents unregisteredApplies when prediction endpoints and output tables require granted access per consumer. measuresscores accumulatedThe rubric must be applied repeatedly to the same system to track change over time. requiresrequires evidence fromHolds for the infrastructure and pipeline sections of the rubric. specializesis a kind ofHolds when training uses input-output demonstrations as targets. implementsimplementsApplies when a pretrained model is adapted to a new task or behavior. requiresstarts from a policy produced byHolds for the standard InstructGPT-style pipeline; some recipes skip or shorten the SFT stage. enablessupplies instruction data forApplies when generated demonstrations are filtered for quality and diversity. causescan causeHolds when new data differs from the original distribution and old data is not replayed. usescommonly mixes inApplies when earlier data is retained and permitted for reuse. limitslimits which past examples are available toHolds when retention or deletion policies remove raw training data after a fixed period. specializesis a step-level kind ofHolds when the scorer is trained to rate intermediate reasoning steps. specializesis a kind ofHolds when rewards come from programmatic outcome checks. requiresrequires many sampled responses fromHolds for policy-gradient training where rewards are computed on generated outputs. causesimperfect checks can causeHolds when a policy or search is optimized against a verifier with exploitable false positives. increasesmultiplies generation cost inGeneration cost scales roughly linearly with n unless prefixes are shared. alternative-tois an alternative toBoth spend parallel samples; best-of-n needs a scorer while self-consistency needs answers that can be voted on. requiresrequires candidate diversity fromHolds because greedy decoding would return n identical candidates. specializesis a kind ofApplies when outputs have an automatic correctness check, as in code generation. determinestemperature shapesThe best temperature rises with k because more diverse samples help coverage. increasesmultiplies per-query compute inHolds in proportion to the number of expanded and scored nodes. usesshares partial-solution state throughApplies when branches share prompt and partial-solution prefixes on one server. increasesmultiplies generation cost inCost scales with the number of sampled reasoning paths. usesuses search guided byApplies to the search-based way of spending inference compute, as opposed to sequential revision. usescan use as the judgeApplies when pairwise comparisons are graded by a model rather than people. usesaggregates votes fromHolds for crowd-sourced arenas where people choose between two outputs. measuresmeasures reliability ofRequires several raters to judge overlapping items. usescan score open-ended outputs withApplies when tasks lack exact-match references. detectscan flagHolds when the judge sees the source evidence and its error rate has been measured against people. causescan causeHolds when missingness patterns or imputation differ between offline data and live requests. detectsflags unexpectedRequires declared expectations on completeness per feature. measureschecksCovers only the properties the checks encode, not semantic correctness. detectsflags undeclaredHolds when incoming data is compared against the expected schema version. causesuncoordinated changes can causeHolds when training and serving read different schema versions or interpret a changed field differently. mitigatesguards againstHolds when splits respect time, entity, and duplicate boundaries. mitigateskeeps a final test partition free ofHolds only if the final test partition is not used for any selection decision. specializesis domain-levelApplies when sampling weights are set per source rather than per example. specializesis label-acquisitionApplies when the cost being allocated is annotation effort. alternative-tois an alternative toBoth reduce manual labeling cost; one writes labeling heuristics, the other chooses what humans label. increasesraises averageHolds when the filter's notion of quality matches the target task; filters can also remove useful diversity. usesruns at corpus scale onApplies to web-scale pretraining corpora. specializesacts as a form ofHolds when transformations are label-preserving. mitigatescan reduce sensitivity toHolds only for shifts that the chosen transformations actually model. alternative-tois an alternative toBoth enlarge training data; augmentation transforms real examples while synthetic data generates new ones. usesgenerates variedApplies to simulator-generated training data for transfer to the real world. causessampling bias can causeHolds when the collected population differs from the deployment population. alternative-tois an alternative toChoose by whether the data is unbounded and how fresh outputs must be. usesconsumes examples fromApplies when updates happen as events arrive. mitigatesadapts toHolds when labels arrive quickly enough to track the change. mitigatesrealigns history to reduceHolds when the backfill uses the same definition that serving computes. part-ofis part ofHolds when a release records the exact dataset versions it was trained and evaluated on. requiresreferences immutableLineage is only replayable if the referenced datasets are versioned. usesprunes reads throughHolds for columnar, partitioned layouts with statistics. limitsbounds queue growth andHolds when upstream producers honor the signal rather than buffering unboundedly. mitigatesreduces the overfitting exploited byReduces but does not bound leakage; formal bounds need differential privacy. limitsbounds the usefulHolds for a given target loss and training phase; it grows as loss falls. increasesadds about one extra forward pass toHolds for full recomputation; selective schemes recompute less. alternative-tois an alternative toBoth relieve accelerator memory; one spends compute, the other interconnect bandwidth. reducesreduces target-taskHolds when a suitable pretrained model exists for the target domain. enablespretrains representations forHolds when the pretraining data covers the target domain. producesproducesApplies to contrastive and masked-prediction objectives. usesclusters errors withApplies to methods that group examples without predefined slice labels. usessearches for coherent slices inHolds for embedding-based methods such as Domino. specializesis a kind ofHolds for PPO-style optimization against a learned reward. alternative-tois an alternative toBoth align to preference data; DPO skips the separate reward model and RL loop. reducesavoids reward and value models inHolds relative to PPO-based RLHF; a frozen reference model is still needed. requiresrequires reliable preferenceHolds because reward accuracy is bounded by comparison label agreement. usesis commonly trained byApplies when robot demonstrations supply the action targets. determinesshifts the best peak rate ofHolds below the critical batch size, where learning rate is typically scaled with batch. determinescarries tuned settings intoHolds under a parameterization such as muP that keeps optima stable across width. usesruns on small proxiesHolds when proxy models share the target's parameterization. usescompares trials throughApplies whenever more than a handful of configurations are evaluated. reducesreduces per-step activationHolds because activation memory scales with micro-batch, not effective batch. reducesreduces the frequency ofHolds when gradient all-reduce is deferred to the last micro-batch. usesparameterizes updates byHolds when the update rank is much smaller than the frozen weight dimensions. usesbroadcasts weights throughApplies when training and rollout workers sit on separate device groups. reducessync time reducesHolds when transfers are not overlapped with rollout or training. increasesasynchronous generation increasesHolds when rollouts overlap with policy updates. usesupdates sparse parameter subsets viaApplies under tight device memory where full backpropagation does not fit. requiresrequiresHolds for cross-device federation; cross-silo clients train on servers. specializesis a rotatingApplies when data is too scarce for a single fixed validation split and examples are exchangeable. requiresrequires a held-out calibrationCoverage holds only under exchangeability with that calibration split. enablesenables early stopping of a harmful stage inHolds when the test's error guarantees allow continuous monitoring. causesshared retraining data can causeHolds when control and treatment models learn from logs that both arms influenced. increasesinflatesHolds when the evaluated items or close variants were seen in training. usesbuilds invariance and directional tests fromApplies to CheckList-style INV and DIR tests. usescan block on failures ofApplies when capability tests are part of the predeclared checks. usescan block on thresholds crossed inApplies to frontier-model releases governed by a scaling policy. limitsbounds by shedding excess loadHolds when rejected work is not retried immediately into the same queue. implementsimplements load shedding forHolds when shedding low-priority work keeps the remaining requests within their SLO. increasesraises averageHolds when demand varies and scale-down is permitted. enablesenables block sharing forHolds when the allocator supports copy-on-write sharing of identical prefix blocks. increasesraises batch size and soHolds when KV memory fragmentation was limiting concurrency. reducesreducesHolds when the draft acceptance rate is high and decode is memory-bound. increasesverifies several tokens per weight read, raisingHolds at small batch sizes where decode is memory-bound. reducesskips recomputed prefill to reduceHolds when requests share long identical prefixes and the cache hits. increasesincreasesHolds under load when requests wait for a prefill slot. increaseslengthens prompts and soHolds because retrieved passages add prefill tokens and retrieval adds a serial step. alternative-tois an alternative toChoose by required freshness, latency, and cost per prediction. usesusesApplies when predictions are served per request behind a network interface. reducesreducesHolds when predictions can be precomputed for a known input set. requiresloads and executesHolds for any server that loads serialized weights and graph at startup or on version change. increasesincreasesHolds when individual models have low or bursty traffic. alternative-tois an alternative toRouting decides before calling a model; a cascade escalates after a cheap model's answer. increasesadds duplicate load toExtra load stays small when hedges fire only after a high-percentile delay. mitigatesmitigatesHolds when slowness is replica-specific rather than shared. usesrelies on dynamic allocation fromHolds for vLLM-style engines where sequences join and leave each iteration. reducesreducesHolds when load signals are fresh and replicas are interchangeable. reducesreducesHolds when imbalance, not total capacity, causes the tail. alternative-tois an alternative toBoth limit prefill-decode interference; one colocates with chunking, the other separates pools. usespiggybacks prefill chunks onHolds when the scheduler admits work at each iteration under a per-iteration token budget. enablesguarantees well-formed arguments forHolds when the tool schema is compiled to a grammar. usesconstrains arguments withApplies when tool arguments must match a declared schema. implementschecks after generationApplies when outputs are parsed against the schema after generation instead of masked during it. increasesretries increaseHolds when invalid outputs trigger regeneration. limitscaps the reasoning share ofHolds when reasoning tokens are billed or provisioned like output tokens. limitsbounds growth ofHolds because reasoning tokens stay in context until the answer. usesusesHolds for bi-encoder retrieval where queries and records are embedded separately. specializesis a kind ofApplies when relevance is scored by similarity of learned vectors. specializesis a kind ofApplies when relevance is scored by weighted term overlap over an inverted index. part-ofis a late stage ofHolds when a cheap candidate stage precedes a costlier scoring model under one latency budget. usescommonly uses cross-encoderApplies to neural rerankers that jointly encode query and candidate. requiresrequiresHolds because runtimes accept only their own formats and operator sets. usesdelegates supported operators toApplies when the device has an accelerator and a delegate for it. requiresmust verifyHolds for any fleet where tampered models could be pushed. usesrelies on on-deviceRequires the previous model to stay on the device. reducescan reduce on-deviceHolds when uplink cost is lower than the compute it moves off device. determinesshifts the best split inHolds because device speed and network differ widely across the fleet. enableskeeps prior versions available forHolds when prior versions and their serving configuration remain registered and deployable. detectscan exposeHolds when shadow outputs are compared with offline predictions on the same inputs. increasesduplicates inference cost inHolds for the shadow period because every mirrored request runs on both models. usesrunsApplies on each change to code, data checks, or model artifacts. part-ofmarks higher levels ofHolds for maturity models in which automated pipeline delivery defines the top level. part-ofis part ofApplies when distribution checks run alongside service and quality signals in production. requirescompares distributions fromHolds because production distributions are reconstructed from logged inputs and predictions. limitslimits outcome-basedHolds when labels arrive late; input-distribution checks remain available. usesfires on thresholds overApplies when alerts are defined over user-facing indicators rather than raw resource metrics. specializesis the input-target form ofHolds when P(y|x) changes even if the input distribution does not. implementsruns the pipelines ofHolds when retraining, validation, and registration steps run as an orchestrated pipeline. usesusesApplies to ML incidents whose cause could lie in data, model, code, or infrastructure. usestraces data and artifact versions throughHolds when the failing prediction can be tied to the data and artifact versions that produced it. enableslocalizes the component behindHolds when every component a request traverses propagates the trace context. enablesattributes token spend per step forHolds when spans record token counts and model identity per call. requiresfinds downstream consumers throughHolds when lineage records which downstream jobs and services read the model's outputs. part-ofis the final stage ofApplies when a retired model must stay reproducible or auditable after removal from service. enablespreserves evidence forHolds when evaluation records and data references are kept for the audit horizon. specializesis a kind ofApplies when KV blocks move between accelerator memory, host memory, and SSD by reuse likelihood. usesusesApplies in disaggregated clusters with RDMA fabrics. useshides latency withHolds when layer-wise KV transfer can proceed while later layers still compute. requiresoptimizes against an end-to-endHolds because the optimizer can only improve what the declared metric scores. reducesreducesHolds relative to one dedicated deployment per fine-tuned model. usesoften uses stepApplies to progressive and consistency distillation, not to training-free fast samplers. reducesreduces per-sampleHolds roughly in proportion to the number of network evaluations removed. increasesincreases cached-activationHolds because cached features from earlier steps must stay resident. reducesreducesHolds when adjacent-step features change little, so skipped blocks lose little quality. reducescompresses vectors to reduceHolds with recall loss that grows as code size shrinks. usesextends traversal ofApplies to predicate-agnostic methods that densify a proximity graph to keep filtered recall. limitsbounds recall on recently changed records inHolds when inserts and deletes reach the index later than the source of truth. usescan grade trajectories withApplies where task success lacks a programmatic check. enablesmakes repeatableHolds when tool responses and model outputs are recorded so a run can be re-executed exactly. limitsbounds the data-parallel speedup ofHolds because beyond it more examples per step no longer cut the steps needed to reach the target loss. reducesreduces the serving memory footprint by producing a smaller studentWhen the student architecture is smaller than the teacher and reaches acceptable quality on the target metric. enablesproduces well-aligned draft models forWhen the draft is distilled from the target model so its token distribution matches closely enough to raise acceptance rates. reducesreducesWhen the pruned weights are stored in a compressed or smaller dense form rather than as explicit zeros. reducesreduces dense-hardwareWhen whole channels, filters, or heads are removed so the remaining layers are smaller dense operations. reducesreducesWhen the chosen rank is much smaller than the matrix dimensions and the factored model keeps acceptable accuracy. alternative-tois an alternative toQAT recovers more accuracy at low bit widths but needs a training run; PTQ needs only a calibration set. enablesenables integer-only execution onWhen the simulated quantization matches the accelerator's integer arithmetic and requantization rules. determinessearches for points onWhen the search objective combines quality and a stated cost measure on the target platform. usesuses latency predictors fromWhen on-device measurement of every candidate is too slow, so a calibrated cost model ranks candidates. enablesenablesWhen the sparsity pattern is chosen to match a sparse matrix-engine instruction, as with 2:4 sparsity on sparse tensor cores. reducesreducesWhen small-batch decoding is memory-bandwidth-bound and dequantization is fused into the matrix multiply. mitigatessidestepsBecause activations stay in higher precision, outlier channels never enter a low-bit activation range. producesdevelops, at scale,In large transformers, roughly beyond a few billion parameters, where a few hidden dimensions carry systematic large-magnitude activations. reducesreducesWhen long-context decoding time is dominated by reading the KV cache from memory. enablesraises the feasible batch size ofWhen serving is KV-memory-bound, so freed capacity admits more concurrent sequences. reducesreducesFor a fixed model, data, and quality target, since lost and redone work lengthens wall-clock time. determinesdetermines the admissibleOffline scenarios permit large batches; server and single-stream scenarios limit batching by their latency constraint. usesspecifies the queryIn server scenarios, where queries arrive by a stated random process at a target rate. reduceslowers measuredBecause recomputed FLOPs are excluded from MFU's model-required count, so MFU falls while hardware FLOPs utilization rises. detectsidentifies the operatingWhen per-kernel time, FLOPs, and bytes are measured, so each kernel can be placed relative to the ridge point. reducesreducesWhen preempted work since the last checkpoint is lost and restart costs are paid on resumption. usesusesFor energy or power benchmarks, which must declare sensor class, sampling, and system boundary to be comparable. enablesenablesA sharded checkpoint can be remapped to a new layout only when all shards were saved from the same training step. requiresrequiresWhen every rank snapshots its shard at the same step boundary before background persistence begins. increasesincreasesWhen the blocking snapshot is much cheaper than the durable write, so training stalls less per checkpoint. enablesenablesWhen an open breaker routes requests to a fallback response rather than returning errors. reducescuts the request rate reaching a failing dependency, lowering itsWhile the breaker is open, which gives the overloaded dependency time to recover. requiresrequiresWhen a failed-over primary may still be alive, so stale writers must be fenced to avoid split brain. increasesincreasesRedundant replicas add capital and operating cost in proportion to the replication factor. limitsboundsFor large synchronous jobs, where job MTBF shrinks with component count and each failure costs rework plus restart. detectsdetectsWhen traces record the latency and errors that callers actually experience rather than health-check results. determinesdeterminesComponents behind a shared switch, rack, or power feed fail together, so the topology fixes the failure domains. usescarries checksums throughFor linear-algebra kernels whose input checksums predict output checksums. increasesadds checksum overhead toThe overhead is small relative to large matrix multiplications but nonzero. measurestests whether alerts fire and reach responders throughWhen the exercise uses real paging and runbook paths. specializesis a kind ofWhen the injected faults are bit flips or stuck-at errors in model state rather than service-level faults. reducesreducesWhen each stage exposes a small fraction of traffic and gates promotion on health signals. reducesreducesWhen a compromised or faulty component holds only the permissions its task requires. increasesincreases exposure toWhen pretrained weights come from untrusted parties, since backdoors can survive fine-tuning on clean data. mitigatespartially mitigatesInput and output classifiers catch known patterns but can be evaded by novel or obfuscated prompts. mitigatespartially mitigatesSafety training reduces disallowed outputs but fails under competing objectives and mismatched generalization. increasesincreasesEach training step runs an inner attack, multiplying compute by roughly the number of attack steps. limitscan strip parameter-embeddedFor watermarks embedded in weights rather than in behavior on trigger inputs, since a distilled student has new weights. mitigatesmitigatesWhen consumers verify signatures and provenance against trusted identities before loading artifacts. usesusesWhen the registry signs artifacts on promotion and verifies them before deployment. mitigatescontains the effects ofWhen untrusted model files are loaded inside a sandbox with no credentials or network access. reduceslowers the advantage ofAttack advantage grows with the train-test generalization gap, so better-generalizing models leak less membership signal. mitigatesmitigatesFor reconstruction of individual training records; inference of population-level attributes is not bounded by DP. mitigatesblocks gradient-basedWhen the server could otherwise reconstruct a client's data from its individual update. requiresrequiresWhen deletion obligations extend to models trained on the deleted records. detectsdetectsFor platform firmware or software whose measured state differs from trusted reference values. increasesincreases exposure toWhen tenants share accelerators, caches, memory, or interconnect whose contention is observable. causescan causeWhen observed timing, power, or memory behavior correlates with secret inputs or model parameters. increasesmultipliesHomomorphic evaluation costs orders of magnitude more compute than plaintext inference, especially for nonlinear layers. increasesincreasesLatency is dominated by communication rounds and garbled-circuit evaluation of nonlinear layers. increasesmultiplies IT energy intoFacility overhead for cooling and power conversion scales IT energy by PUE before carbon intensity applies. determinesclassifies purchased-electricity emissions (Scope 2) inFor an operator buying grid electricity under the GHG Protocol. measuresmeasuresWhen the assessment's inventory includes on-site cooling and off-site generation water. measuresaccounts end-of-life impacts ofWhen the system boundary includes disposal and recycling stages. reducesreducesWhen workloads tolerate preemption through checkpointing or elasticity so the discount is not lost to rework. reduceslowers the averageWhen a router sends requests a cheaper model can handle to that model while meeting the quality target. requiresrequiresWhen shared infrastructure must be charged back to the teams or workloads that consume it. usesallocates shared idle cost usingWhen provisioned but idle capacity must be charged to owners by an explicit rule. determinesdeterminesAt low load, idle power dominates, so poorly energy-proportional hardware loses efficiency. usesmodels the system asSTPA represents the system as controllers, control actions, and feedback paths. detectsdetectsWhen hazards arise from unsafe control actions or flawed process models without any component fault. usesenumerates component failure modes fromFMEA is bottom-up over declared component failure modes. requiresrequiresISO 26262 requires hazard analysis and risk assessment before safety goals are set. usesusesSafety goals are assigned an integrity level, such as an ASIL, that sets development rigor. usesusesWhen component failure modes must be traced to effects on safety goals. usesusesFor predictive safety filters that solve a constrained finite-horizon problem to find the nearest safe action. requiresrequires a model ofThe barrier condition holds only under the modeled dynamics and bounded disturbance. requiresrequiresThe verified safety controller and decision module must keep their timing even when the complex controller overruns. specializesis a kind ofFor physical systems whose degraded mode is a predefined reduced-capability safe state. alternative-tois an alternative toWhen the system can no longer operate safely, it can either hand control to a human or reach a minimal-risk state itself. determinesscopes the scenario catalog ofScenarios and parameter ranges are drawn from the declared operational design domain. usesdraws coverage evidence fromWhen scenario coverage is traced to hazards and operating-domain conditions. enablesenablesFailure regions found by falsification can seed importance sampling for failure-probability estimates. alternative-tois an alternative toFor systems too complex for formal proof, where searching for counterexamples replaces proving their absence. alternative-tois a certified alternative to empiricalFor properties over input regions small enough for the verifier to scale. limitsbounds failure rates withWhen the number of observed failures is small, so confidence bounds dominate the estimate. measuresmeasuresWhen the real compute and hardware run in the loop, exposing timing that pure simulation omits. specializesis a kind ofEach rung tests the assembled software or hardware against a simulated or recorded environment. enablesenablesWhen auditors review the assurance argument, not only its evidence artifacts. detectsdetectsWhen indicators count field events, such as near misses or perception misses, inside the operating domain. alternative-tois an alternative toRollback restores software or configuration state; compensation remedies effects when such a reversal is unavailable. useschecks the current state againstWhen the fallback is a stop and the monitor runs fast enough to trigger it before the state leaves the envelope. measuresmeasures the detection coverage ofWhen injected faults hit the operations the checksums protect, at realistic rates and locations. specializesis a kind ofWhen the fairness criterion is stated over pairs of similar individuals rather than groups. limitsconstrainsWhen a task-specific similarity metric is declared and accepted by stakeholders. limitslimits per-groupWhen base rates differ and error-rate parity is also required. limitslimits which criteria can jointly hold inWhen base rates differ across groups and the predictor is imperfect. implementsimplementsWhen the intervention targets a declared criterion and is re-evaluated on held-out data. usesuses reweighting and resampling fromFor pre-processing methods that change the training distribution across groups. detectsdetectsWhen a feature unavailable at prediction time dominates attributions. alternative-tois an alternative toBoth explain one decision post hoc; counterfactuals give actionable changes, attributions give per-feature contributions. enablesenablesWhen affected people or reviewers can read the decision logic that produced their outcome. enablesenablesWhen auditors can inspect the model directly instead of relying on post-hoc approximations. usesstores and versionsWhen documentation is required for promotion alongside the model artifact. usesdraws evidence fromDatasheets document data provenance and composition but do not replace access to the data. limitscan reduce to rubber stampsWhen reviewers see many correct proposals, weak error cues, and time pressure. limitscan degradeWhen operators lose vigilance and manual skill during long automated stretches. usesusesWhen red-team findings feed the measure and manage functions of the risk process. requiresrequiresFor deployers of high-risk systems in categories where a fundamental-rights impact assessment is required. determinesidentifies the affected groups and criteria forWhen the assessment precedes deployment and is revisited after material changes. usesusesWhen the agent expands and scores alternative partial plans rather than committing to one chain. measuresmeasures plan quality throughWhen intermediate steps, not only final outcomes, are scored across the trajectory. limitsbounds the search depth ofWhen turn, token, or time limits cap how many alternatives a plan can explore. determinesdeclares whether retries are safe throughWhen the contract states an operation's idempotency key or its at-most-once semantics. requiresrequires calibratedDetections and depth are only as accurate as the intrinsic, extrinsic, and timing calibration of the sensors. increasesincreasesWhen an added modality measures a state the others cannot distinguish, such as scale for monocular vision. limitsbounds drift correction inAbsolute pose and heading are unobservable from relative measurements alone, so drift persists without loop closure or external references. limitscan widen the gap inContact forces depend on stiffness, friction, and geometry that simulators approximate poorly. increaseswidens the surface forWhen tool descriptions and results from third-party servers enter the model's context as if trusted. usesusesWhen servers require scoped, delegated credentials before exposing tools or resources. measuresmeasuresWhen tasks run in real applications and success is checked from the resulting system state. usesdispatches tool actions throughWhen the model emits structured calls that the harness validates and executes. usesenforces stopping rules fromThe harness, not the model, applies turn, token, time, and spend limits. limitsbounds the length ofWhen limits on turns, tokens, or wall-clock time stop the run before completion. specializesis a kind ofWhen each step's intent and result are logged so the workflow resumes after a crash without redoing completed steps. requiresrequires recorded model outputs and tool results fromBecause re-sampling the model or re-calling tools would not reproduce the original trajectory. enablesenablesWhen a failing trajectory can be re-run step by step with its original inputs. usespersists waiting state throughWhen a trajectory must survive process restarts while waiting on a human or external event. part-ofis part ofWhen several agents read and write task state under one orchestration topology. causescan causeWhen concurrent writes or stale reads go uncontrolled, so agents act on conflicting views of the task. detectsdetectsWhen handoffs, delegated instructions, and per-agent outputs are recorded as linked spans. limitslimits the benefit ofWhen the sampled answers or voting agents share a base model and prompt, so their errors are not independent. limitsweakens the independence ofWhen the judge shares a base model with the system it grades. causescan causeWhen perception limits, such as missed detections in rare conditions, cause hazards without any fault. requiresmakes measurableObservation age can be computed only when capture and use timestamps share a synchronized clock. usespredicts state with a model ofThe prediction step propagates the estimate through modeled dynamics between measurements. usesusesWhen fused measurements arrive asynchronously and are combined through prediction and correction steps. enablesenables sample-efficientWhen policies are trained on imagined rollouts from the learned model rather than only on environment interaction. requiresrequiresPlans are valid only relative to the estimated pose and map they start from. usesplans over maps fromWhen the environment is not known in advance and the map is built online. usesusesSchedulability is analyzed with criticality-dependent WCET estimates for each task. limitsboundsRate-monotonic scheduling guarantees deadlines only when processor utilization stays below the Liu-Layland bound. usestransfers web-scale vision-language knowledge throughWhen the policy is fine-tuned from a pretrained vision-language backbone. usesis fine-tuned onWhen robot action labels come from human-driven demonstrations. alternative-tois an alternative toSimulated demonstrations are cheaper to scale but carry a sim-to-real gap that teleoperated demonstrations avoid. increasesincreases, for actions late in a chunk,When chunks execute open loop, so later actions rest on older observations. usesreduces compounding error withWhen the policy predicts short action sequences rather than single steps. specializesis a kind ofRandomized simulator parameters act as label-preserving transformations of training scenes. mitigatesmitigatesWhen the randomized ranges cover the real-world parameters without making the task unlearnable. limitsconstrains the commanded actions ofWhen the filter runs on every action at control rate with authority over the learned policy. usesis trained to predict future observations byWhen the model learns from action-conditioned sequences without task labels. part-ofis part ofSensing establishes how observations relate to time as well as to physical quantities and coordinates. specializesis a kind ofDiffusion models are neural networks trained to reverse a gradual noising process over many denoising steps. reduceslowers by computing where weights are storedWeight-stationary matrix-vector workloads whose weights fit in the memory arrays. limitscaps the effective bit width ofAnalog implementations, where device noise, drift, and converter resolution bound precision. enablesprovides programmable analog cells forDevice conductance can be programmed to represent weight values. alternative-tocompetes for on-chip weight storage withDensity and standby power outweigh write endurance and write latency. producesbuildsStaged or traced execution rather than immediate evaluation. enablesexposes whole programs toThe DSL restricts host-language features that cannot be staged. specializesis a kernel-level case ofKernel DSLs embedded in a host language requiresrelies onDerivatives must propagate through control flow and data structures. implementsrealizes through a composable grad transformationPrograms written as pure array functions. reducesremoves per-operator launch overhead fromSmall-batch, latency-bound workloads made of many short kernels. alternative-tois an alternative way to remove launch overhead toPersistent kernels also overlap work across operator boundaries inside one launch. specializesis an extreme form ofFusion extended across operator and layer boundaries into a single launch. part-ofis an optimization stage withinRewrites are verified equivalent before code generation. usessearches rewrites expressed inGraph- or kernel-level IR with defined operator semantics. enablesdiscovers new opportunities forSearch spans kernel and thread-block levels as well as the graph level. implementspackagesParallelism degrees are set by configuration and mapped to the device topology. usescommonly appliesOptimizer and parameter state exceed per-device memory. usescommonly appliesActivation memory limits sequence length or micro-batch size. determinesshifts the cost structure ofComparing rented infrastructure usesrelies on resource pooling throughProviders serve many customers from shared physical resources. usesstores table files inLakehouse designs over open file formats. enablesprovides snapshot time travel forTable formats that retain prior snapshots. usesstores data in columnar layouts such asOpen columnar formats such as Parquet. limitsremoves headroom assumed bySwings are coincident across accelerators, defeating non-coincident-peak assumptions. mitigatesbuffersStorage sized and controlled to respond at the swing frequency. mitigatessmoothsDevice power caps and floors applied to bound ramp rates. increasesraisesCarrying load through utility outages until generators take over. limitscaps the power available toSites whose deliverable grid power is below planned IT load. alternative-tosubstitutes for or supplementsInterconnection queues delay or cap grid supply. determinessets the emissions of supply forOn-site sources supply a material share of load. usesrepairs violations found byValidation rules or statistics flag the records to repair. mitigatesimputes or removesThe missingness mechanism is understood well enough to choose a treatment. increasesraisesRepairs are validated and do not introduce label leakage. mitigatescompensates forResampling or weighting is applied to training data only. determinesshapes the choice ofPer-class, precision-recall, or cost-weighted metrics replace accuracy. trades-offtrades off againstRebalanced training shifts predicted probabilities away from true base rates unless corrected. specializesis a technique withinCategorical inputs with large or open vocabularies. reducesboundsWeight or embedding tables indexed by hash bucket. specializesis a technique withinInteractions among categorical inputs. usesis often bucketed byCrosses of high-cardinality features whose joint vocabulary would explode. usesrepresents interactions throughInteractions not otherwise expressible by an additive model. determinessets policy forRetention periods follow purpose and legal obligations. requiresdepends on traceability fromAccountability for which data reached which model must be demonstrable. usesenforces access decisions throughAccess policies are encoded as enforceable permissions. part-ofis a policy element ofPersonal data is processed. limitsrestrictsPersonal data for which a lawful basis or consent is required. detectsreveals patterns ofProfiles are computed per feature and per slice. enablesinforms constraints forObserved statistics seed a schema or expectation set. usesplots metrics recorded byRuns log curves and artifacts in a queryable store. enablessupportsErrors are displayed alongside inputs and slices. usesestimates candidate performance withData is too small for a dedicated validation split. causescan inflate estimates throughSelection and final assessment reuse the same validation data. usescompares candidates againstThe baseline is evaluated under the same protocol. part-ofis a stage withinHyperparameters are chosen on validation data. usesoptimizes an objective fromUnlabeled corpora supply the training targets. enablessupplies base models forDownstream tasks share structure with the pretraining corpus. usesdraws its corpus fromCorpora combine multiple domains or sources. usescommonly optimizesDecoder-only language-model pretraining. specializesis a text case ofTargets are the next tokens of the input itself. alternative-tocompetes for sample generation withSingle-step sampling versus many denoising steps, with different stability and diversity. causesis prone toThe generator outpaces the discriminator during adversarial training. specializesis a diversity-loss form ofCollapse of output diversity rather than divergence of the loss. specializesis a likelihood-based case ofModels that assign token probabilities. determinessets the unit ofPer-token perplexities are comparable only under one tokenizer; bits per byte normalizes. measuresevaluates held-out performance ofHeld-out text is free of training contamination. specializesis a reference-based case ofTasks with canonical reference answers. alternative-tocompetes for scoring open outputs withReproducibility and cost weighed against tolerance of paraphrase. usesgroups mistakes withError categories are not known in advance. detectsrevealsGradient and loss statistics are logged at step granularity. usesreads plots fromLearning curves and activation statistics are recorded during training. usesjoins outputs to inputs withDistributed batch jobs return outputs out of order. usesjoins predictions to later outcomes withOutcomes arrive after the prediction and must be matched. enableslets replicas be added or removed forNo request depends on replica-local state. usesexposes its interface asInputs and outputs are declared at export. usesdirects requests bySeveral models or providers serve the same application. enablesrecords usage forAll model calls pass through the gateway. usescan serve repeated requests fromResponses are reusable across users and sessions. usesfinds similar past requests withRequests are embedded and indexed. reducesavoids model calls that driveWorkloads with many near-duplicate requests. specializesis a form ofGeneral-purpose packages are wired in without a common interface. mitigatesreduces the need forThe platform wraps packages behind common APIs. specializesis a form ofData preparation evolves without holistic redesign. requiresraises the cost ofFailures can only be detected by expensive end-to-end tests. determinesshapes the composition ofPlatform components are chosen one capability at a time. usescompares options byEngineering, licensing, and operating costs are counted over the same horizon. usesattaches signals to records inFeedback is joined to the prediction that elicited it. enablessupplies labels forFeedback is filtered for noise and position or exposure bias. causescan driveFeedback is gathered only on outputs the model chose to show. mitigatescountersInterfaces convey uncertainty and make errors visible. enablesprovides interfaces forUsers can see why an output was produced and correct or override it. usessheds load throughCaps are applied in response to grid signals. usesdefers flexible work throughInterruptible training or batch jobs checkpoint before suspension. alternative-tocompetes for task adaptation withIteration speed and no training cost weighed against reliability, context cost, and control. useselicits behavior throughThe model follows instructions or demonstrations in its input. implementsautomatesAn end-to-end metric and training examples are available. limitscaps the demonstrations available toDemonstrations and instructions share one input budget. enablesgives rise toSufficient model and data scale. part-ofis the input stage ofCascaded designs that transcribe speech before the language model. usesruns inOn-device recognizers for privacy and latency. limitsbounds the stages ofTurn latency must stay within conversational expectations. reducesreduces the KV portion ofGeneration is long enough that cached keys and values dominate accelerator memory. alternative-tocompetes as a KV-compression option withBoth target KV memory; they can also be combined. usesselects entries using scores fromImportance is estimated from accumulated or recent attention weights. specializesis a runtime-sparsified form ofOnly predicted-important query-key blocks are computed. alternative-toreplaces fixed sparsity patterns ofSparsity is chosen per input rather than fixed in advance. reducesshortens long-promptPrefill attention over long contexts dominates first-token latency. limitscaps the transfer rate available toActivations or KV state cross the host link at each step. alternative-tocomputes on the host instead of moving state back as inHost memory holds state that exceeds accelerator capacity. usesassigns unequal stages throughStage sizes are matched to each device type's compute and memory. usessearches placements and plans withDevice types differ in throughput, memory, and interconnect. specializesis a graph-structured form ofComputation follows the input graph's edges. usesaggregates neighbor features withGraph adjacency is stored sparsely. usestrains on large graphs withFull multi-hop neighborhoods exceed device memory. reducesbounds the per-batchFanout per hop is capped. increasesadds sampling variance to gradient estimates and thusSampled neighborhoods approximate full aggregation. usesmoves weight shards withLayouts differ between phases. implementsrealizesTraining and generation share the same devices. usesremaps shards acrossEach phase uses a different mesh layout. specializesis an input-dependent form ofThe zero pattern changes per token. alternative-toskips work per input rather than permanently as inStatic pruning removes the same weights for every input. reducesshortensDecode is memory-bound and skipped weights need not be read. alternative-togeneralizes beyond the exact-prefix matching ofShared chunks appear at different positions across requests. reducesshortensReused chunks such as retrieved documents dominate prompt length. usesresizes jobs throughJobs tolerate changes in worker count. usesco-allocates workers withDistributed jobs need all workers simultaneously. increasesraises cluster-wideJobs' scaling behavior is measured or predicted. reducesshortens collectiveFew large synchronized flows collide under static hashing. usesspreads traffic over the multiple paths ofThe fabric offers path diversity between endpoints. causescan triggerMany ranks send to the same receiver simultaneously. mitigatesthrottles senders duringSenders react to congestion signals before buffers overflow. increasesraisesBuffers fill and packets queue or drop. determinessets the token budget forCandidate content exceeds the input budget. usesgathers candidates withRelevant items are fetched on demand rather than preloaded. mitigateslimits distractor content that drivesIrrelevant tokens otherwise dilute attention. causescan retain facts that produceStored observations are not revalidated after the environment changes. measuresquantifies one dimension ofObservations carry acquisition timestamps. causescan lead toThe agent reports or acts on state it observed before later changes. specializesis an iterative form ofLater queries depend on earlier results. usesanswers multi-step questions withNo single query retrieves all needed evidence. increasesadds retrieval rounds toEach hop requires another query and model call. specializesis a reference-based form ofRights travel with unforgeable tokens. implementsrealizesDelegated capabilities are attenuated to the task's needs. implementsprovides per-taskEach task or tenant receives a fresh environment. mitigatesconfines local damage that would otherwise becomeEffects stay inside the sandbox rather than reaching external systems. implementsapplies to network accessOnly destinations the task needs are allowed. mitigatesblocks exfiltration channels forLeakage would otherwise travel over outbound connections. usesrestricts outbound network access withThe sandbox has network access. mitigatesprevents untrusted data from steering actions inData from tools and documents carries low-integrity labels. mitigatesblocks labeled secrets fromConfidential data is tagged at its source. causesis a self-induced form ofThe policy's own outputs determine its future inputs. causescan sufferTraining sees only expert-visited states. mitigatesshortens the effective horizon ofActions are predicted in multi-step chunks. specializesis an interactive form ofAn expert can label states the learner visits. mitigatestrains on learner-visited states to counterExpert corrections are available during data collection. usescollects corrections asHuman operators supply corrective actions. part-ofis a configuration step inExamples contain non-target tokens such as prompts or tool output. determinesdefines which tokens contribute toSequences and turns vary in length. part-ofis a core problem withinRewards are delayed or episode-level. implementsprovides step-level signals forIntermediate steps can be judged. determinessets which motions satisfyPlanned motions demand force or speed near the actuator's capability. usesencodes as input constraintsLimits are known and modeled. part-ofis a liveness mechanism withinThe monitored component can hang or stall. usestriggers on expiryA safe fallback state is defined. alternative-toexpires authority by time likeAuthority must lapse when renewal stops. detectsrevealsErrors alter later observations, unlike open-loop replay. implementsprovides simulated loops forThe environment model reacts to the system's actions. enablessupplies reconstruction evidence forRecords cover the interval before and after the event. usesis computed from events logged byField events are recorded consistently. enablesprovides traceable decisions forRecords name the approver and evidence. enablesidentifies whom to challenge inAffected people can see who decided. producesrecords approvals asEach approval is logged with approver identity. part-ofis an intervention mechanism withinOperators can act during execution. usesundoes partial work withCancelled work already produced side effects. part-ofis a progress monitor withinThe harness can inspect the trajectory between steps. detectsflags looping instances ofAgents repeat steps or hand work back and forth. alternative-tostops stuck runs earlier thanProgress can be judged before a budget is exhausted. useschecks each symbolic step withActions have geometric preconditions. requiresaccepts plans only withSymbolic plans can be geometrically infeasible. specializesis a multi-target form ofSeveral objects are observed with unknown correspondence. usespropagates each track withObject motion is modeled between frames. requirestraces affected claims throughClaims and evidence are explicitly linked. determinessets the re-verification scope forA model, software, data, or ODD change is proposed. usesturns unstructured zeros into speedup forSparsity is high enough to overcome irregular access. usesstores nonzeros in compressed formats described byNonzeros are kept in formats such as CSR or coordinate lists. specializesis a periodically synchronized form ofWorkers hold full model replicas. alternative-toreduces communication by frequency rather than size as inBandwidth between workers is scarce. specializesis a federated instance ofClients run several local epochs before averaging. part-ofis a placement decision withinExperts are distributed across devices. mitigatesrelieves overloaded devices that becomeRouting concentrates tokens on a few experts. alternative-tobalances load in the system rather than the model as inRouting remains skewed despite training-time balancing. specializesis a reduced-state form ofFull adaptive state exceeds memory. reducesshrinks the optimizer-state share ofOptimizer state dominates training memory. alternative-toshrinks state rather than distributing it as inFew devices are available to shard across. producesrecordsOperator dependencies are captured along with timings. usesreplays workloads fromTraces are in a standardized format. specializesis a fixed-dataset form ofNo further environment interaction is allowed. alternative-tocan improve on demonstrations unlikeLogged data carries reward labels. limitsbounds what can be learned inThe learned policy visits states the dataset does not cover. mitigatesbounds the influence ofOnly a minority of participants are malicious. usescombines client updates withClients are untrusted. usesplaces tasks withMachines have multidimensional resource capacity. usesenforcesMultiple teams share the cluster. useshonors priorities throughHigher-priority work arrives while capacity is full. causescan triggerDemand for the service is price-elastic. increasesraises totalDemand growth outpaces per-unit savings. mitigatescatches early defects that startChecks run at ingestion and between stages. causescan propagateDownstream consumers depend on undeclared upstream signals. enablesranks candidates forValue estimates are cheaper than retraining on each subset. measuresestimates the contribution side ofA utility metric and baseline model are fixed. usesgenerates actions withAction distributions are multimodal. usespredicts action sequences as inActions are generated over a short horizon. specializesis a physical-channel form ofThe attacker controls signals reaching the sensor. mitigatescross-checks modalities againstThe attacker cannot spoof all modalities consistently. reducescancels shared item variance fromBoth systems are scored on the same items. usestests candidates againstA noninferiority margin is declared in advance. usesdecides promotion withCandidate and incumbent are evaluated on the same set. alternative-tois a competing execution target toThe workload is small, sparse, or latency-bound enough that host execution is viable. usesrelies on caches and DRAM levels ofPerformance depends on working-set reuse in the cache levels. usesoffloads selected operators toModel or KV state exceeds accelerator memory. usesmaps operator work onto the cores and caches ofInference or preprocessing runs on multicore hosts. determinessets achieved core utilization throughThread counts and affinity are configured relative to available cores and NUMA domains. part-ofis a concern withinInput dimensions vary across calls to a compiled function. limitsrestricts reuse ofCaptured graphs are valid only for the shapes recorded at capture time. part-ofis a pass family withinThe compiler applies pattern-based passes before lowering. alternative-tois the hand-written counterpart toBoth seek semantics-preserving graph transformations; one uses fixed rules, the other searched rewrites. enablesexposes fusible patterns forCanonicalization places compatible operators adjacent in the graph. specializesis a compressed form ofStorage records only nonzero entries or actual sequence extents. reducesshrinksDensity is low or lengths vary widely, so padding would dominate storage. usesoperates on operands stored inKernels read compressed index structures rather than dense arrays. usespartitions graphs expressed inThe runtime loads a framework-neutral operator graph. usesdelegates subgraphs throughThe target device exposes an NPU, GPU, or DSP backend with partial operator coverage. increasesadds host orchestration throughOperator fallback splits the graph and forces cross-backend copies. usesis usually built onMost current language models use decoder-only Transformer stacks. producesyields the base weights ofTraining uses broad corpora and a self-supervised language-modeling objective. specializesis a large-scale instance ofThe model is a parameterized composition of differentiable operations. measuressizesDense attention with known layers, KV heads, head dimension, context, and precision. usessizes accelerators fromPlanning precedes measurement on the target system. specializesis the per-token inference case ofThe model is a dense transformer served autoregressively. part-ofis a design choice withinReplicas train on different examples and must coordinate updates. enablesmakes asynchronous updates practical forWorkers push and pull parameters without a global barrier. part-ofis a stage withinBatches are prepared ahead of consumption. increasesraises accelerator utilization by hiding input latency behindInput preparation would otherwise stall compute. part-ofis a mechanism withinEvents carry timestamps and can arrive out of order. determinessets the staleness of streaming features throughWindow size and allowed lateness bound when an aggregate is emitted. usesrelies on flow control fromClients consume output more slowly than the server generates it. measuresquantifies the delivery cadence ofOutput is returned token by token after the first. usespackages models asAny instance must be able to serve any request. limitsbounds the latency attainable byReplicas scale from zero and must load weights before serving. usesreorders requests insideScheduling decisions occur at each generation iteration. determinesshapesLong and short generations share capacity under load. implementspackagesThe image pins packages, native libraries, and runtime configuration. enablesmakes reruns practical forImages are versioned and run unchanged across machines. detectsidentifiesA slow or stuck rank stalls collectives without crashing. usesflags stalls withCollectives and ranks are expected to make progress within a timeout. requiresdepends on an accurateThe benchmark is used to predict deployed behavior. determinessets how closely measured results match deployment forArrival pattern and latency constraint are chosen to mirror the target service. limitsrestricts which sources can enterDatasets or weights carry use, redistribution, or field-of-use terms. usesenforces terms recorded inAccess and use decisions must respect license obligations. usesrelies on cryptographic signatures likeProvenance manifests are bound to content by signatures. requirescan mandateRegulation requires machine-readable marking of synthetic content. usesselects uploads withTriggers prioritize uncertain or rare events for labeling. mitigatescountersMined data covers conditions the deployed fleet now encounters. determinescontributes transport and queuing delay toSensing, inference, and control run as separate communicating nodes. usesreceives synchronized sensor streams throughSensors publish on separate topics consumed by a fusion node. usesgathers experience fromTraining runs in simulation before physical deployment. usessamples randomized environments fromSimulator parameters can be varied per environment instance. implementsrealizesCredentials are scoped to the task and expire after it. usesissues attenuated rights throughDelegated tokens carry narrower scope than the user holds. usesscales training across workers withThe graph and its features exceed one device or machine. usesdraws mini-batches acrossDistributed training samples neighborhoods that span partitions. part-ofis the single-machine stage ofExamples are decoded and collated on the training host. determinesbounds accelerator utilization throughPer-batch preprocessing time approaches the accelerator step time. part-ofoverlaps loading with compute inWorker processes prepare batches ahead of the training step. specializesis streaming inference over audioAudio arrives in chunks and partial transcripts are emitted before the utterance ends. usesrelies onInput or output rates can outpace the other side of the stream. specializesstreams outputThe request is complete but output is delivered incrementally. part-ofis the request interface ofClients call the model over a network protocol. increasesadds serialization time toPayloads are large part-ofis the horizontal axis ofLatency is measured across offered loads. limitsis an upper bound onGoodput counts only completed requests that meet their latency targets. measuresquantifiesHigh percentiles such as p99 are reported. implementsis a common form ofThe SLO is stated as a percentile under a threshold. usesincreases hit rates ofRequests share prompt prefixes or belong to multi-turn sessions. trades-offtrades cache reuse againstPopular prefixes concentrate traffic on a few replicas. enablesestablishes the communicators used byEvery rank must join before the first collective. usesre-runsWorkers join or leave and groups are rebuilt. increasesgrows the size and reuse ofEach step resends a growing shared context to the same model. measuresquantifies the demands ofThe workload is a multi-step agent trajectory rather than independent requests. usesrepeatedly runsEach change is benchmarked against a stored baseline. part-ofis a gate inPerformance thresholds can block a merge or release. limitslimits the precision ofDifferences between systems are comparable to run-to-run spread. limitsconfoundsAsynchronous kernels are timed without synchronization or warm-up. mitigatesreduces the risk ofIdentifiers are removed before training or logging. requiresrequiresData contains personal information subject to regulation. mitigatessuppressesInstability arises from occasional very large gradients. part-ofis applied in the update step ofGradients are clipped before the optimizer step. mitigatesprevents earlyNetworks are deep and use rectifier or residual blocks. determinesprescribes width-dependent scales forHyperparameters tuned on a small model are reused at a larger width. usesachieves effectively-once results withRecords may be redelivered after failures. part-ofis a guarantee ofThe pipeline retries or replays records after failures. mitigatespreventsOnline and offline stores are populated by separate jobs. trades-offtrades againstFeatures are updated faster than they can be materialized consistently. detectsflags individual inputs affected byInputs fall outside the training distribution. enablessupplies abstention signals toThe system may defer uncertain inputs. limitsconstrainsModel limitscan breakA schema change is not mirrored in the model's feature transformations. implementsenforcesCredentials are scoped per workload and rotated. usesdraws onAgents act with credentials issued by the platform. requiresextends to models throughDeleted data has already been used for training. useslocates copies throughDerived datasets and features record their sources. causescan causeTensor sizes vary across steps or requests. increasesinflates reservedFreed blocks cannot be reused for new allocation sizes. part-ofis the sensing share ofPerception runs inside the control loop. limitsbounds model size and resolution forPerception runs on fixed onboard hardware. usesrespectsSafety-critical and best-effort tasks share hardware. specializesis execution placement on a robotWorkloads can run on several onboard processors or be offloaded. enablestells which replicas may receive traffic fromReplicas can be unhealthy or still loading weights. mitigateskeeps traffic away from replicas inA replica reports ready only after its model is loaded and warmed. increasesserializes and adds toSeveral Python threads issue work concurrently. limitspushes preprocessing into worker processes inPreprocessing runs in Python limitsbounds the frequency ofUpdates and logs are written to on-device flash. limitslimits the write budget ofThe device stores weights and data in flash. specializesis a power regime ofThe device has no battery and runs on harvested energy. requiresrequires fine-grainedPower can fail mid-inference. reducesremoves duplicate copies fromSeveral processes serve the same read-only weights. reducesshortensOnly the pages touched by the first requests must load. limitsenforcesThe OS kills or pages out apps that exceed their memory limit. reducesmakes weight pages reclaimable underFile-backed pages can be dropped and reread instead of killing the app. limitsadds copy and pause overhead toThe app calls native inference from a managed language. increasesadds garbage-collection pauses toLarge buffers are allocated on the managed heap per request. alternative-toreplaces CPU copying inThe sensor or peripheral can write into a buffer the model reads. reducesremoves copy time fromSensor data feeds a latency-critical loop. limitscorrupts measurements inSeveral active sensors emit into the same scene. mitigatescan outvote corrupted readings fromIndependent modalities observe the same scene. limitssets a floor underRequests cross regions or continents. part-ofcontributes to the latency term ofMessages travel long physical distances. limitsforces a minority partition to stop answering inA partition separates replicas. determinesdecides whether stale values are servedFeature or model replicas span regions that can partition. alternative-toremoves the central server fromParticipants can exchange updates with peers. alternative-toreplaces global collectives with neighbor exchangeGlobal synchronization is too slow or unreliable. part-ofis part of the timing budget alongsideInference and interrupt handlers share a processor. increasesadds toSensor events are delivered through interrupts. specializesare arrangements ofModels are built from Transformer blocks. determinesdecides whether serving has separateThe model generates output autoregressively. increasesadds toData part-ofis a component ofWorkloads span zones usesexploits the gap betweenSleep power is far below active power. reducescuts averageSensing and inference are needed only intermittently. limitsbreaks the guarantees ofTasks of different priorities share a lock or resource. increasescan unboundedly extendNo priority inheritance protocol is in place. mitigatesprovably boundsPerturbations are within the certified radius. alternative-totrades inference cost for certified guarantees againstA provable bound is required rather than empirical robustness. limitsbounds the data rate ofSensors stream raw data into an embedded processor. usesmoves data arriving overThe peripheral supports direct memory access. specializesis queueing with bounded capacityThe buffer cannot grow without limit. mitigatesprevents overflow ofUpstream producers can slow down when the buffer fills. usesschedules calls intoThe app has a main thread that must stay responsive. reducesremoves inference time from interactiveInference takes longer than a frame interval.