{ "name": "ML Systems Taxonomy", "version": "0.4-draft", "status": "editorial-draft", "url": "https://mlsysbook.ai/taxonomy/", "license": { "name": "CC BY 4.0", "url": "https://creativecommons.org/licenses/by/4.0/" }, "citation": "Vijay Janapa Reddi. ML Systems Taxonomy, version 0.4-draft. Machine Learning Systems, 2026. https://mlsysbook.ai/taxonomy/", "note": "Editorial draft for expert review; not a claim of completeness or consensus.", "tagging_rules": [ "Tag an item with the concepts it centrally teaches or tests: prefer one, and add a second or third only when the item cannot be answered without it. Do not tag prerequisites or passing mentions.", "Tag a subject alone when the item covers the subject broadly or no concept is specific enough. A subject tag is a valid, complete tag.", "When no concept fits, tag the closest subject and record the missing idea as a proposed concept; proposals feed the next version.", "Store concept identities (mlsys.kv-cache) and subject identities, not codes; codes can be renumbered between versions and identities cannot.", "When a concept seems to belong to two subjects, use its home and follow the home rule and its worked examples; related subjects are navigation, not alternative tags.", "Do not tag a workload-class concept (Large language model, Transformer) only because the item involves one; tag it when the item is about that workload class.", "Tag the mechanism, not the platform. Cloud, edge, mobile, and TinyML are contexts. Use Edge & embedded hardware or On-device deployment as the subject only when the item is about the device class itself or about converting, packaging, or updating models for it, not when a device is merely where a memory, interconnect, or scheduling question arises.", "Tag a calculation by what it is about. An estimate of KV-cache size is about the KV cache. Use Performance models, Workload characterization, or Profiling as the subject only when the item teaches the modeling or measurement method itself.", "When an item asks for the effect of a technique (the energy saved by INT8, the throughput gained by batching), the technique is the subject and the affected quantity may be a second concept.", "Depth is recorded separately (for example, StaffML levels L1 to L6+). A competency is a concept or subject paired with a depth.", "When a concept is retired, its identity maps to the concept named in replaced_by. When a concept is split, the old identity is retired and existing tags are re-reviewed; until then they resolve to the replacement." ], "placement_examples": [ { "concept": "mlsys.kv-cache", "home": "llm-inference", "why": "Implements generative serving; it exists for speed but is part of how decoding works." }, { "concept": "mlsys.kv-cache-quantization", "home": "model-compression", "why": "Its only purpose is efficiency, and the same technique applies to weights and activations across layers." }, { "concept": "mlsys.speculative-decoding", "home": "llm-inference", "why": "Changes how the serving stage generates tokens." }, { "concept": "mlsys.activation-recompute", "home": "training-memory", "why": "Implements a training-stage memory plan." }, { "concept": "mlsys.carbon-aware-scheduling", "home": "energy-carbon", "why": "Its only purpose is lower carbon, and it applies to training, batch inference, and data jobs alike." }, { "concept": "mlsys.fusion", "home": "ml-compilers", "why": "Implements the compiler layer." }, { "concept": "mlsys.dp-sgd", "home": "privacy-dp", "why": "Its only purpose is privacy." }, { "concept": "mlsys.embedding-cache", "home": "storage-systems", "why": "A software tiering scheme that specializes storage tiering and serves a workload; the hardware memory hierarchy is where it spills, not what it implements." }, { "concept": "mlsys.irreversible-effects", "home": "failure-semantics", "why": "A general failure class that applies to tool actions and to actuation alike; it lives with the failure semantics it defeats, beside compensating actions and idempotency." }, { "concept": "mlsys.retrieval-metrics", "home": "metrics-uncertainty", "why": "The subject's own scope claims ranking and retrieval metrics; retrieval mechanisms stay in Retrieval & vector search." } ], "groups": [ { "code": "F", "id": "foundations", "label": "Foundations", "question": "What does the system compute, what must it achieve, and how do we reason about its cost?", "areas": [ { "code": "F1", "id": "workloads-requirements", "label": "Workloads & requirements", "question": "What must the system achieve, for which workload, and how is success decided?", "scope": "Targets, workload descriptions, and objectives that every later engineering choice is judged against.", "subjects": [ { "code": "F1.1", "id": "requirements-slos", "label": "Requirements, SLOs & baselines", "definition": "Stating measurable targets for a system's behavior and service, and the baselines and budgets against which they are judged.", "includes": [ "Acceptance criteria and baselines", "SLIs, SLOs, error budgets", "Latency budgets across a request path" ], "excludes": [ "Measured latency behavior and goodput ([[latency-goodput]])", "Statistical evaluation of model quality ([[evaluation]])" ], "contexts": [ "interactive", "batch" ], "concepts": [ "mlsys.requirements", "mlsys.baseline", "mlsys.slo", "mlsys.sli", "mlsys.error-budget", "mlsys.latency-budget", "mlsys.recovery-objectives", "mlsys.service-level-agreement" ], "related_concepts": [ "mlsys.tail-latency", "mlsys.slo-goodput" ], "taught_in": [ { "volume": "vol1", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol1/01_introduction/01_introduction.html" }, { "volume": "vol1", "chapter": "ML Lifecycle", "url": "https://mlsysbook.ai/vol1/03_ml_lifecycle/03_ml_lifecycle.html" }, { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html" }, { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" } ] }, { "code": "F1.2", "id": "workload-characterization", "label": "Workload characterization", "definition": "Describing a workload's compute, memory, data, and traffic demands, and where it runs, so it can be engineered and provisioned for.", "includes": [ "FLOP, byte, and memory accounting", "Deployment tiers from cloud to TinyML", "Workload traces" ], "excludes": [ "Node-level job placement ([[cluster-scheduling]])", "Measured profiles ([[profiling-utilization]])" ], "contexts": [ "edge", "distributed", "deep-learning" ], "concepts": [ "mlsys.workload", "mlsys.placement", "mlsys.flop-accounting", "mlsys.memory-footprint", "mlsys.workload-trace", "mlsys.inference-compute-estimate" ], "related_concepts": [ "mlsys.arrival-rate", "mlsys.capacity-planning" ], "taught_in": [ { "volume": "vol1", "chapter": "System Paradigms", "url": "https://mlsysbook.ai/vol1/02_system_paradigms/02_system_paradigms.html" }, { "volume": "vol1", "chapter": "Neural Computation", "url": "https://mlsysbook.ai/vol1/05_nn_computation/05_nn_computation.html" }, { "volume": "vol1", "chapter": "Network Architectures", "url": "https://mlsysbook.ai/vol1/06_nn_architectures/06_nn_architectures.html" }, { "volume": "vol2", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol2/01_introduction/01_introduction.html" } ] }, { "code": "F1.3", "id": "objectives-decisions", "label": "Objectives, proxies & decision policies", "definition": "How learned outputs become decisions, how optimized proxies relate to intended outcomes, and how competing objectives are traded.", "includes": [ "Proxy objectives and reward hacking", "Cost-sensitive thresholds", "Multi-objective tradeoffs" ], "excludes": [ "Fairness criteria ([[responsible-ai]])", "Optimizer and loss mechanics ([[training]])" ], "contexts": [ "classical-ml", "agentic" ], "concepts": [ "mlsys.proxy-objective", "mlsys.decision-policy", "mlsys.cost-sensitive-decision", "mlsys.pareto-frontier" ], "related_concepts": [ "mlsys.calibration", "mlsys.selective-prediction", "mlsys.reward-hacking" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Lifecycle", "url": "https://mlsysbook.ai/vol1/03_ml_lifecycle/03_ml_lifecycle.html" }, { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html" }, { "volume": "vol3", "chapter": "Reinforcement Learning from Verifiable Rewards", "url": "https://mlsysbook.ai/vol3/17_rlvr/17_rlvr.html" } ] } ] }, { "code": "F2", "id": "models-architectures", "label": "Models & architectures", "question": "What does the model compute, and how does its structure set system cost?", "scope": "Model families, building blocks, and architectural choices viewed as workloads, plus the artifacts that carry them.", "subjects": [ { "code": "F2.1", "id": "model-families", "label": "Model families", "definition": "The classical and neural model regimes, with their inductive assumptions, viewed as workloads with characteristic compute and memory profiles.", "includes": [ "Classical models and neural networks", "Transformers, recommendation models, diffusion models" ], "excludes": [ "Learning paradigms ([[training]])", "Layer-level building blocks ([[architecture-blocks]])" ], "contexts": [ "classical-ml", "deep-learning", "generative", "recommendation", "multimodal" ], "concepts": [ "mlsys.linear-model", "mlsys.tree-ensemble", "mlsys.kernel-method", "mlsys.nearest-neighbor", "mlsys.neural-network", "mlsys.inductive-bias", "mlsys.transformer", "mlsys.recommendation-model", "mlsys.diffusion-model", "mlsys.multimodal-model", "mlsys.gan", "mlsys.graph-neural-network", "mlsys.large-language-model", "mlsys.transformer-variants" ], "related_concepts": [], "taught_in": [ { "volume": "vol1", "chapter": "Neural Computation", "url": "https://mlsysbook.ai/vol1/05_nn_computation/05_nn_computation.html" }, { "volume": "vol1", "chapter": "Network Architectures", "url": "https://mlsysbook.ai/vol1/06_nn_architectures/06_nn_architectures.html" } ] }, { "code": "F2.2", "id": "architecture-blocks", "label": "Architecture building blocks", "definition": "The layer-level operations from which neural models are composed and whose shapes determine compute, memory, and data movement.", "includes": [ "Embeddings, attention, position encoding, tokenization", "Convolution, recurrence, normalization" ], "excludes": [ "Kernel implementations of these operations ([[tensor-operators]])", "Efficient attention variants ([[efficient-sequence-models]])" ], "contexts": [ "deep-learning", "generative", "recommendation", "multimodal" ], "concepts": [ "mlsys.embeddings", "mlsys.attention", "mlsys.position-encoding", "mlsys.tokenization", "mlsys.convolution", "mlsys.recurrence", "mlsys.normalization-layer", "mlsys.depthwise-separable-convolution", "mlsys.modality-encoder", "mlsys.residual-connection", "mlsys.feed-forward-block", "mlsys.activation-function" ], "related_concepts": [ "mlsys.kv-cache" ], "taught_in": [ { "volume": "vol1", "chapter": "Network Architectures", "url": "https://mlsysbook.ai/vol1/06_nn_architectures/06_nn_architectures.html" }, { "volume": "vol3", "chapter": "The Foundation Model", "url": "https://mlsysbook.ai/vol3/02_foundation_model/02_foundation_model.html" } ] }, { "code": "F2.3", "id": "conditional-computation", "label": "Sparse & conditional computation", "definition": "Model designs in which the computation performed varies per input, such as mixture-of-experts routing and early exit.", "includes": [ "Mixture of experts, routing, capacity, load balancing", "Active vs. total parameters", "Early exit" ], "excludes": [ "Weight pruning and static sparsity ([[model-compression]])", "Cascades across separate models ([[serving-architecture]])" ], "contexts": [ "deep-learning", "generative", "distributed" ], "concepts": [ "mlsys.mixture-of-experts", "mlsys.expert-routing", "mlsys.expert-capacity", "mlsys.load-balancing-loss", "mlsys.active-parameters", "mlsys.early-exit" ], "related_concepts": [ "mlsys.expert-parallel", "mlsys.sparsity" ], "taught_in": [ { "volume": "vol1", "chapter": "Network Architectures", "url": "https://mlsysbook.ai/vol1/06_nn_architectures/06_nn_architectures.html" }, { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html" }, { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" } ] }, { "code": "F2.4", "id": "efficient-sequence-models", "label": "Long-context & efficient sequence models", "definition": "Sequence architectures that change how cost scales with context length, co-designed with the memory and compute they demand.", "includes": [ "Sparse, sliding-window, and linear attention", "State-space models; grouped-query attention", "Context-length extension" ], "excludes": [ "Exact attention kernels such as FlashAttention ([[tensor-operators]])", "KV-cache management ([[llm-inference]])" ], "contexts": [ "generative", "deep-learning" ], "concepts": [ "mlsys.sliding-window-attention", "mlsys.linear-attention", "mlsys.state-space-model", "mlsys.grouped-query-attention", "mlsys.context-extension", "mlsys.context", "mlsys.context-degradation", "mlsys.dynamic-sparse-attention", "mlsys.multi-head-latent-attention" ], "related_concepts": [ "mlsys.kv-cache", "mlsys.io-aware-attention", "mlsys.sequence-parallel" ], "taught_in": [ { "volume": "vol1", "chapter": "Network Architectures", "url": "https://mlsysbook.ai/vol1/06_nn_architectures/06_nn_architectures.html" } ] }, { "code": "F2.5", "id": "model-artifacts", "label": "Model artifacts & formats", "definition": "The weights, graphs, metadata, and interface contracts that make a trained model a portable, loadable, deployable artifact.", "includes": [ "Weight serialization and exchange formats", "Model export and signatures" ], "excludes": [ "Versioning and registries ([[release-management]])", "Training checkpoints for recovery ([[reliability]])" ], "contexts": [ "edge" ], "concepts": [ "mlsys.model-artifact", "mlsys.weight-serialization", "mlsys.model-exchange-format", "mlsys.model-export", "mlsys.model-signature", "mlsys.chat-template" ], "related_concepts": [ "mlsys.artifact-version", "mlsys.checkpoint" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Frameworks", "url": "https://mlsysbook.ai/vol1/07_frameworks/07_frameworks.html" }, { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" } ] } ] }, { "code": "F3", "id": "system-models", "label": "Quantitative system models", "question": "Which simple models predict where time, memory, and money go?", "scope": "Analytical and empirical models used to bound performance, capacity, and training cost before or instead of measuring.", "subjects": [ { "code": "F3.1", "id": "roofline-locality", "label": "Roofline, performance models & simulation", "definition": "Bound-based reasoning about whether computation is limited by arithmetic or data movement, and how reuse changes that limit.", "includes": [ "Ridge point and compute- vs. memory-bound regimes", "Energy cost of data movement" ], "excludes": [ "Measured profiles and MFU ([[profiling-utilization]])", "Hardware memory organization ([[memory-systems]])" ], "contexts": [ "deep-learning" ], "concepts": [ "mlsys.arithmetic-intensity", "mlsys.roofline", "mlsys.locality", "mlsys.ridge-point", "mlsys.bound-regime", "mlsys.data-movement-energy", "mlsys.analytical-performance-model", "mlsys.performance-simulation", "mlsys.model-fidelity", "mlsys.execution-trace", "mlsys.communication-intensity" ], "related_concepts": [ "mlsys.tiling", "mlsys.memory-bandwidth", "mlsys.compute-throughput" ], "taught_in": [ { "volume": "vol1", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol1/01_introduction/01_introduction.html" }, { "volume": "vol1", "chapter": "Network Architectures", "url": "https://mlsysbook.ai/vol1/06_nn_architectures/06_nn_architectures.html" }, { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html" } ] }, { "code": "F3.2", "id": "speedup-limits", "label": "Amdahl's law & scaling efficiency", "definition": "Models of how much speedup added parallel or accelerated resources can deliver, given serial work and fixed or growing problem sizes.", "includes": [ "Amdahl's law and serial fractions", "Strong and weak scaling", "Parallel efficiency" ], "excludes": [ "Communication cost models ([[collectives]])" ], "contexts": [ "distributed" ], "concepts": [ "mlsys.serial-fraction", "mlsys.amdahls-law", "mlsys.strong-scaling", "mlsys.weak-scaling", "mlsys.parallel-efficiency" ], "related_concepts": [ "mlsys.stragglers", "mlsys.communication-overlap" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol2", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol2/01_introduction/01_introduction.html" }, { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" } ] }, { "code": "F3.3", "id": "queueing-models", "label": "Queueing & capacity models", "definition": "Models relating arrivals, service rates, utilization, and waiting, used to predict latency and capacity under load.", "includes": [ "Little's law", "Latency-throughput curves", "Burstiness and variability" ], "excludes": [ "Batching and admission mechanisms ([[request-scheduling]])", "Measuring a specific serving system's latency and throughput ([[latency-goodput]])" ], "contexts": [ "interactive", "streaming" ], "concepts": [ "mlsys.arrival-rate", "mlsys.service-capacity", "mlsys.queueing", "mlsys.littles-law", "mlsys.latency-throughput-curve", "mlsys.arrival-burstiness", "mlsys.finite-buffer", "mlsys.bandwidth-delay-product", "mlsys.head-of-line-blocking", "mlsys.open-closed-loop-load" ], "related_concepts": [ "mlsys.tail-latency", "mlsys.batching", "mlsys.utilization" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" }, { "volume": "vol3", "chapter": "KV Cache Management", "url": "https://mlsysbook.ai/vol3/05_kv_cache/05_kv_cache.html" } ] }, { "code": "F3.4", "id": "scaling-laws", "label": "Scaling laws & compute-optimal allocation", "definition": "Empirical relationships among parameters, data, and compute and observed quality, used to allocate training budgets.", "includes": [ "Loss scaling laws", "Compute-optimal model/data allocation", "Training compute estimates" ], "excludes": [ "Hyperparameter search ([[experiment-management]])", "Test-time compute scaling ([[test-time-compute]])" ], "contexts": [ "generative", "deep-learning" ], "concepts": [ "mlsys.scaling-laws", "mlsys.compute-optimal-allocation", "mlsys.training-compute-estimate" ], "related_concepts": [ "mlsys.experiment-budget" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol1", "chapter": "Data Selection", "url": "https://mlsysbook.ai/vol1/09_data_selection/09_data_selection.html" }, { "volume": "vol2", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol2/01_introduction/01_introduction.html" } ] } ] } ] }, { "code": "S", "id": "stack", "label": "Stack", "question": "What does the system run on, from silicon to the fleet?", "areas": [ { "code": "S1", "id": "hardware", "label": "Hardware & accelerators", "question": "How do processors, memory, and chip-level power shape what ML workloads can achieve?", "scope": "Compute, memory, interconnect, and power characteristics of ML hardware from datacenter accelerators to microcontrollers.", "subjects": [ { "code": "S1.1", "id": "accelerator-architecture", "label": "Accelerator architecture", "definition": "The organization of processors specialized for ML: execution units, threading, dataflow, and the throughput they provide.", "includes": [ "GPUs, TPUs, NPUs; SIMT and matrix engines", "Systolic arrays and dataflows", "Domain-specific architectures" ], "excludes": [ "On-device and embedded accelerators ([[edge-embedded-hardware]])", "Compiler mapping onto hardware ([[ml-compilers]])" ], "contexts": [ "deep-learning" ], "concepts": [ "mlsys.accelerator-execution-architecture", "mlsys.compute-throughput", "mlsys.systolic-array", "mlsys.matrix-unit", "mlsys.simt", "mlsys.dataflow", "mlsys.domain-specific-architecture", "mlsys.compute-in-memory", "mlsys.cpu-execution-architecture", "mlsys.advanced-accelerator-packaging", "mlsys.fpga-acceleration", "mlsys.wafer-scale-acceleration" ], "related_concepts": [ "mlsys.matrix-multiply" ], "taught_in": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html" }, { "volume": "vol2", "chapter": "Compute Infrastructure", "url": "https://mlsysbook.ai/vol2/02_compute_infrastructure/02_compute_infrastructure.html" } ] }, { "code": "S1.2", "id": "memory-systems", "label": "Memory hierarchy & bandwidth", "definition": "The storage levels of an ML processor, from registers and SRAM to HBM, and the capacity and bandwidth limits they impose.", "includes": [ "On-chip SRAM and scratchpads", "HBM", "The memory wall" ], "excludes": [ "Persistent and distributed storage ([[storage-systems]])", "Software memory allocation ([[runtime-dispatch]])" ], "contexts": [ "deep-learning", "generative", "recommendation" ], "concepts": [ "mlsys.memory-hierarchy", "mlsys.memory-capacity", "mlsys.memory-bandwidth", "mlsys.hbm", "mlsys.on-chip-sram", "mlsys.memory-wall", "mlsys.emerging-nvm", "mlsys.host-memory" ], "related_concepts": [ "mlsys.locality", "mlsys.tiling" ], "taught_in": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html" }, { "volume": "vol2", "chapter": "Compute Infrastructure", "url": "https://mlsysbook.ai/vol2/02_compute_infrastructure/02_compute_infrastructure.html" } ] }, { "code": "S1.3", "id": "node-interconnect", "label": "Host-device & scale-up interconnect", "definition": "Links that move data between hosts and accelerators and among accelerators within a node or rack-scale domain.", "includes": [ "PCIe host links", "NVLink-class scale-up fabrics", "CXL and peer-to-peer transfers" ], "excludes": [ "Datacenter scale-out networks ([[network-fabrics]])" ], "contexts": [ "distributed" ], "concepts": [ "mlsys.host-device-interconnect", "mlsys.scale-up-interconnect", "mlsys.coherent-interconnect", "mlsys.peer-to-peer-transfer" ], "related_concepts": [ "mlsys.network-topology", "mlsys.collectives" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol2", "chapter": "Compute Infrastructure", "url": "https://mlsysbook.ai/vol2/02_compute_infrastructure/02_compute_infrastructure.html" }, { "volume": "vol2", "chapter": "Network Fabrics", "url": "https://mlsysbook.ai/vol2/03_network_fabrics/03_network_fabrics.html" } ] }, { "code": "S1.4", "id": "chip-power-thermal", "label": "Chip power & thermal limits", "definition": "Electrical and thermal limits at the chip and board level, and the runtime controls that keep execution within them.", "includes": [ "TDP and thermal throttling", "DVFS and power capping", "Dark silicon" ], "excludes": [ "Facility power and cooling ([[datacenter-facilities]])", "Carbon accounting ([[sustainability-cost]])" ], "contexts": [ "edge", "physical" ], "concepts": [ "mlsys.power-delivery", "mlsys.thermal-dynamics", "mlsys.tdp", "mlsys.dvfs", "mlsys.dark-silicon", "mlsys.power-capping", "mlsys.duty-cycling", "mlsys.technology-scaling" ], "related_concepts": [ "mlsys.energy" ], "taught_in": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html" }, { "volume": "vol2", "chapter": "Compute Infrastructure", "url": "https://mlsysbook.ai/vol2/02_compute_infrastructure/02_compute_infrastructure.html" }, { "volume": "vol4", "chapter": "The Physical Body", "url": "https://mlsysbook.ai/vol4/02_body/02_body.html" }, { "volume": "vol4", "chapter": "Workload Placement", "url": "https://mlsysbook.ai/vol4/13_placement/13_placement.html" } ] }, { "code": "S1.5", "id": "edge-embedded-hardware", "label": "Edge & embedded hardware", "definition": "Hardware for on-device ML, from mobile SoCs to microcontrollers, where memory and energy per inference bind.", "includes": [ "Microcontrollers and TinyML devices", "Heterogeneous mobile SoCs and NPUs", "Energy and memory budgets" ], "excludes": [ "On-device runtimes and OTA updates ([[on-device-deployment]])", "Robot-specific real-time platforms ([[real-time-systems]])" ], "contexts": [ "edge", "physical" ], "concepts": [ "mlsys.microcontroller", "mlsys.heterogeneous-soc", "mlsys.edge-npu", "mlsys.on-device-memory-budget", "mlsys.intermittent-computing", "mlsys.sensor-interconnect", "mlsys.execute-in-place", "mlsys.always-on-cascade", "mlsys.dma-cache-maintenance" ], "related_concepts": [ "mlsys.quantization", "mlsys.placement", "mlsys.energy-per-inference" ], "taught_in": [ { "volume": "vol1", "chapter": "System Paradigms", "url": "https://mlsysbook.ai/vol1/02_system_paradigms/02_system_paradigms.html" }, { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html" }, { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html" }, { "volume": "vol4", "chapter": "Workload Placement", "url": "https://mlsysbook.ai/vol4/13_placement/13_placement.html" } ] } ] }, { "code": "S2", "id": "software-stack", "label": "Frameworks, compilers & runtimes", "question": "How is a model program expressed, transformed, and executed efficiently on hardware?", "scope": "The software layers between model code and hardware: programming models, operators, compilers, runtimes, and numerics.", "subjects": [ { "code": "S2.1", "id": "programming-models", "label": "ML frameworks & programming models", "definition": "The abstractions with which ML programs are written: tensors, computation graphs, execution modes, and automatic differentiation.", "includes": [ "Tensors and computation graphs", "Eager vs. graph execution; tracing JIT", "Automatic differentiation" ], "excludes": [ "Compiler optimization passes ([[ml-compilers]])", "Framework product catalogs" ], "contexts": [ "deep-learning" ], "concepts": [ "mlsys.autodiff", "mlsys.tensor", "mlsys.computation-graph", "mlsys.eager-execution", "mlsys.tracing-jit", "mlsys.function-transformations", "mlsys.task-actor-runtime", "mlsys.single-vs-multi-controller", "mlsys.embedded-dsl", "mlsys.differentiable-programming", "mlsys.llm-training-framework", "mlsys.python-gil" ], "related_concepts": [ "mlsys.tensor-layout" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Frameworks", "url": "https://mlsysbook.ai/vol1/07_frameworks/07_frameworks.html" } ] }, { "code": "S2.2", "id": "tensor-operators", "label": "Tensor operators & memory layout", "definition": "Implementations of core tensor operations, such as matrix multiply, convolution, and attention, and the layouts that govern their efficiency.", "includes": [ "GEMM and convolution lowering", "IO-aware attention kernels", "Vendor kernel libraries" ], "excludes": [ "Compiler-generated kernels and autotuning ([[ml-compilers]])", "Architectural changes to attention ([[efficient-sequence-models]])" ], "contexts": [ "deep-learning", "generative" ], "concepts": [ "mlsys.tensor-layout", "mlsys.matrix-multiply", "mlsys.io-aware-attention", "mlsys.convolution-lowering", "mlsys.kernel-library", "mlsys.sparse-tensor-kernels", "mlsys.sparse-ragged-tensors" ], "related_concepts": [ "mlsys.fusion", "mlsys.tiling" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html" } ] }, { "code": "S2.3", "id": "gpu-programming", "label": "Kernel & GPU programming model", "definition": "How kernels are written for parallel accelerators: the thread hierarchy, memory access patterns, occupancy, asynchronous data movement, and programming the matrix units.", "includes": [ "Writing and tuning kernels in CUDA-class or tile-level languages such as Triton", "Shared memory, synchronization, and latency hiding" ], "excludes": [ "Choosing kernels for operators ([[tensor-operators]])", "Automatic kernel generation and fusion ([[ml-compilers]])" ], "contexts": [ "deep-learning" ], "concepts": [ "mlsys.thread-hierarchy", "mlsys.memory-coalescing", "mlsys.occupancy", "mlsys.async-data-movement", "mlsys.matrix-engine-programming", "mlsys.register-pressure", "mlsys.kernel-synchronization", "mlsys.shared-memory-bank-conflict" ], "related_concepts": [ "mlsys.simt", "mlsys.tiling", "mlsys.io-aware-attention" ], "taught_in": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html" }, { "volume": "vol2", "chapter": "Fleet Efficiency", "url": "https://mlsysbook.ai/vol2/09_performance_engineering/09_performance_engineering.html" } ] }, { "code": "S2.4", "id": "ml-compilers", "label": "ML compilers & kernel optimization", "definition": "Transforming model programs through intermediate representations into optimized kernels for target hardware.", "includes": [ "Graph and loop IRs; lowering", "Fusion, tiling, autotuning, kernel DSLs", "Static memory planning" ], "excludes": [ "Quantization as a compression technique ([[model-compression]])" ], "contexts": [ "deep-learning", "edge" ], "concepts": [ "mlsys.compilation", "mlsys.fusion", "mlsys.tiling", "mlsys.intermediate-representation", "mlsys.autotuning", "mlsys.kernel-dsl", "mlsys.memory-planning", "mlsys.persistent-kernel", "mlsys.graph-superoptimization", "mlsys.dynamic-shapes", "mlsys.rule-based-graph-rewriting", "mlsys.software-pipelining", "mlsys.schedule-separation" ], "related_concepts": [ "mlsys.dispatch" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Frameworks", "url": "https://mlsysbook.ai/vol1/07_frameworks/07_frameworks.html" }, { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html" }, { "volume": "vol2", "chapter": "Fleet Efficiency", "url": "https://mlsysbook.ai/vol2/09_performance_engineering/09_performance_engineering.html" } ] }, { "code": "S2.5", "id": "runtime-dispatch", "label": "Runtime execution & dispatch", "definition": "How a runtime launches and orders device work and manages device memory during execution.", "includes": [ "Launch overhead and graph replay", "Streams and asynchronous execution", "Caching allocators" ], "excludes": [ "Request-level serving schedulers ([[request-scheduling]])" ], "contexts": [ "deep-learning", "interactive" ], "concepts": [ "mlsys.dispatch", "mlsys.device-streams", "mlsys.caching-allocator", "mlsys.graph-replay", "mlsys.cpu-threading-numa", "mlsys.execution-providers", "mlsys.memory-fragmentation", "mlsys.mmap-weights", "mlsys.os-memory-pressure", "mlsys.managed-runtime-interop" ], "related_concepts": [ "mlsys.communication-overlap" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Frameworks", "url": "https://mlsysbook.ai/vol1/07_frameworks/07_frameworks.html" }, { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol2", "chapter": "Fleet Efficiency", "url": "https://mlsysbook.ai/vol2/09_performance_engineering/09_performance_engineering.html" } ] }, { "code": "S2.6", "id": "data-loading", "label": "Data loading & input pipelines", "definition": "Delivering training and inference inputs to accelerators fast enough: host-side loading and preprocessing, prefetching, and sampling minibatches from large or irregular data such as graphs.", "includes": [ "Data-loader workers, decoding and augmentation on the host, pinned memory, and prefetching", "Minibatch construction for irregular data, such as graph neighborhood sampling" ], "excludes": [ "Building datasets and transformation pipelines ([[data-pipelines]])", "Storage systems and parallel file systems that hold the data ([[storage-systems]])" ], "contexts": [ "deep-learning", "interactive" ], "concepts": [ "mlsys.input-pipeline", "mlsys.input-prefetching", "mlsys.neighbor-sampling", "mlsys.host-data-loader", "mlsys.zero-copy-dma", "mlsys.sequence-packing", "mlsys.dataset-shuffling" ], "related_concepts": [ "mlsys.dataset-storage-layout" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html" } ] }, { "code": "S2.7", "id": "numerics", "label": "Numerical formats & precision", "definition": "Number formats used in ML computation and the techniques that keep training and inference accurate at reduced precision.", "includes": [ "FP32/BF16/FP8/INT formats and block scaling", "Mixed precision and loss scaling", "Accumulation precision" ], "excludes": [ "Post-training quantization as compression ([[model-compression]])" ], "contexts": [ "deep-learning", "generative" ], "concepts": [ "mlsys.precision", "mlsys.mixed-precision", "mlsys.numerical-stability", "mlsys.loss-scaling", "mlsys.block-scaled-format", "mlsys.accumulation-precision", "mlsys.numerical-nondeterminism", "mlsys.batch-invariance" ], "related_concepts": [ "mlsys.quantization" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html" }, { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" } ] } ] }, { "code": "S3", "id": "distributed", "label": "Parallelism & distributed execution", "question": "How is one workload split across many devices, and what does coordination cost?", "scope": "Partitioning ML work across devices and participants, the communication it requires, and the coordination that keeps it consistent.", "subjects": [ { "code": "S3.1", "id": "parallelism", "label": "Parallelism strategies", "definition": "Ways of partitioning a model's computation and state across devices, and how those dimensions are combined.", "includes": [ "Data, tensor, pipeline, sequence, expert parallelism", "ZeRO/FSDP state sharding", "Hybrid 3D layouts and pipeline bubbles" ], "excludes": [ "Collective algorithms ([[collectives]])", "Job placement on the cluster ([[cluster-scheduling]])", "Searching for and placing a parallel plan ([[parallelism-planning]])" ], "contexts": [ "distributed", "generative", "recommendation" ], "concepts": [ "mlsys.data-parallel", "mlsys.tensor-parallel", "mlsys.pipeline-parallel", "mlsys.state-sharding", "mlsys.expert-parallel", "mlsys.sequence-parallel", "mlsys.hybrid-parallel", "mlsys.pipeline-bubble", "mlsys.embedding-table-sharding", "mlsys.parameter-server", "mlsys.local-sgd", "mlsys.sync-async-training", "mlsys.gnn-graph-partitioning", "mlsys.pipeline-schedule", "mlsys.graph-pipeline-parallelism" ], "related_concepts": [ "mlsys.collectives", "mlsys.training-state", "mlsys.neighbor-sampling" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" } ] }, { "code": "S3.2", "id": "parallelism-planning", "label": "Parallelization planning & placement", "definition": "Choosing and changing how a model is partitioned across devices: describing devices as a mesh, searching or propagating sharding decisions, placing experts and stages on heterogeneous hardware, and resharding at run time.", "includes": [ "Automatic parallelization and sharding propagation", "Device meshes, expert placement, and dynamic resharding" ], "excludes": [ "The parallelism strategies themselves ([[parallelism]])" ], "contexts": [ "distributed", "generative", "recommendation" ], "concepts": [ "mlsys.auto-parallelization", "mlsys.sharding-propagation", "mlsys.device-mesh", "mlsys.heterogeneous-accelerator-execution", "mlsys.dynamic-resharding", "mlsys.expert-placement", "mlsys.variable-length-parallelism", "mlsys.geo-distributed-training" ], "related_concepts": [ "mlsys.hybrid-parallel" ], "taught_in": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" } ] }, { "code": "S3.3", "id": "collectives", "label": "Collective communication", "definition": "Group communication primitives and algorithms used by parallel ML, their cost models, and techniques to reduce or hide their cost.", "includes": [ "All-reduce, all-gather, all-to-all", "Ring and tree algorithms; alpha-beta cost model", "Overlap and gradient compression" ], "excludes": [ "Physical fabric design ([[network-fabrics]])" ], "contexts": [ "distributed", "recommendation" ], "concepts": [ "mlsys.collectives", "mlsys.communication-overlap", "mlsys.stragglers", "mlsys.ring-allreduce", "mlsys.tree-allreduce", "mlsys.all-to-all", "mlsys.alpha-beta-model", "mlsys.gradient-compression", "mlsys.process-group-rendezvous", "mlsys.device-initiated-communication", "mlsys.collective-algorithm-selection", "mlsys.all-reduce", "mlsys.reduce-scatter-all-gather" ], "related_concepts": [ "mlsys.network-topology", "mlsys.bisection-bandwidth", "mlsys.in-network-reduction" ], "taught_in": [ { "volume": "vol2", "chapter": "Collective Communication", "url": "https://mlsysbook.ai/vol2/05_collective_communication/05_collective_communication.html" }, { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" } ] }, { "code": "S3.4", "id": "coordination", "label": "Coordination & consensus", "definition": "Mechanisms by which distributed components agree on authoritative control state and tolerate stale or failed participants.", "includes": [ "Consensus and leader election", "Leases and fencing", "Bounded-staleness consistency" ], "excludes": [ "Checkpoint and recovery protocols ([[reliability]])", "Retry and compensation semantics ([[failure-semantics]])" ], "contexts": [ "distributed", "agentic" ], "concepts": [ "mlsys.consensus", "mlsys.fencing", "mlsys.lease", "mlsys.leader-election", "mlsys.bounded-staleness", "mlsys.cap-tradeoff" ], "related_concepts": [ "mlsys.checkpoint", "mlsys.consistent-cut", "mlsys.idempotency" ], "taught_in": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" }, { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html" }, { "volume": "vol3", "chapter": "Durable Execution", "url": "https://mlsysbook.ai/vol3/11_durable_execution/11_durable_execution.html" }, { "volume": "vol3", "chapter": "Multi-Agent Coordination", "url": "https://mlsysbook.ai/vol3/18_multi_agent/18_multi_agent.html" } ] }, { "code": "S3.5", "id": "federated-learning", "label": "Federated & decentralized learning", "definition": "Training across devices or organizations that keep their data local, coordinated through a shared model.", "includes": [ "Federated averaging", "Client heterogeneity and selection", "Cross-device vs. cross-silo" ], "excludes": [ "Secure aggregation and differential privacy ([[security-privacy]])" ], "contexts": [ "federated", "edge" ], "concepts": [ "mlsys.federated-learning", "mlsys.federated-averaging", "mlsys.non-iid-data", "mlsys.client-sampling", "mlsys.federation-setting", "mlsys.gossip-averaging" ], "related_concepts": [ "mlsys.differential-privacy", "mlsys.privacy" ], "taught_in": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html" }, { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html" } ] } ] }, { "code": "S4", "id": "infrastructure-fleet", "label": "Datacenter infrastructure & fleet", "question": "How are AI datacenters built, connected, fed, and shared across many jobs?", "scope": "Facilities, networks, storage, and cluster management that turn many accelerators into a shared ML fleet.", "subjects": [ { "code": "S4.1", "id": "datacenter-facilities", "label": "Datacenter power, cooling & facilities", "definition": "Engineering of the buildings that house ML fleets: power provisioning, cooling, density, and facility efficiency.", "includes": [ "Facility power and oversubscription", "Air and liquid cooling; rack density", "PUE" ], "excludes": [ "Carbon and water accounting ([[sustainability-cost]])", "Chip-level power limits ([[chip-power-thermal]])" ], "contexts": [ "distributed" ], "concepts": [ "mlsys.pue", "mlsys.liquid-cooling", "mlsys.rack-power-density", "mlsys.power-oversubscription", "mlsys.warehouse-scale-computer", "mlsys.training-power-swings", "mlsys.facility-energy-storage", "mlsys.grid-interconnection", "mlsys.on-site-generation", "mlsys.demand-response" ], "related_concepts": [ "mlsys.energy", "mlsys.carbon", "mlsys.power-delivery", "mlsys.thermal-dynamics" ], "taught_in": [ { "volume": "vol2", "chapter": "Compute Infrastructure", "url": "https://mlsysbook.ai/vol2/02_compute_infrastructure/02_compute_infrastructure.html" }, { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html" } ] }, { "code": "S4.2", "id": "network-fabrics", "label": "Network fabrics", "definition": "Datacenter networks that connect accelerator nodes: transports, topologies, flow and congestion control, and in-network compute.", "includes": [ "RDMA over InfiniBand and RoCE; lossless flow control", "Fat-tree, rail-optimized, torus topologies", "Congestion control; in-network reduction" ], "excludes": [ "Intra-node links ([[node-interconnect]])", "Collective algorithms ([[collectives]])" ], "contexts": [ "distributed" ], "concepts": [ "mlsys.network-topology", "mlsys.rdma", "mlsys.lossless-fabric", "mlsys.congestion-control", "mlsys.fat-tree", "mlsys.rail-optimized", "mlsys.bisection-bandwidth", "mlsys.in-network-reduction", "mlsys.adaptive-routing", "mlsys.incast", "mlsys.network-propagation-delay", "mlsys.optical-circuit-switching", "mlsys.co-packaged-optics", "mlsys.forward-error-correction", "mlsys.traffic-classes", "mlsys.torus-topology", "mlsys.network-oversubscription", "mlsys.link-reach", "mlsys.link-signal-integrity" ], "related_concepts": [ "mlsys.collectives", "mlsys.all-to-all" ], "taught_in": [ { "volume": "vol2", "chapter": "Network Fabrics", "url": "https://mlsysbook.ai/vol2/03_network_fabrics/03_network_fabrics.html" } ] }, { "code": "S4.3", "id": "storage-systems", "label": "Storage & data delivery", "definition": "Storage systems that hold training data, checkpoints, and artifacts at scale and deliver them to accelerators fast enough.", "includes": [ "Parallel file systems and object stores", "Input pipelines and data stalls", "Checkpoint I/O and storage tiering" ], "excludes": [ "Dataset formats and schema ([[data-pipelines]])", "Checkpoint frequency and recovery semantics ([[reliability]])", "Host-side loading and prefetching of batches ([[data-loading]])" ], "contexts": [ "distributed", "batch" ], "concepts": [ "mlsys.parallel-file-system", "mlsys.object-storage", "mlsys.checkpoint-io", "mlsys.storage-tiering", "mlsys.embedding-cache", "mlsys.erasure-coding", "mlsys.storage-access-pattern" ], "related_concepts": [ "mlsys.checkpoint", "mlsys.dataset-storage-layout", "mlsys.data-ingestion", "mlsys.input-pipeline" ], "taught_in": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html" }, { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html" } ] }, { "code": "S4.4", "id": "cluster-scheduling", "label": "Cluster scheduling & orchestration", "definition": "Allocating a shared cluster's accelerators to jobs over time: placement, gang allocation, fairness, preemption, and elasticity.", "includes": [ "Gang scheduling and backfill", "Fair sharing, preemption, elastic jobs", "Topology-aware placement and bin packing" ], "excludes": [ "Request-level serving autoscaling ([[request-scheduling]])", "Workflow DAG orchestration ([[pipelines-reproducibility]])", "Packaging code and dependencies as images ([[pipelines-reproducibility]])" ], "contexts": [ "distributed", "batch" ], "concepts": [ "mlsys.gang-scheduling", "mlsys.job-preemption", "mlsys.dominant-resource-fairness", "mlsys.elastic-training", "mlsys.topology-aware-placement", "mlsys.bin-packing", "mlsys.ml-aware-scheduling", "mlsys.cluster-manager", "mlsys.backfill-scheduling" ], "related_concepts": [ "mlsys.admission", "mlsys.utilization", "mlsys.checkpoint" ], "taught_in": [ { "volume": "vol2", "chapter": "Fleet Orchestration", "url": "https://mlsysbook.ai/vol2/08_fleet_orchestration/08_fleet_orchestration.html" } ] }, { "code": "S4.5", "id": "capacity-provisioning", "label": "Capacity planning & multi-tenancy", "definition": "Provisioning fleet capacity and sharing it among tenants through quotas, accelerator sharing, and interference control.", "includes": [ "Capacity planning and quotas", "Accelerator sharing and partitioning", "Spot capacity and interference" ], "excludes": [ "Security isolation guarantees ([[access-isolation]])", "Unit-cost economics ([[sustainability-cost]])" ], "contexts": [ "distributed", "interactive" ], "concepts": [ "mlsys.capacity-planning", "mlsys.multi-tenancy", "mlsys.resource-quota", "mlsys.gpu-partitioning", "mlsys.spot-capacity", "mlsys.performance-interference", "mlsys.cloud-service-model", "mlsys.fine-grained-accelerator-sharing" ], "related_concepts": [ "mlsys.isolation", "mlsys.autoscaling", "mlsys.unit-cost" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol2", "chapter": "Fleet Orchestration", "url": "https://mlsysbook.ai/vol2/08_fleet_orchestration/08_fleet_orchestration.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" } ] } ] } ] }, { "code": "P", "id": "lifecycle", "label": "Lifecycle", "question": "What do engineers do to build, ship, and keep the system working?", "areas": [ { "code": "P1", "id": "data", "label": "Data engineering", "question": "How is data acquired, labeled, curated, validated, stored, and kept trustworthy?", "scope": "Acquisition and labeling, corpus curation, validation and features, pipelines and storage formats, lineage, and temporal correctness of training and serving data.", "subjects": [ { "code": "P1.1", "id": "data-collection", "label": "Data collection & labeling", "definition": "Obtaining observations and supervisory targets whose sampling, coverage, and label quality are known and fit the task.", "includes": [ "Sampling design, human and programmatic labeling, labeling budgets.", "Annotation quality and agreement." ], "excludes": [ "Sensor design and calibration ([[perception-sensing]])", "Corpus filtering and mixtures ([[data-curation]])" ], "contexts": [ "classical-ml", "deep-learning", "physical" ], "concepts": [ "mlsys.data-acquisition", "mlsys.labels", "mlsys.weak-supervision", "mlsys.active-learning", "mlsys.annotator-agreement", "mlsys.class-imbalance", "mlsys.rebalancing", "mlsys.model-assisted-labeling" ], "related_concepts": [ "mlsys.sensing", "mlsys.privacy" ], "taught_in": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html" }, { "volume": "vol3", "chapter": "Trajectory Curation", "url": "https://mlsysbook.ai/vol3/15_trajectory_curation/15_trajectory_curation.html" }, { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html" } ] }, { "code": "P1.2", "id": "data-curation", "label": "Data selection & curation", "definition": "Choosing, filtering, deduplicating, mixing, augmenting, and synthesizing training examples to maximize learning value per unit of compute and labeling cost.", "includes": [ "Web-scale corpus filtering, deduplication, and domain mixtures.", "Augmentation and synthetic or model-generated data.", "Trajectory and demonstration curation for agents and robots." ], "excludes": [ "Labeling processes ([[data-collection]])", "Teacher-student compression ([[model-compression]])" ], "contexts": [ "deep-learning", "generative", "agentic", "multimodal" ], "concepts": [ "mlsys.data-selection", "mlsys.deduplication", "mlsys.quality-filtering", "mlsys.data-mixture", "mlsys.synthetic-data", "mlsys.data-augmentation", "mlsys.data-valuation", "mlsys.model-collapse", "mlsys.data-curriculum" ], "related_concepts": [ "mlsys.scaling-laws", "mlsys.distillation", "mlsys.benchmark-contamination" ], "taught_in": [ { "volume": "vol1", "chapter": "Data Selection", "url": "https://mlsysbook.ai/vol1/09_data_selection/09_data_selection.html" }, { "volume": "vol3", "chapter": "Trajectory Curation", "url": "https://mlsysbook.ai/vol3/15_trajectory_curation/15_trajectory_curation.html" }, { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html" } ] }, { "code": "P1.3", "id": "data-quality", "label": "Data validation & feature engineering", "definition": "Checking that recorded data is fit for use and transforming it into model inputs whose semantics and availability match training and serving.", "includes": [ "Schema and distribution checks before training and serving.", "Feature construction and shared feature stores." ], "excludes": [ "Learned representations inside the model ([[models-architectures]])", "Production drift monitoring ([[monitoring]])" ], "contexts": [ "classical-ml", "batch", "streaming", "recommendation" ], "concepts": [ "mlsys.data-quality", "mlsys.features", "mlsys.missing-data", "mlsys.data-validation", "mlsys.feature-store", "mlsys.data-cleaning", "mlsys.feature-hashing", "mlsys.feature-cross", "mlsys.exploratory-data-analysis", "mlsys.data-cascade" ], "related_concepts": [ "mlsys.data-contract", "mlsys.train-serve-skew" ], "taught_in": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" } ] }, { "code": "P1.4", "id": "data-pipelines", "label": "Data pipelines & storage formats", "definition": "Engineering ingestion, batch and stream transformation, event-time semantics, storage formats, and schemas for training and serving data.", "includes": [ "ETL/ELT, batch and stream processing, columnar formats.", "Data loaders that keep accelerators fed." ], "excludes": [ "Tensor layout in working memory ([[tensor-operators]])", "Delivering batches to accelerators ([[data-loading]])" ], "contexts": [ "batch", "streaming", "distributed" ], "concepts": [ "mlsys.dataset-storage-layout", "mlsys.data-ingestion", "mlsys.schema-evolution", "mlsys.stream-processing", "mlsys.feature-backfill", "mlsys.batch-data-processing", "mlsys.lakehouse", "mlsys.event-time-windowing", "mlsys.delivery-semantics", "mlsys.streaming-dataset-format" ], "related_concepts": [ "mlsys.workflow-orchestration", "mlsys.input-pipeline" ], "taught_in": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html" }, { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" } ] }, { "code": "P1.5", "id": "data-lineage", "label": "Data contracts, versioning & lineage", "definition": "Specifying, versioning, and tracing the datasets a system consumes so that dependencies, provenance, and changes are explicit.", "includes": [ "Dataset versions, snapshots, and provenance records.", "Producer-consumer contracts and dependency hygiene." ], "excludes": [ "Model and code artifact versioning ([[release-management]])", "Dataset documentation such as datasheets ([[transparency-explainability]])" ], "contexts": [], "concepts": [ "mlsys.lineage", "mlsys.data-contract", "mlsys.data-versioning", "mlsys.data-dependency", "mlsys.data-governance", "mlsys.data-consent", "mlsys.data-model-licensing" ], "related_concepts": [ "mlsys.artifact-version", "mlsys.experiment-reproducibility" ], "taught_in": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" }, { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html" } ] }, { "code": "P1.6", "id": "data-leakage-skew", "label": "Leakage & training-serving skew", "definition": "Ensuring examples use only information available at prediction time and that training and serving compute features identically and equally fresh.", "includes": [ "Point-in-time joins and temporal leakage.", "Feature staleness and skew between offline and online paths." ], "excludes": [ "Deployment distribution change ([[drift-retraining]])" ], "contexts": [ "classical-ml", "streaming" ], "concepts": [ "mlsys.leakage", "mlsys.point-in-time", "mlsys.train-serve-skew", "mlsys.feature-freshness", "mlsys.online-feature-consistency" ], "related_concepts": [ "mlsys.feature-store", "mlsys.benchmark-contamination" ], "taught_in": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" }, { "volume": "vol3", "chapter": "Trajectory Curation", "url": "https://mlsysbook.ai/vol3/15_trajectory_curation/15_trajectory_curation.html" } ] } ] }, { "code": "P2", "id": "training", "label": "Training & adaptation", "question": "How are models trained and adapted, and what does that demand from the system?", "scope": "Learning paradigms, optimization dynamics, training memory, experiment management, fine-tuning, post-training, and continual learning.", "subjects": [ { "code": "P2.1", "id": "learning-paradigms", "label": "Learning paradigms", "definition": "The ways supervision, self-generated targets, interaction, and demonstration produce learning signals, each with distinct data and compute demands.", "includes": [ "Supervised, self-supervised, unsupervised, reinforcement, and imitation learning." ], "excludes": [ "Preference and RL post-training of pretrained models ([[post-training]])" ], "contexts": [ "classical-ml", "deep-learning", "generative", "physical" ], "concepts": [ "mlsys.supervised-learning", "mlsys.self-supervised-learning", "mlsys.unsupervised-learning", "mlsys.reinforcement-learning", "mlsys.imitation-learning", "mlsys.pretraining", "mlsys.causal-lm", "mlsys.compounding-error", "mlsys.interactive-imitation", "mlsys.offline-rl", "mlsys.multi-task-learning", "mlsys.semi-supervised-learning", "mlsys.contrastive-learning" ], "related_concepts": [ "mlsys.federated-learning", "mlsys.in-context-learning" ], "taught_in": [ { "volume": "vol1", "chapter": "Neural Computation", "url": "https://mlsysbook.ai/vol1/05_nn_computation/05_nn_computation.html" }, { "volume": "vol1", "chapter": "Data Selection", "url": "https://mlsysbook.ai/vol1/09_data_selection/09_data_selection.html" }, { "volume": "vol4", "chapter": "Policy Training", "url": "https://mlsysbook.ai/vol4/06_training/06_training.html" } ] }, { "code": "P2.2", "id": "training-optimization", "label": "Optimization & training dynamics", "definition": "The objectives, update rules, schedules, and batch regimes that determine whether and how efficiently training converges.", "includes": [ "Optimizers, learning-rate warmup and decay, large-batch scaling.", "Loss spikes, divergence, and their mitigations." ], "excludes": [ "Floating-point formats and overflow ([[numerics]])" ], "contexts": [ "deep-learning", "distributed" ], "concepts": [ "mlsys.loss", "mlsys.optimizer", "mlsys.regularization", "mlsys.batch-size", "mlsys.learning-rate-schedule", "mlsys.critical-batch-size", "mlsys.training-instability", "mlsys.mode-collapse", "mlsys.gradient-clipping", "mlsys.weight-initialization", "mlsys.overfitting", "mlsys.early-stopping", "mlsys.group-dro" ], "related_concepts": [ "mlsys.numerical-stability", "mlsys.mixed-precision", "mlsys.data-parallel" ], "taught_in": [ { "volume": "vol1", "chapter": "Neural Computation", "url": "https://mlsysbook.ai/vol1/05_nn_computation/05_nn_computation.html" }, { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" } ] }, { "code": "P2.3", "id": "training-memory", "label": "Training memory & state", "definition": "Accounting for and reducing the parameters, gradients, optimizer state, and activations a training step must hold.", "includes": [ "Activation recomputation, gradient accumulation, CPU/NVMe offload." ], "excludes": [ "Sharding state across devices ([[parallelism]])", "Persisting state for recovery ([[checkpoint-recovery]])" ], "contexts": [ "deep-learning", "distributed" ], "concepts": [ "mlsys.training-state", "mlsys.activation-recompute", "mlsys.gradient-accumulation", "mlsys.memory-offloading", "mlsys.memory-efficient-optimizer", "mlsys.training-memory-estimate" ], "related_concepts": [ "mlsys.state-sharding", "mlsys.checkpoint", "mlsys.memory-capacity", "mlsys.neighbor-sampling" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" } ] }, { "code": "P2.4", "id": "experiment-management", "label": "Hyperparameter search & experiment management", "definition": "Allocating a finite budget of training runs across configurations, and recording runs so their outcomes can be compared and reproduced.", "includes": [ "Random, Bayesian, and multi-fidelity search; small-scale proxies.", "Run metadata and experiment tracking." ], "excludes": [ "Pipeline orchestration ([[pipelines-reproducibility]])" ], "contexts": [ "deep-learning" ], "concepts": [ "mlsys.hyperparameter-search", "mlsys.experiment-budget", "mlsys.experiment-tracking", "mlsys.hyperparameter-transfer", "mlsys.model-selection" ], "related_concepts": [ "mlsys.selection-bias", "mlsys.experiment-reproducibility", "mlsys.scaling-laws" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" } ] }, { "code": "P2.5", "id": "fine-tuning", "label": "Adaptation: fine-tuning, prompting & transfer", "definition": "Adapting a pretrained model to a new task or behavior, by continued training of all or part of its parameters or by prompting without parameter updates.", "includes": [ "Instruction and supervised fine-tuning, adapters, LoRA, and model merging.", "Prompt engineering and in-context learning without parameter updates." ], "excludes": [ "Preference or reward-driven training ([[post-training]])", "Assembling context for agents ([[context-memory]])" ], "contexts": [ "deep-learning", "generative", "agentic", "edge" ], "concepts": [ "mlsys.transfer-learning", "mlsys.peft", "mlsys.supervised-fine-tuning", "mlsys.lora", "mlsys.prompt-engineering", "mlsys.loss-masking", "mlsys.in-context-learning", "mlsys.model-merging" ], "related_concepts": [ "mlsys.distillation", "mlsys.quantization", "mlsys.catastrophic-forgetting" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html" }, { "volume": "vol3", "chapter": "Trajectory Fine-Tuning", "url": "https://mlsysbook.ai/vol3/16_fine_tuning/16_fine_tuning.html" } ] }, { "code": "P2.6", "id": "post-training", "label": "Preference & RL post-training", "definition": "Training pretrained models against preference, reward-model, or verifiable-reward signals, and the rollout, scoring, and weight-synchronization systems that loop requires.", "includes": [ "RLHF, DPO-style preference optimization, RL with verifiable rewards.", "Rollout generation, reward scoring, colocated vs disaggregated actor/learner, weight sync." ], "excludes": [ "General RL theory ([[learning-paradigms]])", "Reward exploitation as a specification failure ([[objectives-decisions]])" ], "contexts": [ "generative", "agentic", "distributed" ], "concepts": [ "mlsys.rlhf", "mlsys.reward-model", "mlsys.preference-optimization", "mlsys.rlvr", "mlsys.rl-rollout", "mlsys.weight-sync", "mlsys.policy-staleness", "mlsys.credit-assignment", "mlsys.reward-hacking", "mlsys.group-relative-policy-optimization", "mlsys.agent-training-environment", "mlsys.rejection-sampling-fine-tuning", "mlsys.rl-training-colocation" ], "related_concepts": [ "mlsys.reinforcement-learning", "mlsys.verifier", "mlsys.speculative-decoding" ], "taught_in": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html" }, { "volume": "vol3", "chapter": "Trajectory Fine-Tuning", "url": "https://mlsysbook.ai/vol3/16_fine_tuning/16_fine_tuning.html" }, { "volume": "vol3", "chapter": "Reinforcement Learning from Verifiable Rewards", "url": "https://mlsysbook.ai/vol3/17_rlvr/17_rlvr.html" } ] }, { "code": "P2.7", "id": "continual-learning", "label": "Continual & online learning", "definition": "Updating deployed models incrementally from new or streaming data while retaining prior competence.", "includes": [ "Online and incremental updates, forgetting and replay.", "Personalization and adaptation on device." ], "excludes": [ "Deciding when to retrain ([[drift-retraining]])", "Federated aggregation protocols ([[federated-learning]])" ], "contexts": [ "streaming", "edge", "federated" ], "concepts": [ "mlsys.online-learning", "mlsys.incremental-training", "mlsys.catastrophic-forgetting", "mlsys.experience-replay", "mlsys.on-device-training" ], "related_concepts": [ "mlsys.retraining", "mlsys.distribution-shift", "mlsys.federated-learning" ], "taught_in": [ { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html" } ] } ] }, { "code": "P3", "id": "evaluation", "label": "Model evaluation & testing", "question": "How do we know a model and the system around it behave well enough, and for whom?", "scope": "Evaluation design, metrics and uncertainty, slices and robustness, evaluation of generative and agentic outputs, and ML testing and debugging.", "subjects": [ { "code": "P3.1", "id": "evaluation-protocols", "label": "Evaluation design & experiments", "definition": "Constructing offline splits and online experiments whose comparisons support the conclusions drawn from them.", "includes": [ "Holdout, temporal, and cross-validation splits.", "A/B tests, sequential monitoring, interference between units." ], "excludes": [ "Release staging mechanics ([[release-management]])" ], "contexts": [ "classical-ml", "interactive" ], "concepts": [ "mlsys.evaluation-split", "mlsys.selection-bias", "mlsys.online-experiment", "mlsys.cross-validation", "mlsys.sequential-testing", "mlsys.experiment-interference", "mlsys.evaluation-harness", "mlsys.closed-loop-evaluation", "mlsys.paired-comparison", "mlsys.off-policy-evaluation", "mlsys.multiple-comparisons", "mlsys.statistical-power" ], "related_concepts": [ "mlsys.leakage", "mlsys.shadow-deployment" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Lifecycle", "url": "https://mlsysbook.ai/vol1/03_ml_lifecycle/03_ml_lifecycle.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" }, { "volume": "vol4", "chapter": "Closed-Loop Evaluation", "url": "https://mlsysbook.ai/vol4/07_evaluation/07_evaluation.html" } ] }, { "code": "P3.2", "id": "metrics-uncertainty", "label": "Metrics, calibration & uncertainty", "definition": "Specifying what is measured and quantifying how far estimates and predicted probabilities can be trusted.", "includes": [ "Task metrics, calibration, confidence intervals, abstention.", "Distribution-free prediction sets." ], "excludes": [ "Choosing what to optimize ([[objectives-decisions]])" ], "contexts": [], "concepts": [ "mlsys.generalization", "mlsys.metrics", "mlsys.calibration", "mlsys.statistical-uncertainty", "mlsys.selective-prediction", "mlsys.conformal-prediction", "mlsys.ood-detection", "mlsys.predictive-uncertainty", "mlsys.retrieval-metrics", "mlsys.precision-recall" ], "related_concepts": [ "mlsys.proxy-objective" ], "taught_in": [ { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html" }, { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html" }, { "volume": "vol3", "chapter": "Agent Evaluation", "url": "https://mlsysbook.ai/vol3/13_evaluation/13_evaluation.html" }, { "volume": "vol4", "chapter": "Closed-Loop Evaluation", "url": "https://mlsysbook.ai/vol4/07_evaluation/07_evaluation.html" }, { "volume": "vol4", "chapter": "Grounded Intent", "url": "https://mlsysbook.ai/vol4/10_intent/10_intent.html" } ] }, { "code": "P3.3", "id": "robustness-slices", "label": "Sliced & robustness evaluation", "definition": "Measuring behavior separately across populations and conditions, and under deliberate shift or perturbation, rather than only in aggregate.", "includes": [ "Subgroup and condition slices, shift and perturbation tests." ], "excludes": [ "Fairness criteria themselves ([[fairness-bias]])", "Adversarial attack mechanics ([[security-privacy]])" ], "contexts": [], "concepts": [ "mlsys.sliced-evaluation", "mlsys.robustness-evaluation", "mlsys.slice-discovery" ], "related_concepts": [ "mlsys.fairness", "mlsys.distribution-shift", "mlsys.evasion" ], "taught_in": [ { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html" }, { "volume": "vol4", "chapter": "Adversarial Verification", "url": "https://mlsysbook.ai/vol4/15_verification/15_verification.html" } ] }, { "code": "P3.4", "id": "generative-agent-eval", "label": "Evaluating generative & agentic systems", "definition": "Assessing open-ended outputs and multi-step runs where no single reference answer exists, using graders, sampling-aware metrics, and contamination controls.", "includes": [ "Model graders, pairwise and human preference, pass@k.", "Contamination of benchmarks by training data." ], "excludes": [ "Verifiers used to select outputs at inference ([[test-time-compute]])" ], "contexts": [ "generative", "agentic" ], "concepts": [ "mlsys.hallucination", "mlsys.agent-evaluation", "mlsys.llm-judge", "mlsys.pass-at-k", "mlsys.pairwise-preference-eval", "mlsys.human-evaluation", "mlsys.benchmark-contamination", "mlsys.dangerous-capability-eval", "mlsys.perplexity", "mlsys.reference-match-scoring", "mlsys.judge-bias", "mlsys.sycophancy-evaluation" ], "related_concepts": [ "mlsys.verifier", "mlsys.statistical-uncertainty" ], "taught_in": [ { "volume": "vol3", "chapter": "Test-Time Compute", "url": "https://mlsysbook.ai/vol3/03_test_time_compute/03_test_time_compute.html" }, { "volume": "vol3", "chapter": "Agent Evaluation", "url": "https://mlsysbook.ai/vol3/13_evaluation/13_evaluation.html" } ] }, { "code": "P3.5", "id": "ml-testing", "label": "ML testing & debugging", "definition": "Testing data, models, pipelines, and their integration against stated expectations when exact expected outputs are often unknown.", "includes": [ "Oracles, metamorphic and behavioral tests, readiness rubrics.", "Integration testing, training diagnostics, and error analysis." ], "excludes": [ "Fault injection into infrastructure ([[fault-injection]])" ], "contexts": [], "concepts": [ "mlsys.test-oracle", "mlsys.integration-testing", "mlsys.metamorphic-testing", "mlsys.behavioral-testing", "mlsys.production-readiness", "mlsys.model-visualization", "mlsys.error-analysis", "mlsys.training-diagnostics" ], "related_concepts": [ "mlsys.data-validation", "mlsys.failure-injection" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" }, { "volume": "vol4", "chapter": "Adversarial Verification", "url": "https://mlsysbook.ai/vol4/15_verification/15_verification.html" } ] } ] }, { "code": "P4", "id": "serving", "label": "Deployment & serving", "question": "How are trained models delivered to users within latency, throughput, and cost targets?", "scope": "Serving architectures, latency and goodput, request scheduling, LLM inference mechanisms, structured generation, test-time compute, retrieval, and on-device deployment.", "subjects": [ { "code": "P4.1", "id": "serving-architecture", "label": "Serving architectures", "definition": "The structure through which predictions are produced and delivered, from offline batch scoring to online and serverless model servers, multi-model and multi-adapter hosting, and streaming speech services.", "includes": [ "Batch vs online prediction, inference servers, serverless endpoints, and APIs.", "Multi-model and multi-adapter hosting, ranking funnels, and streaming/voice pipelines." ], "excludes": [ "Replica load balancing ([[request-scheduling]])", "Routing and composing calls across models ([[compound-ai]])", "Second-stage reranking inside retrieval ([[retrieval-rag]])" ], "contexts": [ "batch", "interactive", "streaming", "generative", "recommendation" ], "concepts": [ "mlsys.online-inference", "mlsys.batch-inference", "mlsys.model-server", "mlsys.multi-model-serving", "mlsys.ranking-funnel", "mlsys.multi-adapter-serving", "mlsys.keyed-predictions", "mlsys.stateless-serving-function", "mlsys.serverless-inference", "mlsys.streaming-asr", "mlsys.voice-pipeline", "mlsys.streaming-inference", "mlsys.inference-api", "mlsys.model-swapping" ], "related_concepts": [ "mlsys.retrieval", "mlsys.unit-cost", "mlsys.selective-prediction" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" }, { "volume": "vol3", "chapter": "Agent Economics", "url": "https://mlsysbook.ai/vol3/19_agent_economics/19_agent_economics.html" } ] }, { "code": "P4.2", "id": "compound-ai", "label": "Compound AI systems & model routing", "definition": "Composing several models, retrievers, and tools into one application, and routing, cascading, caching, and gating requests across models to meet quality and cost targets.", "includes": [ "Language-model programs and their optimization", "Model routing, cascades, gateways, and semantic caches" ], "excludes": [ "Single-model serving infrastructure ([[serving-architecture]])", "Agent loops that plan and act through tools ([[agent-harness]])" ], "contexts": [ "batch", "interactive", "streaming", "generative", "recommendation" ], "concepts": [ "mlsys.compound-ai-system", "mlsys.lm-program-optimization", "mlsys.model-routing", "mlsys.model-gateway", "mlsys.semantic-cache", "mlsys.model-cascade", "mlsys.compound-workflow-pipelining" ], "related_concepts": [ "mlsys.retrieval", "mlsys.unit-cost", "mlsys.prompt-engineering" ], "taught_in": [ { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" }, { "volume": "vol3", "chapter": "Agent Economics", "url": "https://mlsysbook.ai/vol3/19_agent_economics/19_agent_economics.html" } ] }, { "code": "P4.3", "id": "latency-goodput", "label": "Latency, throughput & goodput", "definition": "Defining and meeting per-request latency and useful-throughput targets, including tail behavior and token-level latency for generative services.", "includes": [ "Latency budgets, tail latency, TTFT and inter-token latency.", "Tail-tolerant request techniques." ], "excludes": [ "Analytic queueing models ([[queueing-models]])", "General measurement methodology and profiling ([[profiling-utilization]]); analytic queueing models ([[queueing-models]])" ], "contexts": [ "interactive", "generative" ], "concepts": [ "mlsys.tail-latency", "mlsys.slo-goodput", "mlsys.ttft", "mlsys.tpot", "mlsys.hedged-requests", "mlsys.request-throughput", "mlsys.latency-percentiles", "mlsys.inter-token-latency", "mlsys.fan-out-tail-amplification" ], "related_concepts": [ "mlsys.slo", "mlsys.queueing", "mlsys.latency-budget" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" } ] }, { "code": "P4.4", "id": "request-scheduling", "label": "Batching, admission & autoscaling", "definition": "Policies that group, admit, distribute, and provision for requests so a service meets its targets under varying load.", "includes": [ "Dynamic and iteration-level batching, admission, load balancing.", "Autoscaling and model cold start." ], "excludes": [ "Cluster-level job scheduling ([[cluster-scheduling]])" ], "contexts": [ "interactive", "generative", "distributed" ], "concepts": [ "mlsys.batching", "mlsys.admission", "mlsys.autoscaling", "mlsys.backpressure", "mlsys.continuous-batching", "mlsys.load-balancing", "mlsys.cold-start", "mlsys.llm-request-scheduling", "mlsys.cache-aware-routing", "mlsys.fidelity-adaptive-serving" ], "related_concepts": [ "mlsys.service-capacity", "mlsys.capacity-planning" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" } ] }, { "code": "P4.5", "id": "llm-inference", "label": "LLM & generative inference", "definition": "Mechanisms specific to generative inference: managing attention state, separating prompt processing from token generation, accelerating autoregressive decoding, and reducing iterative denoising cost.", "includes": [ "KV cache management, prefix reuse, speculative decoding.", "Chunked prefill and prefill/decode disaggregation.", "Denoising-step reduction and feature caching for diffusion models." ], "excludes": [ "Weight quantization ([[model-compression]])", "Context selection and compaction ([[context-memory]])" ], "contexts": [ "generative", "interactive", "distributed", "multimodal" ], "concepts": [ "mlsys.kv-cache", "mlsys.prefill-decode", "mlsys.paged-kv", "mlsys.prefix-cache", "mlsys.speculative-decoding", "mlsys.chunked-prefill", "mlsys.pd-disaggregation", "mlsys.encoder-disaggregation", "mlsys.kv-cache-offload", "mlsys.kv-cache-transfer", "mlsys.kv-cache-eviction", "mlsys.cpu-gpu-hybrid-inference", "mlsys.non-prefix-kv-reuse", "mlsys.token-streaming", "mlsys.denoising-step-reduction", "mlsys.diffusion-feature-caching", "mlsys.masked-diffusion-decoding", "mlsys.expert-offloading", "mlsys.intra-device-overlap", "mlsys.kv-memory-allocation", "mlsys.attention-ffn-disaggregation" ], "related_concepts": [ "mlsys.continuous-batching", "mlsys.quantization", "mlsys.tensor-parallel", "mlsys.attention" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" }, { "volume": "vol3", "chapter": "KV Cache Management", "url": "https://mlsysbook.ai/vol3/05_kv_cache/05_kv_cache.html" } ] }, { "code": "P4.6", "id": "structured-generation", "label": "Decoding, sampling & structured output", "definition": "Controlling how tokens are selected and constraining generated output to required syntax or schemas.", "includes": [ "Temperature and nucleus sampling, grammar-constrained decoding, schema checks." ], "excludes": [ "Tool authority and effects ([[tool-interfaces]])" ], "contexts": [ "generative", "agentic" ], "concepts": [ "mlsys.grammar", "mlsys.sampling-controls", "mlsys.constrained-decoding", "mlsys.output-validation", "mlsys.generation-stopping" ], "related_concepts": [ "mlsys.tool-contract", "mlsys.hallucination" ], "taught_in": [ { "volume": "vol3", "chapter": "The Foundation Model", "url": "https://mlsysbook.ai/vol3/02_foundation_model/02_foundation_model.html" } ] }, { "code": "P4.7", "id": "test-time-compute", "label": "Test-time compute", "definition": "Spending additional inference computation on reasoning, sampling, search, and verification to raise output quality under a budget.", "includes": [ "Reasoning budgets, parallel sampling and selection, verifiers, tree search.", "Trading inference compute against model size." ], "excludes": [ "Agent control loops and tool use ([[agent-harness]])", "Pass@k as an evaluation metric ([[generative-agent-eval]])" ], "contexts": [ "generative", "agentic" ], "concepts": [ "mlsys.reasoning-budget", "mlsys.best-of-n", "mlsys.self-consistency", "mlsys.verifier", "mlsys.process-reward-model", "mlsys.inference-time-search", "mlsys.test-time-scaling", "mlsys.sequential-revision" ], "related_concepts": [ "mlsys.pass-at-k", "mlsys.reward-model", "mlsys.planning", "mlsys.unit-cost" ], "taught_in": [ { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" }, { "volume": "vol3", "chapter": "Test-Time Compute", "url": "https://mlsysbook.ai/vol3/03_test_time_compute/03_test_time_compute.html" } ] }, { "code": "P4.8", "id": "retrieval-rag", "label": "Retrieval & vector search", "definition": "Selecting relevant records from an external corpus by lexical, dense, or hybrid search and supplying them to a model at inference.", "includes": [ "Retrieval-augmented generation, ANN indexes and vector stores, reranking." ], "excludes": [ "Agent memory write and consolidation policy ([[context-memory]])" ], "contexts": [ "generative", "agentic", "interactive" ], "concepts": [ "mlsys.retrieval", "mlsys.ann", "mlsys.rag", "mlsys.dense-retrieval", "mlsys.lexical-retrieval", "mlsys.reranking", "mlsys.graph-vector-index", "mlsys.quantized-vector-index", "mlsys.filtered-vector-search", "mlsys.index-freshness", "mlsys.multi-hop-retrieval", "mlsys.document-chunking", "mlsys.hybrid-retrieval", "mlsys.embedding-migration", "mlsys.rag-pipeline-scheduling" ], "related_concepts": [ "mlsys.embeddings", "mlsys.context" ], "taught_in": [ { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" }, { "volume": "vol3", "chapter": "Long-Term Memory", "url": "https://mlsysbook.ai/vol3/06_long_term_memory/06_long_term_memory.html" } ] }, { "code": "P4.9", "id": "on-device-deployment", "label": "On-device & edge deployment", "definition": "Converting, packaging, executing, and updating models on phones, embedded devices, and microcontrollers under memory, energy, and connectivity limits.", "includes": [ "Export and conversion, on-device runtimes, device/cloud split.", "Over-the-air model updates across heterogeneous fleets." ], "excludes": [ "Edge silicon ([[edge-embedded-hardware]])", "Compression techniques ([[model-compression]])" ], "contexts": [ "edge", "physical" ], "concepts": [ "mlsys.model-conversion", "mlsys.on-device-runtime", "mlsys.ota-model-update", "mlsys.device-cloud-partitioning", "mlsys.device-heterogeneity", "mlsys.ui-thread-inference" ], "related_concepts": [ "mlsys.quantization", "mlsys.placement", "mlsys.on-device-training", "mlsys.memory-capacity", "mlsys.memory-planning" ], "taught_in": [ { "volume": "vol1", "chapter": "System Paradigms", "url": "https://mlsysbook.ai/vol1/02_system_paradigms/02_system_paradigms.html" }, { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html" }, { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" }, { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html" }, { "volume": "vol4", "chapter": "The Cognitive Brain", "url": "https://mlsysbook.ai/vol4/03_brain/03_brain.html" } ] } ] }, { "code": "P5", "id": "operations", "label": "MLOps & operations", "question": "How is a deployed ML system versioned, observed, changed, and eventually retired?", "scope": "Release management, monitoring, drift and retraining, incident response, pipelines and reproducibility, platforms and teams, and retirement.", "subjects": [ { "code": "P5.1", "id": "release-management", "label": "Versioning, rollout & rollback", "definition": "Identifying, gating, and progressively exposing model releases, with the ability to return to a known-good configuration.", "includes": [ "Registries, CI/CD for models, release gates.", "Shadow, canary, and staged rollout; rollback." ], "excludes": [ "Statistical design of experiments ([[evaluation-protocols]])" ], "contexts": [], "concepts": [ "mlsys.artifact-version", "mlsys.rollout", "mlsys.rollback", "mlsys.model-registry", "mlsys.shadow-deployment", "mlsys.ml-ci-cd", "mlsys.release-gate", "mlsys.model-feature-compatibility" ], "related_concepts": [ "mlsys.online-experiment", "mlsys.model-artifact", "mlsys.ota-model-update" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" }, { "volume": "vol3", "chapter": "Operating Agents", "url": "https://mlsysbook.ai/vol3/14_agent_operations/14_agent_operations.html" }, { "volume": "vol4", "chapter": "Deployment Release", "url": "https://mlsysbook.ai/vol4/16_release/16_release.html" } ] }, { "code": "P5.2", "id": "monitoring", "label": "Monitoring & observability", "definition": "Instrumenting deployed systems and detecting, from logged signals, changes in inputs, outputs, quality, and health.", "includes": [ "Prediction logging, drift statistics, alerting and alert fatigue.", "Metrics, logs, traces for model services and agents." ], "excludes": [ "Responding to detected incidents ([[incident-response]])" ], "contexts": [], "concepts": [ "mlsys.monitoring", "mlsys.observability", "mlsys.drift-detection", "mlsys.inference-logging", "mlsys.alerting", "mlsys.distributed-tracing", "mlsys.llm-call-tracing", "mlsys.misuse-monitoring", "mlsys.telemetry-sampling" ], "related_concepts": [ "mlsys.distribution-shift", "mlsys.delayed-outcomes", "mlsys.privacy-leakage" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" }, { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" }, { "volume": "vol3", "chapter": "Operating Agents", "url": "https://mlsysbook.ai/vol3/14_agent_operations/14_agent_operations.html" } ] }, { "code": "P5.3", "id": "drift-retraining", "label": "Drift, feedback loops & retraining", "definition": "Understanding how deployment data and outcomes change, including changes the system itself causes, and deciding when and how to update the model.", "includes": [ "Covariate and concept drift, delayed labels, hidden feedback loops.", "Scheduled and triggered retraining, continuous training." ], "excludes": [ "Incremental learning algorithms ([[continual-learning]])" ], "contexts": [ "classical-ml", "streaming" ], "concepts": [ "mlsys.distribution-shift", "mlsys.concept-drift", "mlsys.delayed-outcomes", "mlsys.feedback-loop", "mlsys.retraining", "mlsys.continuous-training", "mlsys.user-feedback", "mlsys.fleet-data-engine", "mlsys.label-shift-estimation" ], "related_concepts": [ "mlsys.online-learning", "mlsys.drift-detection" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" }, { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html" }, { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html" } ] }, { "code": "P5.4", "id": "incident-response", "label": "Incident response", "definition": "Detecting, triaging, mitigating, and learning from production failures, including attributing them to data, model, or infrastructure causes.", "includes": [ "Triage, mitigation, blameless postmortems, on-call." ], "excludes": [ "Failure semantics and recovery mechanisms ([[reliability]])" ], "contexts": [], "concepts": [ "mlsys.incident-response", "mlsys.postmortem", "mlsys.failure-attribution", "mlsys.runbook" ], "related_concepts": [ "mlsys.rollback", "mlsys.observability", "mlsys.failure-model" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" }, { "volume": "vol3", "chapter": "Operating Agents", "url": "https://mlsysbook.ai/vol3/14_agent_operations/14_agent_operations.html" } ] }, { "code": "P5.5", "id": "pipelines-reproducibility", "label": "ML pipelines & reproducibility", "definition": "Automating the steps that produce models and controlling their environments and configuration so runs can be repeated.", "includes": [ "Orchestrated pipelines, environment pinning, configuration as code." ], "excludes": [ "Experiment search strategy ([[experiment-management]])", "Scheduling containers across a cluster ([[cluster-scheduling]])" ], "contexts": [ "batch" ], "concepts": [ "mlsys.dependency-environment", "mlsys.workflow-orchestration", "mlsys.experiment-reproducibility", "mlsys.configuration-management", "mlsys.infrastructure-as-code", "mlsys.containerization" ], "related_concepts": [ "mlsys.lineage", "mlsys.experiment-tracking" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Lifecycle", "url": "https://mlsysbook.ai/vol1/03_ml_lifecycle/03_ml_lifecycle.html" }, { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" } ] }, { "code": "P5.6", "id": "ml-platforms", "label": "ML platforms, teams & technical debt", "definition": "Shared internal platforms, team structures, and debt management that let an organization operate many models at sustainable cost.", "includes": [ "Platform-as-product, maturity levels, ownership and roles.", "ML-specific technical debt and anti-patterns." ], "excludes": [ "Fleet hardware and scheduling ([[infrastructure-fleet]])", "Governance and regulation ([[governance-regulation]])" ], "contexts": [], "concepts": [ "mlsys.ml-platform", "mlsys.technical-debt", "mlsys.entanglement", "mlsys.undeclared-consumers", "mlsys.mlops-maturity", "mlsys.ml-roles", "mlsys.glue-code", "mlsys.pipeline-jungle", "mlsys.build-vs-buy" ], "related_concepts": [ "mlsys.feature-store", "mlsys.model-registry", "mlsys.data-dependency" ], "taught_in": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" } ] }, { "code": "P5.7", "id": "retirement", "label": "Model retirement", "definition": "Ending a model's service by migrating its consumers, archiving what must be kept, and deleting data and resources that are no longer justified.", "includes": [ "Deprecation with consumer migration, archival for audit, retention and deletion." ], "excludes": [ "Removing learned influence of specific records ([[privacy-dp]])" ], "contexts": [], "concepts": [ "mlsys.model-deprecation", "mlsys.artifact-archival", "mlsys.data-retention", "mlsys.deletion-propagation" ], "related_concepts": [ "mlsys.undeclared-consumers", "mlsys.privacy", "mlsys.lineage" ], "taught_in": [ { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" } ] } ] } ] }, { "code": "Q", "id": "qualities", "label": "Qualities", "question": "By which properties is the system judged, and how are they engineered?", "areas": [ { "code": "Q1", "id": "efficiency", "label": "Performance, efficiency & benchmarking", "question": "How do we measure, compare, and reduce the time, memory, and energy a workload consumes?", "scope": "Measuring and benchmarking where time, memory, and energy go, and the cross-layer techniques whose primary purpose is to reduce them, such as compression and hardware-aware design.", "subjects": [ { "code": "Q1.1", "id": "profiling-utilization", "label": "Profiling & utilization", "definition": "Measuring where execution time and resources go under a stated workload, and expressing how much of the hardware's capacity does useful model work.", "includes": [ "Profilers and execution-time breakdowns", "Utilization metrics such as MFU and training goodput", "Bottleneck attribution across compute, memory, I/O, and communication" ], "excludes": [ "Standardized benchmark suites and reporting rules ([[system-benchmarking]])", "Worst-case execution-time bounds ([[real-time-systems]])", "Serving-specific latency and goodput targets ([[latency-goodput]])" ], "contexts": [ "deep-learning", "distributed" ], "concepts": [ "mlsys.profiling", "mlsys.execution-time", "mlsys.utilization", "mlsys.model-flops-utilization", "mlsys.bottleneck-analysis", "mlsys.training-goodput", "mlsys.cross-stack-profiling", "mlsys.hardware-performance-counters", "mlsys.model-bandwidth-utilization" ], "related_concepts": [ "mlsys.roofline", "mlsys.benchmark", "mlsys.slo-goodput" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html" }, { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html" }, { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol2", "chapter": "Fleet Efficiency", "url": "https://mlsysbook.ai/vol2/09_performance_engineering/09_performance_engineering.html" } ] }, { "code": "Q1.2", "id": "system-benchmarking", "label": "System benchmarking", "definition": "Measuring training and inference system performance under specified workloads, quality targets, load scenarios, and run rules so results are comparable.", "includes": [ "MLPerf-style time-to-quality and scenario-based inference metrics." ], "excludes": [ "Profiling a single workload ([[profiling-utilization]])" ], "contexts": [ "distributed", "edge" ], "concepts": [ "mlsys.benchmark", "mlsys.time-to-train", "mlsys.inference-load-scenario", "mlsys.benchmark-representativeness", "mlsys.performance-regression-testing", "mlsys.measurement-variability" ], "related_concepts": [ "mlsys.profiling", "mlsys.energy", "mlsys.slo-goodput", "mlsys.statistical-uncertainty" ], "taught_in": [ { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html" }, { "volume": "vol2", "chapter": "Fleet Efficiency", "url": "https://mlsysbook.ai/vol2/09_performance_engineering/09_performance_engineering.html" } ] }, { "code": "Q1.3", "id": "model-compression", "label": "Model compression", "definition": "Techniques that shrink a trained or training model's parameters, precision, or rank to reduce memory, compute, and latency at a measured cost in quality.", "includes": [ "Quantization (post-training and quantization-aware)", "Pruning (structured and unstructured) and sparsity exploitation", "Knowledge distillation and low-rank factorization" ], "excludes": [ "Numeric format definitions and mixed-precision training ([[numerics]])", "Parameter-efficient fine-tuning such as LoRA, whose purpose is adaptation ([[fine-tuning]])" ], "contexts": [ "deep-learning", "generative", "edge" ], "concepts": [ "mlsys.quantization", "mlsys.sparsity", "mlsys.distillation", "mlsys.pruning", "mlsys.structured-pruning", "mlsys.low-rank-factorization", "mlsys.quantization-aware-training", "mlsys.post-training-quantization", "mlsys.weight-only-quantization", "mlsys.activation-outliers", "mlsys.kv-cache-quantization", "mlsys.nm-sparsity", "mlsys.activation-sparsity", "mlsys.integer-only-inference", "mlsys.quantization-granularity", "mlsys.compression-composition" ], "related_concepts": [ "mlsys.precision", "mlsys.mixed-precision" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" } ] }, { "code": "Q1.4", "id": "hardware-aware-design", "label": "Hardware-aware model design", "definition": "Choosing or searching model architectures against a target device's measured cost, so that quality and efficiency are optimized together rather than compressed after the fact.", "includes": [ "Neural architecture search, including hardware-aware objectives", "Efficient building blocks and the accuracy-efficiency frontier", "Joint model and hardware co-design" ], "excludes": [ "Mixture-of-experts and early exit ([[conditional-computation]])", "Efficient attention and state-space models ([[efficient-sequence-models]])" ], "contexts": [ "deep-learning", "edge" ], "concepts": [ "mlsys.neural-architecture-search", "mlsys.hardware-aware-nas", "mlsys.efficiency-frontier", "mlsys.model-hardware-co-design" ], "related_concepts": [ "mlsys.accelerator-execution-architecture", "mlsys.arithmetic-intensity", "mlsys.depthwise-separable-convolution" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html" } ] } ] }, { "code": "Q2", "id": "reliability", "label": "Reliability & fault tolerance", "question": "How does the system keep working, or fail safely, when components fail at scale?", "scope": "Infrastructure and service fault tolerance, from failure assumptions and recovery semantics to fleet failure rates, checkpoint economics, and resilience testing.", "subjects": [ { "code": "Q2.1", "id": "failure-semantics", "label": "Failure models, retries & degradation", "definition": "Declared failure assumptions and the service-level mechanisms (retries, idempotency, replication, circuit breaking, degradation) that keep requests correct and available when components fail.", "includes": [ "Failure models and recovery semantics (idempotency, compensation)", "Retries, circuit breakers, replication, and failover", "Graceful degradation of quality or features under failure" ], "excludes": [ "Physical safety fallbacks and stopping envelopes ([[runtime-enforcement]])", "Durable agent execution and replay ([[durable-execution]])" ], "contexts": [ "interactive", "distributed", "agentic" ], "concepts": [ "mlsys.failure-model", "mlsys.idempotency", "mlsys.compensation", "mlsys.irreversible-effects", "mlsys.retry-backoff", "mlsys.circuit-breaker", "mlsys.replication-failover", "mlsys.graceful-degradation", "mlsys.health-checks", "mlsys.metastable-failure", "mlsys.request-timeout", "mlsys.in-doubt-action" ], "related_concepts": [ "mlsys.backpressure", "mlsys.admission", "mlsys.runtime-assurance" ], "taught_in": [ { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html" }, { "volume": "vol3", "chapter": "Failure Recovery", "url": "https://mlsysbook.ai/vol3/12_failure_recovery/12_failure_recovery.html" } ] }, { "code": "Q2.2", "id": "checkpoint-recovery", "label": "Checkpointing & elastic recovery", "definition": "Persisting recoverable state for long-running jobs and choosing checkpoint frequency and recovery strategy to minimize lost work and restart cost under a given failure rate.", "includes": [ "Checkpoints and consistent cuts for distributed state", "Optimal checkpoint interval (Young-Daly) and asynchronous checkpointing", "Elastic recovery on a changed set of workers" ], "excludes": [ "Activation recomputation, sometimes called gradient checkpointing ([[training-memory]])", "Checkpoint storage bandwidth and I/O systems ([[storage-systems]])" ], "contexts": [ "distributed", "deep-learning" ], "concepts": [ "mlsys.checkpoint", "mlsys.consistent-cut", "mlsys.checkpoint-interval", "mlsys.async-checkpointing", "mlsys.checkpoint-resharding", "mlsys.incremental-checkpointing", "mlsys.hot-spare-capacity", "mlsys.restart-free-recovery" ], "related_concepts": [ "mlsys.training-state", "mlsys.training-goodput", "mlsys.elastic-training" ], "taught_in": [ { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html" }, { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html" } ] }, { "code": "Q2.3", "id": "hardware-faults", "label": "Hardware faults & silent data corruption", "definition": "How hardware fails in a fleet, from fail-stop crashes to gray and silent faults, how often it fails, and how such faults are detected and contained.", "includes": [ "Fail-stop, gray, and silent failures and their detection", "Fleet failure rates (MTBF composition) and availability", "Correlated failures and failure domains" ], "excludes": [ "Distribution shift and adversarial robustness ([[drift-retraining]]; [[ml-attacks]])", "Chip power and thermal limits as design constraints ([[chip-power-thermal]])" ], "contexts": [ "distributed" ], "concepts": [ "mlsys.silent-data-corruption", "mlsys.mtbf", "mlsys.availability", "mlsys.gray-failure", "mlsys.failure-domain", "mlsys.abft", "mlsys.distributed-hang-diagnosis", "mlsys.flash-endurance", "mlsys.ecc-memory", "mlsys.fleet-hardware-screening" ], "related_concepts": [ "mlsys.stragglers", "mlsys.thermal-dynamics" ], "taught_in": [ { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html" }, { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html" } ] }, { "code": "Q2.4", "id": "fault-injection", "label": "Fault injection & resilience testing", "definition": "Deliberately introducing software, infrastructure, or hardware faults under controlled scope to verify that detection, recovery, and degradation work as designed.", "includes": [ "Chaos engineering and failure injection in production-like settings", "Game days and recovery rehearsals", "Hardware fault injection to estimate model error resilience" ], "excludes": [ "Functional ML tests and test oracles ([[ml-testing]])", "Scenario-based safety validation ([[verification-validation]])" ], "contexts": [ "distributed", "interactive", "physical" ], "concepts": [ "mlsys.failure-injection", "mlsys.chaos-engineering", "mlsys.game-day", "mlsys.hardware-fault-injection", "mlsys.blast-radius" ], "related_concepts": [ "mlsys.incident-response", "mlsys.integration-testing" ], "taught_in": [ { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html" }, { "volume": "vol4", "chapter": "Adversarial Verification", "url": "https://mlsysbook.ai/vol4/15_verification/15_verification.html" } ] } ] }, { "code": "Q3", "id": "security-privacy", "label": "Security & privacy", "question": "How do we protect ML systems and the data they touch from adversaries and leakage?", "scope": "Adversaries and defenses across data, model, artifact, and execution boundaries, and the leakage of information about people or records through learned systems.", "subjects": [ { "code": "Q3.1", "id": "ml-attacks", "label": "Threat models & adversarial attacks", "definition": "Specifying adversaries against learned systems and the attacks they mount on training data, model inputs, and instruction boundaries, with the defenses aimed at each.", "includes": [ "Threat models for ML assets and interfaces", "Poisoning, backdoors, evasion, jailbreaks, and prompt injection", "Defenses such as adversarial training" ], "excludes": [ "Robustness evaluation methodology ([[robustness-slices]])", "Theft and integrity of model artifacts ([[model-supply-chain]])" ], "contexts": [ "deep-learning", "generative", "agentic" ], "concepts": [ "mlsys.threat-model", "mlsys.poisoning", "mlsys.evasion", "mlsys.prompt-injection", "mlsys.backdoor", "mlsys.jailbreak", "mlsys.adversarial-training", "mlsys.red-teaming", "mlsys.byzantine-robust-aggregation", "mlsys.sensor-spoofing", "mlsys.randomized-smoothing", "mlsys.resource-exhaustion-attack", "mlsys.instruction-hierarchy" ], "related_concepts": [ "mlsys.robustness-evaluation" ], "taught_in": [ { "volume": "vol2", "chapter": "Security & Privacy", "url": "https://mlsysbook.ai/vol2/13_security_privacy/13_security_privacy.html" }, { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html" }, { "volume": "vol3", "chapter": "Agent Security", "url": "https://mlsysbook.ai/vol3/09_security/09_security.html" } ] }, { "code": "Q3.2", "id": "model-supply-chain", "label": "Model & supply-chain security", "definition": "Protecting models as valuable artifacts from theft and tampering, and protecting the chain of datasets, pretrained weights, libraries, and builds that produce them.", "includes": [ "Model extraction and ownership evidence (watermarking)", "Artifact signing and provenance", "Malicious weights, serialization formats, and dependencies" ], "excludes": [ "Artifact versioning and registries as release tooling ([[release-management]])", "Data lineage as a data-engineering practice ([[data-lineage]])" ], "contexts": [ "deep-learning", "generative" ], "concepts": [ "mlsys.model-extraction", "mlsys.model-watermarking", "mlsys.artifact-signing", "mlsys.unsafe-deserialization", "mlsys.supply-chain-compromise", "mlsys.weight-exfiltration", "mlsys.ai-bill-of-materials" ], "related_concepts": [ "mlsys.model-artifact", "mlsys.lineage", "mlsys.dependency-environment" ], "taught_in": [ { "volume": "vol2", "chapter": "Security & Privacy", "url": "https://mlsysbook.ai/vol2/13_security_privacy/13_security_privacy.html" } ] }, { "code": "Q3.3", "id": "privacy-dp", "label": "Privacy leakage & differential privacy", "definition": "Privacy requirements for data about people, the attacks that extract it from models and outputs, and formal or architectural mechanisms that bound or remove that leakage.", "includes": [ "Membership inference and model inversion", "Differential privacy and DP-SGD", "Machine unlearning and secure aggregation" ], "excludes": [ "Federated training protocols and client scheduling ([[federated-learning]])", "Privacy regulation and governance processes ([[governance-regulation]])" ], "contexts": [ "federated", "generative", "edge" ], "concepts": [ "mlsys.privacy", "mlsys.differential-privacy", "mlsys.privacy-leakage", "mlsys.membership-inference", "mlsys.model-inversion", "mlsys.dp-sgd", "mlsys.machine-unlearning", "mlsys.secure-aggregation", "mlsys.he-inference", "mlsys.secure-computation-inference", "mlsys.training-data-extraction", "mlsys.pii-deidentification", "mlsys.privacy-budget" ], "related_concepts": [ "mlsys.federated-learning" ], "taught_in": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html" }, { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html" }, { "volume": "vol2", "chapter": "Security & Privacy", "url": "https://mlsysbook.ai/vol2/13_security_privacy/13_security_privacy.html" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html" } ] }, { "code": "Q3.4", "id": "access-isolation", "label": "Access control & isolation", "definition": "Constraining who may act, what code may reach, and which parties can observe data in use, through authorization, least privilege, isolation, and hardware-rooted trust.", "includes": [ "Authorization, least privilege, and sandboxed execution", "Confidential computing and remote attestation", "Side channels across shared accelerators and caches" ], "excludes": [ "Tenant quotas and resource isolation for performance ([[capacity-provisioning]])", "Agent tool contracts and action approval ([[tool-interfaces]])" ], "contexts": [ "agentic", "distributed" ], "concepts": [ "mlsys.authorization", "mlsys.isolation", "mlsys.least-privilege", "mlsys.confidential-computing", "mlsys.remote-attestation", "mlsys.side-channel", "mlsys.capability-based-security", "mlsys.ephemeral-sandbox", "mlsys.egress-control", "mlsys.information-flow-control", "mlsys.agent-credentials", "mlsys.secrets-management", "mlsys.confused-deputy", "mlsys.quarantined-model-pattern" ], "related_concepts": [ "mlsys.tool-contract" ], "taught_in": [ { "volume": "vol2", "chapter": "Security & Privacy", "url": "https://mlsysbook.ai/vol2/13_security_privacy/13_security_privacy.html" }, { "volume": "vol3", "chapter": "Agent Sandboxes", "url": "https://mlsysbook.ai/vol3/08_sandboxes/08_sandboxes.html" }, { "volume": "vol3", "chapter": "Agent Security", "url": "https://mlsysbook.ai/vol3/09_security/09_security.html" } ] } ] }, { "code": "Q4", "id": "safety-assurance", "label": "Safety & assurance", "question": "How do we argue, enforce, and verify that a system will not cause unacceptable harm?", "scope": "Identifying hazards, bounding the operating domain, enforcing safety at runtime, verifying and validating behavior, and arguing acceptable residual risk against recognized standards.", "subjects": [ { "code": "Q4.1", "id": "hazards-operating-domain", "label": "Hazard analysis & operating domains", "definition": "Identifying how a system can contribute to harm, including hazards from functional insufficiency rather than faults, and fixing the operating conditions within which safety claims are made.", "includes": [ "Hazard identification and analysis (STPA, FMEA, fault trees) applied to learned components.", "Operational design domains and triggering conditions for functional insufficiency (SOTIF)." ], "excludes": [ "Adversary-driven threat modeling ([[ml-attacks]])", "Infrastructure fault models ([[failure-semantics]])" ], "contexts": [ "physical", "agentic", "deep-learning" ], "concepts": [ "mlsys.operating-domain", "mlsys.hazard-analysis", "mlsys.stpa", "mlsys.fmea", "mlsys.functional-insufficiency", "mlsys.fault-tree-analysis", "mlsys.tolerable-risk" ], "related_concepts": [ "mlsys.failure-model", "mlsys.threat-model", "mlsys.distribution-shift" ], "taught_in": [ { "volume": "vol4", "chapter": "The Causal Boundary", "url": "https://mlsysbook.ai/vol4/01_boundary/01_boundary.html" }, { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html" }, { "volume": "vol4", "chapter": "Adversarial Verification", "url": "https://mlsysbook.ai/vol4/15_verification/15_verification.html" } ] }, { "code": "Q4.2", "id": "runtime-enforcement", "label": "Runtime assurance & safety filters", "definition": "Mechanisms that monitor a learned component at run time and override, constrain, or replace its outputs so the system stays within a verified safe set.", "includes": [ "Safety filters, control barrier functions, Simplex switching, and fallback to a minimal-risk condition.", "Input and output guardrails that block or rewrite unsafe generated content or actions." ], "excludes": [ "Human approval of agent actions ([[agent-harness]])", "Retry, failover, and graceful degradation for availability ([[reliability]])" ], "contexts": [ "physical", "agentic", "generative", "interactive" ], "concepts": [ "mlsys.runtime-assurance", "mlsys.stopping-envelope", "mlsys.safety-filter", "mlsys.control-barrier-function", "mlsys.simplex-architecture", "mlsys.minimal-risk-condition", "mlsys.guardrails", "mlsys.over-refusal", "mlsys.reachability-analysis" ], "related_concepts": [ "mlsys.trajectory-feasibility", "mlsys.selective-prediction", "mlsys.approval-gate" ], "taught_in": [ { "volume": "vol4", "chapter": "Safety Enforcement", "url": "https://mlsysbook.ai/vol4/12_enforcement/12_enforcement.html" }, { "volume": "vol4", "chapter": "Workload Placement", "url": "https://mlsysbook.ai/vol4/13_placement/13_placement.html" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html" } ] }, { "code": "Q4.3", "id": "verification-validation", "label": "Safety verification & validation", "definition": "Producing evidence that a system meets its safety requirements across its operating domain, through scenario testing, falsification, formal proof, and statistical validation of rare failures.", "includes": [ "Scenario coverage, simulation-based falsification, and software/hardware-in-the-loop qualification.", "Formal and statistical guarantees, including verification of neural networks and rare-event estimation." ], "excludes": [ "Average-case model metrics and benchmarks ([[evaluation]])" ], "contexts": [ "physical", "deep-learning" ], "concepts": [ "mlsys.scenario-based-testing", "mlsys.falsification", "mlsys.formal-verification", "mlsys.neural-network-verification", "mlsys.rare-event-validation", "mlsys.xil-testing" ], "related_concepts": [ "mlsys.robustness-evaluation", "mlsys.failure-injection", "mlsys.sim-to-real", "mlsys.test-oracle" ], "taught_in": [ { "volume": "vol4", "chapter": "Closed-Loop Evaluation", "url": "https://mlsysbook.ai/vol4/07_evaluation/07_evaluation.html" }, { "volume": "vol4", "chapter": "Adversarial Verification", "url": "https://mlsysbook.ai/vol4/15_verification/15_verification.html" } ] }, { "code": "Q4.4", "id": "safety-cases", "label": "Safety cases, standards & certification", "definition": "Structured arguments that a system is acceptably safe for a defined use, the standards that set what evidence is required, and the field indicators that keep the argument valid after release.", "includes": [ "Safety-case structure and notation (claims, arguments, evidence; GSN).", "Functional-safety and autonomy standards (ISO 26262, IEC 61508, ISO 21448, UL 4600) and integrity levels." ], "excludes": [ "Regulatory compliance not specific to safety, such as the EU AI Act ([[governance-regulation]])" ], "contexts": [ "physical", "agentic" ], "concepts": [ "mlsys.safety-case", "mlsys.goal-structuring-notation", "mlsys.functional-safety", "mlsys.safety-integrity-level", "mlsys.safety-performance-indicator", "mlsys.event-data-recorder", "mlsys.change-impact-analysis", "mlsys.safety-certification" ], "related_concepts": [ "mlsys.operating-domain", "mlsys.rollout", "mlsys.artifact-version" ], "taught_in": [ { "volume": "vol4", "chapter": "Deployment Release", "url": "https://mlsysbook.ai/vol4/16_release/16_release.html" }, { "volume": "vol4", "chapter": "Engineering Physical AI", "url": "https://mlsysbook.ai/vol4/18_frontier/18_frontier.html" }, { "volume": "vol3", "chapter": "Operating Agents", "url": "https://mlsysbook.ai/vol3/14_agent_operations/14_agent_operations.html" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html" } ] } ] }, { "code": "Q5", "id": "responsible-ai", "label": "Responsible AI", "question": "How do we ensure the system's outcomes are fair, accountable, transparent, and governed?", "scope": "Measuring and mitigating unfair outcomes, explaining and documenting models and data, keeping people able to oversee and contest decisions, and governing AI risk under organizational and legal regimes.", "subjects": [ { "code": "Q5.1", "id": "fairness-bias", "label": "Fairness & bias", "definition": "Specifying, measuring, and mitigating how a system's outcomes and errors differ across people and groups.", "includes": [ "Group and individual fairness criteria and the metrics that measure them.", "Pre-, in-, and post-processing mitigation and its accuracy cost." ], "excludes": [ "Per-slice measurement machinery in general ([[robustness-slices]])", "Scheduler fair sharing ([[cluster-scheduling]])" ], "contexts": [ "classical-ml", "deep-learning", "generative" ], "concepts": [ "mlsys.fairness", "mlsys.group-fairness-metric", "mlsys.individual-fairness", "mlsys.fairness-incompatibility", "mlsys.bias-mitigation", "mlsys.representational-bias-evaluation", "mlsys.proxy-discrimination", "mlsys.bias-sources", "mlsys.fairness-monitoring" ], "related_concepts": [ "mlsys.sliced-evaluation", "mlsys.feedback-loop", "mlsys.proxy-objective", "mlsys.data-acquisition" ], "taught_in": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html" }, { "volume": "vol4", "chapter": "Responsible Physical AI", "url": "https://mlsysbook.ai/vol4/17_responsibility/17_responsibility.html" } ] }, { "code": "Q5.2", "id": "transparency-explainability", "label": "Explainability & documentation", "definition": "Making model behavior understandable to the people who build, deploy, and are affected by it, through explanations, interpretable models, and standardized documentation of models and data.", "includes": [ "Post-hoc explanations (feature attribution, counterfactuals) and inherently interpretable models.", "Model cards, datasheets, and the serving cost of producing explanations." ], "excludes": [ "Debugging-oriented model inspection ([[ml-testing]])", "Data provenance records ([[data-lineage]])" ], "contexts": [ "classical-ml", "deep-learning", "generative" ], "concepts": [ "mlsys.feature-attribution", "mlsys.counterfactual-explanation", "mlsys.interpretable-model", "mlsys.model-card", "mlsys.datasheet", "mlsys.mechanistic-interpretability" ], "related_concepts": [ "mlsys.lineage", "mlsys.calibration" ], "taught_in": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html" } ] }, { "code": "Q5.3", "id": "human-oversight", "label": "Human oversight & contestability", "definition": "Arrangements that let people supervise, intervene in, take over from, and challenge consequential system decisions, and the human-factors limits on doing so.", "includes": [ "Human-in-the-loop and on-the-loop control, recourse and appeal, and control handover.", "Automation bias, complacency, and the ironies of automation." ], "excludes": [ "The mechanics of agent approval gates ([[agent-harness]])" ], "contexts": [ "classical-ml", "agentic", "physical", "interactive" ], "concepts": [ "mlsys.human-oversight", "mlsys.contestability", "mlsys.automation-bias", "mlsys.authority-handover", "mlsys.human-ai-interaction" ], "related_concepts": [ "mlsys.approval-gate", "mlsys.selective-prediction", "mlsys.irreversible-effects" ], "taught_in": [ { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html" }, { "volume": "vol4", "chapter": "Supervisory Intervention", "url": "https://mlsysbook.ai/vol4/14_intervention/14_intervention.html" }, { "volume": "vol4", "chapter": "Responsible Physical AI", "url": "https://mlsysbook.ai/vol4/17_responsibility/17_responsibility.html" } ] }, { "code": "Q5.4", "id": "governance-regulation", "label": "AI governance, audit & regulation", "definition": "Organizational and legal mechanisms that identify, assess, and control AI risk across a system's life, including risk-management frameworks, impact assessments, audits, and regulation.", "includes": [ "NIST AI RMF and ISO/IEC 42001 management systems; internal and external audits.", "Risk-tiered regulation (EU AI Act) and frontier-model capability thresholds." ], "excludes": [ "Data-protection mechanisms such as differential privacy ([[security-privacy]])", "Safety-standard certification ([[safety-cases]])" ], "contexts": [ "classical-ml", "deep-learning", "generative", "agentic" ], "concepts": [ "mlsys.ai-risk-management", "mlsys.impact-assessment", "mlsys.algorithmic-audit", "mlsys.risk-tiered-regulation", "mlsys.frontier-safety-framework", "mlsys.decision-accountability", "mlsys.content-provenance", "mlsys.generated-content-watermarking", "mlsys.data-protection-regulation", "mlsys.ai-incident-reporting" ], "related_concepts": [ "mlsys.safety-case", "mlsys.privacy", "mlsys.incident-response" ], "taught_in": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html" }, { "volume": "vol4", "chapter": "Deployment Release", "url": "https://mlsysbook.ai/vol4/16_release/16_release.html" }, { "volume": "vol4", "chapter": "Responsible Physical AI", "url": "https://mlsysbook.ai/vol4/17_responsibility/17_responsibility.html" } ] } ] }, { "code": "Q6", "id": "sustainability-cost", "label": "Sustainability & cost", "question": "What does the system cost in energy, carbon, water, and money over its lifetime?", "scope": "Accounting for environmental and monetary consequences of ML systems under declared boundaries, and the decisions that reduce them per useful outcome.", "subjects": [ { "code": "Q6.1", "id": "energy-carbon", "label": "Energy & carbon accounting", "definition": "Measuring the energy a workload consumes and converting it to operational emissions under declared boundaries, and scheduling work to reduce both.", "includes": [ "Energy measurement and energy efficiency (work per joule)", "Grid carbon intensity and operational carbon", "Carbon-aware placement and scheduling" ], "excludes": [ "Facility power, cooling, and PUE ([[datacenter-facilities]])", "Manufacturing emissions and hardware end of life ([[embodied-lifecycle]])" ], "contexts": [ "distributed", "generative", "batch" ], "concepts": [ "mlsys.energy", "mlsys.carbon", "mlsys.energy-efficiency", "mlsys.carbon-intensity", "mlsys.operational-carbon", "mlsys.emission-scopes", "mlsys.carbon-aware-scheduling", "mlsys.energy-per-inference", "mlsys.energy-measurement", "mlsys.idle-dynamic-power", "mlsys.rebound-effect", "mlsys.carbon-accounting-method" ], "related_concepts": [ "mlsys.pue", "mlsys.power-delivery" ], "taught_in": [ { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html" }, { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html" }, { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html" } ] }, { "code": "Q6.2", "id": "embodied-lifecycle", "label": "Embodied carbon & lifecycle impact", "definition": "Environmental impacts beyond operating energy, from manufacturing emissions and water to hardware lifetime and disposal, assessed over the full hardware lifecycle.", "includes": [ "Embodied carbon and lifecycle assessment", "Water consumption of training and serving", "Hardware lifetime, refresh, and e-waste" ], "excludes": [ "Operational emissions from electricity ([[energy-carbon]])", "Model retirement and decommissioning ([[retirement]])" ], "contexts": [ "edge", "distributed" ], "concepts": [ "mlsys.embodied-carbon", "mlsys.lifecycle-assessment", "mlsys.water-footprint", "mlsys.hardware-lifetime", "mlsys.e-waste" ], "related_concepts": [ "mlsys.energy", "mlsys.carbon", "mlsys.operational-carbon" ], "taught_in": [ { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html" } ] }, { "code": "Q6.3", "id": "cost-economics", "label": "Cost, TCO & unit economics", "definition": "Monetary accounting for ML systems, from total cost of ownership to cost per token or accepted outcome, and the architectural choices that lower cost at fixed quality.", "includes": [ "Total cost of ownership and cost attribution (FinOps)", "Cost per token and cost per accepted outcome", "Model routing and cascades chosen for cost" ], "excludes": [ "Capacity provisioning and quotas ([[capacity-provisioning]])", "Spot and preemption scheduling mechanics ([[cluster-scheduling]])" ], "contexts": [ "generative", "agentic", "interactive" ], "concepts": [ "mlsys.unit-cost", "mlsys.tco", "mlsys.cost-per-token", "mlsys.cost-attribution", "mlsys.data-transfer-cost", "mlsys.compute-cost-estimate" ], "related_concepts": [ "mlsys.capacity-planning", "mlsys.prefix-cache", "mlsys.model-routing" ], "taught_in": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html" }, { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html" }, { "volume": "vol2", "chapter": "Fleet Orchestration", "url": "https://mlsysbook.ai/vol2/08_fleet_orchestration/08_fleet_orchestration.html" }, { "volume": "vol2", "chapter": "Fleet Efficiency", "url": "https://mlsysbook.ai/vol2/09_performance_engineering/09_performance_engineering.html" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html" }, { "volume": "vol3", "chapter": "Agent Economics", "url": "https://mlsysbook.ai/vol3/19_agent_economics/19_agent_economics.html" } ] } ] } ] }, { "code": "T", "id": "system-types", "label": "System types", "question": "What is distinctive when the system acts as an agent or in the physical world?", "areas": [ { "code": "T1", "id": "agentic", "label": "Agentic systems", "question": "What does it take to engineer models that plan and act through tools over many steps?", "scope": "Context engineering and agent memory, tool and action interfaces, the agent harness and its control loop, durable agent state, and multi-agent coordination.", "subjects": [ { "code": "T1.1", "id": "context-memory", "label": "Context engineering & agent memory", "definition": "Deciding what information enters each model call over a long task, and maintaining persistent memory the agent can write, consolidate, and retrieve across calls and sessions.", "includes": [ "Working-set selection, ordering for prefix reuse, compaction, and long-context degradation.", "Long-term memory write paths, consolidation, expiry, and invalidation." ], "excludes": [ "Retrieval and vector indexes in general ([[retrieval-rag]])", "KV-cache management ([[llm-inference]])" ], "contexts": [ "agentic", "generative", "interactive" ], "concepts": [ "mlsys.context-compaction", "mlsys.agent-memory", "mlsys.context-selection", "mlsys.context-staleness" ], "related_concepts": [ "mlsys.retrieval", "mlsys.prefix-cache", "mlsys.kv-cache", "mlsys.prompt-injection", "mlsys.context", "mlsys.context-degradation" ], "taught_in": [ { "volume": "vol3", "chapter": "Context Engineering", "url": "https://mlsysbook.ai/vol3/04_context_engineering/04_context_engineering.html" }, { "volume": "vol3", "chapter": "Long-Term Memory", "url": "https://mlsysbook.ai/vol3/06_long_term_memory/06_long_term_memory.html" } ] }, { "code": "T1.2", "id": "tool-interfaces", "label": "Tool use & action interfaces", "definition": "The interfaces through which a model acts on the world, including structured tool calls, tool discovery protocols, computer use, and the effects and authority those actions carry.", "includes": [ "Tool schemas, preconditions, function calling, and MCP-style discovery.", "Computer-use agents and the handling of irreversible effects." ], "excludes": [ "Sandboxing and capability isolation ([[access-isolation]])", "Prompt injection ([[ml-attacks]])" ], "contexts": [ "agentic", "generative" ], "concepts": [ "mlsys.tool-contract", "mlsys.function-calling", "mlsys.model-context-protocol", "mlsys.computer-use" ], "related_concepts": [ "mlsys.authorization", "mlsys.isolation", "mlsys.idempotency", "mlsys.prompt-injection", "mlsys.grammar" ], "taught_in": [ { "volume": "vol3", "chapter": "The Foundation Model", "url": "https://mlsysbook.ai/vol3/02_foundation_model/02_foundation_model.html" }, { "volume": "vol3", "chapter": "Tool Calling", "url": "https://mlsysbook.ai/vol3/07_tool_calling/07_tool_calling.html" } ] }, { "code": "T1.3", "id": "agent-harness", "label": "Agent harness & control loop", "definition": "The runtime that owns an agent's loop of model calls and tool actions, enforcing budgets, stopping rules, and approval gates and recording the trajectory.", "includes": [ "Reason-act loops, workflows versus model-directed loops, and plan revision.", "Turn, token, time, and spend budgets; stopping rules; human approval gates." ], "excludes": [ "Inference-time search and reasoning budgets per call ([[test-time-compute]])", "Trajectory-level evaluation ([[generative-agent-eval]])" ], "contexts": [ "agentic", "generative", "interactive" ], "concepts": [ "mlsys.trajectory", "mlsys.planning", "mlsys.agent-harness", "mlsys.agent-budget", "mlsys.approval-gate", "mlsys.agent-interruption", "mlsys.agent-loop-detection", "mlsys.agentic-workload", "mlsys.model-directed-control-flow" ], "related_concepts": [ "mlsys.agent-evaluation", "mlsys.human-oversight", "mlsys.unit-cost", "mlsys.decision-policy" ], "taught_in": [ { "volume": "vol3", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol3/01_introduction/01_introduction.html" }, { "volume": "vol3", "chapter": "The Agent Harness", "url": "https://mlsysbook.ai/vol3/10_agent_harness/10_agent_harness.html" } ] }, { "code": "T1.4", "id": "durable-execution", "label": "Durable execution & agent state", "definition": "Persisting an agent's progress so long-running trajectories survive crashes, pauses, and redeployments without repeating completed effects, and can be replayed for debugging.", "includes": [ "Event-sourced trajectory logs, resume protocols, and record-and-replay.", "Pausing for human or external events and resuming safely." ], "excludes": [ "General checkpointing, idempotency, and sagas ([[reliability]])" ], "contexts": [ "agentic", "distributed" ], "concepts": [ "mlsys.durable-execution", "mlsys.deterministic-replay", "mlsys.suspend-resume", "mlsys.trajectory-log" ], "related_concepts": [ "mlsys.checkpoint", "mlsys.idempotency", "mlsys.compensation", "mlsys.fencing", "mlsys.workflow-orchestration" ], "taught_in": [ { "volume": "vol3", "chapter": "Durable Execution", "url": "https://mlsysbook.ai/vol3/11_durable_execution/11_durable_execution.html" }, { "volume": "vol3", "chapter": "Failure Recovery", "url": "https://mlsysbook.ai/vol3/12_failure_recovery/12_failure_recovery.html" } ] }, { "code": "T1.5", "id": "multi-agent", "label": "Multi-agent orchestration", "definition": "Decomposing a task across several model-driven agents and coordinating their delegation, shared state, and results, including when coordination costs outweigh the benefit.", "includes": [ "Orchestrator-worker and peer topologies, typed task envelopes, and delegation.", "Coordination failures and correlated errors among agents sharing a base model." ], "excludes": [ "Distributed consensus for infrastructure state ([[coordination]])" ], "contexts": [ "agentic", "generative", "distributed" ], "concepts": [ "mlsys.multi-agent-orchestration", "mlsys.shared-agent-state", "mlsys.coordination-failure", "mlsys.correlated-agent-errors" ], "related_concepts": [ "mlsys.baseline", "mlsys.agent-evaluation", "mlsys.fencing" ], "taught_in": [ { "volume": "vol3", "chapter": "Multi-Agent Coordination", "url": "https://mlsysbook.ai/vol3/18_multi_agent/18_multi_agent.html" } ] } ] }, { "code": "T2", "id": "embodied", "label": "Embodied & physical systems", "question": "What changes when ML closes the loop with sensors, actuators, and the physical world?", "scope": "Sensing and perception, state estimation and mapping, motion planning and control, real-time and mixed-criticality computing, and robot learning from demonstration and simulation.", "subjects": [ { "code": "T2.1", "id": "perception-sensing", "label": "Sensing, perception & fusion", "definition": "Turning time-stamped measurements from physical sensors into task-relevant estimates of the scene, including calibration, synchronization, and fusion across modalities.", "includes": [ "Camera, LiDAR, radar, event, and tactile sensing; calibration chains.", "Detection, segmentation, learned depth, and multi-sensor fusion." ], "excludes": [ "Sensor spoofing ([[ml-attacks]])" ], "contexts": [ "physical", "edge", "streaming" ], "concepts": [ "mlsys.sensing", "mlsys.robot-perception", "mlsys.sensor-fusion", "mlsys.time-synchronization", "mlsys.perception-latency-budget", "mlsys.sensor-interference", "mlsys.coordinate-frames", "mlsys.sensor-degradation", "mlsys.capture-epoch" ], "related_concepts": [ "mlsys.observation-age", "mlsys.distribution-shift", "mlsys.streaming-asr", "mlsys.voice-pipeline" ], "taught_in": [ { "volume": "vol4", "chapter": "Sensor Perception", "url": "https://mlsysbook.ai/vol4/08_perception/08_perception.html" } ] }, { "code": "T2.2", "id": "state-estimation", "label": "State estimation, localization & mapping", "definition": "Inferring a robot's and its environment's latent state and uncertainty over time, through recursive filtering, localization and mapping, and learned predictive world models.", "includes": [ "Kalman and particle filters, tracking and data association, SLAM, metric and semantic maps.", "Learned world models used for prediction and planning." ], "excludes": [ "Software observability of services ([[monitoring]])" ], "contexts": [ "physical", "edge", "streaming" ], "concepts": [ "mlsys.state-estimation", "mlsys.state-observability", "mlsys.recursive-filtering", "mlsys.slam", "mlsys.world-model", "mlsys.multi-object-tracking", "mlsys.motion-forecasting", "mlsys.odometry", "mlsys.map-representation" ], "related_concepts": [ "mlsys.sensing", "mlsys.calibration" ], "taught_in": [ { "volume": "vol4", "chapter": "Spatial Memory", "url": "https://mlsysbook.ai/vol4/09_memory/09_memory.html" }, { "volume": "vol4", "chapter": "Engineering Physical AI", "url": "https://mlsysbook.ai/vol4/18_frontier/18_frontier.html" } ] }, { "code": "T2.3", "id": "planning-control", "label": "Motion planning & control", "definition": "Generating feasible, collision-free motions and closing feedback loops that drive a physical system along them under dynamics, contact, and actuation limits.", "includes": [ "Sampling- and optimization-based motion planning, task and motion planning.", "Feedback control, MPC, and contact-rich manipulation." ], "excludes": [ "Runtime safety filters over a planner's output ([[runtime-enforcement]])" ], "contexts": [ "physical", "edge" ], "concepts": [ "mlsys.dynamics", "mlsys.feedback-control", "mlsys.trajectory-feasibility", "mlsys.contact-compliance", "mlsys.motion-planning", "mlsys.model-predictive-control", "mlsys.actuator-limits", "mlsys.task-and-motion-planning", "mlsys.system-identification", "mlsys.robot-kinematics", "mlsys.loop-stability-margin", "mlsys.impedance-force-control" ], "related_concepts": [ "mlsys.stopping-envelope", "mlsys.planning" ], "taught_in": [ { "volume": "vol4", "chapter": "The Physical Body", "url": "https://mlsysbook.ai/vol4/02_body/02_body.html" }, { "volume": "vol4", "chapter": "Grounded Intent", "url": "https://mlsysbook.ai/vol4/10_intent/10_intent.html" }, { "volume": "vol4", "chapter": "Trajectory Planning", "url": "https://mlsysbook.ai/vol4/11_planning/11_planning.html" } ] }, { "code": "T2.4", "id": "real-time-systems", "label": "Real-time & mixed-criticality computing", "definition": "Guaranteeing that sensing, inference, and control complete within deadlines on shared embedded hardware, and isolating safety-critical functions from less trusted ones.", "includes": [ "Multi-rate scheduling, response-time analysis, WCET, and real-time buses.", "Mixed-criticality partitioning and hypervisors on shared silicon." ], "excludes": [ "Throughput-oriented profiling and utilization ([[profiling-utilization]])" ], "contexts": [ "physical", "edge", "streaming" ], "concepts": [ "mlsys.observation-age", "mlsys.feedback-delay", "mlsys.wcet", "mlsys.real-time-scheduling", "mlsys.mixed-criticality", "mlsys.watchdog-timer", "mlsys.robotics-middleware", "mlsys.onboard-compute-partitioning", "mlsys.interrupt-latency", "mlsys.priority-inversion", "mlsys.timing-jitter", "mlsys.deterministic-field-network" ], "related_concepts": [ "mlsys.execution-time", "mlsys.placement", "mlsys.thermal-dynamics", "mlsys.isolation" ], "taught_in": [ { "volume": "vol4", "chapter": "The Nervous System", "url": "https://mlsysbook.ai/vol4/04_nervous/04_nervous.html" }, { "volume": "vol4", "chapter": "Workload Placement", "url": "https://mlsysbook.ai/vol4/13_placement/13_placement.html" } ] }, { "code": "T2.5", "id": "robot-learning", "label": "Robot learning & sim-to-real", "definition": "Learning robot policies from teleoperated demonstrations, simulation, and pretrained vision-language backbones, and transferring them to physical hardware.", "includes": [ "Vision-language-action models, action chunking, and cross-embodiment pretraining.", "Teleoperation data collection, domain randomization, and sim-to-real transfer." ], "excludes": [ "Imitation and reinforcement learning as general paradigms ([[learning-paradigms]])", "Closed-loop robot evaluation statistics ([[evaluation]])" ], "contexts": [ "physical", "deep-learning", "edge" ], "concepts": [ "mlsys.sim-to-real", "mlsys.vla-model", "mlsys.teleoperation-data", "mlsys.action-chunking", "mlsys.domain-randomization", "mlsys.diffusion-policy", "mlsys.parallel-simulation", "mlsys.action-tokenization", "mlsys.cross-embodiment-training", "mlsys.dual-rate-hierarchical-policy" ], "related_concepts": [ "mlsys.imitation-learning", "mlsys.reinforcement-learning", "mlsys.distribution-shift" ], "taught_in": [ { "volume": "vol4", "chapter": "The Cognitive Brain", "url": "https://mlsysbook.ai/vol4/03_brain/03_brain.html" }, { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html" }, { "volume": "vol4", "chapter": "Policy Training", "url": "https://mlsysbook.ai/vol4/06_training/06_training.html" } ] } ] } ] } ], "concepts": [ { "id": "mlsys.requirements", "label": "System requirements", "aliases": [ "acceptance criteria", "non-functional requirements" ], "definition": "Measurable desired behavior and constraints for a specified workload.", "background": false, "home": "F1.1", "home_id": "requirements-slos", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 40, "organizational": 2, "user_proxy": 19 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Lifecycle", "url": "https://mlsysbook.ai/vol1/03_ml_lifecycle/03_ml_lifecycle.html#sec-ml-workflow-balancing-competing-constraints-b92a" } ] }, { "id": "mlsys.baseline", "label": "Baseline", "aliases": [], "definition": "A reference method or system used to justify the value of a proposed change.", "background": false, "home": "F1.1", "home_id": "requirements-slos", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 93, "organizational": 3, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.proxy-objective", "label": "Proxy objective", "aliases": [], "definition": "A measurable target used in place of the outcome actually desired.", "background": false, "home": "F1.3", "home_id": "objectives-decisions", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 2, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.slo", "label": "Service-level objective", "aliases": [ "SLO" ], "definition": "A target for a defined service measure over a specified population and window.", "background": false, "home": "F1.1", "home_id": "requirements-slos", "added_in": "0.1-draft", "warrant": { "literary": 4, "corpus_chapters": 24, "organizational": 4, "user_proxy": 6 }, "book_passages": [ { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html#sec-benchmarking-latency-tail-latency-5cde" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html#sec-ops-scale-end-to-end-evaluation-contract" } ] }, { "id": "mlsys.workload", "label": "Workload characterization", "aliases": [], "definition": "The input, request, model, and execution distributions that a system must support.", "background": false, "home": "F1.2", "home_id": "workload-characterization", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.placement", "label": "Execution placement", "aliases": [], "definition": "The choice of where system components execute, such as device, edge, or datacenter.", "background": false, "home": "F1.2", "home_id": "workload-characterization", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 3, "user_proxy": 48 }, "book_passages": [] }, { "id": "mlsys.data-quality", "label": "Data quality", "aliases": [], "definition": "Fitness of recorded examples for the intended task, including validity, coverage, and error.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 20, "organizational": 1, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.data-acquisition", "label": "Data acquisition and sampling", "aliases": [], "definition": "The observation and selection process that determines which examples and populations enter a dataset.", "background": false, "home": "P1.1", "home_id": "data-collection", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 20 }, "book_passages": [] }, { "id": "mlsys.features", "label": "Feature engineering", "aliases": [], "definition": "Constructing model inputs whose semantics and availability match the task.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.missing-data", "label": "Missing data", "aliases": [], "definition": "Absent observations whose pattern and treatment can change learned behavior.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.leakage", "label": "Data leakage", "aliases": [], "definition": "Information unavailable at the intended prediction boundary entering development or evaluation.", "background": false, "home": "P1.6", "home_id": "data-leakage-skew", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 3, "user_proxy": 4 }, "book_passages": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html#sec-data-engineering-data-engineering-dataset-compilation-0496" } ] }, { "id": "mlsys.point-in-time", "label": "Point-in-time correctness", "aliases": [], "definition": "Constructing examples using only information available at their stated prediction time.", "background": false, "home": "P1.6", "home_id": "data-leakage-skew", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 2, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.lineage", "label": "Data lineage", "aliases": [], "definition": "Recorded dependencies that identify the sources and transformations of a dataset or artifact.", "background": false, "home": "P1.5", "home_id": "data-lineage", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 9, "organizational": 10, "user_proxy": 15 }, "book_passages": [ { "volume": "vol3", "chapter": "Long-Term Memory", "url": "https://mlsysbook.ai/vol3/06_long_term_memory/06_long_term_memory.html#sec-vol3-persistent-result-envelope" } ] }, { "id": "mlsys.train-serve-skew", "label": "Training-serving skew", "aliases": [], "definition": "A mismatch between the data transformations or feature semantics used in training and serving.", "background": false, "home": "P1.6", "home_id": "data-leakage-skew", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 12, "organizational": 4, "user_proxy": 32 }, "book_passages": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html#sec-data-engineering-ensuring-trainingserving-consistency-c683" } ] }, { "id": "mlsys.data-selection", "label": "Data selection", "aliases": [], "definition": "Allocating learning effort to examples according to value, coverage, and acquisition cost.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 18, "organizational": 4, "user_proxy": 56 }, "book_passages": [ { "volume": "vol1", "chapter": "Data Selection", "url": "https://mlsysbook.ai/vol1/09_data_selection/09_data_selection.html#sec-data-selection-informationcompute-ratio-8c0b" }, { "volume": "vol1", "chapter": "Data Selection", "url": "https://mlsysbook.ai/vol1/09_data_selection/09_data_selection.html#sec-data-selection-data-selection-fundamentals-e839" } ] }, { "id": "mlsys.deduplication", "label": "Dataset deduplication", "aliases": [], "definition": "Finding and managing exact or near-duplicate examples within a corpus and overlaps with evaluation data to control sample weighting, redundant compute, memorization, and contamination.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 17 }, "book_passages": [] }, { "id": "mlsys.embeddings", "label": "Embeddings", "aliases": [], "definition": "Vector representations of tokens, items, or inputs, learned or constructed, whose dimension and lookup or encoder implementation determine storage and computation costs.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 33, "organizational": 1, "user_proxy": 23 }, "book_passages": [] }, { "id": "mlsys.distribution-shift", "label": "Distribution shift", "aliases": [ "dataset shift", "covariate shift", "label shift" ], "definition": "A difference between development and deployment data-generating distributions.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 37, "organizational": 6, "user_proxy": 41 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html#sec-ml-operations-model-infrastructure-monitoring-3988" }, { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html#sec-robust-ai-distribution-shift-concept-drift-55e2" }, { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html#sec-data-covariate-shift" } ] }, { "id": "mlsys.linear-model", "label": "Linear models", "aliases": [ "logistic regression", "linear regression" ], "definition": "Models whose score is linear in their supplied features, with a possible output link function.", "background": true, "home": "F2.1", "home_id": "model-families", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.tree-ensemble", "label": "Tree ensembles", "aliases": [ "gradient-boosted decision trees", "GBDT", "random forest", "XGBoost" ], "definition": "Collections of decision trees whose branch traversal and aggregation determine predictions.", "background": true, "home": "F2.1", "home_id": "model-families", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.kernel-method", "label": "Kernel methods", "aliases": [ "support vector machine", "SVM" ], "definition": "Models that score an input by a kernel similarity to retained training examples or to an explicit feature approximation, so storage and evaluation cost grow with the number of retained examples unless the kernel is approximated.", "background": true, "home": "F2.1", "home_id": "model-families", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.nearest-neighbor", "label": "Nearest-neighbor models", "aliases": [ "k-nearest neighbors", "kNN" ], "definition": "Predictions or retrieval based on similarity to stored examples or vectors.", "background": true, "home": "F2.1", "home_id": "model-families", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.neural-network", "label": "Neural networks", "aliases": [], "definition": "Parameterized compositions of differentiable operations used to learn representations and predictions.", "background": true, "home": "F2.1", "home_id": "model-families", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 37, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.inductive-bias", "label": "Inductive bias", "aliases": [], "definition": "Structural assumptions that favor some learned functions over others.", "background": true, "home": "F2.1", "home_id": "model-families", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 2, "user_proxy": 7 }, "book_passages": [ { "volume": "vol1", "chapter": "Network Architectures", "url": "https://mlsysbook.ai/vol1/06_nn_architectures/06_nn_architectures.html#sec-network-architectures-architectural-principles-engineering-tradeoffs-362e" } ] }, { "id": "mlsys.attention", "label": "Attention", "aliases": [ "self-attention", "multi-head attention", "cross-attention", "scaled dot-product attention" ], "definition": "Input-dependent aggregation of value vectors using query-key compatibility scores.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 70, "organizational": 2, "user_proxy": 57 }, "book_passages": [ { "volume": "vol1", "chapter": "Network Architectures", "url": "https://mlsysbook.ai/vol1/06_nn_architectures/06_nn_architectures.html#sec-network-architectures-architectural-principles-engineering-tradeoffs-362e" } ] }, { "id": "mlsys.position-encoding", "label": "Position encoding", "aliases": [ "Positional encoding", "rotary position embedding", "RoPE", "ALiBi", "relative position encoding" ], "definition": "Representations that make sequence position available to a model.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.tokenization", "label": "Tokenization", "aliases": [ "subword tokenization", "BPE" ], "definition": "Mapping text into vocabulary-dependent token identifiers, which sets sequence lengths and so influences context usage, embedding size, and processing cost.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 24, "organizational": 1, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.loss", "label": "Training loss", "aliases": [], "definition": "A computable criterion optimized on training examples.", "background": true, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 24, "organizational": 1, "user_proxy": 4 }, "book_passages": [ { "volume": "vol1", "chapter": "Neural Computation", "url": "https://mlsysbook.ai/vol1/05_nn_computation/05_nn_computation.html#sec-neural-computation-minibatch-gradient-updates-4a04" } ] }, { "id": "mlsys.autodiff", "label": "Automatic differentiation", "aliases": [], "definition": "Computing derivatives by composing local derivative rules through an execution graph.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 4, "user_proxy": 45 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Frameworks", "url": "https://mlsysbook.ai/vol1/07_frameworks/07_frameworks.html#sec-ml-frameworks-differentiation-problem-8b8a" } ] }, { "id": "mlsys.optimizer", "label": "Optimization algorithm", "aliases": [], "definition": "A rule that uses training information to update model parameters and optimizer state.", "background": true, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 1, "user_proxy": 7 }, "book_passages": [ { "volume": "vol1", "chapter": "Neural Computation", "url": "https://mlsysbook.ai/vol1/05_nn_computation/05_nn_computation.html#sec-neural-computation-minibatch-gradient-updates-4a04" } ] }, { "id": "mlsys.regularization", "label": "Regularization", "aliases": [], "definition": "Constraints or training procedures used to control sensitivity and generalization behavior.", "background": true, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 18, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.batch-size", "label": "Training batch size", "aliases": [], "definition": "The number of examples contributing to an optimization update or microstep.", "background": false, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 1, "user_proxy": 17 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html#sec-model-training-batch-size-parameter-updates-4d0b" } ] }, { "id": "mlsys.training-state", "label": "Training state", "aliases": [], "definition": "The parameters, gradients, optimizer variables, and other state needed to continue training.", "background": false, "home": "P2.3", "home_id": "training-memory", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 30, "organizational": 6, "user_proxy": 111 }, "book_passages": [ { "volume": "vol1", "chapter": "Neural Computation", "url": "https://mlsysbook.ai/vol1/05_nn_computation/05_nn_computation.html#sec-neural-computation-memory-training-vs-inference" } ] }, { "id": "mlsys.activation-recompute", "label": "Activation recomputation", "aliases": [ "Activation checkpointing", "Gradient checkpointing" ], "definition": "Recreating forward intermediates during backward execution to reduce retained memory.", "background": false, "home": "P2.3", "home_id": "training-memory", "added_in": "0.1-draft", "warrant": { "literary": 4, "corpus_chapters": 11, "organizational": 5, "user_proxy": 94 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html#sec-model-training-gradient-accumulation-checkpointing-0c47" } ] }, { "id": "mlsys.scaling-laws", "label": "Empirical scaling laws", "aliases": [ "neural scaling laws", "power-law scaling", "Kaplan scaling laws" ], "definition": "Fitted relationships between resources, data, model size, and observed model performance.", "background": false, "home": "F3.4", "home_id": "scaling-laws", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 9, "organizational": 3, "user_proxy": 1 }, "book_passages": [ { "volume": "vol2", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol2/01_introduction/01_introduction.html#sec-vol2-intro-ai-scaling-laws-a043" } ] }, { "id": "mlsys.hyperparameter-search", "label": "Hyperparameter search", "aliases": [ "Hyperparameter optimization", "HPO" ], "definition": "Choosing model or training settings through a sequence of evaluations.", "background": false, "home": "P2.4", "home_id": "experiment-management", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 6, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.experiment-budget", "label": "Experiment-budget allocation", "aliases": [], "definition": "Dividing a finite evaluation budget across configurations, repetitions, and fidelities.", "background": false, "home": "P2.4", "home_id": "experiment-management", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.transfer-learning", "label": "Transfer learning", "aliases": [], "definition": "Adapting representations or parameters learned from one setting to another.", "background": false, "home": "P2.5", "home_id": "fine-tuning", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 9, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.peft", "label": "Parameter-efficient adaptation", "aliases": [ "PEFT", "adapter tuning", "prefix tuning", "prompt tuning", "soft prompts" ], "definition": "Adapting a model by changing a limited parameter subset or added parameterization.", "background": false, "home": "P2.5", "home_id": "fine-tuning", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 2, "organizational": 2, "user_proxy": 17 }, "book_passages": [] }, { "id": "mlsys.tensor-layout", "label": "Tensor layout", "aliases": [ "memory alignment", "alignment padding" ], "definition": "The mapping from tensor indices to storage locations, including strides and materialization.", "background": false, "home": "S2.2", "home_id": "tensor-operators", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 0, "user_proxy": 51 }, "book_passages": [] }, { "id": "mlsys.matrix-multiply", "label": "Matrix multiplication", "aliases": [], "definition": "The dense product of two matrices, a central arithmetic workload in many neural networks, whose operand shapes, tiling, and precision determine execution cost on a specified processor.", "background": false, "home": "S2.2", "home_id": "tensor-operators", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 26, "organizational": 3, "user_proxy": 16 }, "book_passages": [ { "volume": "vol1", "chapter": "Neural Computation", "url": "https://mlsysbook.ai/vol1/05_nn_computation/05_nn_computation.html#sec-neural-computation-matrix-multiplication-formulation-417c" } ] }, { "id": "mlsys.precision", "label": "Numerical precision", "aliases": [ "number format", "floating-point format", "FP32", "FP16", "BF16", "bfloat16", "FP8", "FP4" ], "definition": "The representable values and arithmetic behavior of a numerical format.", "background": false, "home": "S2.7", "home_id": "numerics", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 64, "organizational": 1, "user_proxy": 58 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html#sec-model-compression-performance-gains-0656" } ] }, { "id": "mlsys.mixed-precision", "label": "Mixed-precision training", "aliases": [], "definition": "Training with multiple numerical formats and explicit handling of their range and accumulation behavior.", "background": false, "home": "S2.7", "home_id": "numerics", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 19, "organizational": 2, "user_proxy": 42 }, "book_passages": [] }, { "id": "mlsys.numerical-stability", "label": "Numerical stability", "aliases": [], "definition": "Sensitivity of a computed result to finite-precision operations and intermediate representations.", "background": false, "home": "S2.7", "home_id": "numerics", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 17, "organizational": 0, "user_proxy": 32 }, "book_passages": [] }, { "id": "mlsys.quantization", "label": "Quantization", "aliases": [], "definition": "Representing values using a reduced discrete set with a specified encoding and approximation rule.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.1-draft", "warrant": { "literary": 9, "corpus_chapters": 55, "organizational": 2, "user_proxy": 424 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html#sec-model-compression-quantization-precision-cd46" } ] }, { "id": "mlsys.sparsity", "label": "Sparsity", "aliases": [], "definition": "A pattern of absent or zero values whose execution benefit depends on representation and support.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.1-draft", "warrant": { "literary": 4, "corpus_chapters": 26, "organizational": 2, "user_proxy": 38 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html#sec-model-compression-pruning-d1cb" } ] }, { "id": "mlsys.distillation", "label": "Knowledge distillation", "aliases": [], "definition": "Training a student to match output distributions or internal signals from a teacher model or ensemble, potentially reducing serving cost when the student is cheaper while requiring validation of retained quality.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 11, "organizational": 1, "user_proxy": 85 }, "book_passages": [] }, { "id": "mlsys.compilation", "label": "Graph compilation", "aliases": [ "compilation breakeven", "compile-time amortization" ], "definition": "Transforming a model or execution graph into an executable representation.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.1-draft", "warrant": { "literary": 6, "corpus_chapters": 6, "organizational": 0, "user_proxy": 64 }, "book_passages": [] }, { "id": "mlsys.fusion", "label": "Kernel fusion", "aliases": [], "definition": "Combining compatible operations to reduce intermediate materialization or dispatch.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.1-draft", "warrant": { "literary": 9, "corpus_chapters": 23, "organizational": 3, "user_proxy": 260 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Frameworks", "url": "https://mlsysbook.ai/vol1/07_frameworks/07_frameworks.html#sec-ml-frameworks-hybrid-approaches-jit-compilation-8954" } ] }, { "id": "mlsys.dispatch", "label": "Host dispatch overhead", "aliases": [ "kernel launch overhead" ], "definition": "Host-side work required to launch or orchestrate device operations.", "background": false, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 14, "organizational": 2, "user_proxy": 67 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Frameworks", "url": "https://mlsysbook.ai/vol1/07_frameworks/07_frameworks.html#sec-ml-frameworks-dispatch-tax" } ] }, { "id": "mlsys.memory-hierarchy", "label": "Memory hierarchy", "aliases": [], "definition": "Arrangement of storage levels and transfer paths with differing capacity, latency, bandwidth, access control, and management mechanisms. The available levels depend on the platform.", "background": false, "home": "S1.2", "home_id": "memory-systems", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 31, "organizational": 2, "user_proxy": 151 }, "book_passages": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html#sec-hardware-acceleration-calculating-memory-bandwidth-bounds-7fc0" } ] }, { "id": "mlsys.memory-capacity", "label": "Memory capacity", "aliases": [], "definition": "Available storage for concurrently resident state and intermediates.", "background": false, "home": "S1.2", "home_id": "memory-systems", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 52, "organizational": 5, "user_proxy": 91 }, "book_passages": [ { "volume": "vol1", "chapter": "System Paradigms", "url": "https://mlsysbook.ai/vol1/02_system_paradigms/02_system_paradigms.html#sec-ml-systems-deployment-paradigm-framework-0d25" }, { "volume": "vol2", "chapter": "Compute Infrastructure", "url": "https://mlsysbook.ai/vol2/02_compute_infrastructure/02_compute_infrastructure.html#sec-compute-memory-wall" } ] }, { "id": "mlsys.accelerator-execution-architecture", "label": "Accelerator execution architecture", "aliases": [ "GPU architecture", "streaming multiprocessor", "SM" ], "definition": "Organization of arithmetic units, execution lanes or threads, scheduling resources, and local storage that determines how supported operations run. SIMD/SIMT, matrix units, and systolic arrays are distinct examples.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 14, "organizational": 2, "user_proxy": 18 }, "book_passages": [] }, { "id": "mlsys.arithmetic-intensity", "label": "Arithmetic intensity", "aliases": [ "operational intensity", "FLOPs per byte" ], "definition": "Arithmetic work per byte transferred at a specified memory boundary.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 38, "organizational": 2, "user_proxy": 283 }, "book_passages": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html#sec-hardware-acceleration-roofline-model-3e1a" } ] }, { "id": "mlsys.roofline", "label": "Roofline bound", "aliases": [ "roofline model", "roofline analysis" ], "definition": "A performance bound combining a compute ceiling with an intensity-dependent bandwidth ceiling.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.1-draft", "warrant": { "literary": 4, "corpus_chapters": 37, "organizational": 5, "user_proxy": 153 }, "book_passages": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html#sec-hardware-acceleration-roofline-model-3e1a" } ] }, { "id": "mlsys.locality", "label": "Data locality", "aliases": [ "temporal locality", "spatial locality", "data reuse" ], "definition": "Reuse of data near its consumers across time or execution locations.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 22, "organizational": 1, "user_proxy": 126 }, "book_passages": [ { "volume": "vol1", "chapter": "System Paradigms", "url": "https://mlsysbook.ai/vol1/02_system_paradigms/02_system_paradigms.html#sec-ml-systems-data-locality-invariant" } ] }, { "id": "mlsys.tiling", "label": "Tiling", "aliases": [], "definition": "Partitioning computation so working data can be reused within a chosen storage level.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 21, "organizational": 2, "user_proxy": 115 }, "book_passages": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html#sec-hardware-acceleration-tiling-fundamentals-e9e6" } ] }, { "id": "mlsys.profiling", "label": "Profiling", "aliases": [], "definition": "Measuring execution to attribute time, work, stalls, and resource use.", "background": false, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 39, "organizational": 1, "user_proxy": 178 }, "book_passages": [] }, { "id": "mlsys.execution-time", "label": "Execution time", "aliases": [], "definition": "Elapsed time for a specified execution boundary, workload, and operating condition.", "background": false, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 69, "organizational": 0, "user_proxy": 48 }, "book_passages": [] }, { "id": "mlsys.serial-fraction", "label": "Serial fraction", "aliases": [], "definition": "The portion of execution that a proposed parallel or accelerated transformation leaves unchanged.", "background": false, "home": "F3.2", "home_id": "speedup-limits", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 13, "organizational": 2, "user_proxy": 1 }, "book_passages": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html#sec-hardware-acceleration-ai-hardware-acceleration-fundamentals-9b28" } ] }, { "id": "mlsys.collectives", "label": "Collective communication", "aliases": [], "definition": "Coordinated data exchange and reduction across a process group.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.1-draft", "warrant": { "literary": 4, "corpus_chapters": 25, "organizational": 4, "user_proxy": 101 }, "book_passages": [ { "volume": "vol2", "chapter": "Collective Communication", "url": "https://mlsysbook.ai/vol2/05_collective_communication/05_collective_communication.html#sec-collective-communication-six-core-primitives-8a66" } ] }, { "id": "mlsys.data-parallel", "label": "Data parallelism", "aliases": [], "definition": "Executing model replicas on different examples while coordinating learning state.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 24, "organizational": 4, "user_proxy": 17 }, "book_passages": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html#sec-distributed-training-systems-systems-data-parallelism-6132" } ] }, { "id": "mlsys.tensor-parallel", "label": "Tensor parallelism", "aliases": [ "intra-layer model parallelism" ], "definition": "Partitioning a layer's tensors and operations across devices so each computes a slice, reducing per-device state while adding communication along the partitioned dimensions, usually within a fast scale-up domain.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.1-draft", "warrant": { "literary": 5, "corpus_chapters": 22, "organizational": 4, "user_proxy": 98 }, "book_passages": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html#sec-distributed-training-systems-systems-tensor-parallelism-d76e" } ] }, { "id": "mlsys.pipeline-parallel", "label": "Pipeline parallelism", "aliases": [ "inter-layer model parallelism" ], "definition": "Partitioning a computation into stages that process different microbatches concurrently.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.1-draft", "warrant": { "literary": 6, "corpus_chapters": 21, "organizational": 2, "user_proxy": 85 }, "book_passages": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html#sec-distributed-training-systems-systems-pipeline-parallelism-8748" } ] }, { "id": "mlsys.state-sharding", "label": "Training-state sharding", "aliases": [ "ZeRO", "FSDP", "fully sharded data parallel" ], "definition": "Partitioning parameters, gradients, or optimizer state across workers.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.1-draft", "warrant": { "literary": 5, "corpus_chapters": 89, "organizational": 3, "user_proxy": 49 }, "book_passages": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html#sec-distributed-training-systems-systems-memoryefficient-data-parallelism-zero-fsdp-0e69" } ] }, { "id": "mlsys.communication-overlap", "label": "Communication overlap", "aliases": [], "definition": "Scheduling communication concurrently with independent computation.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.1-draft", "warrant": { "literary": 14, "corpus_chapters": 7, "organizational": 3, "user_proxy": 109 }, "book_passages": [ { "volume": "vol2", "chapter": "Collective Communication", "url": "https://mlsysbook.ai/vol2/05_collective_communication/05_collective_communication.html#sec-communication-overlap-limits" } ] }, { "id": "mlsys.network-topology", "label": "Network topology", "aliases": [], "definition": "The connectivity and capacity structure that constrains communication paths.", "background": false, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 17, "organizational": 3, "user_proxy": 32 }, "book_passages": [ { "volume": "vol2", "chapter": "Network Fabrics", "url": "https://mlsysbook.ai/vol2/03_network_fabrics/03_network_fabrics.html#sec-network-fabrics-topology" } ] }, { "id": "mlsys.stragglers", "label": "Stragglers", "aliases": [], "definition": "Workers or tasks whose delayed completion holds back coordinated work.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.1-draft", "warrant": { "literary": 5, "corpus_chapters": 19, "organizational": 3, "user_proxy": 35 }, "book_passages": [] }, { "id": "mlsys.checkpoint", "label": "Recovery checkpoint", "aliases": [], "definition": "A persisted state from which a specified computation can resume.", "background": false, "home": "Q2.2", "home_id": "checkpoint-recovery", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 5, "organizational": 3, "user_proxy": 91 }, "book_passages": [ { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html#sec-storage-checkpointing" }, { "volume": "vol3", "chapter": "Durable Execution", "url": "https://mlsysbook.ai/vol3/11_durable_execution/11_durable_execution.html#sec-vol3-persistence-checkpointing" } ] }, { "id": "mlsys.consistent-cut", "label": "Consistent checkpoint cut", "aliases": [], "definition": "A set of per-worker checkpoint states taken so that no saved state depends on an update or message that another worker's saved state has not yet recorded, so restoring all of them together yields one coherent global state.", "background": false, "home": "Q2.2", "home_id": "checkpoint-recovery", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 4, "user_proxy": 4 }, "book_passages": [ { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html#sec-fault-tolerance-reliability-checkpoint-coordination-patterns-f670" } ] }, { "id": "mlsys.consensus", "label": "Replicated-log consensus", "aliases": [ "Paxos", "Raft" ], "definition": "Agreement on an ordered state history under specified timing and failure assumptions.", "background": true, "home": "S3.4", "home_id": "coordination", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.fencing", "label": "Fencing", "aliases": [], "definition": "Rejecting effects from an actor whose epoch or authority is stale.", "background": true, "home": "S3.4", "home_id": "coordination", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 13, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html#sec-fault-tolerance-reliability-reliability-recovery-procedures-002d" } ] }, { "id": "mlsys.federated-learning", "label": "Federated learning", "aliases": [], "definition": "Coordinating model learning across separately controlled data holders.", "background": false, "home": "S3.5", "home_id": "federated-learning", "added_in": "0.1-draft", "warrant": { "literary": 7, "corpus_chapters": 14, "organizational": 3, "user_proxy": 85 }, "book_passages": [ { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html#sec-edge-intelligence-learning-protocols-139a" } ] }, { "id": "mlsys.arrival-rate", "label": "Arrival rate", "aliases": [], "definition": "The frequency and distribution of requests reaching a service boundary.", "background": false, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 32, "organizational": 1, "user_proxy": 18 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-utilizationlatency-relationship-a2f0" } ] }, { "id": "mlsys.service-capacity", "label": "Service capacity", "aliases": [ "service rate" ], "definition": "The completion rate supported by a service under a stated workload and operating point.", "background": false, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 26, "organizational": 1, "user_proxy": 74 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-capacity-planning-96a3" } ] }, { "id": "mlsys.queueing", "label": "Queueing delay", "aliases": [], "definition": "Time a request waits for a resource before receiving service.", "background": false, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 15, "organizational": 4, "user_proxy": 123 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-utilizationlatency-relationship-a2f0" } ] }, { "id": "mlsys.tail-latency", "label": "Tail latency", "aliases": [], "definition": "A high quantile of a defined request-latency distribution.", "background": false, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 46, "organizational": 3, "user_proxy": 71 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-request-lifecycle-d9c6" } ] }, { "id": "mlsys.batching", "label": "Request batching", "aliases": [], "definition": "Executing requests together to share work or improve resource use.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.1-draft", "warrant": { "literary": 5, "corpus_chapters": 6, "organizational": 6, "user_proxy": 227 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-batching-helps-f1dc" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html#sec-inference-scale-physics-batching-efficiency-curve-021c" }, { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-batching-tax" } ] }, { "id": "mlsys.admission", "label": "Admission control", "aliases": [ "load shedding", "rate limiting" ], "definition": "Accepting, delaying, or rejecting work according to capacity, policy, and service constraints.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 27, "organizational": 5, "user_proxy": 83 }, "book_passages": [] }, { "id": "mlsys.autoscaling", "label": "Autoscaling", "aliases": [], "definition": "Changing provisioned service resources in response to measured or predicted demand.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.1-draft", "warrant": { "literary": 7, "corpus_chapters": 9, "organizational": 0, "user_proxy": 25 }, "book_passages": [] }, { "id": "mlsys.kv-cache", "label": "KV cache", "aliases": [ "Key-value cache" ], "definition": "Retained attention keys and values reused during autoregressive execution.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.1-draft", "warrant": { "literary": 5, "corpus_chapters": 52, "organizational": 9, "user_proxy": 307 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-memory-kv-cache-d1ea" }, { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html#sec-inference-scale-kv-cache-wall" }, { "volume": "vol3", "chapter": "KV Cache Management", "url": "https://mlsysbook.ai/vol3/05_kv_cache/05_kv_cache.html#sec-vol3-kvcache-geometry" } ] }, { "id": "mlsys.prefill-decode", "label": "Prefill and decode", "aliases": [], "definition": "The prompt-processing and incremental token-generation phases of autoregressive inference.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.1-draft", "warrant": { "literary": 4, "corpus_chapters": 10, "organizational": 5, "user_proxy": 73 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-performance-metrics-ttft-tpot-b009" }, { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-memory-kv-cache-d1ea" } ] }, { "id": "mlsys.paged-kv", "label": "Paged KV allocation", "aliases": [ "PagedAttention" ], "definition": "Managing attention state in blocks to reduce allocation waste and support sharing.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.1-draft", "warrant": { "literary": 5, "corpus_chapters": 5, "organizational": 5, "user_proxy": 101 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-pagedattention-b8d4" } ] }, { "id": "mlsys.prefix-cache", "label": "Prefix caching", "aliases": [ "prompt caching", "context caching" ], "definition": "Reusing compatible computed state for an identical request prefix.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.1-draft", "warrant": { "literary": 9, "corpus_chapters": 14, "organizational": 9, "user_proxy": 32 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-prefix-caching-offloading" }, { "volume": "vol3", "chapter": "KV Cache Management", "url": "https://mlsysbook.ai/vol3/05_kv_cache/05_kv_cache.html#sec-vol3-kv-cache-prefix-compatibility" } ] }, { "id": "mlsys.speculative-decoding", "label": "Speculative decoding", "aliases": [ "draft-and-verify", "EAGLE", "self-speculative decoding", "multi-token draft heads" ], "definition": "Using proposed tokens and a validating procedure to accelerate generation while preserving a specified output distribution.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.1-draft", "warrant": { "literary": 10, "corpus_chapters": 9, "organizational": 0, "user_proxy": 82 }, "book_passages": [] }, { "id": "mlsys.slo-goodput", "label": "SLO goodput", "aliases": [], "definition": "The rate of useful completed work that satisfies the specified quality and service criteria.", "background": false, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 4, "organizational": 3, "user_proxy": 3 }, "book_passages": [ { "volume": "vol2", "chapter": "Serving at Scale", "url": "https://mlsysbook.ai/vol2/10_inference/10_inference.html#sec-inference-scale-fundamental-inversion-training-vs-inference-8834" } ] }, { "id": "mlsys.generalization", "label": "Generalization", "aliases": [], "definition": "Performance on relevant examples or conditions beyond those used to fit the system.", "background": true, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 27, "organizational": 1, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.evaluation-split", "label": "Evaluation split", "aliases": [], "definition": "A partitioning protocol that separates learning, selection, and final assessment.", "background": false, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 3, "user_proxy": 2 }, "book_passages": [ { "volume": "vol3", "chapter": "Trajectory Curation", "url": "https://mlsysbook.ai/vol3/15_trajectory_curation/15_trajectory_curation.html#sec-vol3-trajectory-curation-heldout-protection" } ] }, { "id": "mlsys.selection-bias", "label": "Model-selection optimism", "aliases": [], "definition": "Overstated performance caused by choosing a winner using noisy evaluation results.", "background": false, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.metrics", "label": "Evaluation metrics", "aliases": [], "definition": "Specified measurements of predictive or task behavior, with explicit populations and aggregation.", "background": false, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 13, "organizational": 1, "user_proxy": 7 }, "book_passages": [ { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html#sec-benchmarking-machine-learning-benchmarking-framework-70b8" } ] }, { "id": "mlsys.calibration", "label": "Probability calibration", "aliases": [], "definition": "Agreement between predicted probabilities and observed event frequencies under an evaluation distribution.", "background": false, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 2, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.statistical-uncertainty", "label": "Statistical uncertainty", "aliases": [], "definition": "Uncertainty in an estimated quantity due to finite and dependent observations.", "background": false, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 7, "organizational": 5, "user_proxy": 23 }, "book_passages": [] }, { "id": "mlsys.benchmark", "label": "Benchmark protocol", "aliases": [], "definition": "A reproducible specification of workload, measurement boundary, run rules, and comparison procedure that lets results from different systems be compared and lets a system be judged against a baseline.", "background": false, "home": "Q1.2", "home_id": "system-benchmarking", "added_in": "0.1-draft", "warrant": { "literary": 5, "corpus_chapters": 1, "organizational": 6, "user_proxy": 24 }, "book_passages": [ { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html#sec-benchmarking-benchmark-components-97cc" } ] }, { "id": "mlsys.online-experiment", "label": "Online experimentation", "aliases": [ "A/B testing", "A/B test", "controlled online experiment" ], "definition": "A controlled intervention used to estimate effects on deployed-system outcomes.", "background": false, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 9, "organizational": 0, "user_proxy": 20 }, "book_passages": [] }, { "id": "mlsys.sliced-evaluation", "label": "Sliced evaluation", "aliases": [], "definition": "Measuring behavior separately on relevant populations, conditions, or workload groups.", "background": false, "home": "P3.3", "home_id": "robustness-slices", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 3, "user_proxy": 38 }, "book_passages": [ { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html#sec-benchmarking-machine-learning-benchmarking-framework-70b8" } ] }, { "id": "mlsys.selective-prediction", "label": "Selective prediction", "aliases": [], "definition": "Decision-gating mechanism that emits model outputs only for admitted cases and abstains on or defers the remaining cases.", "background": false, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 4, "user_proxy": 3 }, "book_passages": [ { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html#sec-robust-ai-selective-prediction" } ] }, { "id": "mlsys.artifact-version", "label": "Artifact versioning", "aliases": [ "prompt versioning", "prompt as artifact", "system-prompt version" ], "definition": "Identifying the exact model, code, data, configuration, and dependencies in an execution or release.", "background": false, "home": "P5.1", "home_id": "release-management", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 2, "user_proxy": 11 }, "book_passages": [ { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html#sec-storage-data-versioning-orchestration" }, { "volume": "vol4", "chapter": "Deployment Release", "url": "https://mlsysbook.ai/vol4/16_release/16_release.html#sec-release-manifest" } ] }, { "id": "mlsys.monitoring", "label": "Production monitoring", "aliases": [], "definition": "Collecting and interpreting signals about deployed system behavior and health.", "background": false, "home": "P5.2", "home_id": "monitoring", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 14, "organizational": 4, "user_proxy": 44 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html#sec-ml-operations-production-monitoring-5e1b" } ] }, { "id": "mlsys.rollout", "label": "Staged rollout", "aliases": [ "canary release", "canary deployment", "progressive delivery" ], "definition": "Increasing exposure to a candidate system through bounded release stages.", "background": false, "home": "P5.1", "home_id": "release-management", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 12, "organizational": 5, "user_proxy": 74 }, "book_passages": [ { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html#sec-ml-operations-scale-staged-rollout-strategies-2d1f" }, { "volume": "vol3", "chapter": "Operating Agents", "url": "https://mlsysbook.ai/vol3/14_agent_operations/14_agent_operations.html#sec-vol3-evaluation-releases" } ] }, { "id": "mlsys.rollback", "label": "Rollback", "aliases": [], "definition": "Returning to a previously accepted deployment configuration with compatible state.", "background": false, "home": "P5.1", "home_id": "release-management", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 38, "organizational": 3, "user_proxy": 62 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html#sec-ml-operations-rollback-strategies-safety-mechanisms-7707" } ] }, { "id": "mlsys.concept-drift", "label": "Concept drift", "aliases": [], "definition": "Change over time in the relationship between inputs and targets, so the same input warrants a different prediction.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 3, "user_proxy": 12 }, "book_passages": [ { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html#sec-robust-ai-distribution-shift-concept-drift-55e2" } ] }, { "id": "mlsys.delayed-outcomes", "label": "Delayed outcomes", "aliases": [], "definition": "Target outcomes that become observable only after a delay or for a selected subset of cases.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 2, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.incident-response", "label": "Incident response", "aliases": [], "definition": "Detecting, containing, diagnosing, and recovering from unexpected system behavior.", "background": false, "home": "P5.4", "home_id": "incident-response", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 8, "organizational": 2, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.observability", "label": "System observability", "aliases": [], "definition": "Instrumentation that permits investigation of internal execution from recorded outputs, traces, and metrics.", "background": false, "home": "P5.2", "home_id": "monitoring", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 32 }, "book_passages": [] }, { "id": "mlsys.threat-model", "label": "Threat modeling", "aliases": [], "definition": "Specifying adversaries, assets, interfaces, capabilities, and failure consequences.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 2, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.poisoning", "label": "Data poisoning", "aliases": [], "definition": "Adversarial manipulation of learning data or signals to alter learned behavior.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 8, "organizational": 4, "user_proxy": 13 }, "book_passages": [ { "volume": "vol2", "chapter": "Security & Privacy", "url": "https://mlsysbook.ai/vol2/13_security_privacy/13_security_privacy.html#sec-security-privacy-data-poisoning-351f" } ] }, { "id": "mlsys.evasion", "label": "Adversarial evasion", "aliases": [ "adversarial example", "evasion attack", "adversarial perturbation" ], "definition": "Inputs chosen or perturbed at inference time so that a model misclassifies or misbehaves, within an adversary's stated knowledge of the model and perturbation budget, without any change to the model or its training data.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 1, "user_proxy": 37 }, "book_passages": [] }, { "id": "mlsys.privacy", "label": "Privacy requirements", "aliases": [], "definition": "Requirements governing the collection, use, retention, and disclosure of information about people or records.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 1, "user_proxy": 39 }, "book_passages": [] }, { "id": "mlsys.differential-privacy", "label": "Differential privacy", "aliases": [ "DP" ], "definition": "A parameterized bound on changes in output distributions under a specified neighboring-dataset relation.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 19, "organizational": 3, "user_proxy": 37 }, "book_passages": [ { "volume": "vol2", "chapter": "Security & Privacy", "url": "https://mlsysbook.ai/vol2/13_security_privacy/13_security_privacy.html#sec-security-privacy-differential-privacy-8c2b" } ] }, { "id": "mlsys.fairness", "label": "Fairness criteria", "aliases": [], "definition": "Explicit requirements for how system outcomes or errors are distributed across affected groups.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 4, "user_proxy": 3 }, "book_passages": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html#sec-responsible-engineering-testing-across-populations-9f20" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html#sec-responsible-ai-fairness-machine-learning-2ba4" } ] }, { "id": "mlsys.authorization", "label": "Authorization", "aliases": [], "definition": "Determining whether an actor may perform a specified action on a resource.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 40, "organizational": 4, "user_proxy": 6 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Sandboxes", "url": "https://mlsysbook.ai/vol3/08_sandboxes/08_sandboxes.html#sec-vol3-sandboxes-scoped-grants" } ] }, { "id": "mlsys.isolation", "label": "Execution isolation", "aliases": [], "definition": "Restricting which resources and effects a computation can access.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 4, "organizational": 6, "user_proxy": 22 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Sandboxes", "url": "https://mlsysbook.ai/vol3/08_sandboxes/08_sandboxes.html#sec-vol3-sandboxes-microvms" }, { "volume": "vol4", "chapter": "Workload Placement", "url": "https://mlsysbook.ai/vol4/13_placement/13_placement.html#sec-placement-isolation-decisions" } ] }, { "id": "mlsys.human-oversight", "label": "Human oversight", "aliases": [ "human-on-the-loop", "meaningful human control" ], "definition": "An arrangement in which people can understand, intervene in, and contest consequential system behavior.", "background": false, "home": "Q5.3", "home_id": "human-oversight", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 10, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Responsible Physical AI", "url": "https://mlsysbook.ai/vol4/17_responsibility/17_responsibility.html#sec-responsibility-meaningful-oversight" } ] }, { "id": "mlsys.energy", "label": "Operating energy", "aliases": [], "definition": "Energy consumed by a specified system boundary over a defined workload or interval.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 2, "user_proxy": 280 }, "book_passages": [ { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html#sec-sustainable-ai-power-modeling-fundamentals-67cb-67cb" } ] }, { "id": "mlsys.carbon", "label": "Lifecycle carbon", "aliases": [], "definition": "Emissions attributed to operation and embodied system resources under declared accounting boundaries.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 4, "user_proxy": 4 }, "book_passages": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html#sec-responsible-engineering-environmental-impact-0292" }, { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html#sec-sustainable-ai-embodied-carbon-assessment-9de0" }, { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html#sec-sustainable-ai-operational-carbon-calculation-4627-calculation-4627" } ] }, { "id": "mlsys.unit-cost", "label": "Cost per useful outcome", "aliases": [], "definition": "Total relevant expenditure divided by outcomes meeting a specified acceptance criterion.", "background": false, "home": "Q6.3", "home_id": "cost-economics", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 11, "user_proxy": 34 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Economics", "url": "https://mlsysbook.ai/vol3/19_agent_economics/19_agent_economics.html#sec-vol3-agent-economics-cohort-ratio" } ] }, { "id": "mlsys.capacity-planning", "label": "Capacity planning", "aliases": [], "definition": "Provisioning resources for a workload distribution and service objectives.", "background": false, "home": "S4.5", "home_id": "capacity-provisioning", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 28, "organizational": 1, "user_proxy": 65 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-capacity-planning-96a3" } ] }, { "id": "mlsys.trajectory", "label": "Agent trajectory", "aliases": [], "definition": "A sequence of observations, model proposals, tool actions, and external outcomes for a task.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 15, "organizational": 4, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol3/01_introduction/01_introduction.html#sec-vol3-intro-trajectory-systems-boundary" } ] }, { "id": "mlsys.planning", "label": "Agent planning", "aliases": [], "definition": "Constructing and revising prospective actions in response to goals, dependencies, and observations.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.1-draft", "warrant": { "literary": 4, "corpus_chapters": 2, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.sliding-window-attention", "label": "Sparse and sliding-window attention", "aliases": [ "local attention" ], "definition": "Attention restricted to local windows or predetermined sparse patterns, reducing attended query-key pairs; cost is linear in sequence length when the attended positions per query remain bounded.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 1, "user_proxy": 24 }, "book_passages": [] }, { "id": "mlsys.retrieval", "label": "Retrieval", "aliases": [], "definition": "Selecting relevant records from an external collection using an explicit query and ranking process.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 49, "organizational": 1, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Long-Term Memory", "url": "https://mlsysbook.ai/vol3/06_long_term_memory/06_long_term_memory.html#sec-vol3-persistent-governance" } ] }, { "id": "mlsys.ann", "label": "Approximate nearest-neighbor search", "aliases": [ "ANN" ], "definition": "Trading exact neighbor recovery for reduced search cost using an index or approximation.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 4, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.hallucination", "label": "Hallucination", "aliases": [ "unsupported generated claims" ], "definition": "Generated factual assertions or summaries not supported by the relevant source or evidence.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 10, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.tool-contract", "label": "Tool contract", "aliases": [], "definition": "A specification of tool inputs, outputs, authority, effects, and failure behavior.", "background": false, "home": "T1.2", "home_id": "tool-interfaces", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.failure-model", "label": "Failure model", "aliases": [], "definition": "The declared classes of faults and timing assumptions under which a mechanism is analyzed.", "background": false, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 14, "organizational": 4, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.idempotency", "label": "Idempotency", "aliases": [], "definition": "A specified operation's repeated execution having the same intended effect as one execution.", "background": false, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 18, "organizational": 5, "user_proxy": 7 }, "book_passages": [ { "volume": "vol3", "chapter": "Tool Calling", "url": "https://mlsysbook.ai/vol3/07_tool_calling/07_tool_calling.html#sec-vol3-tool-calling-operation-ids" } ] }, { "id": "mlsys.prompt-injection", "label": "Prompt injection", "aliases": [ "indirect prompt injection" ], "definition": "Untrusted content attempting to redirect model-driven behavior across an instruction or authority boundary.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 12, "organizational": 6, "user_proxy": 10 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Security", "url": "https://mlsysbook.ai/vol3/09_security/09_security.html#sec-vol3-security-injection-entry-points" } ] }, { "id": "mlsys.agent-evaluation", "label": "Trajectory evaluation", "aliases": [], "definition": "Assessing task outcomes, intermediate effects, and constraint compliance across a complete agent run.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 1, "organizational": 6, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Evaluation", "url": "https://mlsysbook.ai/vol3/13_evaluation/13_evaluation.html#sec-vol3-evaluation-contract" } ] }, { "id": "mlsys.rlhf", "label": "RL from human feedback", "aliases": [ "RLHF" ], "definition": "Optimizing a model's policy against a reward model trained on human preference comparisons, usually with a penalty for drifting from a reference model.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 4, "organizational": 0, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.dynamics", "label": "Physical dynamics", "aliases": [], "definition": "The evolution of a physical system under inputs, disturbances, and constraints.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 13, "organizational": 1, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "The Physical Body", "url": "https://mlsysbook.ai/vol4/02_body/02_body.html#sec-body-kinetic-momentum" } ] }, { "id": "mlsys.sensing", "label": "Sensor calibration", "aliases": [], "definition": "Establishing how observations relate to physical quantities, coordinates, and time.", "background": false, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 11, "organizational": 2, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.state-estimation", "label": "State estimation", "aliases": [], "definition": "Inferring latent physical state and uncertainty from observations and a model.", "background": true, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 22, "organizational": 3, "user_proxy": 1 }, "book_passages": [ { "volume": "vol4", "chapter": "Spatial Memory", "url": "https://mlsysbook.ai/vol4/09_memory/09_memory.html#sec-perception-state-estimation" } ] }, { "id": "mlsys.state-observability", "label": "State observability", "aliases": [], "definition": "Whether distinct internal states can be distinguished through the available observations and model.", "background": true, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 2, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Engineering Physical AI", "url": "https://mlsysbook.ai/vol4/18_frontier/18_frontier.html#sec-frontier-observational-indistinguishability-sub" } ] }, { "id": "mlsys.observation-age", "label": "Observation age", "aliases": [], "definition": "Elapsed time between a physical observation and its use in a decision.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 19, "organizational": 9, "user_proxy": 10 }, "book_passages": [ { "volume": "vol4", "chapter": "Sensor Perception", "url": "https://mlsysbook.ai/vol4/08_perception/08_perception.html#sec-perception-age-budget" } ] }, { "id": "mlsys.feedback-delay", "label": "Feedback delay", "aliases": [], "definition": "The total delay around the sense-compute-communicate-actuate loop, which erodes a controller's stability margin and adds motion before any response can begin, so it must be budgeted end to end rather than stage by stage.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 3, "user_proxy": 2 }, "book_passages": [ { "volume": "vol4", "chapter": "The Nervous System", "url": "https://mlsysbook.ai/vol4/04_nervous/04_nervous.html#sec-nervous-time-constants-phase-lag" } ] }, { "id": "mlsys.wcet", "label": "Worst-case execution time", "aliases": [ "WCET" ], "definition": "A justified execution-time upper bound under specified hardware and execution assumptions.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 14, "organizational": 4, "user_proxy": 49 }, "book_passages": [ { "volume": "vol4", "chapter": "Workload Placement", "url": "https://mlsysbook.ai/vol4/13_placement/13_placement.html#sec-placement-testing-separation" } ] }, { "id": "mlsys.feedback-control", "label": "Feedback control", "aliases": [ "PID control", "PID controller", "proportional-integral-derivative control" ], "definition": "Selecting inputs using measured or estimated deviation from desired behavior.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "The Nervous System", "url": "https://mlsysbook.ai/vol4/04_nervous/04_nervous.html#sec-nervous-time-constants-phase-lag" } ] }, { "id": "mlsys.trajectory-feasibility", "label": "Trajectory feasibility", "aliases": [], "definition": "Satisfaction of motion, state, input, and dynamic constraints along a proposed trajectory.", "background": false, "home": "T2.3", "home_id": "planning-control", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 5, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Trajectory Planning", "url": "https://mlsysbook.ai/vol4/11_planning/11_planning.html#sec-planning-kinodynamic-feasibility" } ] }, { "id": "mlsys.contact-compliance", "label": "Contact and compliance", "aliases": [], "definition": "Mechanical interaction whose forces depend on geometry, motion, stiffness, and damping.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 1, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Safety Enforcement", "url": "https://mlsysbook.ai/vol4/12_enforcement/12_enforcement.html#sec-enforcement-intentional-contact" } ] }, { "id": "mlsys.sim-to-real", "label": "Simulation-to-real transfer", "aliases": [ "sim-to-real", "sim2real" ], "definition": "Transferring a policy trained in simulation to physical hardware, where differences in dynamics, sensing, contact, and timing can degrade behavior and motivate randomization, system identification, or adaptation on real experience.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 2, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Policy Training", "url": "https://mlsysbook.ai/vol4/06_training/06_training.html#sec-training-simulator-component" } ] }, { "id": "mlsys.runtime-assurance", "label": "Runtime assurance", "aliases": [], "definition": "Monitoring and constraining execution using explicitly stated safety premises and fallback mechanisms.", "background": false, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 8, "user_proxy": 17 }, "book_passages": [ { "volume": "vol4", "chapter": "Safety Enforcement", "url": "https://mlsysbook.ai/vol4/12_enforcement/12_enforcement.html#sec-enforcement-minimal-intervention" } ] }, { "id": "mlsys.stopping-envelope", "label": "Stopping envelope", "aliases": [ "safe stopping region" ], "definition": "The state and clearance conditions within which a specified stopping response remains feasible.", "background": false, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 14, "organizational": 10, "user_proxy": 1 }, "book_passages": [ { "volume": "vol4", "chapter": "Safety Enforcement", "url": "https://mlsysbook.ai/vol4/12_enforcement/12_enforcement.html#sec-enforcement-stopping-envelopes" }, { "volume": "vol4", "chapter": "The Causal Boundary", "url": "https://mlsysbook.ai/vol4/01_boundary/01_boundary.html#sec-boundary-kinematic-roofline" } ] }, { "id": "mlsys.operating-domain", "label": "Operational design domain", "aliases": [ "ODD" ], "definition": "The declared set of conditions, covering environment, inputs, actors, and system state, within which a system is designed to perform and outside which its safety claims and evidence do not apply.", "background": false, "home": "Q4.1", "home_id": "hazards-operating-domain", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 16, "organizational": 5, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html#sec-data-policy-coverage" } ] }, { "id": "mlsys.safety-case", "label": "Safety case", "aliases": [ "assurance case", "UL 4600 safety case" ], "definition": "A structured, reviewable argument, with its evidence, assumptions, and defeaters, that a defined system is acceptably safe for a stated use within a stated operating domain.", "background": false, "home": "Q4.4", "home_id": "safety-cases", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 21, "organizational": 10, "user_proxy": 3 }, "book_passages": [ { "volume": "vol4", "chapter": "Deployment Release", "url": "https://mlsysbook.ai/vol4/16_release/16_release.html#sec-release-assurance-grammar" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html#sec-responsible-ai-governance-assurance-lifecycle" } ] }, { "id": "mlsys.grammar", "label": "Structured output constraint", "aliases": [], "definition": "A contract on the admissible syntax or schema of model outputs, enforced during generation by constrained decoding or checked afterward by output validation, which establishes form but not factual correctness or authority.", "background": false, "home": "P4.6", "home_id": "structured-generation", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.labels", "label": "Label quality", "aliases": [ "label noise" ], "definition": "The correctness, consistency, and coverage of recorded supervision relative to the intended task, which bound the training signal and the trustworthiness of evaluation.", "background": false, "home": "P1.1", "home_id": "data-collection", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 2, "user_proxy": 28 }, "book_passages": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html#sec-data-engineering-data-labeling-6836" } ] }, { "id": "mlsys.data-contract", "label": "Data contract", "aliases": [], "definition": "A specification of schema, semantics, validity rules, and availability for data exchanged between components.", "background": false, "home": "P1.5", "home_id": "data-lineage", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 20 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Lifecycle", "url": "https://mlsysbook.ai/vol1/03_ml_lifecycle/03_ml_lifecycle.html#sec-ml-workflow-stage-interface-specification-ae3c" } ] }, { "id": "mlsys.feedback-loop", "label": "Deployment feedback loop", "aliases": [], "definition": "A deployed system's outputs changing the data or behavior encountered by future learning or evaluation.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 3, "user_proxy": 10 }, "book_passages": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html#sec-responsible-engineering-monitoring-incident-response-54f4" } ] }, { "id": "mlsys.decision-policy", "label": "Decision policy", "aliases": [], "definition": "The rule that converts model outputs and system state into actions, thresholds, abstentions, or escalation.", "background": false, "home": "F1.3", "home_id": "objectives-decisions", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.robustness-evaluation", "label": "Robustness evaluation", "aliases": [], "definition": "Testing behavior under explicitly chosen shifts, perturbations, faults, or operating stresses.", "background": false, "home": "P3.3", "home_id": "robustness-slices", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 4, "organizational": 1, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.supervised-learning", "label": "Supervised learning", "aliases": [], "definition": "Learning a prediction rule from examples with target outputs.", "background": true, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 11, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.self-supervised-learning", "label": "Self-supervised learning", "aliases": [], "definition": "Learning from targets constructed from the observed data itself.", "background": true, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.unsupervised-learning", "label": "Unsupervised learning", "aliases": [], "definition": "Learning structure in data without externally supplied example-level targets.", "background": true, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.reinforcement-learning", "label": "Reinforcement learning", "aliases": [], "definition": "Learning a policy from action-dependent observations and reward signals.", "background": true, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 31, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.imitation-learning", "label": "Imitation learning", "aliases": [], "definition": "Learning a policy from demonstrated behavior and possibly corrective interaction.", "background": true, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 15, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-artifact", "label": "Model artifact", "aliases": [], "definition": "The learned parameters, structures, example stores, and metadata required to execute a model.", "background": false, "home": "F2.5", "home_id": "model-artifacts", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 8, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.compensation", "label": "Compensating action", "aliases": [], "definition": "An explicit action intended to remedy an earlier effect whose direct reversal is unavailable.", "background": false, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 6, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Failure Recovery", "url": "https://mlsysbook.ai/vol3/12_failure_recovery/12_failure_recovery.html#sec-vol3-failure-recovery-semantic-compensation" } ] }, { "id": "mlsys.backpressure", "label": "Backpressure", "aliases": [], "definition": "A mechanism that signals downstream saturation to limit or slow upstream production.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 19, "organizational": 1, "user_proxy": 36 }, "book_passages": [] }, { "id": "mlsys.privacy-leakage", "label": "Privacy leakage", "aliases": [], "definition": "Information about people or records revealed through data, artifacts, or model interaction.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.utilization", "label": "Resource utilization", "aliases": [], "definition": "The fraction of a resource's defined service capacity that is occupied over a stated interval.", "background": false, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 7, "organizational": 4, "user_proxy": 81 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-utilizationlatency-relationship-a2f0" } ] }, { "id": "mlsys.irreversible-effects", "label": "Irreversible effects", "aliases": [], "definition": "External consequences that cannot be fully undone by restoring software state, retrying, or performing a compensating action.", "background": false, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "The Causal Boundary", "url": "https://mlsysbook.ai/vol4/01_boundary/01_boundary.html#sec-boundary-thermodynamic-irreversibility" } ] }, { "id": "mlsys.dataset-storage-layout", "label": "Dataset storage layout", "aliases": [], "definition": "Organization of stored dataset records, columns, partitions, and metadata that determines which data an access reads or decodes.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 1, "user_proxy": 85 }, "book_passages": [] }, { "id": "mlsys.data-ingestion", "label": "Data ingestion", "aliases": [], "definition": "Reading externally produced records into a processing or learning system under declared delivery, ordering, and timing assumptions.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.1-draft", "warrant": { "literary": 3, "corpus_chapters": 14, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.schema-evolution", "label": "Schema evolution", "aliases": [], "definition": "Changing a data schema while explicitly managing how existing data and producers or consumers remain interpretable and compatible.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.test-oracle", "label": "Test oracle", "aliases": [], "definition": "A specification, procedure, reference, or informed judgment used to assess whether observed behavior satisfies stated expectations for a test.", "background": false, "home": "P3.5", "home_id": "ml-testing", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 2, "user_proxy": 2 }, "book_passages": [ { "volume": "vol4", "chapter": "Adversarial Verification", "url": "https://mlsysbook.ai/vol4/15_verification/15_verification.html#sec-verification-oracles-acceptance" } ] }, { "id": "mlsys.integration-testing", "label": "Integration testing", "aliases": [], "definition": "Testing interactions among assembled components against stated interface and behavior expectations.", "background": false, "home": "P3.5", "home_id": "ml-testing", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.failure-injection", "label": "Failure injection", "aliases": [], "definition": "Deliberately introducing selected faults or disruptive conditions to observe system response under controlled experimental boundaries.", "background": false, "home": "Q2.4", "home_id": "fault-injection", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 3, "user_proxy": 2 }, "book_passages": [ { "volume": "vol4", "chapter": "Adversarial Verification", "url": "https://mlsysbook.ai/vol4/15_verification/15_verification.html#sec-verification-fault-injection" } ] }, { "id": "mlsys.dependency-environment", "label": "Dependency environment", "aliases": [], "definition": "The software packages, versions, native libraries, and runtime configuration on which an execution depends.", "background": false, "home": "P5.5", "home_id": "pipelines-reproducibility", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.workflow-orchestration", "label": "Workflow orchestration", "aliases": [], "definition": "Coordinating dependent computation steps, their data or control flow, execution conditions, and failure handling.", "background": false, "home": "P5.5", "home_id": "pipelines-reproducibility", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 2, "organizational": 0, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.experiment-reproducibility", "label": "Experiment reproducibility", "aliases": [], "definition": "The ability to repeat an experimental procedure and obtain results consistent with declared criteria when its inputs, environment, randomness, and execution conditions are specified.", "background": false, "home": "P5.5", "home_id": "pipelines-reproducibility", "added_in": "0.1-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 2, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.compute-throughput", "label": "Compute throughput", "aliases": [], "definition": "The sustainable rate of a specified arithmetic workload on an execution resource.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 25, "organizational": 2, "user_proxy": 211 }, "book_passages": [] }, { "id": "mlsys.memory-bandwidth", "label": "Memory bandwidth", "aliases": [], "definition": "The sustainable rate at which bytes move across a specified memory interface.", "background": false, "home": "S1.2", "home_id": "memory-systems", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 67, "organizational": 6, "user_proxy": 314 }, "book_passages": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html#sec-hardware-acceleration-calculating-memory-bandwidth-bounds-7fc0" }, { "volume": "vol4", "chapter": "The Cognitive Brain", "url": "https://mlsysbook.ai/vol4/03_brain/03_brain.html#sec-brain-memory-wall" } ] }, { "id": "mlsys.host-device-interconnect", "label": "Host-device interconnect", "aliases": [], "definition": "Communication path through which a host and accelerator exchange data and control, with transport, topology, and memory-access semantics that affect transfers.", "background": false, "home": "S1.3", "home_id": "node-interconnect", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 0, "user_proxy": 279 }, "book_passages": [] }, { "id": "mlsys.power-delivery", "label": "Power delivery and integrity", "aliases": [], "definition": "Supply paths, conversion stages, limits, and transient behavior that must maintain required voltages while workloads and actuators draw or return power.", "background": false, "home": "S1.4", "home_id": "chip-power-thermal", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 2, "user_proxy": 22 }, "book_passages": [ { "volume": "vol4", "chapter": "The Physical Body", "url": "https://mlsysbook.ai/vol4/02_body/02_body.html#sec-body-power-integrity" } ] }, { "id": "mlsys.thermal-dynamics", "label": "Thermal behavior and throttling", "aliases": [], "definition": "Evolution of temperature under heat generation and removal, and the performance controls or operating restrictions that follow under a platform thermal policy.", "background": false, "home": "S1.4", "home_id": "chip-power-thermal", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 4, "user_proxy": 199 }, "book_passages": [ { "volume": "vol4", "chapter": "Workload Placement", "url": "https://mlsysbook.ai/vol4/13_placement/13_placement.html#sec-placement-thermal-timing-coupling" } ] }, { "id": "mlsys.sli", "label": "Service-level indicator", "aliases": [ "SLI" ], "definition": "A quantitative measure of one aspect of delivered service, such as latency, availability, or accepted-output rate, computed over a stated population and window.", "background": false, "home": "F1.1", "home_id": "requirements-slos", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.error-budget", "label": "Error budget", "aliases": [], "definition": "The amount of service-level-objective violation permitted over a window, used to pace changes and risk against a reliability target.", "background": false, "home": "F1.1", "home_id": "requirements-slos", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 9, "organizational": 1, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.latency-budget", "label": "Latency budget", "aliases": [], "definition": "An allocation of an end-to-end latency target across the stages of a request path, such as network, preprocessing, inference, and postprocessing.", "background": false, "home": "F1.1", "home_id": "requirements-slos", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 49, "organizational": 7, "user_proxy": 111 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-request-lifecycle-d9c6" } ] }, { "id": "mlsys.flop-accounting", "label": "Operation and byte accounting", "aliases": [ "FLOP counting", "resource accounting", "MAC counting" ], "definition": "Estimating a workload's arithmetic operations and bytes moved per example, token, or step from model shapes, to predict cost before measurement.", "background": false, "home": "F1.2", "home_id": "workload-characterization", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 15, "organizational": 3, "user_proxy": 42 }, "book_passages": [] }, { "id": "mlsys.memory-footprint", "label": "Memory footprint", "aliases": [], "definition": "The peak memory a workload occupies, summed over parameters, activations, optimizer or cache state, and runtime buffers.", "background": false, "home": "F1.2", "home_id": "workload-characterization", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 42, "organizational": 3, "user_proxy": 441 }, "book_passages": [] }, { "id": "mlsys.workload-trace", "label": "Workload trace", "aliases": [], "definition": "A recorded sequence of requests or jobs with arrival times, sizes, and durations, replayed to evaluate capacity, scheduling, or serving policies.", "background": false, "home": "F1.2", "home_id": "workload-characterization", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.cost-sensitive-decision", "label": "Cost-sensitive decision", "aliases": [], "definition": "Choosing actions or thresholds to minimize expected cost when different error types carry unequal consequences.", "background": false, "home": "F1.3", "home_id": "objectives-decisions", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.pareto-frontier", "label": "Pareto frontier", "aliases": [], "definition": "The set of configurations for which no objective, such as accuracy, latency, or cost, can improve without worsening another.", "background": false, "home": "F1.3", "home_id": "objectives-decisions", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 13, "organizational": 1, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.transformer", "label": "Transformer", "aliases": [], "definition": "A neural architecture of stacked attention and feed-forward blocks with residual connections and normalization, processing all sequence positions in parallel during training.", "background": false, "home": "F2.1", "home_id": "model-families", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 59, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.recommendation-model", "label": "Deep recommendation model", "aliases": [ "DLRM" ], "definition": "Models combining large sparse embedding tables for categorical features with dense layers, whose cost is often dominated by embedding lookups and memory capacity.", "background": false, "home": "F2.1", "home_id": "model-families", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 33, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.diffusion-model", "label": "Diffusion model", "aliases": [], "definition": "A generative model that learns to reverse a gradual noising process, producing samples through many iterative denoising network evaluations.", "background": false, "home": "F2.1", "home_id": "model-families", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 3, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.convolution", "label": "Convolution", "aliases": [ "convolutional layer" ], "definition": "A locally connected operation applying shared filters across spatial or temporal positions, giving high weight reuse and arithmetic intensity.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 21, "organizational": 0, "user_proxy": 21 }, "book_passages": [] }, { "id": "mlsys.recurrence", "label": "Recurrent computation", "aliases": [ "RNN", "LSTM" ], "definition": "Computation that updates a hidden state step by step over a sequence, creating sequential dependencies that limit parallelism across time.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 5, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.normalization-layer", "label": "Normalization layers", "aliases": [ "BatchNorm", "LayerNorm", "RMSNorm" ], "definition": "Layers that rescale activations using batch, layer, or root-mean-square statistics to stabilize and speed training of deep networks.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 11, "organizational": 0, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.mixture-of-experts", "label": "Mixture of experts", "aliases": [ "MoE" ], "definition": "An architecture in which a learned router activates a small subset of expert subnetworks per token, decoupling total parameter count from per-token compute.", "background": false, "home": "F2.3", "home_id": "conditional-computation", "added_in": "0.2-draft", "warrant": { "literary": 9, "corpus_chapters": 18, "organizational": 0, "user_proxy": 24 }, "book_passages": [] }, { "id": "mlsys.expert-routing", "label": "Expert routing", "aliases": [ "gating", "router" ], "definition": "The gating function that assigns each token to one or more experts, such as top-k softmax gating or expert-choice routing.", "background": false, "home": "F2.3", "home_id": "conditional-computation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 33, "organizational": 1, "user_proxy": 43 }, "book_passages": [] }, { "id": "mlsys.expert-capacity", "label": "Expert capacity", "aliases": [ "capacity factor" ], "definition": "The maximum number of tokens an expert processes per batch, set by a capacity factor; tokens beyond it are dropped or rerouted.", "background": false, "home": "F2.3", "home_id": "conditional-computation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 0, "user_proxy": 15 }, "book_passages": [] }, { "id": "mlsys.load-balancing-loss", "label": "Load-balancing loss", "aliases": [ "auxiliary loss" ], "definition": "An auxiliary training objective that penalizes uneven token assignment across experts so routing stays balanced.", "background": false, "home": "F2.3", "home_id": "conditional-computation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.active-parameters", "label": "Active parameters", "aliases": [], "definition": "The parameters actually used to process one token or example, as distinct from total stored parameters in conditionally computed models.", "background": false, "home": "F2.3", "home_id": "conditional-computation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 24 }, "book_passages": [] }, { "id": "mlsys.early-exit", "label": "Early exit", "aliases": [ "adaptive computation" ], "definition": "Terminating inference at an intermediate layer once a confidence criterion is met, so per-input compute varies with input difficulty.", "background": false, "home": "F2.3", "home_id": "conditional-computation", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 6, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.linear-attention", "label": "Linear attention", "aliases": [], "definition": "Attention reformulated with kernel feature maps so it is computed in linear time with a fixed-size recurrent state.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.state-space-model", "label": "State-space sequence model", "aliases": [ "SSM", "Mamba", "S4" ], "definition": "Sequence models built on discretized linear state-space recurrences, trainable in parallel and run with constant-size state per step at inference.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 12, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.grouped-query-attention", "label": "Grouped-query attention", "aliases": [ "GQA", "multi-query attention", "MQA" ], "definition": "Attention in which groups of query heads share key and value heads, shrinking the key-value cache and decode memory traffic.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 10, "organizational": 1, "user_proxy": 50 }, "book_passages": [] }, { "id": "mlsys.context-extension", "label": "Context-length extension", "aliases": [ "RoPE scaling" ], "definition": "Techniques that extend a trained model's usable sequence length, such as position-encoding interpolation followed by long-sequence fine-tuning.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.context", "label": "Context window", "aliases": [ "context length" ], "definition": "The maximum token span a model supports in a stated execution mode, bounding the history it can use and the longest sequences the system must process.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 38, "organizational": 3, "user_proxy": 6 }, "book_passages": [ { "volume": "vol3", "chapter": "Context Engineering", "url": "https://mlsysbook.ai/vol3/04_context_engineering/04_context_engineering.html#sec-vol3-context-engineering-decision" } ] }, { "id": "mlsys.weight-serialization", "label": "Weight serialization format", "aliases": [ "safetensors", "GGUF", "pickle-based weights" ], "definition": "The on-disk encoding of a model's parameter tensors and metadata, governing load time, memory mapping, and safety of deserialization.", "background": false, "home": "F2.5", "home_id": "model-artifacts", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.model-exchange-format", "label": "Model exchange format", "aliases": [ "ONNX" ], "definition": "A framework-neutral representation of a model's operator graph and weights used to move models between training frameworks and inference runtimes.", "background": false, "home": "F2.5", "home_id": "model-artifacts", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.model-export", "label": "Model export", "aliases": [ "TorchScript", "torch.export", "SavedModel" ], "definition": "Converting a framework model into an artifact with an explicit computation and input contract that a compatible runtime can execute.", "background": false, "home": "F2.5", "home_id": "model-artifacts", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.model-signature", "label": "Model signature", "aliases": [], "definition": "The declared names, types, and shapes of a model artifact's inputs and outputs, forming its interface contract with callers.", "background": false, "home": "F2.5", "home_id": "model-artifacts", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.ridge-point", "label": "Ridge point", "aliases": [], "definition": "The arithmetic intensity at which a roofline's bandwidth and compute ceilings meet, equal to peak compute throughput divided by peak memory bandwidth.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 30, "organizational": 1, "user_proxy": 105 }, "book_passages": [] }, { "id": "mlsys.bound-regime", "label": "Compute-bound vs. memory-bound", "aliases": [ "compute-bound", "memory-bound", "bandwidth-bound", "overhead-bound" ], "definition": "Classifying a kernel or phase by what limits it: the roofline's arithmetic ceiling, its bandwidth ceiling, or, for operations too small to approach either, fixed per-operation overhead, indicating whether more throughput, more bandwidth, or fewer and larger operations would help.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 59, "organizational": 7, "user_proxy": 429 }, "book_passages": [] }, { "id": "mlsys.data-movement-energy", "label": "Data-movement energy", "aliases": [], "definition": "The energy cost of moving operands between storage levels, which grows with distance and typically exceeds the energy of the arithmetic performed on them.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 1, "user_proxy": 122 }, "book_passages": [] }, { "id": "mlsys.amdahls-law", "label": "Amdahl's law", "aliases": [], "definition": "The bound that overall speedup from accelerating part of an execution is limited by the fraction left unaccelerated.", "background": false, "home": "F3.2", "home_id": "speedup-limits", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 29, "organizational": 6, "user_proxy": 25 }, "book_passages": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html#sec-hardware-acceleration-ai-hardware-acceleration-fundamentals-9b28" } ] }, { "id": "mlsys.strong-scaling", "label": "Strong scaling", "aliases": [], "definition": "How time to solution changes as resources increase for a fixed total problem size.", "background": false, "home": "F3.2", "home_id": "speedup-limits", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 1, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.weak-scaling", "label": "Weak scaling", "aliases": [ "Gustafson's law" ], "definition": "How time to solution changes as resources and problem size grow proportionally, keeping work per resource fixed.", "background": false, "home": "F3.2", "home_id": "speedup-limits", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.parallel-efficiency", "label": "Parallel scaling efficiency", "aliases": [ "scaling efficiency" ], "definition": "Achieved speedup divided by the factor of added resources, measuring how much additional hardware turns into useful progress.", "background": false, "home": "F3.2", "home_id": "speedup-limits", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 19, "organizational": 4, "user_proxy": 11 }, "book_passages": [ { "volume": "vol2", "chapter": "Introduction", "url": "https://mlsysbook.ai/vol2/01_introduction/01_introduction.html#sec-vol2-introduction-fleet-law" } ] }, { "id": "mlsys.littles-law", "label": "Little's law", "aliases": [ "L = lambda W" ], "definition": "The relation that the mean number of items in a stable system equals mean arrival rate times mean time in the system.", "background": false, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 20, "organizational": 4, "user_proxy": 28 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-utilizationlatency-relationship-a2f0" } ] }, { "id": "mlsys.latency-throughput-curve", "label": "Latency-throughput curve", "aliases": [ "load-latency curve" ], "definition": "The relationship between offered load and latency, in which queueing delay grows sharply as utilization approaches capacity.", "background": false, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 2, "user_proxy": 27 }, "book_passages": [] }, { "id": "mlsys.arrival-burstiness", "label": "Arrival burstiness", "aliases": [], "definition": "Variability of request inter-arrival times or sizes beyond a Poisson baseline, which raises queueing delay at a given mean utilization.", "background": false, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.compute-optimal-allocation", "label": "Compute-optimal allocation", "aliases": [ "Chinchilla scaling" ], "definition": "Choosing model size and training data quantity to minimize loss under a fixed training compute budget, using fitted scaling laws.", "background": false, "home": "F3.4", "home_id": "scaling-laws", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 10, "organizational": 2, "user_proxy": 11 }, "book_passages": [ { "volume": "vol1", "chapter": "Data Selection", "url": "https://mlsysbook.ai/vol1/09_data_selection/09_data_selection.html#sec-data-selection-computeoptimal-frontier-24ba" } ] }, { "id": "mlsys.training-compute-estimate", "label": "Training compute estimate", "aliases": [ "6ND" ], "definition": "Approximating dense transformer training compute as roughly six floating-point operations per parameter per training token.", "background": false, "home": "F3.4", "home_id": "scaling-laws", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 2, "user_proxy": 32 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html#sec-model-training-time-to-quality" } ] }, { "id": "mlsys.systolic-array", "label": "Systolic array", "aliases": [], "definition": "A grid of processing elements that pass operands rhythmically to neighbors, performing matrix multiplication with high operand reuse and little register-file traffic.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 9, "organizational": 1, "user_proxy": 21 }, "book_passages": [] }, { "id": "mlsys.matrix-unit", "label": "Matrix engine", "aliases": [ "Tensor Core", "MXU" ], "definition": "A functional unit that executes small dense matrix multiply-accumulate tiles per instruction, supplying most of an accelerator's peak low-precision throughput.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 20, "organizational": 2, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.simt", "label": "SIMT execution", "aliases": [ "SIMT" ], "definition": "An execution model in which groups of threads issue common instructions to active lanes, with divergent paths executed under lane masks and architecture-specific reconvergence.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.dataflow", "label": "Accelerator dataflow", "aliases": [ "stationarity" ], "definition": "The mapping that decides which operands stay resident in local storage while others stream, such as weight-, output-, or row-stationary schedules.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 4, "organizational": 2, "user_proxy": 66 }, "book_passages": [ { "volume": "vol1", "chapter": "Hardware Acceleration", "url": "https://mlsysbook.ai/vol1/11_hw_acceleration/11_hw_acceleration.html#sec-hardware-acceleration-matrix-operations-neural-networks-527a" } ] }, { "id": "mlsys.domain-specific-architecture", "label": "Domain-specific architecture", "aliases": [ "DSA" ], "definition": "Hardware specialized to a class of workloads, trading programmability for higher performance per watt on its target operations.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.2-draft", "warrant": { "literary": 5, "corpus_chapters": 1, "organizational": 1, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.hbm", "label": "High-bandwidth memory", "aliases": [ "HBM" ], "definition": "Stacked DRAM placed beside the processor on an interposer, providing very high bandwidth with limited capacity per stack.", "background": false, "home": "S1.2", "home_id": "memory-systems", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 49, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.on-chip-sram", "label": "On-chip SRAM and scratchpads", "aliases": [], "definition": "Small, fast on-die storage, such as caches, shared memory, and register files, used to stage reused operands near compute units.", "background": false, "home": "S1.2", "home_id": "memory-systems", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 36 }, "book_passages": [] }, { "id": "mlsys.memory-wall", "label": "Memory wall", "aliases": [], "definition": "The widening gap between processor arithmetic throughput and memory bandwidth or latency, which makes data movement the limiting factor.", "background": true, "home": "S1.2", "home_id": "memory-systems", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 30, "organizational": 2, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.scale-up-interconnect", "label": "Scale-up interconnect", "aliases": [], "definition": "High-bandwidth links and switches, such as NVLink-class fabrics, joining accelerators within a node or rack into one tightly coupled domain.", "background": false, "home": "S1.3", "home_id": "node-interconnect", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 1, "user_proxy": 90 }, "book_passages": [] }, { "id": "mlsys.coherent-interconnect", "label": "Cache-coherent interconnect", "aliases": [ "CXL" ], "definition": "A link protocol that lets hosts and devices share or pool memory with hardware coherence, such as CXL over a PCIe physical layer.", "background": false, "home": "S1.3", "home_id": "node-interconnect", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 6, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.peer-to-peer-transfer", "label": "Peer-to-peer device transfer", "aliases": [ "GPUDirect" ], "definition": "Direct memory access between accelerators, or between accelerators and NICs or storage, without staging data through host memory.", "background": false, "home": "S1.3", "home_id": "node-interconnect", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 6, "organizational": 2, "user_proxy": 42 }, "book_passages": [] }, { "id": "mlsys.tdp", "label": "Thermal design power", "aliases": [ "TDP" ], "definition": "The sustained heat dissipation a cooling solution is designed to remove, used as a chip's nominal power envelope.", "background": true, "home": "S1.4", "home_id": "chip-power-thermal", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 20, "organizational": 1, "user_proxy": 18 }, "book_passages": [] }, { "id": "mlsys.dvfs", "label": "Dynamic voltage and frequency scaling", "aliases": [ "DVFS" ], "definition": "Adjusting supply voltage and clock frequency at runtime to trade performance for power, since dynamic power grows superlinearly with voltage.", "background": true, "home": "S1.4", "home_id": "chip-power-thermal", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 9, "organizational": 1, "user_proxy": 70 }, "book_passages": [] }, { "id": "mlsys.dark-silicon", "label": "Dark silicon", "aliases": [], "definition": "The fraction of a chip that cannot be powered at full frequency simultaneously within its power and thermal budget.", "background": true, "home": "S1.4", "home_id": "chip-power-thermal", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.power-capping", "label": "Power capping", "aliases": [], "definition": "Enforcing a maximum power draw on a device or group by limiting clocks or activity, trading peak throughput for bounded power.", "background": false, "home": "S1.4", "home_id": "chip-power-thermal", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 63 }, "book_passages": [] }, { "id": "mlsys.microcontroller", "label": "Microcontroller-class device", "aliases": [ "MCU", "TinyML device" ], "definition": "A low-power processor with on-chip flash and kilobytes to a few megabytes of SRAM, typically without an operating system or virtual memory.", "background": false, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 25, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.heterogeneous-soc", "label": "Heterogeneous mobile SoC", "aliases": [], "definition": "A system-on-chip combining CPU, GPU, DSP, and NPU blocks under shared memory and power limits, across which model operators are partitioned.", "background": false, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.2-draft", "warrant": { "literary": 5, "corpus_chapters": 0, "organizational": 0, "user_proxy": 29 }, "book_passages": [] }, { "id": "mlsys.edge-npu", "label": "On-device neural accelerator", "aliases": [ "NPU" ], "definition": "An accelerator integrated in mobile or embedded silicon for low-precision neural inference, usually supporting a restricted set of operators.", "background": false, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.2-draft", "warrant": { "literary": 5, "corpus_chapters": 17, "organizational": 0, "user_proxy": 18 }, "book_passages": [] }, { "id": "mlsys.energy-efficiency", "label": "Energy efficiency", "aliases": [ "performance per watt" ], "definition": "Useful work completed per unit energy, such as inferences or tokens per joule, measured within a stated system boundary.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 18, "organizational": 1, "user_proxy": 137 }, "book_passages": [] }, { "id": "mlsys.on-device-memory-budget", "label": "On-device memory budget", "aliases": [ "app memory limit", "mobile memory budget" ], "definition": "The flash and SRAM limits that bound a deployed model's stored weights and peak activation memory on embedded hardware.", "background": false, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 1, "user_proxy": 43 }, "book_passages": [] }, { "id": "mlsys.tensor", "label": "Tensor", "aliases": [], "definition": "A multidimensional array carrying shape, element type, device, and layout metadata, the core data abstraction of ML frameworks.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 69, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.computation-graph", "label": "Computation graph", "aliases": [], "definition": "A directed graph of tensor operations and their data dependencies representing a model program, used for differentiation, optimization, and scheduling.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 12, "organizational": 0, "user_proxy": 10 }, "book_passages": [] }, { "id": "mlsys.eager-execution", "label": "Eager vs. graph execution", "aliases": [ "define-by-run" ], "definition": "The choice between running each operation as the host program issues it and building a whole graph first for later optimization and execution.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 2, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.tracing-jit", "label": "Graph tracing", "aliases": [], "definition": "Recording supported operations during a program execution with concrete or abstract inputs to construct an intermediate graph for transformation, compilation, or export.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 1, "user_proxy": 23 }, "book_passages": [] }, { "id": "mlsys.io-aware-attention", "label": "IO-aware attention kernel", "aliases": [ "FlashAttention" ], "definition": "An exact attention implementation that tiles queries, keys, and values through on-chip memory to avoid materializing the full score matrix in off-chip memory.", "background": false, "home": "S2.2", "home_id": "tensor-operators", "added_in": "0.2-draft", "warrant": { "literary": 10, "corpus_chapters": 10, "organizational": 1, "user_proxy": 115 }, "book_passages": [] }, { "id": "mlsys.convolution-lowering", "label": "Convolution lowering", "aliases": [ "im2col" ], "definition": "Implementing convolution by rearranging input patches into matrices for GEMM, or by direct, FFT, or Winograd algorithms with different memory costs.", "background": false, "home": "S2.2", "home_id": "tensor-operators", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 1, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.kernel-library", "label": "Vendor kernel library", "aliases": [ "cuBLAS", "cuDNN" ], "definition": "Tuned implementations of common operators supplied for a hardware platform, which frameworks call rather than generating their own kernels.", "background": false, "home": "S2.2", "home_id": "tensor-operators", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 7, "organizational": 1, "user_proxy": 20 }, "book_passages": [] }, { "id": "mlsys.intermediate-representation", "label": "ML compiler IR", "aliases": [ "graph IR", "MLIR" ], "definition": "Program representations at graph, tensor, and loop levels on which compilers apply lowering and optimization passes.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.autotuning", "label": "Kernel autotuning", "aliases": [], "definition": "Searching kernel schedule parameters, such as tile sizes and loop orders, by measurement or learned cost models to find fast implementations for a target.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 0, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.kernel-dsl", "label": "Kernel DSL", "aliases": [ "Triton" ], "definition": "A domain-specific language for writing tiled tensor kernels above the level of vendor assembly, with the compiler handling memory and thread mapping.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 5, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.memory-planning", "label": "Static memory planning", "aliases": [], "definition": "Compile-time assignment of tensor buffers to memory with lifetime-based reuse, reducing peak memory and runtime allocation.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 128 }, "book_passages": [] }, { "id": "mlsys.device-streams", "label": "Device streams", "aliases": [ "CUDA streams" ], "definition": "Ordered queues of device work that the host enqueues asynchronously, with events expressing dependencies and allowing overlap across queues.", "background": false, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 9, "organizational": 0, "user_proxy": 49 }, "book_passages": [] }, { "id": "mlsys.caching-allocator", "label": "Caching memory allocator", "aliases": [], "definition": "A runtime allocator that retains and reuses freed device memory blocks to avoid costly allocation calls, at the risk of fragmentation.", "background": false, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.graph-replay", "label": "Captured graph replay", "aliases": [ "CUDA Graphs" ], "definition": "Recording a fixed sequence of kernel launches once and replaying it as a unit to remove per-launch host overhead.", "background": false, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 6, "organizational": 1, "user_proxy": 22 }, "book_passages": [] }, { "id": "mlsys.loss-scaling", "label": "Loss scaling", "aliases": [], "definition": "Multiplying the loss before backpropagation so small gradients remain representable in reduced-precision formats, then unscaling before the optimizer step.", "background": false, "home": "S2.7", "home_id": "numerics", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 10, "organizational": 1, "user_proxy": 16 }, "book_passages": [] }, { "id": "mlsys.block-scaled-format", "label": "Block-scaled number formats", "aliases": [ "MX", "microscaling", "MXFP8", "FP8 with block scaling" ], "definition": "Low-precision formats in which small blocks of values share a scale factor, extending dynamic range for FP8, FP6, FP4, or INT8 elements.", "background": false, "home": "S2.7", "home_id": "numerics", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 1, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.accumulation-precision", "label": "Accumulation precision", "aliases": [], "definition": "The numerical format used to sum products in dot products and reductions, which may differ from input precision and governs rounding-error growth.", "background": false, "home": "S2.7", "home_id": "numerics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 16 }, "book_passages": [] }, { "id": "mlsys.expert-parallel", "label": "Expert parallelism", "aliases": [ "EP" ], "definition": "Placing different experts of a mixture-of-experts layer on different devices and exchanging routed tokens among them.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 7, "corpus_chapters": 7, "organizational": 0, "user_proxy": 28 }, "book_passages": [] }, { "id": "mlsys.sequence-parallel", "label": "Sequence and context parallelism", "aliases": [ "context parallelism", "ring attention", "sequence parallelism" ], "definition": "Partitioning activations along the sequence dimension across devices, exchanging keys, values, or partial results so long sequences fit.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 7, "corpus_chapters": 3, "organizational": 0, "user_proxy": 17 }, "book_passages": [] }, { "id": "mlsys.hybrid-parallel", "label": "Hybrid parallelism", "aliases": [ "3D parallelism" ], "definition": "Combining data, tensor, pipeline, and other parallelism dimensions, mapped onto the hardware hierarchy so the most communication-intensive dimension uses the fastest links.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 12, "organizational": 2, "user_proxy": 62 }, "book_passages": [] }, { "id": "mlsys.pipeline-bubble", "label": "Pipeline bubble", "aliases": [], "definition": "Idle device time during pipeline fill and drain, whose fraction shrinks as the number of microbatches grows relative to pipeline stages.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 16, "organizational": 2, "user_proxy": 69 }, "book_passages": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html#sec-distributed-training-systems-systems-pipeline-parallelism-8748" } ] }, { "id": "mlsys.ring-allreduce", "label": "Ring all-reduce", "aliases": [], "definition": "An all-reduce performed as reduce-scatter then all-gather around a ring, bandwidth-optimal for large messages but with latency linear in participants.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 4, "user_proxy": 97 }, "book_passages": [] }, { "id": "mlsys.tree-allreduce", "label": "Tree and hierarchical all-reduce", "aliases": [], "definition": "All-reduce schedules over trees or hierarchical groups, with latency logarithmic in participants, suited to small messages and multi-level networks.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 2, "user_proxy": 51 }, "book_passages": [] }, { "id": "mlsys.all-to-all", "label": "All-to-all exchange", "aliases": [ "AllToAll" ], "definition": "A collective in which every participant sends distinct data to every other, as in expert routing; its throughput is sensitive to bisection bandwidth.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 14, "organizational": 3, "user_proxy": 41 }, "book_passages": [] }, { "id": "mlsys.alpha-beta-model", "label": "Alpha-beta communication model", "aliases": [ "latency-bandwidth model", "network transfer time", "transfer time" ], "definition": "A cost model estimating message time as a fixed latency term plus message size times a per-byte cost, used to compare collective algorithms.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 38, "organizational": 3, "user_proxy": 49 }, "book_passages": [ { "volume": "vol2", "chapter": "Network Fabrics", "url": "https://mlsysbook.ai/vol2/03_network_fabrics/03_network_fabrics.html#sec-network-fabrics-performance-model" } ] }, { "id": "mlsys.gradient-compression", "label": "Gradient compression", "aliases": [], "definition": "Reducing gradient communication through quantization, sparsification, or low-rank approximation, often with error feedback to preserve convergence.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 8, "organizational": 3, "user_proxy": 67 }, "book_passages": [ { "volume": "vol2", "chapter": "Collective Communication", "url": "https://mlsysbook.ai/vol2/05_collective_communication/05_collective_communication.html#sec-communication-collective-operations-collective-operations-gradient-compression-7a5c" } ] }, { "id": "mlsys.lease", "label": "Lease", "aliases": [], "definition": "A time-bounded grant of authority or cached state that expires unless renewed, letting a system recover from failed holders without explicit release.", "background": true, "home": "S3.4", "home_id": "coordination", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 46, "organizational": 7, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "The Nervous System", "url": "https://mlsysbook.ai/vol4/04_nervous/04_nervous.html#sec-nervous-lease-sizing" } ] }, { "id": "mlsys.leader-election", "label": "Leader election", "aliases": [], "definition": "Selecting a single coordinator among replicas, typically through consensus and leases, so control decisions have one authoritative source.", "background": true, "home": "S3.4", "home_id": "coordination", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.bounded-staleness", "label": "Bounded staleness", "aliases": [ "stale synchronous parallel" ], "definition": "A consistency model in which workers may compute on parameters up to a fixed number of updates old, trading freshness for less waiting.", "background": false, "home": "S3.4", "home_id": "coordination", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 1, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.federated-averaging", "label": "Federated averaging", "aliases": [ "FedAvg" ], "definition": "A federated algorithm in which sampled clients run several local optimization steps and a server averages their model updates, weighted by data size.", "background": false, "home": "S3.5", "home_id": "federated-learning", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 5, "organizational": 2, "user_proxy": 29 }, "book_passages": [ { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html#sec-edge-intelligence-learning-protocols-139a" } ] }, { "id": "mlsys.non-iid-data", "label": "Statistical heterogeneity", "aliases": [ "non-IID data" ], "definition": "Differences in data distribution and quantity across federated clients, which slow convergence and can bias a shared model.", "background": false, "home": "S3.5", "home_id": "federated-learning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 2, "user_proxy": 24 }, "book_passages": [] }, { "id": "mlsys.client-sampling", "label": "Client selection", "aliases": [], "definition": "Choosing which available devices participate in each federated round, subject to availability, connectivity, and power constraints.", "background": false, "home": "S3.5", "home_id": "federated-learning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 2, "user_proxy": 22 }, "book_passages": [ { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html#sec-edge-intelligence-client-scheduling-f675" } ] }, { "id": "mlsys.federation-setting", "label": "Cross-device vs. cross-silo federation", "aliases": [], "definition": "The distinction between federations of many intermittently available devices and federations of a few reliable organizational data holders.", "background": false, "home": "S3.5", "home_id": "federated-learning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.pue", "label": "Power usage effectiveness", "aliases": [ "PUE" ], "definition": "Total facility energy divided by IT equipment energy, measuring overhead from cooling, power conversion, and other non-compute loads.", "background": true, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 11, "organizational": 3, "user_proxy": 59 }, "book_passages": [ { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html#sec-sustainable-ai-operational-carbon-calculation-4627-calculation-4627" } ] }, { "id": "mlsys.liquid-cooling", "label": "Liquid cooling", "aliases": [], "definition": "Removing heat with liquid through cold plates or immersion, enabling rack power densities beyond the practical limits of air cooling.", "background": true, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 9, "organizational": 0, "user_proxy": 22 }, "book_passages": [] }, { "id": "mlsys.rack-power-density", "label": "Rack power density", "aliases": [], "definition": "Electrical power drawn per rack, which determines cooling technology, floor layout, and how many accelerators fit within a facility's power envelope.", "background": false, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 2, "user_proxy": 35 }, "book_passages": [ { "volume": "vol2", "chapter": "Compute Infrastructure", "url": "https://mlsysbook.ai/vol2/02_compute_infrastructure/02_compute_infrastructure.html#sec-compute-power-wall" } ] }, { "id": "mlsys.power-oversubscription", "label": "Power oversubscription", "aliases": [], "definition": "Deploying more equipment than the sum of nameplate power a feed supports, relying on non-coincident peaks and power capping to stay within limits.", "background": true, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.warehouse-scale-computer", "label": "Warehouse-scale computer", "aliases": [ "WSC" ], "definition": "A datacenter treated as a single computer whose power, cooling, network, and software are co-designed for a common workload.", "background": true, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.rdma", "label": "Remote direct memory access", "aliases": [ "RDMA", "InfiniBand", "RoCE" ], "definition": "Network transfers that read or write remote memory directly through the network interface without remote CPU involvement, giving low latency and high throughput.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 24, "organizational": 1, "user_proxy": 106 }, "book_passages": [] }, { "id": "mlsys.lossless-fabric", "label": "Lossless network fabric", "aliases": [ "PFC" ], "definition": "A network that avoids packet drops through credit-based or priority flow control, as RDMA transports expect, at the risk of head-of-line blocking.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 2, "user_proxy": 47 }, "book_passages": [] }, { "id": "mlsys.congestion-control", "label": "Datacenter congestion control", "aliases": [ "DCQCN" ], "definition": "End-host and switch mechanisms, such as ECN-driven rate control, that limit queue buildup from incast and synchronized collective traffic.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 4, "user_proxy": 58 }, "book_passages": [ { "volume": "vol2", "chapter": "Network Fabrics", "url": "https://mlsysbook.ai/vol2/03_network_fabrics/03_network_fabrics.html#sec-network-fabrics-behavior" } ] }, { "id": "mlsys.fat-tree", "label": "Fat-tree (Clos) topology", "aliases": [ "Clos" ], "definition": "A multistage switch topology providing many equal-cost paths and, without oversubscription, full bisection bandwidth between endpoints.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 1, "user_proxy": 19 }, "book_passages": [] }, { "id": "mlsys.rail-optimized", "label": "Rail-optimized topology", "aliases": [], "definition": "A fabric design connecting the same-index accelerator of every node to a common leaf switch, keeping most collective traffic within one switch hop.", "background": false, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 1, "user_proxy": 15 }, "book_passages": [] }, { "id": "mlsys.bisection-bandwidth", "label": "Bisection bandwidth", "aliases": [], "definition": "The minimum total bandwidth across any cut dividing a network into two equal halves, bounding throughput of all-to-all and global exchanges.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 12, "organizational": 3, "user_proxy": 31 }, "book_passages": [ { "volume": "vol2", "chapter": "Network Fabrics", "url": "https://mlsysbook.ai/vol2/03_network_fabrics/03_network_fabrics.html#sec-network-fabrics-topology" } ] }, { "id": "mlsys.in-network-reduction", "label": "In-network reduction", "aliases": [ "SHARP" ], "definition": "Performing collective reductions inside network switches so data crosses each link fewer times, lowering all-reduce traffic and latency.", "background": false, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.parallel-file-system", "label": "Parallel file system", "aliases": [ "Lustre", "GPFS" ], "definition": "A distributed file system that stripes data across many storage servers to deliver aggregate bandwidth to many concurrent clients.", "background": true, "home": "S4.3", "home_id": "storage-systems", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 2, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.object-storage", "label": "Object storage", "aliases": [], "definition": "Storage of immutable named objects behind an HTTP-style interface, offering scalable capacity and throughput with higher per-request latency than file systems.", "background": true, "home": "S4.3", "home_id": "storage-systems", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 10, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.input-pipeline", "label": "Training input pipeline", "aliases": [ "data echoing" ], "definition": "The read, decode, transform, shuffle, and prefetch stages that deliver batches to accelerators; when slower than compute, they cause data stalls.", "background": false, "home": "S2.6", "home_id": "data-loading", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 1, "organizational": 5, "user_proxy": 43 }, "book_passages": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html#sec-data-engineering-ml-storage-requirements-performance-1b0d" }, { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html#sec-storage-pipeline-equation" } ] }, { "id": "mlsys.checkpoint-io", "label": "Checkpoint I/O", "aliases": [], "definition": "Writing and reading large training-state snapshots, where storage bandwidth, asynchronous persistence, and sharded layout determine checkpoint cost.", "background": false, "home": "S4.3", "home_id": "storage-systems", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 7, "organizational": 1, "user_proxy": 53 }, "book_passages": [] }, { "id": "mlsys.storage-tiering", "label": "Storage tiering and caching", "aliases": [], "definition": "Placing data across tiers, from memory and local NVMe caches to shared and archival storage, according to access frequency and cost.", "background": false, "home": "S4.3", "home_id": "storage-systems", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 3, "user_proxy": 40 }, "book_passages": [] }, { "id": "mlsys.gang-scheduling", "label": "Gang scheduling", "aliases": [], "definition": "Allocating all resources a parallel job needs at once, or none, so tightly coupled workers start together and partial allocations cannot deadlock.", "background": true, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 5, "organizational": 2, "user_proxy": 25 }, "book_passages": [ { "volume": "vol2", "chapter": "Fleet Orchestration", "url": "https://mlsysbook.ai/vol2/08_fleet_orchestration/08_fleet_orchestration.html#sec-fleet-orchestration-gang-scheduling" } ] }, { "id": "mlsys.job-preemption", "label": "Job preemption", "aliases": [], "definition": "Suspending or evicting a running job to reclaim its resources for higher-priority work, at a cost set by checkpointing and lost progress.", "background": true, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 2, "organizational": 1, "user_proxy": 14 }, "book_passages": [] }, { "id": "mlsys.dominant-resource-fairness", "label": "Dominant resource fairness", "aliases": [ "DRF" ], "definition": "A multi-resource fair-sharing policy that equalizes each user's largest share across resource types such as accelerators, CPUs, and memory.", "background": true, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 2, "user_proxy": 1 }, "book_passages": [ { "volume": "vol2", "chapter": "Fleet Orchestration", "url": "https://mlsysbook.ai/vol2/08_fleet_orchestration/08_fleet_orchestration.html#sec-fleet-orchestration-core-allocation" } ] }, { "id": "mlsys.elastic-training", "label": "Elastic training", "aliases": [], "definition": "Training that continues while the number of workers changes, resharding state and adjusting batch size or learning rate as resources join or leave.", "background": false, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 7, "organizational": 2, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.topology-aware-placement", "label": "Topology-aware placement", "aliases": [], "definition": "Assigning a job's workers to devices and nodes so that heavily communicating groups share fast links or switches.", "background": false, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 3, "user_proxy": 37 }, "book_passages": [ { "volume": "vol2", "chapter": "Fleet Orchestration", "url": "https://mlsysbook.ai/vol2/08_fleet_orchestration/08_fleet_orchestration.html#sec-fleet-orchestration-topology-aware" } ] }, { "id": "mlsys.bin-packing", "label": "Bin packing", "aliases": [], "definition": "Placing jobs onto nodes to minimize fragmentation of resources such as accelerators, trading packing density against interference and locality.", "background": true, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 2, "user_proxy": 21 }, "book_passages": [] }, { "id": "mlsys.multi-tenancy", "label": "Multi-tenancy", "aliases": [], "definition": "Sharing infrastructure among teams, jobs, or customers under policies for allocation, isolation, priority, and accounting.", "background": false, "home": "S4.5", "home_id": "capacity-provisioning", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 6, "organizational": 1, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.resource-quota", "label": "Resource quota", "aliases": [], "definition": "A limit on the resources a tenant may hold or request, sometimes with guaranteed minimums and borrowing of idle capacity.", "background": false, "home": "S4.5", "home_id": "capacity-provisioning", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.gpu-partitioning", "label": "Accelerator sharing and partitioning", "aliases": [ "MIG", "GPU sharing", "SR-IOV" ], "definition": "Running several workloads on one accelerator by time-slicing, concurrent contexts, or hardware partitions with dedicated memory and compute slices.", "background": false, "home": "S4.5", "home_id": "capacity-provisioning", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 6, "organizational": 0, "user_proxy": 30 }, "book_passages": [] }, { "id": "mlsys.spot-capacity", "label": "Spot and preemptible capacity", "aliases": [], "definition": "Discounted compute that the provider may reclaim at short notice, usable by workloads that tolerate interruption.", "background": false, "home": "S4.5", "home_id": "capacity-provisioning", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 1, "user_proxy": 17 }, "book_passages": [] }, { "id": "mlsys.performance-interference", "label": "Performance interference", "aliases": [ "noisy neighbor" ], "definition": "Slowdown of one workload caused by co-located workloads contending for shared caches, memory bandwidth, network, or power.", "background": false, "home": "S4.5", "home_id": "capacity-provisioning", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 7, "organizational": 2, "user_proxy": 111 }, "book_passages": [] }, { "id": "mlsys.weak-supervision", "label": "Weak supervision", "aliases": [], "definition": "Producing training labels from noisy programmatic sources such as heuristics, rules, or distant supervision, then modeling and combining their errors.", "background": false, "home": "P1.1", "home_id": "data-collection", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.active-learning", "label": "Active learning", "aliases": [], "definition": "Selecting which unlabeled examples to annotate next according to their expected value to the model.", "background": false, "home": "P1.1", "home_id": "data-collection", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 0, "user_proxy": 56 }, "book_passages": [] }, { "id": "mlsys.annotator-agreement", "label": "Inter-annotator agreement", "aliases": [], "definition": "The degree to which independent annotators assign the same label to the same item, used to measure label reliability and task ambiguity.", "background": false, "home": "P1.1", "home_id": "data-collection", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 5, "organizational": 2, "user_proxy": 17 }, "book_passages": [ { "volume": "vol1", "chapter": "Data Engineering", "url": "https://mlsysbook.ai/vol1/04_data_engineering/04_data_engineering.html#sec-data-engineering-data-labeling-6836" } ] }, { "id": "mlsys.quality-filtering", "label": "Quality filtering", "aliases": [], "definition": "Removing or down-weighting training examples judged low quality by heuristics, classifiers, or model scores.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 2, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.data-mixture", "label": "Data mixture", "aliases": [ "domain mixture" ], "definition": "The proportions in which source domains or datasets are sampled during training.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.synthetic-data", "label": "Synthetic data", "aliases": [], "definition": "Training or evaluation examples generated by simulators, programs, or models rather than collected from the target environment.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 9, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.data-augmentation", "label": "Data augmentation", "aliases": [], "definition": "Creating additional training examples by applying label-preserving transformations to existing ones.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 10, "organizational": 0, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.data-validation", "label": "Data validation", "aliases": [], "definition": "Automated checks of incoming data against expected schema, value ranges, and distributions before it is used for training or serving.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 7, "organizational": 3, "user_proxy": 63 }, "book_passages": [] }, { "id": "mlsys.feature-store", "label": "Feature store", "aliases": [], "definition": "A system that defines, computes, stores, and serves features consistently for offline training and online inference.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 11, "organizational": 1, "user_proxy": 29 }, "book_passages": [] }, { "id": "mlsys.stream-processing", "label": "Stream processing", "aliases": [], "definition": "Continuously transforming unbounded event streams with low latency and incrementally maintained state.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 0, "user_proxy": 20 }, "book_passages": [] }, { "id": "mlsys.feature-backfill", "label": "Feature backfill", "aliases": [], "definition": "Recomputing a feature's historical values under a new or corrected definition so training data matches what serving will produce.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.data-versioning", "label": "Dataset versioning", "aliases": [], "definition": "Recording immutable, addressable versions of datasets so that a training run or evaluation can identify exactly which data it used.", "background": false, "home": "P1.5", "home_id": "data-lineage", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 3, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.data-dependency", "label": "Data dependency debt", "aliases": [], "definition": "Undeclared, unstable, or underused input signals on which a model depends, which can change or break its behavior silently.", "background": false, "home": "P1.5", "home_id": "data-lineage", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.feature-freshness", "label": "Feature freshness", "aliases": [], "definition": "How recent a feature's value is at the moment it is used for a prediction, relative to the recency assumed when training examples were built; stale values shift inputs away from the training distribution.", "background": false, "home": "P1.6", "home_id": "data-leakage-skew", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 2, "user_proxy": 18 }, "book_passages": [] }, { "id": "mlsys.learning-rate-schedule", "label": "Learning-rate schedule", "aliases": [], "definition": "The planned variation of the optimizer step size over training, such as warmup followed by decay.", "background": false, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.critical-batch-size", "label": "Critical batch size", "aliases": [], "definition": "The batch size beyond which increasing it further yields diminishing reductions in optimization steps to reach a target loss.", "background": false, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 2, "user_proxy": 7 }, "book_passages": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html#sec-distributed-training-systems-systems-critical-batch-size-parallelism-hurt-4961" } ] }, { "id": "mlsys.training-instability", "label": "Training instability", "aliases": [], "definition": "Loss spikes, divergence, or collapse during training, often mitigated by gradient clipping, schedule changes, normalization, or restarting from a checkpoint.", "background": false, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 15 }, "book_passages": [] }, { "id": "mlsys.gradient-accumulation", "label": "Gradient accumulation", "aliases": [], "definition": "Summing gradients over several micro-batches before one optimizer update to emulate a larger batch within limited memory.", "background": false, "home": "P2.3", "home_id": "training-memory", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 14, "organizational": 1, "user_proxy": 32 }, "book_passages": [] }, { "id": "mlsys.memory-offloading", "label": "Training-state offloading", "aliases": [], "definition": "Moving optimizer state, parameters, or activations from accelerator memory to host memory or storage between uses.", "background": false, "home": "P2.3", "home_id": "training-memory", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 0, "organizational": 1, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.experiment-tracking", "label": "Experiment tracking", "aliases": [], "definition": "Recording each run's code, data, configuration, metrics, and artifacts in a queryable store for comparison and reproduction.", "background": false, "home": "P2.4", "home_id": "experiment-management", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.hyperparameter-transfer", "label": "Hyperparameter transfer", "aliases": [ "muP" ], "definition": "Tuning hyperparameters on smaller proxy models and reusing them at larger scale under a parameterization designed to keep them stable.", "background": false, "home": "P2.4", "home_id": "experiment-management", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.supervised-fine-tuning", "label": "Supervised fine-tuning", "aliases": [ "SFT", "instruction tuning" ], "definition": "Continued training of a pretrained model on input-output demonstrations, such as instruction-response pairs, to shape task behavior.", "background": false, "home": "P2.5", "home_id": "fine-tuning", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 11, "organizational": 3, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.lora", "label": "Low-rank adaptation", "aliases": [ "LoRA", "QLoRA", "low-rank adapters" ], "definition": "Fine-tuning by learning low-rank update matrices added to frozen pretrained weights.", "background": false, "home": "P2.5", "home_id": "fine-tuning", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 10, "organizational": 2, "user_proxy": 54 }, "book_passages": [] }, { "id": "mlsys.reward-model", "label": "Reward model", "aliases": [], "definition": "A learned model that scores outputs or trajectories to supply the training signal for policy optimization.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.preference-optimization", "label": "Direct preference optimization", "aliases": [ "DPO" ], "definition": "Training a policy directly on preferred-versus-rejected output pairs with a classification-style loss, without a separate reward model or RL loop.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 1, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.rlvr", "label": "RL with verifiable rewards", "aliases": [ "RLVR" ], "definition": "Reinforcement learning whose rewards come from programmatic outcome checks, making verifier correctness and reward computation part of the training system.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Reinforcement Learning from Verifiable Rewards", "url": "https://mlsysbook.ai/vol3/17_rlvr/17_rlvr.html#sec-vol3-rlvr-score-intended-outcome" } ] }, { "id": "mlsys.rl-rollout", "label": "RL rollout generation", "aliases": [], "definition": "Sampling responses or trajectories from the current policy, often with an inference engine, to produce experience for policy updates.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.2-draft", "warrant": { "literary": 5, "corpus_chapters": 0, "organizational": 3, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.weight-sync", "label": "Policy weight synchronization", "aliases": [], "definition": "Transferring updated policy weights from training workers to rollout workers, whose cost and frequency shape RL post-training throughput.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.policy-staleness", "label": "Policy staleness", "aliases": [ "off-policy lag" ], "definition": "The lag between the policy version that generated rollouts and the version being updated, introduced by asynchronous rollout and training.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.credit-assignment", "label": "Credit assignment", "aliases": [ "temporal credit assignment", "multi-turn credit assignment" ], "definition": "Attributing a delayed or episode-level reward to the individual actions, turns, or tokens that produced it, so updates reinforce the responsible decisions.", "background": true, "home": "P2.6", "home_id": "post-training", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 2, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Reinforcement Learning from Verifiable Rewards", "url": "https://mlsysbook.ai/vol3/17_rlvr/17_rlvr.html#sec-vol3-rlvr-credit" } ] }, { "id": "mlsys.online-learning", "label": "Online learning", "aliases": [], "definition": "Updating a model incrementally as each example or small batch arrives, rather than retraining on a fixed dataset.", "background": false, "home": "P2.7", "home_id": "continual-learning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.incremental-training", "label": "Incremental training", "aliases": [ "stateful training", "warm-start training", "continued pretraining", "domain-adaptive pretraining", "continual pretraining" ], "definition": "Continuing training of an existing model from its current weights on new data instead of training from scratch.", "background": false, "home": "P2.7", "home_id": "continual-learning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.catastrophic-forgetting", "label": "Catastrophic forgetting", "aliases": [], "definition": "Loss of previously learned capability when a model is trained on new data or tasks.", "background": false, "home": "P2.7", "home_id": "continual-learning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 15 }, "book_passages": [] }, { "id": "mlsys.experience-replay", "label": "Experience replay", "aliases": [], "definition": "Mixing stored past examples into updates on new data to reduce forgetting or reuse experience.", "background": true, "home": "P2.7", "home_id": "continual-learning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.on-device-training", "label": "On-device training", "aliases": [], "definition": "Updating model parameters on the deployment device using local data under its memory, energy, and compute limits.", "background": false, "home": "P2.7", "home_id": "continual-learning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 3, "user_proxy": 38 }, "book_passages": [] }, { "id": "mlsys.cross-validation", "label": "Cross-validation", "aliases": [], "definition": "Estimating performance by repeatedly training on some data partitions and evaluating on the held-out remainder.", "background": true, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.sequential-testing", "label": "Sequential testing", "aliases": [ "SPRT", "always-valid inference" ], "definition": "Experimental analysis that permits repeated looks at accumulating data while controlling error rates, unlike fixed-horizon tests.", "background": false, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.experiment-interference", "label": "Experiment interference", "aliases": [ "SUTVA violation" ], "definition": "Treatment of one experimental unit affecting outcomes of another, violating the no-interference assumption (SUTVA) of standard A/B analysis.", "background": false, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.conformal-prediction", "label": "Conformal prediction", "aliases": [], "definition": "Constructing prediction sets from conformity scores with finite-sample coverage guarantees under stated assumptions, such as exchangeability.", "background": true, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol2", "chapter": "Robust AI", "url": "https://mlsysbook.ai/vol2/14_robust_ai/14_robust_ai.html#sec-robust-ai-selective-prediction" } ] }, { "id": "mlsys.slice-discovery", "label": "Slice discovery", "aliases": [], "definition": "Automatically finding coherent data subsets on which a model underperforms, without predefined slice labels.", "background": false, "home": "P3.3", "home_id": "robustness-slices", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.llm-judge", "label": "Model-based grading", "aliases": [ "LLM-as-judge" ], "definition": "Using a language model to score or compare generated outputs against a rubric, with its own biases and error rates to be measured.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.pass-at-k", "label": "pass@k", "aliases": [], "definition": "The probability that at least one of k sampled outputs satisfies a correctness check.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Test-Time Compute", "url": "https://mlsysbook.ai/vol3/03_test_time_compute/03_test_time_compute.html#sec-vol3-test-time-compute-candidate-diversity" }, { "volume": "vol3", "chapter": "Agent Evaluation", "url": "https://mlsysbook.ai/vol3/13_evaluation/13_evaluation.html#sec-vol3-evaluation-pass-k" } ] }, { "id": "mlsys.pairwise-preference-eval", "label": "Pairwise preference evaluation", "aliases": [], "definition": "Ranking models from judgments of which of two outputs is better, typically aggregated with a Bradley-Terry or Elo-style model.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.human-evaluation", "label": "Human evaluation", "aliases": [], "definition": "Assessment of system outputs by people following a protocol, used where automatic metrics or model graders are not trusted.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.benchmark-contamination", "label": "Benchmark contamination", "aliases": [], "definition": "Presence of evaluation items, or close variants, in a model's training data, inflating measured performance.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.time-to-train", "label": "Time to train", "aliases": [], "definition": "Wall-clock time for a system to train a specified model to a specified quality target under fixed run rules.", "background": false, "home": "Q1.2", "home_id": "system-benchmarking", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 1, "user_proxy": 7 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html#sec-model-training-time-to-quality" } ] }, { "id": "mlsys.inference-load-scenario", "label": "Inference load scenario", "aliases": [], "definition": "A specified query-arrival pattern and latency constraint, such as single-stream, server, or offline, under which inference performance is measured.", "background": false, "home": "Q1.2", "home_id": "system-benchmarking", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 2, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.metamorphic-testing", "label": "Metamorphic testing", "aliases": [], "definition": "Testing by checking that known relations hold between outputs for related inputs, when correct outputs for single inputs are unknown.", "background": false, "home": "P3.5", "home_id": "ml-testing", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.behavioral-testing", "label": "Behavioral testing", "aliases": [], "definition": "Testing model capabilities with targeted input suites for invariance, directional change, and minimum functionality, independent of aggregate accuracy.", "background": false, "home": "P3.5", "home_id": "ml-testing", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.production-readiness", "label": "Production-readiness assessment", "aliases": [], "definition": "A structured rubric of data, model, infrastructure, and monitoring tests used to judge whether an ML system is ready to operate.", "background": false, "home": "P3.5", "home_id": "ml-testing", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 2, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.online-inference", "label": "Online inference", "aliases": [ "real-time prediction" ], "definition": "Producing predictions on demand for individual requests within a latency target.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 6, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.batch-inference", "label": "Batch inference", "aliases": [ "offline prediction" ], "definition": "Computing predictions for many inputs in scheduled jobs and storing them for later lookup.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 13, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.model-server", "label": "Model server", "aliases": [ "inference server" ], "definition": "A service that loads model artifacts and executes inference requests behind a network interface, managing versions, batching, and resources.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 14, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.multi-model-serving", "label": "Multi-model serving", "aliases": [], "definition": "Hosting many models or adapters on shared serving resources, trading isolation and load time against utilization.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 38 }, "book_passages": [] }, { "id": "mlsys.compound-ai-system", "label": "Compound AI system", "aliases": [ "LM program", "LLM pipeline" ], "definition": "An application composing multiple model calls, retrievers, and tools into one program whose components and control flow are tuned and served jointly for quality and cost.", "background": false, "home": "P4.2", "home_id": "compound-ai", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 0, "organizational": 0, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.lm-program-optimization", "label": "LM program optimization", "aliases": [ "prompt optimization" ], "definition": "Automatically tuning the prompts, demonstrations, or weights of each model call in a declared pipeline against an end-to-end metric rather than per call.", "background": false, "home": "P4.2", "home_id": "compound-ai", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.ttft", "label": "Time to first token", "aliases": [ "TTFT" ], "definition": "Latency from request arrival until the first generated token is returned, dominated by queueing and prefill.", "background": false, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 13, "organizational": 3, "user_proxy": 90 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-performance-metrics-ttft-tpot-b009" } ] }, { "id": "mlsys.tpot", "label": "Time per output token", "aliases": [ "TPOT" ], "definition": "For a request producing multiple tokens, its elapsed decoding time after the first output divided by its remaining generated-token count, summarizing average decoding speed at a declared measurement boundary.", "background": false, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 5, "organizational": 2, "user_proxy": 39 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-performance-metrics-ttft-tpot-b009" } ] }, { "id": "mlsys.hedged-requests", "label": "Hedged requests", "aliases": [], "definition": "Issuing a duplicate request to another replica when the first has not responded within a delay, and using whichever returns first.", "background": false, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 16 }, "book_passages": [] }, { "id": "mlsys.continuous-batching", "label": "Continuous batching", "aliases": [ "iteration-level batching", "in-flight batching" ], "definition": "Scheduling at each generation iteration so finished sequences leave and new requests join a running batch.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 10, "organizational": 3, "user_proxy": 77 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-continuous-batching-8bb6" } ] }, { "id": "mlsys.load-balancing", "label": "Request load balancing", "aliases": [], "definition": "Distributing requests across replicas according to load, locality, or cached state.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 1, "user_proxy": 58 }, "book_passages": [] }, { "id": "mlsys.cold-start", "label": "Model cold start", "aliases": [], "definition": "Delay before a newly started replica can serve, from provisioning, loading weights, and warming runtime state.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 0, "user_proxy": 60 }, "book_passages": [] }, { "id": "mlsys.chunked-prefill", "label": "Chunked prefill", "aliases": [], "definition": "Splitting a long prompt's prefill into chunks interleaved with decode steps to bound per-iteration latency.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 4, "organizational": 2, "user_proxy": 27 }, "book_passages": [] }, { "id": "mlsys.pd-disaggregation", "label": "Prefill-decode disaggregation", "aliases": [ "disaggregated serving" ], "definition": "Running prefill and decode on separate resource pools, transferring KV state between them, so each phase is provisioned and batched independently.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 5, "corpus_chapters": 3, "organizational": 1, "user_proxy": 26 }, "book_passages": [] }, { "id": "mlsys.sampling-controls", "label": "Decoding and sampling controls", "aliases": [ "temperature", "top-k sampling", "nucleus sampling", "top-p", "min-p", "repetition penalty", "greedy decoding" ], "definition": "Parameters and rules, such as greedy, temperature, top-k, and nucleus sampling, that determine how tokens are chosen from a model's distribution.", "background": false, "home": "P4.6", "home_id": "structured-generation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 51, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.constrained-decoding", "label": "Constrained decoding", "aliases": [ "grammar-guided decoding" ], "definition": "A generation mechanism that restricts token choices to continuations permitted by an output constraint, so a successfully completed output satisfies the supported syntax or schema constraints.", "background": false, "home": "P4.6", "home_id": "structured-generation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 1, "user_proxy": 2 }, "book_passages": [ { "volume": "vol3", "chapter": "The Foundation Model", "url": "https://mlsysbook.ai/vol3/02_foundation_model/02_foundation_model.html#sec-vol3-foundation-model-structural-guarantee-scope" } ] }, { "id": "mlsys.output-validation", "label": "Output validation", "aliases": [], "definition": "Checking generated output against declared schemas or rules and reporting whether it satisfies them.", "background": false, "home": "P4.6", "home_id": "structured-generation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.reasoning-budget", "label": "Reasoning budget", "aliases": [ "thinking budget" ], "definition": "A limit on the intermediate tokens or steps a model may spend before answering, set to trade quality against latency and cost.", "background": false, "home": "P4.7", "home_id": "test-time-compute", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.best-of-n", "label": "Best-of-n sampling", "aliases": [], "definition": "Generating n candidate outputs and returning the one ranked highest by a scorer or verifier.", "background": false, "home": "P4.7", "home_id": "test-time-compute", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 2, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.self-consistency", "label": "Self-consistency voting", "aliases": [ "majority voting" ], "definition": "Sampling multiple reasoning paths and returning the most frequent final answer.", "background": false, "home": "P4.7", "home_id": "test-time-compute", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.verifier", "label": "Verifier", "aliases": [ "outcome reward model" ], "definition": "A model or program that judges whether a candidate output or reasoning step is correct, used for selection, search, or reward.", "background": false, "home": "P4.7", "home_id": "test-time-compute", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 26, "organizational": 6, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Test-Time Compute", "url": "https://mlsysbook.ai/vol3/03_test_time_compute/03_test_time_compute.html#sec-vol3-test-time-compute-check-precision" } ] }, { "id": "mlsys.process-reward-model", "label": "Process reward model", "aliases": [ "PRM" ], "definition": "A verifier that scores intermediate reasoning steps rather than only final answers.", "background": false, "home": "P4.7", "home_id": "test-time-compute", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.inference-time-search", "label": "Inference-time search", "aliases": [], "definition": "Exploring a tree of partial solutions, as in beam or tree search, guided by a scorer, to find a higher-quality output.", "background": false, "home": "P4.7", "home_id": "test-time-compute", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 2, "user_proxy": 1 }, "book_passages": [ { "volume": "vol3", "chapter": "Test-Time Compute", "url": "https://mlsysbook.ai/vol3/03_test_time_compute/03_test_time_compute.html#sec-vol3-test-time-compute-computation-allocation" } ] }, { "id": "mlsys.test-time-scaling", "label": "Test-time scaling", "aliases": [], "definition": "The empirical relation between inference compute spent per query and output quality, used to allocate compute between model size and inference.", "background": false, "home": "P4.7", "home_id": "test-time-compute", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.rag", "label": "Retrieval-augmented generation", "aliases": [ "RAG" ], "definition": "Conditioning generation on documents retrieved for the request so outputs can draw on information outside the model's parameters.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 8, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.dense-retrieval", "label": "Dense retrieval", "aliases": [ "two-tower model", "dual encoder", "bi-encoder" ], "definition": "Retrieving records by similarity between learned embeddings of query and records.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.lexical-retrieval", "label": "Lexical retrieval", "aliases": [ "sparse retrieval", "BM25" ], "definition": "Retrieving records by weighted term overlap with the query using an inverted index, as in BM25.", "background": true, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.reranking", "label": "Reranking", "aliases": [], "definition": "Reordering a retrieved candidate set with a more expensive relevance model before use.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.model-conversion", "label": "Model conversion", "aliases": [], "definition": "Translating a trained model into the format and operator set of a target runtime, possibly changing numerics and supported operations.", "background": false, "home": "P4.9", "home_id": "on-device-deployment", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 74 }, "book_passages": [] }, { "id": "mlsys.on-device-runtime", "label": "On-device inference runtime", "aliases": [ "browser inference runtime", "WebGPU inference", "WebAssembly inference" ], "definition": "An interpreter or compiled engine that executes models locally on personal or embedded devices, including inside browsers, under the target environment's memory, operator, and execution limits.", "background": false, "home": "P4.9", "home_id": "on-device-deployment", "added_in": "0.2-draft", "warrant": { "literary": 5, "corpus_chapters": 0, "organizational": 0, "user_proxy": 28 }, "book_passages": [] }, { "id": "mlsys.ota-model-update", "label": "Over-the-air model update", "aliases": [ "OTA update", "on-demand model download", "model download on first use", "deferred model download" ], "definition": "Delivering model weights to deployed devices remotely, whether the first download after installation or a later version, with staging, verification, and rollback across the device fleet.", "background": false, "home": "P4.9", "home_id": "on-device-deployment", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 130 }, "book_passages": [] }, { "id": "mlsys.device-cloud-partitioning", "label": "Device-cloud partitioning", "aliases": [ "split computing", "offloading" ], "definition": "Splitting inference between device and server, by layer or by model, to balance latency, energy, bandwidth, and privacy.", "background": false, "home": "P4.9", "home_id": "on-device-deployment", "added_in": "0.2-draft", "warrant": { "literary": 5, "corpus_chapters": 22, "organizational": 2, "user_proxy": 46 }, "book_passages": [] }, { "id": "mlsys.device-heterogeneity", "label": "Device heterogeneity", "aliases": [], "definition": "Variation in hardware, operating system, and runtime capability across a deployed device fleet that a single model release must accommodate.", "background": false, "home": "P4.9", "home_id": "on-device-deployment", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 32 }, "book_passages": [] }, { "id": "mlsys.model-registry", "label": "Model registry", "aliases": [], "definition": "A versioned store of model artifacts with metadata, lineage, and lifecycle stage used to govern promotion and deployment.", "background": false, "home": "P5.1", "home_id": "release-management", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 11, "organizational": 2, "user_proxy": 12 }, "book_passages": [ { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html#sec-ml-operations-scale-model-registries-scale-ccba" } ] }, { "id": "mlsys.shadow-deployment", "label": "Shadow deployment", "aliases": [ "shadow mode" ], "definition": "Running a candidate model on live traffic without serving its outputs, to compare it against the production model.", "background": false, "home": "P5.1", "home_id": "release-management", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 3, "user_proxy": 35 }, "book_passages": [] }, { "id": "mlsys.ml-ci-cd", "label": "CI/CD for ML", "aliases": [], "definition": "Automated building, testing, and deployment of ML code, data checks, and model artifacts on each change.", "background": false, "home": "P5.1", "home_id": "release-management", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 31 }, "book_passages": [] }, { "id": "mlsys.release-gate", "label": "Release gate", "aliases": [ "validation gate" ], "definition": "A predeclared set of checks a candidate model must pass before it advances to a wider deployment stage.", "background": false, "home": "P5.1", "home_id": "release-management", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 17, "organizational": 10, "user_proxy": 44 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Lifecycle", "url": "https://mlsysbook.ai/vol1/03_ml_lifecycle/03_ml_lifecycle.html#sec-ml-workflow-validation-deployment-readiness-207c" }, { "volume": "vol2", "chapter": "ML Operations at Scale", "url": "https://mlsysbook.ai/vol2/12_ops_scale/12_ops_scale.html#sec-ops-scale-end-to-end-evaluation-contract" } ] }, { "id": "mlsys.drift-detection", "label": "Drift detection", "aliases": [], "definition": "Statistical tests or distances comparing production and reference distributions of inputs, predictions, or outcomes to flag change.", "background": false, "home": "P5.2", "home_id": "monitoring", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 11, "organizational": 5, "user_proxy": 100 }, "book_passages": [ { "volume": "vol1", "chapter": "ML Operations", "url": "https://mlsysbook.ai/vol1/14_ml_ops/14_ml_ops.html#sec-ml-operations-model-infrastructure-monitoring-3988" } ] }, { "id": "mlsys.inference-logging", "label": "Prediction logging", "aliases": [], "definition": "Recording requests, model versions, outputs, and later outcomes so production behavior can be analyzed, debugged, and reused as data.", "background": false, "home": "P5.2", "home_id": "monitoring", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.alerting", "label": "Alerting", "aliases": [], "definition": "Rules that turn monitored signals into notifications for responders, tuned to balance missed incidents against alert fatigue.", "background": false, "home": "P5.2", "home_id": "monitoring", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 1, "user_proxy": 23 }, "book_passages": [] }, { "id": "mlsys.retraining", "label": "Retraining", "aliases": [ "retraining trigger", "retraining cadence" ], "definition": "Training a replacement version of a deployed model on updated or revised data, with timing and data chosen to balance staleness against training and validation cost.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.1-draft", "warrant": { "literary": 1, "corpus_chapters": 34, "organizational": 1, "user_proxy": 14 }, "book_passages": [] }, { "id": "mlsys.continuous-training", "label": "Continuous training", "aliases": [ "CT" ], "definition": "An automated pipeline that retrains, validates, and proposes models on a schedule or when triggered by data or performance signals.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 5, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.postmortem", "label": "Blameless postmortem", "aliases": [], "definition": "A written analysis after an incident of its timeline, causes, and corrective actions, focused on system fixes rather than blame.", "background": true, "home": "P5.4", "home_id": "incident-response", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.failure-attribution", "label": "ML failure attribution", "aliases": [], "definition": "Determining whether a production failure stems from data, model, code, configuration, or infrastructure, or from an adversary.", "background": false, "home": "P5.4", "home_id": "incident-response", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 4, "user_proxy": 18 }, "book_passages": [ { "volume": "vol3", "chapter": "Operating Agents", "url": "https://mlsysbook.ai/vol3/14_agent_operations/14_agent_operations.html#sec-vol3-postmortem-ablation" } ] }, { "id": "mlsys.configuration-management", "label": "ML configuration management", "aliases": [], "definition": "Treating model, feature, and pipeline configuration as versioned, reviewed, and tested code.", "background": false, "home": "P5.5", "home_id": "pipelines-reproducibility", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.infrastructure-as-code", "label": "Infrastructure as code", "aliases": [ "IaC" ], "definition": "Declaring compute, storage, and service infrastructure in versioned machine-readable files from which environments are provisioned.", "background": true, "home": "P5.5", "home_id": "pipelines-reproducibility", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.ml-platform", "label": "ML platform", "aliases": [], "definition": "Shared internal infrastructure and services that standardize how teams train, deploy, and operate models.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 5, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.technical-debt", "label": "ML technical debt", "aliases": [], "definition": "Accumulated system costs from shortcuts in ML code, data dependencies, configuration, and pipelines that make future change slower and riskier.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 2, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.entanglement", "label": "Entanglement", "aliases": [ "CACE" ], "definition": "The property that changing any input, feature, or hyperparameter of a model can change its behavior on everything else.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 1, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.undeclared-consumers", "label": "Undeclared consumers", "aliases": [], "definition": "Systems that use a model's outputs without the model owner's knowledge, creating hidden coupling.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.mlops-maturity", "label": "MLOps maturity", "aliases": [], "definition": "Staged levels of automation, from manual training and deployment to automated continuous training and delivery.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.ml-roles", "label": "ML team roles", "aliases": [], "definition": "The division of responsibilities among data, modeling, platform, and operations roles that owns an ML system across its lifecycle.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-deprecation", "label": "Model deprecation", "aliases": [], "definition": "Planned withdrawal of a model from service, including notifying and migrating its consumers before removing it.", "background": false, "home": "P5.7", "home_id": "retirement", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.artifact-archival", "label": "Artifact archival", "aliases": [], "definition": "Preserving retired models, data references, and evaluation records needed for audit, reproduction, or rollback after service ends.", "background": false, "home": "P5.7", "home_id": "retirement", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.data-retention", "label": "Data retention", "aliases": [], "definition": "Policy and mechanisms that set how long data, logs, and derived artifacts are kept and when they are deleted.", "background": false, "home": "P5.7", "home_id": "retirement", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 10, "organizational": 3, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.model-flops-utilization", "label": "Model FLOPs utilization", "aliases": [ "MFU" ], "definition": "Model-required floating-point operations per second achieved, divided by the aggregate peak FLOP rate of the hardware used, excluding recomputation.", "background": false, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 25, "organizational": 6, "user_proxy": 35 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Training", "url": "https://mlsysbook.ai/vol1/08_training/08_training.html#sec-model-training-identifying-bottlenecks-f57f" }, { "volume": "vol2", "chapter": "Fleet Efficiency", "url": "https://mlsysbook.ai/vol2/09_performance_engineering/09_performance_engineering.html#sec-performance-engineering-program-goodput" } ] }, { "id": "mlsys.bottleneck-analysis", "label": "Bottleneck analysis", "aliases": [ "optimization prioritization", "optimization planning" ], "definition": "Identifying the resource or stage, such as compute, memory bandwidth, I/O, or communication, whose capacity limits end-to-end throughput or latency for a workload.", "background": false, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 3, "user_proxy": 257 }, "book_passages": [] }, { "id": "mlsys.training-goodput", "label": "Training goodput", "aliases": [ "ML goodput" ], "definition": "The fraction of allocated accelerator time that produces retained training progress, after losses to scheduling, failures, restarts, and lost work.", "background": false, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 7, "user_proxy": 4 }, "book_passages": [ { "volume": "vol2", "chapter": "Fleet Efficiency", "url": "https://mlsysbook.ai/vol2/09_performance_engineering/09_performance_engineering.html#sec-performance-engineering-goodput-factors" } ] }, { "id": "mlsys.pruning", "label": "Pruning", "aliases": [], "definition": "Removing weights, neurons, channels, or other structures judged unimportant from a network, usually followed by fine-tuning to recover quality.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 41, "organizational": 1, "user_proxy": 50 }, "book_passages": [] }, { "id": "mlsys.structured-pruning", "label": "Structured pruning", "aliases": [], "definition": "Pruning whole channels, filters, heads, or layers so that the remaining model stays dense and runs faster without sparse-execution support.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 5, "organizational": 2, "user_proxy": 62 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html#sec-model-compression-pruning-d1cb" } ] }, { "id": "mlsys.low-rank-factorization", "label": "Low-rank factorization", "aliases": [ "low-rank approximation" ], "definition": "Approximating a weight matrix or tensor by a product of smaller factors of reduced rank to cut parameters and operations.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.quantization-aware-training", "label": "Quantization-aware training", "aliases": [ "QAT" ], "definition": "Training or fine-tuning with simulated quantization in the forward pass so the model adapts to reduced precision, in contrast to post-training quantization.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 2, "user_proxy": 85 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html#sec-model-compression-quantization-precision-cd46" } ] }, { "id": "mlsys.neural-architecture-search", "label": "Neural architecture search", "aliases": [ "NAS" ], "definition": "Automated search over a defined space of network architectures for designs that optimize a stated objective such as accuracy under a cost budget.", "background": false, "home": "Q1.4", "home_id": "hardware-aware-design", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 7, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.hardware-aware-nas", "label": "Hardware-aware NAS", "aliases": [ "platform-aware NAS" ], "definition": "Architecture search whose objective includes latency, energy, or memory measured or predicted on the target device rather than proxy counts such as FLOPs.", "background": false, "home": "Q1.4", "home_id": "hardware-aware-design", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 1, "organizational": 0, "user_proxy": 47 }, "book_passages": [] }, { "id": "mlsys.depthwise-separable-convolution", "label": "Depthwise separable convolution", "aliases": [], "definition": "Factoring a standard convolution into a per-channel spatial convolution followed by a pointwise convolution, reducing operations and parameters.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 129 }, "book_passages": [] }, { "id": "mlsys.efficiency-frontier", "label": "Accuracy-efficiency frontier", "aliases": [ "accuracy-cost frontier" ], "definition": "The set of model configurations for which no alternative improves quality without worsening a stated cost measure, such as latency, memory, or energy, on a stated platform.", "background": false, "home": "Q1.4", "home_id": "hardware-aware-design", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 1, "user_proxy": 20 }, "book_passages": [] }, { "id": "mlsys.model-hardware-co-design", "label": "Model-hardware co-design", "aliases": [], "definition": "Jointly choosing model architecture and hardware or kernel design so that each exploits the other's structure.", "background": false, "home": "Q1.4", "home_id": "hardware-aware-design", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 0, "organizational": 0, "user_proxy": 82 }, "book_passages": [] }, { "id": "mlsys.retry-backoff", "label": "Retry with backoff", "aliases": [], "definition": "Reissuing a failed request after increasing, typically jittered, delays and within a retry budget, so that transient faults recover without overloading dependencies.", "background": true, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.circuit-breaker", "label": "Circuit breaker", "aliases": [], "definition": "A client-side guard that stops sending requests to a failing dependency after an error threshold and probes it before resuming traffic.", "background": true, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 10, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.replication-failover", "label": "Replication and failover", "aliases": [], "definition": "Maintaining redundant replicas of a service or state and redirecting work to a healthy replica when one fails.", "background": false, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 1, "organizational": 0, "user_proxy": 27 }, "book_passages": [] }, { "id": "mlsys.graceful-degradation", "label": "Graceful degradation", "aliases": [], "definition": "Continuing to serve with reduced quality, features, or freshness under failure or overload instead of failing outright.", "background": false, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 14, "organizational": 0, "user_proxy": 101 }, "book_passages": [] }, { "id": "mlsys.checkpoint-interval", "label": "Optimal checkpoint interval", "aliases": [ "Young-Daly interval" ], "definition": "The checkpoint period that minimizes expected lost time, balancing checkpoint write cost against expected rework, approximately the square root of twice the write cost times MTBF.", "background": false, "home": "Q2.2", "home_id": "checkpoint-recovery", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 10, "organizational": 1, "user_proxy": 46 }, "book_passages": [] }, { "id": "mlsys.async-checkpointing", "label": "Asynchronous checkpointing", "aliases": [], "definition": "Snapshotting training state quickly to host memory and persisting it in the background so computation resumes before the durable write completes.", "background": false, "home": "Q2.2", "home_id": "checkpoint-recovery", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 8, "organizational": 1, "user_proxy": 48 }, "book_passages": [] }, { "id": "mlsys.silent-data-corruption", "label": "Silent data corruption", "aliases": [ "SDC" ], "definition": "Hardware producing wrong computational results without raising an error or crash, so the corruption propagates undetected.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 1, "user_proxy": 20 }, "book_passages": [] }, { "id": "mlsys.mtbf", "label": "Mean time between failures", "aliases": [ "MTBF" ], "definition": "The expected operating time between failures of a component or system; for independent components, system MTBF shrinks roughly in inverse proportion to component count.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 20, "organizational": 4, "user_proxy": 30 }, "book_passages": [ { "volume": "vol2", "chapter": "Fault Tolerance", "url": "https://mlsysbook.ai/vol2/07_fault_tolerance/07_fault_tolerance.html#sec-fault-tolerance-reliability-reliability-mathematics-inevitable-failure-93ef" } ] }, { "id": "mlsys.availability", "label": "Availability", "aliases": [], "definition": "The fraction of time or requests for which a system performs its specified function, approximately MTBF divided by MTBF plus mean time to repair.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 52, "organizational": 1, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.gray-failure", "label": "Gray failure", "aliases": [], "definition": "A partial or degraded failure that some observers detect and others do not, such as a slow or lossy node that still passes health checks.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 5, "organizational": 3, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.failure-domain", "label": "Failure domain", "aliases": [], "definition": "A set of components that can fail together from a shared cause, such as a rack, power feed, switch, or software release.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 23, "organizational": 3, "user_proxy": 15 }, "book_passages": [] }, { "id": "mlsys.abft", "label": "Algorithm-based fault tolerance", "aliases": [ "ABFT" ], "definition": "Detecting or correcting computational errors by carrying checksums through an algorithm, such as matrix multiplication, and verifying invariants on its outputs.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.chaos-engineering", "label": "Chaos engineering", "aliases": [], "definition": "Running controlled experiments that inject real-world failures into production-like systems to test a hypothesis about steady-state behavior.", "background": true, "home": "Q2.4", "home_id": "fault-injection", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.game-day", "label": "Game day", "aliases": [ "disaster recovery exercise" ], "definition": "A scheduled exercise in which teams deliberately trigger failure scenarios to rehearse detection, response, and recovery procedures.", "background": true, "home": "Q2.4", "home_id": "fault-injection", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.hardware-fault-injection", "label": "Hardware fault injection", "aliases": [], "definition": "Emulating hardware faults such as bit flips in weights, activations, or registers to measure how errors propagate to model outputs.", "background": false, "home": "Q2.4", "home_id": "fault-injection", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 8, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.blast-radius", "label": "Blast radius", "aliases": [], "definition": "The extent of users, components, or effects that a single fault, change, or experiment can impact.", "background": false, "home": "Q2.4", "home_id": "fault-injection", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 27, "organizational": 3, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.backdoor", "label": "Backdoor attack", "aliases": [ "trojan attack" ], "definition": "Implanting a hidden behavior in a model, via poisoned data or modified weights, that activates only on inputs containing an attacker-chosen trigger.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.jailbreak", "label": "Jailbreak", "aliases": [], "definition": "A user-crafted input that induces a model to bypass its safety training or usage policy and produce disallowed output.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.adversarial-training", "label": "Adversarial training", "aliases": [], "definition": "Training on adversarially perturbed inputs generated during training to increase robustness within a specified perturbation set.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 6, "organizational": 1, "user_proxy": 17 }, "book_passages": [] }, { "id": "mlsys.model-extraction", "label": "Model extraction", "aliases": [ "model stealing" ], "definition": "Reconstructing a functionally equivalent copy or the parameters of a model by querying its prediction interface.", "background": false, "home": "Q3.2", "home_id": "model-supply-chain", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 1, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.model-watermarking", "label": "Model watermarking", "aliases": [], "definition": "Embedding a verifiable signature in a model's parameters or its behavior on trigger inputs so that ownership of copies or derivatives can later be demonstrated, with survival under fine-tuning, pruning, and distillation deciding its value.", "background": false, "home": "Q3.2", "home_id": "model-supply-chain", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.artifact-signing", "label": "Model artifact signing", "aliases": [], "definition": "Cryptographically signing model files and their provenance metadata so that consumers can verify integrity and origin before loading.", "background": false, "home": "Q3.2", "home_id": "model-supply-chain", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 2, "user_proxy": 10 }, "book_passages": [] }, { "id": "mlsys.unsafe-deserialization", "label": "Unsafe model deserialization", "aliases": [], "definition": "Code execution triggered by loading a model stored in an executable serialization format, such as pickle, from an untrusted source.", "background": false, "home": "Q3.2", "home_id": "model-supply-chain", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.supply-chain-compromise", "label": "ML supply-chain compromise", "aliases": [], "definition": "Introducing malicious or vulnerable components into an ML system through third-party datasets, pretrained weights, libraries, build tools, or hardware.", "background": false, "home": "Q3.2", "home_id": "model-supply-chain", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.membership-inference", "label": "Membership inference", "aliases": [], "definition": "Inferring whether a specific record was in a model's training data from the model's outputs or parameters.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 1, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.model-inversion", "label": "Model inversion", "aliases": [ "gradient inversion", "attribute inference" ], "definition": "Reconstructing sensitive attributes or representative training inputs from a model's outputs, confidence scores, or shared parameter gradients.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 1, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.dp-sgd", "label": "DP-SGD", "aliases": [ "differentially private SGD" ], "definition": "Stochastic gradient descent with per-example gradient clipping and calibrated noise, with privacy loss tracked across steps by an accountant.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 1, "user_proxy": 39 }, "book_passages": [ { "volume": "vol2", "chapter": "Security & Privacy", "url": "https://mlsysbook.ai/vol2/13_security_privacy/13_security_privacy.html#sec-security-privacy-privacy-budget-composition-edbe" } ] }, { "id": "mlsys.machine-unlearning", "label": "Machine unlearning", "aliases": [], "definition": "Removing the influence of specified training records from a trained model, either exactly by retraining or approximately at lower cost.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 1, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.secure-aggregation", "label": "Secure aggregation", "aliases": [], "definition": "A cryptographic protocol that lets a server compute the sum of client updates without seeing any individual client's update.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 6, "organizational": 2, "user_proxy": 19 }, "book_passages": [ { "volume": "vol2", "chapter": "Edge Intelligence", "url": "https://mlsysbook.ai/vol2/11_edge_intelligence/11_edge_intelligence.html#sec-edge-intelligence-federated-privacy-a1ed" } ] }, { "id": "mlsys.least-privilege", "label": "Least privilege", "aliases": [], "definition": "Granting each component or actor only the permissions its task requires, for only as long as it requires them.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 16, "organizational": 2, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.confidential-computing", "label": "Confidential computing", "aliases": [ "trusted execution environment", "TEE" ], "definition": "Protecting data and models while in use by running computation inside hardware-isolated, attestable trusted execution environments.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 2, "organizational": 0, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.remote-attestation", "label": "Remote attestation", "aliases": [], "definition": "A procedure by which a platform proves to a remote verifier the identity and measured state of its hardware and software before receiving secrets.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.side-channel", "label": "Side-channel attack", "aliases": [], "definition": "Inferring secrets from observable resource behavior, such as timing, power, memory, or cache contention, across an isolation boundary.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 25 }, "book_passages": [] }, { "id": "mlsys.carbon-intensity", "label": "Grid carbon intensity", "aliases": [], "definition": "Greenhouse-gas emissions per unit of electricity supplied, varying by region, time, and generation mix, in grams CO2-equivalent per kilowatt-hour.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 1, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.operational-carbon", "label": "Operational carbon", "aliases": [], "definition": "Emissions from electricity consumed while operating a system, computed from energy use and the carbon intensity of its supply.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 4, "user_proxy": 44 }, "book_passages": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html#sec-responsible-engineering-environmental-impact-0292" }, { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html#sec-sustainable-ai-operational-carbon-calculation-4627-calculation-4627" } ] }, { "id": "mlsys.emission-scopes", "label": "Emission scopes", "aliases": [ "Scope 1, 2, and 3 emissions", "GHG scopes" ], "definition": "The GHG Protocol classification of emissions into direct (Scope 1), purchased energy (Scope 2), and value-chain (Scope 3) categories.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.carbon-aware-scheduling", "label": "Carbon-aware scheduling", "aliases": [], "definition": "Shifting flexible workloads in time or location toward periods and regions of lower grid carbon intensity, subject to deadlines and capacity.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 2, "user_proxy": 28 }, "book_passages": [ { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html#sec-sustainable-ai-carbonaware-scheduling-c2db" } ] }, { "id": "mlsys.energy-per-inference", "label": "Energy per inference", "aliases": [], "definition": "Energy consumed to process one input end to end on a device, the binding metric for battery-powered and energy-harvesting deployments.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 11, "organizational": 2, "user_proxy": 85 }, "book_passages": [ { "volume": "vol1", "chapter": "System Paradigms", "url": "https://mlsysbook.ai/vol1/02_system_paradigms/02_system_paradigms.html#sec-ml-systems-mobile-ml-benefits-resource-constraints-c568" } ] }, { "id": "mlsys.embodied-carbon", "label": "Embodied carbon", "aliases": [], "definition": "Emissions from manufacturing, transporting, and disposing of hardware and facilities, attributed to a system independent of its operation.", "background": false, "home": "Q6.2", "home_id": "embodied-lifecycle", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 4, "user_proxy": 14 }, "book_passages": [ { "volume": "vol2", "chapter": "Sustainable AI", "url": "https://mlsysbook.ai/vol2/15_sustainable_ai/15_sustainable_ai.html#sec-sustainable-ai-embodied-carbon-assessment-9de0" } ] }, { "id": "mlsys.lifecycle-assessment", "label": "Life cycle assessment", "aliases": [ "LCA" ], "definition": "A standardized method for compiling and evaluating a product's environmental inputs, outputs, and impacts from raw materials through disposal.", "background": false, "home": "Q6.2", "home_id": "embodied-lifecycle", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.water-footprint", "label": "Water footprint", "aliases": [ "water usage" ], "definition": "Water withdrawn or consumed by a system, on site for cooling and off site for electricity generation, per unit of work or over its lifetime.", "background": false, "home": "Q6.2", "home_id": "embodied-lifecycle", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.hardware-lifetime", "label": "Hardware service lifetime", "aliases": [ "refresh cycle" ], "definition": "The period a hardware asset stays in service before replacement, over which its embodied impacts and capital cost are amortized.", "background": false, "home": "Q6.2", "home_id": "embodied-lifecycle", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.e-waste", "label": "Electronic waste", "aliases": [ "e-waste" ], "definition": "Discarded electronic equipment and components, including retired accelerators and servers, and the hazards and lost materials of their disposal.", "background": false, "home": "Q6.2", "home_id": "embodied-lifecycle", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.tco", "label": "Total cost of ownership", "aliases": [ "TCO" ], "definition": "The full cost of acquiring and operating a system over its life, including capital, energy, facilities, networking, staff, and amortization.", "background": false, "home": "Q6.3", "home_id": "cost-economics", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 14, "organizational": 6, "user_proxy": 256 }, "book_passages": [] }, { "id": "mlsys.cost-per-token", "label": "Cost per token", "aliases": [], "definition": "Serving or API expenditure divided by tokens processed or generated, often priced separately for input and output tokens.", "background": false, "home": "Q6.3", "home_id": "cost-economics", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 4, "organizational": 2, "user_proxy": 24 }, "book_passages": [] }, { "id": "mlsys.cost-attribution", "label": "Cost attribution", "aliases": [ "FinOps", "chargeback" ], "definition": "Allocating shared infrastructure spend to teams, models, or workloads so that owners see and act on their costs.", "background": false, "home": "Q6.3", "home_id": "cost-economics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 1, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.hazard-analysis", "label": "Hazard analysis", "aliases": [], "definition": "Systematically identifying conditions under which a system could contribute to harm, with their causes, severity, and required controls.", "background": false, "home": "Q4.1", "home_id": "hazards-operating-domain", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.stpa", "label": "System-theoretic process analysis", "aliases": [ "STPA" ], "definition": "Hazard analysis that models a system as a control structure and identifies unsafe control actions and the flawed feedback or process models that cause them.", "background": false, "home": "Q4.1", "home_id": "hazards-operating-domain", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.fmea", "label": "Failure modes and effects analysis", "aliases": [ "FMEA" ], "definition": "Enumerating each component's possible failure modes and tracing their effects on system behavior to rank and mitigate them.", "background": false, "home": "Q4.1", "home_id": "hazards-operating-domain", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.functional-insufficiency", "label": "Functional insufficiency", "aliases": [], "definition": "An insufficiency in the specification or performance of an intended function that can contribute to hazardous behavior without a fault covered by functional-safety analysis.", "background": false, "home": "Q4.1", "home_id": "hazards-operating-domain", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.safety-filter", "label": "Safety filter", "aliases": [ "shield", "shielding", "action shield" ], "definition": "A runtime component that passes a proposed action when it is certified safe and otherwise replaces it with a minimally different safe action.", "background": false, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 19, "organizational": 4, "user_proxy": 1 }, "book_passages": [ { "volume": "vol4", "chapter": "Safety Enforcement", "url": "https://mlsysbook.ai/vol4/12_enforcement/12_enforcement.html#sec-enforcement-minimal-intervention" } ] }, { "id": "mlsys.control-barrier-function", "label": "Control barrier function", "aliases": [ "CBF" ], "definition": "A function whose sign defines a safe set and whose derivative condition yields input constraints that keep the system state inside that set.", "background": false, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 10, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.simplex-architecture", "label": "Simplex architecture", "aliases": [], "definition": "Pairing an unverified high-performance controller with a verified safety controller and a decision module that switches to the safety controller before safety is lost.", "background": false, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 3, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.minimal-risk-condition", "label": "Minimal-risk condition", "aliases": [ "MRC", "safe fallback", "fallback ladder" ], "definition": "A predefined reduced-capability state the system reaches when it can no longer operate safely, often via an ordered ladder of fallback behaviors.", "background": false, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 16, "organizational": 4, "user_proxy": 16 }, "book_passages": [] }, { "id": "mlsys.guardrails", "label": "Guardrails", "aliases": [ "content safety filter", "safety classifier" ], "definition": "Runtime checks on model inputs and outputs that detect and block, rewrite, or escalate content or actions violating a stated safety policy.", "background": false, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 18, "organizational": 0, "user_proxy": 27 }, "book_passages": [] }, { "id": "mlsys.scenario-based-testing", "label": "Scenario-based testing", "aliases": [], "definition": "Validating a system against a structured catalog of operating scenarios and their parameter ranges, with coverage measured over the operating domain.", "background": false, "home": "Q4.3", "home_id": "verification-validation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.falsification", "label": "Simulation-based falsification", "aliases": [], "definition": "Searching scenario parameters, often guided by temporal-logic robustness, to find counterexamples that violate a safety specification.", "background": false, "home": "Q4.3", "home_id": "verification-validation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.formal-verification", "label": "Formal verification", "aliases": [], "definition": "Mathematically proving that a model of a system satisfies a formal specification for all behaviors within stated assumptions.", "background": false, "home": "Q4.3", "home_id": "verification-validation", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 7, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.neural-network-verification", "label": "Neural network verification", "aliases": [], "definition": "Proving or refuting input-output properties of a trained network over a bounded input region, such as local robustness, using exact or bound-propagation methods.", "background": false, "home": "Q4.3", "home_id": "verification-validation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.rare-event-validation", "label": "Rare-event safety validation", "aliases": [], "definition": "Estimating or bounding failure probabilities too small to observe in feasible exposure, using importance sampling, extreme-value models, or zero-failure statistical bounds.", "background": false, "home": "Q4.3", "home_id": "verification-validation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 4, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Closed-Loop Evaluation", "url": "https://mlsysbook.ai/vol4/07_evaluation/07_evaluation.html#sec-evaluation-closed-loop-survival" } ] }, { "id": "mlsys.xil-testing", "label": "X-in-the-loop testing", "aliases": [ "HIL", "software-in-the-loop", "hardware-in-the-loop" ], "definition": "A qualification ladder that runs the real software, then real hardware, against simulated or recorded environments before physical trials.", "background": false, "home": "Q4.3", "home_id": "verification-validation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 13, "organizational": 3, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.goal-structuring-notation", "label": "Goal Structuring Notation", "aliases": [ "GSN" ], "definition": "A graphical notation for safety cases that links goals, strategies, solutions (evidence), context, and assumptions into an explicit argument structure.", "background": true, "home": "Q4.4", "home_id": "safety-cases", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.functional-safety", "label": "Functional safety", "aliases": [ "ISO 26262", "IEC 61508", "safety lifecycle" ], "definition": "Absence of unreasonable risk from malfunctioning electrical or electronic systems, achieved through a standardized safety lifecycle such as IEC 61508 or ISO 26262.", "background": false, "home": "Q4.4", "home_id": "safety-cases", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 15, "organizational": 3, "user_proxy": 16 }, "book_passages": [ { "volume": "vol4", "chapter": "Safety Enforcement", "url": "https://mlsysbook.ai/vol4/12_enforcement/12_enforcement.html#sec-enforcement-certified-functions" } ] }, { "id": "mlsys.safety-integrity-level", "label": "Safety integrity level", "aliases": [ "SIL", "ASIL" ], "definition": "A graded risk class assigned to a safety function that sets the rigor of development, verification, and tolerable failure rate required.", "background": true, "home": "Q4.4", "home_id": "safety-cases", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 3, "user_proxy": 2 }, "book_passages": [ { "volume": "vol4", "chapter": "Safety Enforcement", "url": "https://mlsysbook.ai/vol4/12_enforcement/12_enforcement.html#sec-enforcement-certified-functions" } ] }, { "id": "mlsys.safety-performance-indicator", "label": "Safety performance indicator", "aliases": [ "SPI" ], "definition": "A field-measured metric tied to a safety-case claim whose threshold violation signals that the argument may no longer hold.", "background": false, "home": "Q4.4", "home_id": "safety-cases", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.group-fairness-metric", "label": "Group fairness metrics", "aliases": [ "demographic parity", "equalized odds", "predictive parity", "disparate impact ratio", "equal opportunity" ], "definition": "Statistics comparing selection rates, error rates, or calibration of a model's decisions across protected groups.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 4, "user_proxy": 91 }, "book_passages": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html#sec-responsible-engineering-testing-across-populations-9f20" } ] }, { "id": "mlsys.individual-fairness", "label": "Individual fairness", "aliases": [], "definition": "A requirement that individuals who are similar with respect to the task receive similar outcomes under a stated similarity metric.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.fairness-incompatibility", "label": "Fairness criteria incompatibility", "aliases": [], "definition": "The result that calibration and error-rate parity across groups cannot generally hold together when base rates differ, forcing an explicit choice among criteria.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 4, "user_proxy": 6 }, "book_passages": [ { "volume": "vol1", "chapter": "Responsible Engineering", "url": "https://mlsysbook.ai/vol1/15_responsible_engr/15_responsible_engr.html#sec-responsible-engineering-testing-across-populations-9f20" }, { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html#sec-responsible-ai-fairness-machine-learning-2ba4" } ] }, { "id": "mlsys.bias-mitigation", "label": "Bias mitigation", "aliases": [], "definition": "Interventions on data (pre-processing), training objective (in-processing), or decisions (post-processing) that reduce measured disparities, usually at some cost in other metrics.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 1, "user_proxy": 31 }, "book_passages": [] }, { "id": "mlsys.feature-attribution", "label": "Feature attribution", "aliases": [ "SHAP", "LIME", "integrated gradients", "saliency" ], "definition": "Post-hoc explanation assigning each input feature a contribution to a specific prediction.", "background": false, "home": "Q5.2", "home_id": "transparency-explainability", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 0, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.counterfactual-explanation", "label": "Counterfactual explanation", "aliases": [], "definition": "An explanation stating the smallest feasible change to an input that would change the model's decision.", "background": false, "home": "Q5.2", "home_id": "transparency-explainability", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.interpretable-model", "label": "Inherently interpretable model", "aliases": [], "definition": "A model whose structure, such as a sparse linear model, short rule list, or shallow tree, can be directly inspected to understand its predictions.", "background": false, "home": "Q5.2", "home_id": "transparency-explainability", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.model-card", "label": "Model card", "aliases": [ "model documentation" ], "definition": "A standardized document reporting a model's intended use, evaluation results across conditions and groups, limitations, and ethical considerations.", "background": false, "home": "Q5.2", "home_id": "transparency-explainability", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 1, "user_proxy": 15 }, "book_passages": [] }, { "id": "mlsys.datasheet", "label": "Datasheet for datasets", "aliases": [ "data card" ], "definition": "A standardized document recording a dataset's motivation, composition, collection process, preprocessing, intended uses, and maintenance.", "background": false, "home": "Q5.2", "home_id": "transparency-explainability", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.contestability", "label": "Contestability and recourse", "aliases": [], "definition": "The ability of an affected person to understand, challenge, and obtain review or reversal of a consequential automated decision.", "background": false, "home": "Q5.3", "home_id": "human-oversight", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 2, "user_proxy": 1 }, "book_passages": [ { "volume": "vol2", "chapter": "Responsible AI", "url": "https://mlsysbook.ai/vol2/16_responsible_ai/16_responsible_ai.html#sec-responsible-ai-transparency-contestability-2f2c" } ] }, { "id": "mlsys.automation-bias", "label": "Automation bias", "aliases": [ "automation complacency", "ironies of automation" ], "definition": "The tendency of human supervisors to over-rely on automated outputs and to lose the vigilance and skill needed to catch their errors.", "background": false, "home": "Q5.3", "home_id": "human-oversight", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 5, "organizational": 3, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.authority-handover", "label": "Authority handover", "aliases": [ "takeover", "shared autonomy", "bumpless transfer" ], "definition": "Transfer of control between an automated system and a human operator, bounded by human reaction time, situational awareness, and mode clarity.", "background": false, "home": "Q5.3", "home_id": "human-oversight", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 16, "organizational": 8, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Supervisory Intervention", "url": "https://mlsysbook.ai/vol4/14_intervention/14_intervention.html#sec-intervention-reaction-drift" }, { "volume": "vol4", "chapter": "Supervisory Intervention", "url": "https://mlsysbook.ai/vol4/14_intervention/14_intervention.html#sec-intervention-handshake-protocol" } ] }, { "id": "mlsys.ai-risk-management", "label": "AI risk management", "aliases": [ "NIST AI RMF", "ISO/IEC 42001" ], "definition": "An organizational process to govern, map, measure, and manage risks of AI systems throughout their lifecycle.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 10 }, "book_passages": [] }, { "id": "mlsys.impact-assessment", "label": "Algorithmic impact assessment", "aliases": [ "data protection impact assessment", "DPIA" ], "definition": "A structured pre-deployment analysis of who an AI system could affect, how, and with what severity, used to decide mitigations or whether to deploy.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.algorithmic-audit", "label": "Algorithmic audit", "aliases": [], "definition": "An internal or independent examination of an AI system's development records and behavior against stated requirements, standards, or law.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.risk-tiered-regulation", "label": "Risk-tiered AI regulation", "aliases": [ "EU AI Act" ], "definition": "Law that classifies AI uses by risk level and imposes obligations, such as conformity assessment, documentation, and human oversight, scaled to that level.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 1, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.frontier-safety-framework", "label": "Frontier safety framework", "aliases": [ "responsible scaling policy", "preparedness framework" ], "definition": "A developer policy that ties dangerous-capability evaluation thresholds to required safeguards before further training or deployment of frontier models.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.context-compaction", "label": "Context compaction", "aliases": [], "definition": "Replacing older parts of an agent's context with summaries or structured records so a long task fits the model's input budget.", "background": false, "home": "T1.1", "home_id": "context-memory", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 10, "organizational": 3, "user_proxy": 3 }, "book_passages": [ { "volume": "vol3", "chapter": "Context Engineering", "url": "https://mlsysbook.ai/vol3/04_context_engineering/04_context_engineering.html#sec-vol3-context-engineering-compaction" } ] }, { "id": "mlsys.context-degradation", "label": "Long-context degradation", "aliases": [ "lost in the middle", "context rot" ], "definition": "Decline in a model's use of relevant information as context grows or as that information sits away from the context's start or end.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.agent-memory", "label": "Long-term agent memory", "aliases": [], "definition": "A persistent store an agent writes to and retrieves from across calls and sessions, with policies for consolidation, expiry, and invalidation.", "background": false, "home": "T1.1", "home_id": "context-memory", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 4, "user_proxy": 1 }, "book_passages": [ { "volume": "vol3", "chapter": "Long-Term Memory", "url": "https://mlsysbook.ai/vol3/06_long_term_memory/06_long_term_memory.html#sec-vol3-long-term-memory-record-roles" } ] }, { "id": "mlsys.function-calling", "label": "Function calling", "aliases": [ "tool calling", "tool use" ], "definition": "A model emitting a structured invocation of a declared tool, which the harness validates, executes, and returns as an observation.", "background": false, "home": "T1.2", "home_id": "tool-interfaces", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 24, "organizational": 4, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-context-protocol", "label": "Tool discovery protocol", "aliases": [ "Model Context Protocol", "MCP" ], "definition": "A shared client-server interface through which agent hosts discover the tools, resources, and prompts external services expose, with their schemas, reducing custom integration.", "background": false, "home": "T1.2", "home_id": "tool-interfaces", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.computer-use", "label": "Computer use", "aliases": [ "GUI agent", "browser agent" ], "definition": "An agent operating graphical software through screenshots or accessibility trees and synthesized keyboard and pointer input.", "background": false, "home": "T1.2", "home_id": "tool-interfaces", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 2, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.agent-harness", "label": "Agent harness", "aliases": [ "agent loop", "agent runtime" ], "definition": "The runtime that repeatedly calls a model, dispatches its tool actions, appends observations, and decides whether the trajectory continues, pauses, or ends.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 25, "organizational": 7, "user_proxy": 1 }, "book_passages": [ { "volume": "vol3", "chapter": "The Agent Harness", "url": "https://mlsysbook.ai/vol3/10_agent_harness/10_agent_harness.html#sec-vol3-agent-harness-transition-guards" } ] }, { "id": "mlsys.agent-budget", "label": "Agent execution budget", "aliases": [], "definition": "Limits on turns, tokens, wall-clock time, or spend for a trajectory, with stopping rules applied when a limit or completion criterion is reached.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 8, "user_proxy": 1 }, "book_passages": [ { "volume": "vol3", "chapter": "The Agent Harness", "url": "https://mlsysbook.ai/vol3/10_agent_harness/10_agent_harness.html#sec-vol3-harness-multidim-budgets" } ] }, { "id": "mlsys.approval-gate", "label": "Action approval gate", "aliases": [ "human-in-the-loop approval" ], "definition": "A harness checkpoint that holds a specific proposed action until an authorized human or policy approves it, binding the approval to that exact action.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 4, "user_proxy": 1 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Security", "url": "https://mlsysbook.ai/vol3/09_security/09_security.html#sec-vol3-security-action-bound-approval" } ] }, { "id": "mlsys.durable-execution", "label": "Durable execution", "aliases": [], "definition": "Execution that persists workflow progress and reuses recorded results on recovery, with external-effect guarantees determined by endpoint idempotency, transactional boundaries, and reconciliation.", "background": false, "home": "T1.4", "home_id": "durable-execution", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 14, "organizational": 7, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Durable Execution", "url": "https://mlsysbook.ai/vol3/11_durable_execution/11_durable_execution.html#sec-vol3-persistence-event-sourcing" } ] }, { "id": "mlsys.deterministic-replay", "label": "Record and replay", "aliases": [ "deterministic replay" ], "definition": "Re-executing a trajectory from its log by substituting recorded model outputs and tool results, to reproduce, debug, or test a divergence.", "background": false, "home": "T1.4", "home_id": "durable-execution", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 3, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.suspend-resume", "label": "Suspend and resume", "aliases": [], "definition": "Persisting a trajectory while it waits on a human, timer, or external event, then resuming with a validated state and authority.", "background": false, "home": "T1.4", "home_id": "durable-execution", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.multi-agent-orchestration", "label": "Multi-agent orchestration", "aliases": [ "orchestrator-worker", "delegation" ], "definition": "Structuring a task as delegated subtasks among agents under a topology, such as orchestrator-worker or peer, with typed handoffs and result merging.", "background": false, "home": "T1.5", "home_id": "multi-agent", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 15, "organizational": 7, "user_proxy": 1 }, "book_passages": [ { "volume": "vol3", "chapter": "Multi-Agent Coordination", "url": "https://mlsysbook.ai/vol3/18_multi_agent/18_multi_agent.html#sec-vol3-multiagent-contracts" } ] }, { "id": "mlsys.shared-agent-state", "label": "Shared agent state", "aliases": [], "definition": "Task state read and written by several agents, requiring concurrency control such as leases or optimistic versioning to avoid lost or conflicting updates.", "background": false, "home": "T1.5", "home_id": "multi-agent", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.coordination-failure", "label": "Multi-agent coordination failure", "aliases": [], "definition": "Task failure arising from interaction among agents, such as mis-specified delegation, lost context at handoff, conflicting actions, or premature termination.", "background": false, "home": "T1.5", "home_id": "multi-agent", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.correlated-agent-errors", "label": "Correlated agent errors", "aliases": [], "definition": "Errors shared across agents built on the same model or prompts, which make voting, debate, or cross-checking less independent than assumed.", "background": false, "home": "T1.5", "home_id": "multi-agent", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.robot-perception", "label": "Robot perception", "aliases": [], "definition": "Models that convert raw sensor streams into task-relevant scene estimates, such as detections, segmentation, depth, and pose, under latency limits.", "background": true, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.sensor-fusion", "label": "Sensor fusion", "aliases": [], "definition": "Combining measurements from multiple sensors or modalities, at raw, feature, or decision level, into one estimate more accurate or robust than any single source.", "background": false, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 0, "user_proxy": 10 }, "book_passages": [] }, { "id": "mlsys.time-synchronization", "label": "Sensor time synchronization", "aliases": [], "definition": "Aligning clocks and timestamps across sensors, compute, and actuators so measurements can be associated with the physical instant they describe.", "background": false, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 5, "user_proxy": 16 }, "book_passages": [] }, { "id": "mlsys.recursive-filtering", "label": "Recursive state filtering", "aliases": [ "Kalman filter", "EKF", "particle filter", "filter consistency", "NEES", "NIS" ], "definition": "Updating a state estimate and its uncertainty by alternating model-based prediction with measurement correction at each step.", "background": true, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 2, "user_proxy": 1 }, "book_passages": [ { "volume": "vol4", "chapter": "Spatial Memory", "url": "https://mlsysbook.ai/vol4/09_memory/09_memory.html#sec-perception-state-estimation" } ] }, { "id": "mlsys.slam", "label": "Simultaneous localization and mapping", "aliases": [ "SLAM" ], "definition": "Jointly estimating a platform's pose and a map of an unknown environment from its own sensor observations.", "background": true, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.world-model", "label": "Learned world model", "aliases": [], "definition": "A learned model that predicts future observations or latent states conditioned on actions, used for planning, policy training, or simulation.", "background": false, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.motion-planning", "label": "Motion planning", "aliases": [], "definition": "Computing a collision-free path or trajectory from a start to a goal configuration using search, sampling, or trajectory optimization.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-predictive-control", "label": "Model predictive control", "aliases": [ "MPC (control)" ], "definition": "Repeatedly solving a finite-horizon constrained optimal control problem from the current state and applying only the first input.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.real-time-scheduling", "label": "Real-time scheduling", "aliases": [ "rate-monotonic", "EDF", "response-time analysis", "hard real-time deadline" ], "definition": "Assigning processor time to periodic and sporadic tasks so that each meets its deadline, with schedulability established by analysis.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 12, "organizational": 3, "user_proxy": 94 }, "book_passages": [ { "volume": "vol4", "chapter": "The Nervous System", "url": "https://mlsysbook.ai/vol4/04_nervous/04_nervous.html#sec-nervous-deterministic-fieldbuses" } ] }, { "id": "mlsys.mixed-criticality", "label": "Mixed-criticality partitioning", "aliases": [], "definition": "Running functions of different assurance levels on shared hardware while preventing lower-criticality work from delaying or corrupting higher-criticality functions.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 5, "user_proxy": 24 }, "book_passages": [] }, { "id": "mlsys.vla-model", "label": "Vision-language-action model", "aliases": [ "VLA", "robot foundation model" ], "definition": "A policy built on a pretrained vision-language backbone that maps images and language instructions to robot actions.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 15, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.teleoperation-data", "label": "Teleoperation data", "aliases": [], "definition": "Robot demonstrations collected by humans driving the robot remotely, whose cost, coverage, and action-label fidelity bound imitation-learned policies.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.action-chunking", "label": "Action chunking", "aliases": [], "definition": "Predicting a short sequence of future actions per policy inference, decoupling model invocation rate from control rate and reducing compounding error.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 12, "organizational": 6, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "The Cognitive Brain", "url": "https://mlsysbook.ai/vol4/03_brain/03_brain.html#sec-brain-action-chunking" } ] }, { "id": "mlsys.domain-randomization", "label": "Domain randomization", "aliases": [], "definition": "Training in simulation over randomized visual and physical parameters so the policy treats the real world as one more variation.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 2, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Policy Training", "url": "https://mlsysbook.ai/vol4/06_training/06_training.html#sec-training-simulator-component" } ] }, { "id": "mlsys.thread-hierarchy", "label": "Thread hierarchy", "aliases": [ "grid/block/warp" ], "definition": "The decomposition of a kernel into grids of thread blocks and warps that share on-chip memory and synchronize at defined scopes.", "background": false, "home": "S2.3", "home_id": "gpu-programming", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.memory-coalescing", "label": "Memory coalescing", "aliases": [ "coalesced access" ], "definition": "Combining memory requests from threads in a warp into fewer global-memory transactions, reducing wasted memory traffic.", "background": false, "home": "S2.3", "home_id": "gpu-programming", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 18 }, "book_passages": [] }, { "id": "mlsys.occupancy", "label": "Occupancy and latency hiding", "aliases": [ "occupancy" ], "definition": "The ratio of resident to maximum warps, and how oversubscribing threads hides memory latency at the cost of registers and shared memory per thread.", "background": false, "home": "S2.3", "home_id": "gpu-programming", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 46, "organizational": 0, "user_proxy": 26 }, "book_passages": [] }, { "id": "mlsys.async-data-movement", "label": "Asynchronous data movement", "aliases": [ "TMA" ], "definition": "Transferring data into on-chip memory while independent computation proceeds, with explicit synchronization on the transfer's completion; hardware copy engines and producer-consumer warp roles are its implementations.", "background": false, "home": "S2.3", "home_id": "gpu-programming", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 0, "organizational": 0, "user_proxy": 16 }, "book_passages": [] }, { "id": "mlsys.matrix-engine-programming", "label": "Matrix-engine programming", "aliases": [ "tensor core programming", "MMA" ], "definition": "Issuing tile-level matrix-multiply-accumulate instructions with the layouts, fragment shapes, and precisions a tensor-core-class unit requires.", "background": false, "home": "S2.3", "home_id": "gpu-programming", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 0, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.embedding-table-sharding", "label": "Embedding-table sharding", "aliases": [ "DLRM sharding" ], "definition": "Partitioning very large sparse embedding tables row-, column-, or table-wise across devices, usually model-parallel beside data-parallel dense layers.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 21 }, "book_passages": [] }, { "id": "mlsys.embedding-cache", "label": "Embedding cache and tiering", "aliases": [ "embedding offload" ], "definition": "Keeping frequently accessed embedding rows in accelerator memory and cold rows in host memory or SSD, exploiting the skew of lookups.", "background": false, "home": "S4.3", "home_id": "storage-systems", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 1, "user_proxy": 18 }, "book_passages": [] }, { "id": "mlsys.parameter-server", "label": "Parameter server", "aliases": [ "PS architecture" ], "definition": "An architecture in which workers pull and push model parameters from sharded server processes, common for sparse recommendation models and asynchronous training.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.ranking-funnel", "label": "Multi-stage retrieval and ranking", "aliases": [ "ranking funnel", "candidate generation" ], "definition": "Candidate generation followed by progressively costlier ranking stages under one latency budget, the dominant architecture for recommendation and search.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 7, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.multimodal-model", "label": "Multimodal model", "aliases": [ "vision-language model", "VLM" ], "definition": "A model that consumes or produces several modalities, such as text, images, audio, or video, through modality encoders feeding a shared backbone.", "background": false, "home": "F2.1", "home_id": "model-families", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.modality-encoder", "label": "Modality encoder", "aliases": [ "vision encoder" ], "definition": "A network that turns images, audio, or video into token sequences for a shared backbone; visual tokens often dominate prefill cost and KV size.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.encoder-disaggregation", "label": "Multimodal encoder disaggregation", "aliases": [ "encode-prefill-decode disaggregation" ], "definition": "Scheduling modality encoders separately from prefill and decode and caching encoder outputs so image or audio inputs do not stall token generation.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.auto-parallelization", "label": "Automatic parallelization", "aliases": [ "auto-sharding", "automatic parallelism" ], "definition": "Compiler-derived multi-device partitioning of a single-device program, found by propagating sharding annotations or by searching parallelism strategies under a cost model.", "background": false, "home": "S3.2", "home_id": "parallelism-planning", "added_in": "0.2-draft", "warrant": { "literary": 8, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.sharding-propagation", "label": "Sharding annotation and propagation", "aliases": [ "SPMD partitioning", "GSPMD-style sharding" ], "definition": "Annotating a few tensors with layouts over a device mesh and letting the compiler infer the remaining layouts and insert required collectives, yielding one SPMD program.", "background": false, "home": "S3.2", "home_id": "parallelism-planning", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.device-mesh", "label": "Device mesh", "aliases": [], "definition": "A logical multidimensional arrangement of devices whose axes are assigned to parallelism dimensions, usually aligned so the busiest axis maps to the fastest interconnect.", "background": false, "home": "S3.2", "home_id": "parallelism-planning", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 2, "organizational": 1, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.function-transformations", "label": "Composable function transformations", "aliases": [ "grad/vmap/jit", "functional transformations" ], "definition": "Higher-order operations that map pure array functions to new functions, such as differentiation, vectorization, compilation, and sharding, and compose subject to each transformation's input, output, and tracing contract.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.task-actor-runtime", "label": "Distributed task and actor runtime", "aliases": [ "actor model runtime", "Ray-style runtime" ], "definition": "A runtime exposing stateless remote tasks and stateful actors as future-returning calls scheduled across a cluster; implementations add a shared object store and lineage-based recovery.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.single-vs-multi-controller", "label": "Single- vs. multi-controller execution", "aliases": [], "definition": "Whether one coordinating process issues work to all devices or every host runs the same program independently, trading orchestration flexibility against dispatch scalability.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.analytical-performance-model", "label": "Analytical performance model", "aliases": [], "definition": "A closed-form model composing compute, memory, communication, and queueing terms to predict a workload's time, memory, and cost on a configuration without running it.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 0, "organizational": 2, "user_proxy": 30 }, "book_passages": [] }, { "id": "mlsys.performance-simulation", "label": "Performance simulation", "aliases": [ "system simulation" ], "definition": "Predicting system behavior by executing a model of compute, memory, network, or scheduling events at a chosen granularity, trading simulation speed against detail.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.2-draft", "warrant": { "literary": 9, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-fidelity", "label": "Simulator fidelity", "aliases": [ "validation error" ], "definition": "Agreement between a performance model's or simulator's predictions and measured behavior on reference workloads, reported as validation error over stated configurations.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 6, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.post-training-quantization", "label": "Post-training quantization", "aliases": [ "PTQ", "SmoothQuant" ], "definition": "Quantizing a trained model without retraining, choosing scales and rounding from statistics of the weights and, for many methods, a small calibration set of representative inputs.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 2, "user_proxy": 156 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Compression", "url": "https://mlsysbook.ai/vol1/10_model_compression/10_model_compression.html#sec-model-compression-quantization-precision-cd46" } ] }, { "id": "mlsys.weight-only-quantization", "label": "Weight-only quantization", "aliases": [ "W4A16", "GPTQ", "AWQ", "codebook weight quantization", "vector quantization of weights", "additive quantization" ], "definition": "Storing weights at low bit width while computing in higher-precision activations, cutting memory traffic for bandwidth-bound decoding, with error-compensating or activation-aware rounding.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 5, "organizational": 2, "user_proxy": 136 }, "book_passages": [] }, { "id": "mlsys.activation-outliers", "label": "Activation outliers", "aliases": [ "outlier features" ], "definition": "Rare, large-magnitude activation channels in large transformers that inflate quantization ranges, handled by mixed-precision decomposition or by migrating scale into weights.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 2, "organizational": 0, "user_proxy": 52 }, "book_passages": [] }, { "id": "mlsys.kv-cache-quantization", "label": "KV-cache quantization", "aliases": [ "KV quantization" ], "definition": "Storing attention keys and values at reduced precision, often with per-channel keys and per-token values, to raise batch size or context length under fixed memory.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 2, "organizational": 1, "user_proxy": 66 }, "book_passages": [] }, { "id": "mlsys.nm-sparsity", "label": "N:M structured sparsity", "aliases": [ "2:4 sparsity", "semi-structured sparsity" ], "definition": "A fine-grained pattern keeping at most N nonzero values in every group of M, such as 2:4, that matrix engines can skip for real speedup.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 3, "organizational": 1, "user_proxy": 26 }, "book_passages": [] }, { "id": "mlsys.kv-cache-offload", "label": "KV-cache offloading and tiering", "aliases": [ "KV offload", "KV tiering" ], "definition": "Moving KV-cache blocks from accelerator memory to host memory or SSD and reloading them on reuse, extending cache capacity at transfer bandwidth and latency cost.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 8, "corpus_chapters": 0, "organizational": 2, "user_proxy": 30 }, "book_passages": [ { "volume": "vol3", "chapter": "KV Cache Management", "url": "https://mlsysbook.ai/vol3/05_kv_cache/05_kv_cache.html#sec-vol3-kv-cache-keep-move-drop" } ] }, { "id": "mlsys.kv-cache-transfer", "label": "KV-cache transfer", "aliases": [ "KV migration" ], "definition": "Shipping KV state between instances or nodes, as in disaggregated prefill and decode or cross-instance prefix reuse, overlapped with compute to hide network latency.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 1, "user_proxy": 31 }, "book_passages": [] }, { "id": "mlsys.model-routing", "label": "Model routing", "aliases": [], "definition": "Choosing which model serves each request from request features, predicted quality, and cost.", "background": false, "home": "P4.2", "home_id": "compound-ai", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 2, "organizational": 0, "user_proxy": 22 }, "book_passages": [] }, { "id": "mlsys.model-gateway", "label": "Model gateway", "aliases": [ "LLM gateway", "AI gateway" ], "definition": "A proxy in front of one or more model providers that centralizes authentication, routing, rate limits, fallbacks, and usage logging for model calls.", "background": false, "home": "P4.2", "home_id": "compound-ai", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.multi-adapter-serving", "label": "Multi-adapter serving", "aliases": [ "multi-LoRA serving", "multi-tenant LoRA" ], "definition": "Serving many low-rank adapters over one shared base model by batching requests for different adapters with heterogeneous kernels and paging adapter weights on demand.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 18 }, "book_passages": [] }, { "id": "mlsys.kv-cache-eviction", "label": "KV-cache eviction", "aliases": [ "KV token eviction", "KV-cache sparsification", "heavy-hitter retention", "attention sink retention", "sink-aware KV eviction" ], "definition": "Retaining only a selected subset of cached key-value entries, chosen by attention scores, recency, or learned importance, to bound KV memory during long generation.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 1, "organizational": 0, "user_proxy": 27 }, "book_passages": [] }, { "id": "mlsys.cpu-gpu-hybrid-inference", "label": "CPU-GPU hybrid inference", "aliases": [ "host-assisted inference", "CPU offloaded attention" ], "definition": "Splitting inference so the host CPU computes some operations, such as attention or rarely active neurons, from host memory when model or KV state exceeds accelerator memory.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 6, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.graph-vector-index", "label": "Graph-based vector index", "aliases": [ "HNSW", "proximity-graph index" ], "definition": "An approximate nearest-neighbor index linking vectors into a proximity graph, often hierarchical, searched by greedy traversal for high recall at high memory cost.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 2, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.quantized-vector-index", "label": "Partitioned and quantized vector index", "aliases": [ "IVF-PQ" ], "definition": "An index that clusters vectors into partitions probed selectively and compresses them with product quantization, trading recall for memory footprint and scan cost.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.filtered-vector-search", "label": "Filtered vector search", "aliases": [ "predicate-constrained ANN" ], "definition": "Nearest-neighbor search restricted to records satisfying metadata predicates, where naive pre- or post-filtering loses recall or cost under selective filters.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.index-freshness", "label": "Vector index freshness", "aliases": [ "streaming ANN updates" ], "definition": "The delay and recall cost of reflecting inserts, updates, and deletes in an ANN index, managed by incremental insertion, delta indexes, tombstones, and periodic rebuilds.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 1, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.numerical-nondeterminism", "label": "Numerical nondeterminism", "aliases": [], "definition": "Run-to-run variation in floating-point results because non-associative reductions change order with atomics, kernel selection, parallel layout, batch composition, or hardware.", "background": false, "home": "S2.7", "home_id": "numerics", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 1, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.batch-invariance", "label": "Batch invariance", "aliases": [ "batch-invariant kernels" ], "definition": "A kernel property that each request's output is bitwise independent of batch size and co-batched requests, one way to remove a source of run-to-run numerical variation in serving.", "background": false, "home": "S2.7", "home_id": "numerics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.checkpoint-resharding", "label": "Checkpoint resharding", "aliases": [ "universal checkpoint", "distributed checkpoint" ], "definition": "Loading a checkpoint saved under one parallel layout or world size into another by remapping sharded tensors, needed for elastic restarts and training-to-serving handoff.", "background": false, "home": "Q2.2", "home_id": "checkpoint-recovery", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.batch-data-processing", "label": "Distributed batch data processing", "aliases": [ "MapReduce", "Spark-style dataflow" ], "definition": "Transforming bounded datasets with dataflow engines that partition records across workers, shuffle between stages, and re-execute failed tasks, for large-scale preprocessing and feature generation.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 3, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.distributed-tracing", "label": "Distributed tracing", "aliases": [ "request tracing", "spans" ], "definition": "Recording request-scoped, causally linked spans across services so that one request's latency and failures can be attributed to the components it traversed.", "background": false, "home": "P5.2", "home_id": "monitoring", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 73, "organizational": 2, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.llm-call-tracing", "label": "Model and agent call tracing", "aliases": [ "LLM observability", "GenAI tracing" ], "definition": "Tracing model calls, retrievals, and tool invocations as spans annotated with prompts, token counts, latency, and cost, so quality and spend are attributable per step.", "background": false, "home": "P5.2", "home_id": "monitoring", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 2, "user_proxy": 1 }, "book_passages": [ { "volume": "vol3", "chapter": "Operating Agents", "url": "https://mlsysbook.ai/vol3/14_agent_operations/14_agent_operations.html#sec-vol3-evaluation-tracing-fields" } ] }, { "id": "mlsys.evaluation-harness", "label": "Evaluation harness", "aliases": [ "eval harness", "hidden tests", "grader isolation" ], "definition": "A system running versioned benchmark suites against model checkpoints with fixed prompts, decoding settings, and scorers, sharding and caching work so results stay reproducible and comparable.", "background": false, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 16, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.energy-measurement", "label": "Energy measurement methodology", "aliases": [ "power measurement" ], "definition": "The declared sensor sources, sampling rate, measurement boundary, and attribution rules by which a workload's energy is measured, from on-chip counters to facility meters.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 1, "user_proxy": 21 }, "book_passages": [] }, { "id": "mlsys.idle-dynamic-power", "label": "Idle vs. dynamic power", "aliases": [ "static vs. dynamic power" ], "definition": "Separating power drawn by provisioned but idle hardware from the incremental power of active work, which decides how shared energy is attributed to workloads.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 132 }, "book_passages": [] }, { "id": "mlsys.he-inference", "label": "Homomorphic-encryption inference", "aliases": [ "HE inference", "encrypted inference" ], "definition": "Evaluating a model on encrypted inputs with homomorphic encryption so the server never sees plaintext, at large compute overhead and with polynomial approximations of nonlinearities.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.secure-computation-inference", "label": "Secure multi-party computation inference", "aliases": [ "secure two-party inference", "SMPC inference", "2PC inference" ], "definition": "Jointly evaluating a model over secret-shared inputs or weights so no party learns another's data, with cost dominated by communication rounds for nonlinear layers.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.red-teaming", "label": "Red teaming", "aliases": [ "AI red teaming", "automated red teaming" ], "definition": "Structured adversarial probing of a model or system, by people or automated attacker models, to elicit harmful, insecure, or policy-violating behavior before and after release.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 5, "organizational": 1, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.dangerous-capability-eval", "label": "Dangerous-capability evaluation", "aliases": [ "frontier capability evaluation", "extreme-risk evaluation" ], "definition": "Evaluating whether a model has capabilities, such as cyberoffense, biological uplift, or autonomous replication, that cross thresholds requiring added safeguards.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.compute-in-memory", "label": "Compute-in-memory", "aliases": [ "in-memory computing", "processing-in-memory", "PIM", "analog in-memory computing" ], "definition": "Performing matrix-vector operations inside or beside memory arrays, often in the analog domain, so weights need not move to separate compute units.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.emerging-nvm", "label": "Emerging non-volatile memory", "aliases": [ "ReRAM", "phase-change memory", "MRAM", "resistive memory" ], "definition": "Resistive, phase-change, and magnetic memory devices that retain state without power and can store weights densely or act as analog compute elements.", "background": false, "home": "S1.2", "home_id": "memory-systems", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.embedded-dsl", "label": "Embedded ML DSL", "aliases": [ "host-embedded DSL", "staged array language" ], "definition": "A domain-specific language hosted inside a general language such as Python, whose operations build a program representation that a separate compiler or runtime executes.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.differentiable-programming", "label": "Differentiable programming", "aliases": [ "differentiable programming language" ], "definition": "Writing general programs, including control flow, recursion, and data structures, that the language or framework differentiates end to end by automatic differentiation.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.persistent-kernel", "label": "Persistent kernel", "aliases": [ "megakernel", "mega-kernel", "persistent megakernel" ], "definition": "A long-running kernel that stays resident on the device and executes many operators, or a whole forward pass, from an in-kernel work queue.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 1, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.graph-superoptimization", "label": "Tensor graph superoptimization", "aliases": [ "graph substitution", "TASO-style superoptimization" ], "definition": "Automatically generating and verifying semantics-preserving rewrites of tensor computation graphs or kernels, then searching the rewritten space under a cost model.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.llm-training-framework", "label": "Large-model training framework", "aliases": [ "Megatron-LM", "DeepSpeed", "distributed training library" ], "definition": "A library that composes data, tensor, pipeline, and sharded-state parallelism with memory optimizations behind one training configuration for models exceeding one device.", "background": false, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 11, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.cloud-service-model", "label": "Cloud service models", "aliases": [ "IaaS", "PaaS", "SaaS", "managed ML service" ], "definition": "Divisions of responsibility between cloud provider and customer, from rented virtual infrastructure to managed platforms to hosted model APIs.", "background": false, "home": "S4.5", "home_id": "capacity-provisioning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.lakehouse", "label": "Data warehouse and lakehouse", "aliases": [ "data warehouse", "data lake", "lakehouse" ], "definition": "Analytical stores of governed tables queried with SQL; a lakehouse adds transactional table formats over open files in object storage.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.training-power-swings", "label": "Synchronous training power swings", "aliases": [ "training power fluctuations", "training load swings" ], "definition": "Rapid cluster-wide swings in power draw as synchronized accelerators alternate between compute and communication or stall together at checkpoints and failures.", "background": false, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.facility-energy-storage", "label": "UPS and battery energy storage", "aliases": [ "UPS", "uninterruptible power supply", "BESS" ], "definition": "Facility batteries that carry load through outages until generators start and can also buffer fast load swings or shave demand peaks.", "background": true, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.grid-interconnection", "label": "Grid interconnection", "aliases": [ "grid connection", "utility interconnection" ], "definition": "The utility connection that fixes how much power a site can draw; its capacity and multi-year lead times increasingly decide where AI datacenters are built.", "background": true, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.on-site-generation", "label": "On-site power generation", "aliases": [ "behind-the-meter generation", "backup generators" ], "definition": "Generators, turbines, fuel cells, or renewables at the datacenter that provide backup power or supply load alongside or instead of the grid.", "background": true, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.data-cleaning", "label": "Data cleaning", "aliases": [ "data cleansing", "data repair" ], "definition": "Detecting and correcting erroneous, inconsistent, duplicated, or outlying records with rules, statistics, or learned repair models before the data is used for training.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 6, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.class-imbalance", "label": "Class imbalance", "aliases": [ "imbalanced data", "imbalanced classes" ], "definition": "A skewed label distribution in which some classes are rare, so aggregate accuracy misleads and training favors the majority classes.", "background": false, "home": "P1.1", "home_id": "data-collection", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.rebalancing", "label": "Class rebalancing", "aliases": [ "oversampling", "undersampling", "class weighting", "SMOTE" ], "definition": "Changing per-class sampling rates, synthesizing minority examples, or weighting the loss so rare classes contribute adequately to training.", "background": false, "home": "P1.1", "home_id": "data-collection", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.feature-hashing", "label": "Feature hashing", "aliases": [ "hashing trick", "hashed features" ], "definition": "Mapping high-cardinality or open-vocabulary categorical values into a fixed number of buckets with a hash function, bounding table size at the cost of collisions.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.feature-cross", "label": "Feature cross", "aliases": [ "crossed features", "feature conjunction" ], "definition": "A synthetic feature formed from the conjunction of two or more categorical features so a model, especially a linear one, can learn their interaction.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.data-governance", "label": "Data governance", "aliases": [ "data stewardship" ], "definition": "Policies, roles, and controls that decide who may collect, access, use, and share which data, for which purposes, and under what accountability.", "background": false, "home": "P1.5", "home_id": "data-lineage", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.data-consent", "label": "Consent and purpose limitation", "aliases": [ "purpose limitation", "data-use consent" ], "definition": "Restricting collection and use of personal data to purposes the data subject agreed to or a legal basis allows, recorded so downstream uses can be checked.", "background": false, "home": "P1.5", "home_id": "data-lineage", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 1, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.exploratory-data-analysis", "label": "Exploratory data analysis", "aliases": [ "EDA", "data profiling" ], "definition": "Summarizing and plotting distributions, correlations, and anomalies in a dataset before modeling to inform cleaning, feature design, and validation rules.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-visualization", "label": "Model and prediction visualization", "aliases": [ "training dashboards", "embedding projection", "confusion-matrix plots" ], "definition": "Plotting training curves, activations, embeddings, and prediction errors such as confusion matrices to inspect what a model learned and where it fails.", "background": false, "home": "P3.5", "home_id": "ml-testing", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-selection", "label": "Model selection", "aliases": [ "model development" ], "definition": "Choosing among candidate model families, architectures, and hyperparameters by validated performance under the system's accuracy, cost, and latency constraints.", "background": false, "home": "P2.4", "home_id": "experiment-management", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 23, "organizational": 0, "user_proxy": 33 }, "book_passages": [] }, { "id": "mlsys.pretraining", "label": "Pretraining", "aliases": [ "base-model pretraining", "pretraining run" ], "definition": "Training a model from scratch on a large, broad corpus with a self-supervised objective to produce a general base model for later adaptation.", "background": false, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 32, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.causal-lm", "label": "Causal language modeling", "aliases": [ "next-token prediction", "autoregressive language modeling" ], "definition": "Training a model to predict each token from the tokens before it, maximizing the likelihood of text under a left-to-right factorization.", "background": false, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.gan", "label": "Generative adversarial network", "aliases": [ "GAN" ], "definition": "A generator and a discriminator trained against each other, so the generator learns to produce samples the discriminator cannot tell apart from data.", "background": false, "home": "F2.1", "home_id": "model-families", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.mode-collapse", "label": "Mode collapse", "aliases": [], "definition": "A generative-training failure in which the model covers only a few modes of the data distribution and produces low-diversity samples.", "background": false, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.perplexity", "label": "Perplexity", "aliases": [], "definition": "The exponentiated average negative log-likelihood per token of held-out text, measuring how well a language model predicts that text.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 8, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.reference-match-scoring", "label": "Reference-match scoring", "aliases": [ "exact match", "token-overlap F1", "BLEU", "ROUGE" ], "definition": "Scoring a generated output by exact or token-overlap agreement with reference answers; cheap and reproducible but blind to valid answers phrased differently.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.error-analysis", "label": "Error analysis", "aliases": [], "definition": "Inspecting a model's mistakes, grouping them into categories, and ranking categories by frequency and fixability to decide the next data or model change.", "background": false, "home": "P3.5", "home_id": "ml-testing", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.training-diagnostics", "label": "Training diagnostics", "aliases": [ "overfit-one-batch test", "learning-curve diagnosis" ], "definition": "Checks such as overfitting a tiny batch, reading learning curves, and tracking gradient and activation statistics that localize bugs in training code, data, and configuration.", "background": false, "home": "P3.5", "home_id": "ml-testing", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.keyed-predictions", "label": "Request correlation keys", "aliases": [ "keyed predictions", "pass-through keys" ], "definition": "Passing a client-supplied key through inference unchanged so outputs of batch or asynchronous requests can be matched to their inputs regardless of order.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.stateless-serving-function", "label": "Stateless model serving", "aliases": [ "stateless serving function" ], "definition": "Exporting a model as a stateless function from request inputs to prediction, so any replica can serve any request and replicas scale horizontally.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.semantic-cache", "label": "Semantic cache", "aliases": [ "semantic caching" ], "definition": "Returning a stored model response when a new request is close in embedding space to an earlier one, trading answer fidelity for latency and cost.", "background": false, "home": "P4.2", "home_id": "compound-ai", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.model-cascade", "label": "Model cascade", "aliases": [], "definition": "Sending a request to cheaper models first and escalating to more capable models only when a confidence or quality criterion is not met.", "background": false, "home": "P4.2", "home_id": "compound-ai", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 2, "user_proxy": 24 }, "book_passages": [] }, { "id": "mlsys.glue-code", "label": "Glue code", "aliases": [], "definition": "Supporting code written to move data into and out of general-purpose ML packages, often far larger than the model code and binding the system to package specifics.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.pipeline-jungle", "label": "Pipeline jungle", "aliases": [], "definition": "Data preparation grown incrementally into a tangle of scrapes, joins, and intermediate outputs that is hard to test, trace, and change.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.build-vs-buy", "label": "Build-versus-buy decision", "aliases": [ "build vs. buy", "MLOps tool selection" ], "definition": "Choosing whether to build an ML platform component in house, adopt open source, or buy a managed service, weighing fit, cost, lock-in, and team capacity.", "background": false, "home": "P5.6", "home_id": "ml-platforms", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 10 }, "book_passages": [] }, { "id": "mlsys.user-feedback", "label": "User feedback collection", "aliases": [ "explicit feedback", "implicit feedback" ], "definition": "Capturing explicit signals such as ratings and corrections and implicit signals such as clicks, edits, and abandonment, for use as labels or quality signals.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.human-ai-interaction", "label": "Human-AI interaction design", "aliases": [ "human-centered AI", "human-AI guidelines" ], "definition": "Designing how a system conveys its capabilities, uncertainty, and errors to users and lets them correct, override, or dismiss its outputs.", "background": false, "home": "Q5.3", "home_id": "human-oversight", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.demand-response", "label": "Datacenter demand response", "aliases": [ "flexible datacenter load", "grid-interactive datacenter" ], "definition": "Shifting, shedding, or capping datacenter power in response to grid signals, made practical by deferrable training and batch work.", "background": false, "home": "S4.1", "home_id": "datacenter-facilities", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.prompt-engineering", "label": "Prompt engineering", "aliases": [ "prompt design" ], "definition": "Designing instructions, examples, and input layout for a model call to elicit desired behavior without changing the model's weights.", "background": false, "home": "P2.5", "home_id": "fine-tuning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.loss-masking", "label": "Loss masking and token normalization", "aliases": [ "prompt masking", "assistant-only loss" ], "definition": "Excluding prompt, tool-output, and other non-target tokens from the training loss, and choosing how remaining token losses are averaged across turns and sequences.", "background": false, "home": "P2.5", "home_id": "fine-tuning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Trajectory Fine-Tuning", "url": "https://mlsysbook.ai/vol3/16_fine_tuning/16_fine_tuning.html#sec-vol3-fine-tuning-loss-masking" } ] }, { "id": "mlsys.serverless-inference", "label": "Serverless inference", "aliases": [ "scale-to-zero", "function-as-a-service inference" ], "definition": "Running models on a managed platform that provisions replicas on demand, bills per use, and can scale to zero, exposing requests to cold-start latency.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 2, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.streaming-asr", "label": "Streaming speech recognition", "aliases": [ "streaming ASR", "on-device speech recognition" ], "definition": "Serving speech recognition incrementally: consuming audio in chunks and emitting partial transcripts within a fixed lag, with encoder state carried between chunks.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.non-prefix-kv-reuse", "label": "Non-prefix KV reuse", "aliases": [ "position-independent KV caching", "KV cache blending" ], "definition": "Reusing precomputed KV entries for text chunks appearing at arbitrary positions across requests, selectively recomputing cross-chunk attention instead of the full prefill.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.dynamic-sparse-attention", "label": "Dynamic sparse attention", "aliases": [ "content-adaptive sparse attention", "input-dependent sparse attention" ], "definition": "Attention that computes only the query-key blocks predicted important for each input at runtime, rather than a fixed pattern, to cut long-context cost.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.2-draft", "warrant": { "literary": 7, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.token-streaming", "label": "Token streaming", "aliases": [ "incremental response delivery", "server-sent events" ], "definition": "Delivering generated output to the client incrementally as tokens are produced, holding per-stream state and stopping generation when the client cancels or disconnects.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 7, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.heterogeneous-accelerator-execution", "label": "Heterogeneous-accelerator training and serving", "aliases": [ "heterogeneous GPU clusters", "mixed-generation GPU training" ], "definition": "Partitioning one training or serving job across accelerators of different types or generations, sizing stages and batch shares to each device's compute, memory, and links.", "background": false, "home": "S3.2", "home_id": "parallelism-planning", "added_in": "0.2-draft", "warrant": { "literary": 10, "corpus_chapters": 0, "organizational": 0, "user_proxy": 14 }, "book_passages": [] }, { "id": "mlsys.graph-neural-network", "label": "Graph neural network", "aliases": [ "GNN", "message-passing neural network" ], "definition": "Neural network that computes node, edge, or graph representations by repeatedly aggregating features from each node's neighbors along graph edges.", "background": false, "home": "F2.1", "home_id": "model-families", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 1, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.input-prefetching", "label": "Input prefetching", "aliases": [ "data prefetching", "prefetch buffer" ], "definition": "Preparing upcoming batches in background workers and buffers while the accelerator processes the current batch, overlapping input delivery with computation.", "background": false, "home": "S2.6", "home_id": "data-loading", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 1, "user_proxy": 47 }, "book_passages": [ { "volume": "vol2", "chapter": "Data Storage", "url": "https://mlsysbook.ai/vol2/04_data_storage/04_data_storage.html#sec-storage-pipelining" } ] }, { "id": "mlsys.dynamic-resharding", "label": "Dynamic re-parallelization", "aliases": [ "parameter reallocation", "runtime resharding" ], "definition": "Redistributing model weights among devices at runtime so successive phases, such as generation and training or prefill and decode, each run under a better parallel layout.", "background": false, "home": "S3.2", "home_id": "parallelism-planning", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.activation-sparsity", "label": "Contextual activation sparsity", "aliases": [ "dynamic activation sparsity", "contextual sparsity" ], "definition": "Skipping neurons, heads, or channels whose activations are predicted negligible for the current input, reducing per-token compute and weight reads at inference.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.denoising-step-reduction", "label": "Denoising-step reduction", "aliases": [ "few-step sampling", "step distillation" ], "definition": "Lowering network evaluations per diffusion sample with faster samplers, step distillation, or consistency models, trading sample quality against latency and cost.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.ml-aware-scheduling", "label": "ML-aware cluster scheduling", "aliases": [ "deep-learning cluster scheduling", "goodput-aware scheduling" ], "definition": "Allocating cluster accelerators using ML job properties, such as scaling curves, iteration time, and parallelism choice, to optimize objectives like job completion time or goodput.", "background": false, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 1, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.adaptive-routing", "label": "Adaptive routing", "aliases": [ "flowlet switching", "packet spraying", "ECMP collision avoidance" ], "definition": "Choosing network paths per packet or flowlet from current congestion, avoiding the hash collisions static ECMP suffers with few large, synchronized training flows.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 3, "user_proxy": 35 }, "book_passages": [ { "volume": "vol2", "chapter": "Network Fabrics", "url": "https://mlsysbook.ai/vol2/03_network_fabrics/03_network_fabrics.html#sec-network-fabrics-behavior" } ] }, { "id": "mlsys.incast", "label": "Incast congestion", "aliases": [ "incast", "many-to-one congestion" ], "definition": "Congestion that occurs when many senders transmit to one receiver at once, overflowing its switch port buffer and causing queueing, drops, or pause propagation.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 1, "user_proxy": 37 }, "book_passages": [] }, { "id": "mlsys.context-selection", "label": "Context selection", "aliases": [ "context budget allocation", "tool selection", "working-set selection" ], "definition": "Choosing which instructions, tool descriptions, retrieved items, state, and observations enter a model's limited context at each step, within a token budget.", "background": false, "home": "T1.1", "home_id": "context-memory", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 3, "user_proxy": 3 }, "book_passages": [ { "volume": "vol3", "chapter": "Context Engineering", "url": "https://mlsysbook.ai/vol3/04_context_engineering/04_context_engineering.html#sec-vol3-context-engineering-decision" } ] }, { "id": "mlsys.context-staleness", "label": "Context staleness", "aliases": [ "stale observations", "context invalidation" ], "definition": "Observations held in an agent's context or memory that no longer reflect the environment after later actions or external changes, yet still steer decisions unless invalidated.", "background": false, "home": "T1.1", "home_id": "context-memory", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 10, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Context Engineering", "url": "https://mlsysbook.ai/vol3/04_context_engineering/04_context_engineering.html#sec-vol3-context-engineering-claim-dependencies" } ] }, { "id": "mlsys.multi-hop-retrieval", "label": "Iterative multi-hop retrieval", "aliases": [ "agentic retrieval", "multi-hop RAG" ], "definition": "Retrieval that issues successive queries, each formulated from previously retrieved evidence or reasoning steps, to answer questions no single lookup resolves.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.capability-based-security", "label": "Capability-based authorization", "aliases": [ "capability tokens", "object capabilities", "attenuated credentials" ], "definition": "Authorization in which holding an unforgeable, scoped reference grants a specific right that can be delegated with attenuated scope and later revoked.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Sandboxes", "url": "https://mlsysbook.ai/vol3/08_sandboxes/08_sandboxes.html#sec-vol3-sandboxes-scoped-grants" } ] }, { "id": "mlsys.ephemeral-sandbox", "label": "Ephemeral sandbox", "aliases": [ "copy-on-write workspace", "disposable execution environment", "microVM sandbox" ], "definition": "An isolated execution environment created per task from a snapshot, often copy-on-write, and destroyed or reset afterward so no state carries across tasks or tenants.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.egress-control", "label": "Network egress control", "aliases": [ "egress filtering", "outbound allowlist" ], "definition": "Restricting which external network destinations a workload or sandbox may contact, typically through default-deny allowlists, to block exfiltration and unintended remote actions.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 2, "user_proxy": 1 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Sandboxes", "url": "https://mlsysbook.ai/vol3/08_sandboxes/08_sandboxes.html#sec-vol3-sandboxes-egress-enforcement" } ] }, { "id": "mlsys.information-flow-control", "label": "Information-flow control", "aliases": [ "taint tracking", "data-flow labels" ], "definition": "Labeling data with confidentiality and integrity tags and propagating them through computation so policy can block untrusted inputs from steering privileged actions or secrets from leaking.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Agent Security", "url": "https://mlsysbook.ai/vol3/09_security/09_security.html#sec-vol3-security-integrity-confidentiality" } ] }, { "id": "mlsys.compounding-error", "label": "Compounding error in sequential decisions", "aliases": [ "covariate shift in imitation", "exposure bias", "horizon error compounding" ], "definition": "Per-step errors of a learned sequential policy or generator push later inputs away from the training distribution, so failure probability grows with horizon length.", "background": false, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 7, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Trajectory Fine-Tuning", "url": "https://mlsysbook.ai/vol3/16_fine_tuning/16_fine_tuning.html#sec-vol3-fine-tuning-exposure-bias" }, { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html#sec-data-curse-dimensionality-recoverability" } ] }, { "id": "mlsys.interactive-imitation", "label": "Interactive imitation learning", "aliases": [ "DAgger", "dataset aggregation", "recovery demonstrations" ], "definition": "Imitation learning that queries an expert for corrective labels on states the learner's own policy visits and aggregates them into training data over rounds.", "background": false, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 14, "organizational": 3, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "Physical Data", "url": "https://mlsysbook.ai/vol4/05_data/05_data.html#sec-data-curse-dimensionality-recoverability" } ] }, { "id": "mlsys.in-context-learning", "label": "In-context learning", "aliases": [ "few-shot prompting", "zero-shot prompting" ], "definition": "A pretrained model performing a task from instructions or examples supplied in its input at inference time, with no weight updates.", "background": false, "home": "P2.5", "home_id": "fine-tuning", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.reward-hacking", "label": "Reward hacking", "aliases": [ "specification gaming", "reward overoptimization" ], "definition": "Improving a measured reward through behavior that fails the intended task objective.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.1-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 4, "user_proxy": 3 }, "book_passages": [ { "volume": "vol3", "chapter": "Reinforcement Learning from Verifiable Rewards", "url": "https://mlsysbook.ai/vol3/17_rlvr/17_rlvr.html#sec-vol3-rlvr-score-intended-outcome" } ] }, { "id": "mlsys.actuator-limits", "label": "Actuator limits", "aliases": [ "torque saturation", "thermal duty limits" ], "definition": "Bounds on the force, torque, speed, and sustained duty an actuator can deliver, set by drive current limits, gearing, and motor heating.", "background": false, "home": "T2.3", "home_id": "planning-control", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 19, "organizational": 4, "user_proxy": 0 }, "book_passages": [ { "volume": "vol4", "chapter": "The Physical Body", "url": "https://mlsysbook.ai/vol4/02_body/02_body.html#sec-body-kinetic-momentum" }, { "volume": "vol4", "chapter": "The Physical Body", "url": "https://mlsysbook.ai/vol4/02_body/02_body.html#sec-body-thermal-capacity" } ] }, { "id": "mlsys.watchdog-timer", "label": "Watchdog timer", "aliases": [ "heartbeat timeout", "windowed watchdog", "authority-expiry timer" ], "definition": "A timer that a monitored component must refresh within a deadline or window; expiry triggers a defined safe response such as reset or revoking control authority.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 16, "organizational": 2, "user_proxy": 67 }, "book_passages": [] }, { "id": "mlsys.closed-loop-evaluation", "label": "Closed-loop policy evaluation", "aliases": [ "closed-loop testing", "interactive policy evaluation" ], "definition": "Evaluating a policy by letting its own actions determine subsequent observations, in simulation or deployment, rather than scoring its predictions against logged data.", "background": false, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 4, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Trajectory Fine-Tuning", "url": "https://mlsysbook.ai/vol3/16_fine_tuning/16_fine_tuning.html#sec-vol3-fine-tuning-eval-decoupling" }, { "volume": "vol4", "chapter": "Closed-Loop Evaluation", "url": "https://mlsysbook.ai/vol4/07_evaluation/07_evaluation.html#sec-evaluation-what-measured" } ] }, { "id": "mlsys.event-data-recorder", "label": "Event data recorder", "aliases": [ "black-box recorder", "safety event log", "authority record" ], "definition": "Tamper-evident onboard logging of inputs, decisions, control authority, and safety interventions around events, retained for incident reconstruction and assurance evidence.", "background": false, "home": "Q4.4", "home_id": "safety-cases", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 12, "organizational": 2, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.decision-accountability", "label": "Decision accountability record", "aliases": [ "decision ownership", "accountability assignment", "approval record" ], "definition": "A record of who authorized each consequential decision, policy, or automated action and on what evidence, so responsibility can be assigned and reviewed later.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 10, "organizational": 3, "user_proxy": 1 }, "book_passages": [ { "volume": "vol4", "chapter": "Responsible Physical AI", "url": "https://mlsysbook.ai/vol4/17_responsibility/17_responsibility.html#sec-responsibility-approval-record" } ] }, { "id": "mlsys.agent-interruption", "label": "Agent interruption and cancellation", "aliases": [ "pause and cancel", "operator interrupt", "agent kill switch" ], "definition": "Operator controls that pause, redirect, cancel, or terminate a running agent at defined boundaries while settling in-flight tool calls into a known state.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.agent-loop-detection", "label": "Agent loop detection", "aliases": [ "semantic watchdog", "livelock detection", "repetition detection" ], "definition": "Monitoring an agent trajectory for repeated actions, cycling states, or absent progress, and halting or escalating when the agent appears stuck.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 2, "user_proxy": 0 }, "book_passages": [ { "volume": "vol3", "chapter": "Failure Recovery", "url": "https://mlsysbook.ai/vol3/12_failure_recovery/12_failure_recovery.html#sec-vol3-sagas-watchdogs" } ] }, { "id": "mlsys.task-and-motion-planning", "label": "Task and motion planning", "aliases": [ "TAMP", "affordance grounding", "SayCan-style grounding" ], "definition": "Planning that jointly searches discrete symbolic action sequences and continuous motions, accepting a task plan only when each step admits a feasible motion.", "background": false, "home": "T2.3", "home_id": "planning-control", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.multi-object-tracking", "label": "Multi-object tracking", "aliases": [ "MOT", "data association", "track management" ], "definition": "Maintaining identities and states of multiple moving objects over time by associating new detections with existing tracks and starting or ending tracks.", "background": false, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.change-impact-analysis", "label": "Safety change-impact analysis", "aliases": [ "impact analysis", "safety-case maintenance" ], "definition": "Determining which hazards, safety-case claims, and evidence a change to a model, software, data, or operating domain invalidates, and what must be re-verified.", "background": false, "home": "Q4.4", "home_id": "safety-cases", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.sparse-tensor-kernels", "label": "Sparse tensor kernels", "aliases": [ "SpMM", "SDDMM", "sparse convolution" ], "definition": "Kernels such as sparse-dense matrix multiply, sampled dense-dense multiply, and sparse convolution that compute over compressed nonzero structure, trading irregular access for skipped work.", "background": false, "home": "S2.2", "home_id": "tensor-operators", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 1, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.local-sgd", "label": "Local SGD", "aliases": [ "periodic model averaging", "DiLoCo-style training", "low-communication training" ], "definition": "Data-parallel training in which workers take several local optimizer steps between parameter synchronizations, cutting communication frequency for low-bandwidth or geo-distributed settings.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.expert-placement", "label": "Expert replication and placement", "aliases": [ "expert load balancing", "redundant experts" ], "definition": "Assigning mixture-of-experts experts to devices, and replicating frequently routed ones, to balance per-device token load under skewed routing.", "background": false, "home": "S3.2", "home_id": "parallelism-planning", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 20 }, "book_passages": [] }, { "id": "mlsys.memory-efficient-optimizer", "label": "Memory-efficient optimizer", "aliases": [ "low-rank optimizer state", "8-bit optimizer" ], "definition": "Optimizers that shrink per-parameter state through factorization, low-rank projection, or low-precision storage, cutting the memory adaptive methods such as Adam otherwise require.", "background": false, "home": "P2.3", "home_id": "training-memory", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.execution-trace", "label": "Operator execution trace", "aliases": [ "execution graph trace", "Chakra trace" ], "definition": "A recorded graph of a training or inference step's compute and communication operators with their dependencies and timings, replayable for simulation and benchmarking.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.offline-rl", "label": "Offline reinforcement learning", "aliases": [ "batch reinforcement learning", "offline RL" ], "definition": "Learning a policy from a fixed dataset of previously logged interactions without further environment access, limited by how well the data covers states the policy would visit.", "background": true, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.byzantine-robust-aggregation", "label": "Byzantine-robust aggregation", "aliases": [ "robust gradient aggregation", "Krum" ], "definition": "Combining worker or client updates with robust statistics, such as medians, trimmed means, or Krum, so a bounded fraction of arbitrary malicious updates cannot steer training.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 1, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.cluster-manager", "label": "Cluster manager", "aliases": [ "cluster orchestrator", "cluster control plane", "Kubernetes" ], "definition": "The control plane that admits jobs, places their tasks on machines, enforces quotas and priorities, and reconciles running state toward declared intent across a cluster.", "background": false, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 12, "organizational": 1, "user_proxy": 23 }, "book_passages": [] }, { "id": "mlsys.rebound-effect", "label": "Rebound effect", "aliases": [ "Jevons paradox" ], "definition": "Efficiency gains lower the cost of a service enough that total demand, and with it total energy or resource use, grows, partly or fully offsetting the savings.", "background": false, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 1, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.data-cascade", "label": "Data cascade", "aliases": [ "compounding data defects" ], "definition": "Compounding downstream harm from upstream data problems, such as poor collection or labeling, that surface late, often after deployment, and are costly to trace.", "background": false, "home": "P1.3", "home_id": "data-quality", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.data-valuation", "label": "Data valuation", "aliases": [ "Data Shapley", "data attribution" ], "definition": "Estimating each training example's or data source's contribution to model quality, for example through influence or Shapley-style attribution, to guide selection, pricing, or removal.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.diffusion-policy", "label": "Diffusion or flow-matching policy", "aliases": [ "flow-matching policy", "diffusion policy" ], "definition": "Robot policies that generate continuous actions through iterative denoising or learned-flow integration, representing multiple action modes while spending an inference-iteration budget.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.sensor-spoofing", "label": "Sensor spoofing", "aliases": [ "transduction attack", "physical sensor attack", "LiDAR spoofing" ], "definition": "Attacks that inject false signals into a sensor's physical channel, such as fabricated LiDAR returns or GPS and acoustic spoofing, so perception receives counterfeit measurements.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.paired-comparison", "label": "Paired comparison and noninferiority testing", "aliases": [ "paired difference test", "noninferiority test" ], "definition": "Comparing two systems on the same items or units and testing per-item differences, including whether a candidate is no worse than a baseline by a preset margin.", "background": false, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.cpu-execution-architecture", "label": "CPU execution architecture", "aliases": [ "general-purpose CPU", "host CPU", "CPU vector units" ], "definition": "Organization of general-purpose processor cores, SIMD vector units, and multilevel caches that runs host code, preprocessing, classical models, and CPU inference.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 21, "organizational": 0, "user_proxy": 38 }, "book_passages": [] }, { "id": "mlsys.cpu-threading-numa", "label": "CPU threading and NUMA placement", "aliases": [ "thread affinity", "intra-op and inter-op threads", "NUMA-aware execution" ], "definition": "Configuring thread pools, core affinity, and memory placement across NUMA domains so CPU operators parallelize without oversubscription or remote-memory penalties.", "background": true, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 30 }, "book_passages": [] }, { "id": "mlsys.dynamic-shapes", "label": "Dynamic shapes and shape specialization", "aliases": [ "symbolic shapes", "shape bucketing", "shape recompilation" ], "definition": "Handling inputs whose tensor dimensions vary across calls by specializing compiled code per shape, padding into buckets, or compiling with symbolic dimensions.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 41 }, "book_passages": [] }, { "id": "mlsys.rule-based-graph-rewriting", "label": "Rule-based graph rewriting", "aliases": [ "constant folding", "algebraic simplification", "operator substitution", "graph optimization passes", "graph rewriting" ], "definition": "Applying hand-written semantics-preserving transformations to a computation graph, such as constant folding, dead-code elimination, operator substitution, and algebraic simplification.", "background": false, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 3, "organizational": 0, "user_proxy": 50 }, "book_passages": [] }, { "id": "mlsys.sparse-ragged-tensors", "label": "Sparse and ragged tensor representation", "aliases": [ "ragged tensor", "nested tensor", "jagged tensor", "CSR format" ], "definition": "Storing tensors with mostly zero entries or variable-length dimensions in compressed formats such as COO, CSR, or offset-indexed ragged layouts instead of padded dense arrays.", "background": false, "home": "S2.2", "home_id": "tensor-operators", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.execution-providers", "label": "Execution providers and operator fallback", "aliases": [ "execution provider", "backend delegation", "operator fallback" ], "definition": "Partitioning a model graph across runtime backends by operator support, running supported subgraphs on accelerators and falling back to a CPU backend otherwise.", "background": false, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 266 }, "book_passages": [] }, { "id": "mlsys.large-language-model", "label": "Large language model", "aliases": [ "LLM", "LLM workload" ], "definition": "A language model large enough that parameter memory, attention state that grows with context, and, when serving, token-by-token decoding dominate its resource demands.", "background": false, "home": "F2.1", "home_id": "model-families", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 44, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.inference-compute-estimate", "label": "Inference compute and memory estimate", "aliases": [ "2N FLOPs per token", "KV-cache sizing" ], "definition": "Approximating dense transformer inference as about two FLOPs per parameter per token, plus KV-cache bytes from layers, KV heads, head dimension, context length, batch, and precision.", "background": false, "home": "F1.2", "home_id": "workload-characterization", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 12 }, "book_passages": [] }, { "id": "mlsys.sync-async-training", "label": "Synchronous vs. asynchronous training", "aliases": [ "synchronous SGD", "asynchronous SGD", "async SGD" ], "definition": "Choosing whether workers apply an update only after every gradient for a step is combined, or apply updates independently on possibly stale parameters.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.neighbor-sampling", "label": "Neighbor sampling", "aliases": [ "graph sampling", "mini-batch graph sampling" ], "definition": "Building GNN mini-batches from sampled subsets of each node's multi-hop neighborhood, bounding the exponential neighbor growth that makes full-neighborhood computation exceed memory.", "background": false, "home": "S2.6", "home_id": "data-loading", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.event-time-windowing", "label": "Event-time windowing and watermarks", "aliases": [ "stream watermarks", "late data handling", "windowed aggregation" ], "definition": "Grouping stream events into windows by when they occurred, using watermarks to estimate completeness and explicit policies for events arriving after a window closes.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.diffusion-feature-caching", "label": "Cross-step feature caching", "aliases": [ "diffusion caching" ], "definition": "Reusing intermediate activations that change little between adjacent denoising steps instead of recomputing them, cutting per-step cost at bounded quality loss.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.voice-pipeline", "label": "Voice interaction pipeline", "aliases": [ "voice agent", "ASR-LLM-TTS cascade" ], "definition": "Serving a spoken interaction as one latency-budgeted pipeline of streaming speech recognition, a language model, and speech synthesis, with turn detection and interruption handling between stages.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.llm-request-scheduling", "label": "LLM request scheduling and preemption", "aliases": [ "request preemption", "iteration-level scheduling" ], "definition": "Ordering, prioritizing, and preempting in-flight generation requests at iteration granularity, swapping or recomputing their KV state, to meet latency targets and fairness goals.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 4, "corpus_chapters": 5, "organizational": 0, "user_proxy": 51 }, "book_passages": [] }, { "id": "mlsys.containerization", "label": "Containerization", "aliases": [ "containers", "Docker" ], "definition": "Packaging a model, its code, and its dependencies as an immutable image that runs unchanged across machines, leaving the scheduling of those containers across a cluster to the cluster manager.", "background": false, "home": "P5.5", "home_id": "pipelines-reproducibility", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 24, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.distributed-hang-diagnosis", "label": "Distributed job hang diagnosis", "aliases": [ "collective timeout", "NCCL watchdog", "node health checks" ], "definition": "Detecting and localizing stalled distributed jobs through collective timeouts, heartbeats, per-rank progress traces, and node health checks, so faulty ranks can be isolated.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.2-draft", "warrant": { "literary": 5, "corpus_chapters": 3, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.benchmark-representativeness", "label": "Benchmark representativeness", "aliases": [ "microbenchmark vs. end-to-end benchmark", "benchmark realism" ], "definition": "The degree to which a benchmark's workload, scale, and measurement boundary predict production behavior, from isolated microbenchmarks to end-to-end application runs.", "background": false, "home": "Q1.2", "home_id": "system-benchmarking", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 29 }, "book_passages": [] }, { "id": "mlsys.data-model-licensing", "label": "Data and model licensing", "aliases": [ "dataset license", "model license", "open-weight license" ], "definition": "Legal terms governing how datasets and model weights may be used, modified, redistributed, or deployed, which constrain acquisition, fine-tuning, and release decisions.", "background": false, "home": "P1.5", "home_id": "data-lineage", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.content-provenance", "label": "Content provenance", "aliases": [ "C2PA", "content credentials" ], "definition": "Cryptographically signed metadata attached to media that records its origin and edit history, including model generation, so recipients can verify source claims.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.fleet-data-engine", "label": "Fleet data engine", "aliases": [ "data engine", "triggered data collection", "edge-case mining", "data flywheel" ], "definition": "Collecting training data from deployed devices through on-device triggers that flag rare or uncertain events, uploading selected samples under bandwidth budgets for labeling and retraining.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 18 }, "book_passages": [] }, { "id": "mlsys.robotics-middleware", "label": "Robotics middleware and dataflow", "aliases": [ "ROS 2", "publish-subscribe middleware", "DDS" ], "definition": "Software connecting sensing, perception, planning, and control processes as nodes that exchange typed messages over publish-subscribe topics with configurable delivery and timing policies.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 8, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.parallel-simulation", "label": "Parallel physics simulation", "aliases": [ "GPU-parallel simulation", "simulated environments", "vectorized environments" ], "definition": "Running many simulated robot environments concurrently, often batched on accelerators, so policy learning and data generation gather experience faster than physical collection.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.agent-credentials", "label": "Agent credentials and delegated authority", "aliases": [ "agent identity", "delegated tokens", "secret brokering" ], "definition": "Issuing tool-using agents scoped, short-lived credentials that act on a user's behalf, brokering secrets outside model context so authority stays bounded and revocable.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.gnn-graph-partitioning", "label": "Graph partitioning for GNN training", "aliases": [ "graph partitioning", "METIS partitioning", "distributed GNN training" ], "definition": "Splitting a large graph across workers to balance computation while minimizing cut edges, whose cross-partition neighbor features must be fetched remotely.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.host-data-loader", "label": "Host-side data loading", "aliases": [ "data loader", "DataLoader workers", "pinned memory" ], "definition": "Reading, decoding, augmenting, and collating examples on host CPU worker processes and staging batches in pinned memory so that a single machine's accelerators are not left waiting for input.", "background": false, "home": "S2.6", "home_id": "data-loading", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 17, "organizational": 0, "user_proxy": 66 }, "book_passages": [] }, { "id": "mlsys.streaming-inference", "label": "Streaming inference", "aliases": [ "incremental inference", "online streaming inference" ], "definition": "Serving a model over an open stream of input, producing outputs incrementally from partial input while carrying state between chunks, rather than answering one complete request at a time.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.inference-api", "label": "Inference API and payload serialization", "aliases": [ "prediction API", "gRPC inference", "tensor serialization" ], "definition": "The request and response protocol of a model server, including the transport (HTTP, gRPC, streaming connections), the tensor and payload encoding, and the cost of serializing inputs and outputs.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 23 }, "book_passages": [] }, { "id": "mlsys.request-throughput", "label": "Request throughput", "aliases": [ "queries per second", "QPS", "tokens per second" ], "definition": "The number of requests, or tokens, a serving system completes per unit time at a stated load and latency target.", "background": false, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 23, "organizational": 0, "user_proxy": 29 }, "book_passages": [] }, { "id": "mlsys.latency-percentiles", "label": "Latency percentiles", "aliases": [ "p50", "p99", "percentile latency" ], "definition": "Summarizing a latency distribution by percentiles such as p50, p95, and p99, so that a service is judged by the experience of most requests and of the slow tail rather than by the mean.", "background": false, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 44, "organizational": 1, "user_proxy": 7 }, "book_passages": [ { "volume": "vol1", "chapter": "Benchmarking", "url": "https://mlsysbook.ai/vol1/12_benchmarking/12_benchmarking.html#sec-benchmarking-latency-tail-latency-5cde" } ] }, { "id": "mlsys.cache-aware-routing", "label": "Cache-aware request routing", "aliases": [ "prefix-affinity routing", "KV-aware routing", "session affinity" ], "definition": "Sending each request to the replica that already holds a reusable prefix or KV cache for it, trading load balance against cache reuse.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.2-draft", "warrant": { "literary": 2, "corpus_chapters": 4, "organizational": 0, "user_proxy": 10 }, "book_passages": [] }, { "id": "mlsys.process-group-rendezvous", "label": "Distributed job launch and process groups", "aliases": [ "rendezvous", "ranks", "communicator initialization", "torchrun" ], "definition": "Starting the processes of a distributed job, having them discover each other through a rendezvous, assigning ranks, and forming the process groups and communicators that collectives run over.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 41, "organizational": 0, "user_proxy": 10 }, "book_passages": [] }, { "id": "mlsys.agentic-workload", "label": "Agentic workload", "aliases": [ "agent workload", "multi-turn inference load" ], "definition": "The load an agent places on the serving stack, including context that grows each turn, idle gaps while tools run, many short dependent model calls, and reuse of KV state across steps.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.performance-regression-testing", "label": "Performance regression testing", "aliases": [ "continuous benchmarking", "perf regression gate" ], "definition": "Benchmarking every change to code, kernels, models, or configuration continuously and failing the change when throughput, latency, or memory degrades beyond a set tolerance.", "background": false, "home": "Q1.2", "home_id": "system-benchmarking", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.measurement-variability", "label": "Performance measurement variability", "aliases": [ "benchmark noise", "warm-up effects", "timing variance" ], "definition": "Run-to-run variation in measured time and throughput from warm-up, asynchronous execution, clock and thermal state, and contention, handled by synchronizing, repeating runs, and reporting spread.", "background": false, "home": "Q1.2", "home_id": "system-benchmarking", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.training-data-extraction", "label": "Training data extraction", "aliases": [ "memorization attack", "training data regurgitation" ], "definition": "Recovering particular training examples or their sensitive contents through a trained model's outputs, exposing information that was not intended for release.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.gradient-clipping", "label": "Gradient clipping", "aliases": [ "gradient norm clipping" ], "definition": "Rescaling or truncating gradients whose norm exceeds a threshold so that a single large update cannot destabilize training.", "background": true, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 0, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.weight-initialization", "label": "Weight initialization", "aliases": [ "Xavier initialization", "He initialization" ], "definition": "Choosing the scale and distribution of initial parameters so that activations and gradients keep stable variance through depth at the start of training.", "background": true, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.delivery-semantics", "label": "Processing delivery semantics", "aliases": [ "exactly-once processing", "at-least-once delivery", "duplicate events" ], "definition": "The guarantee a data pipeline gives about how often each record takes effect (at most once, at least once, or effectively once) under retries, replays, and failures, and the idempotent or transactional sinks that achieve it.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.online-feature-consistency", "label": "Online and offline feature consistency", "aliases": [ "feature materialization", "online store consistency" ], "definition": "Keeping the feature values materialized in the low-latency online store consistent with the offline values used for training, as features are recomputed, backfilled, and updated.", "background": false, "home": "P1.6", "home_id": "data-leakage-skew", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.ood-detection", "label": "Out-of-distribution detection", "aliases": [ "OOD detection", "novelty detection" ], "definition": "Flagging individual inputs that differ from the training distribution so the system can abstain, defer, or fall back instead of trusting the prediction.", "background": false, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.model-feature-compatibility", "label": "Model and feature compatibility", "aliases": [ "interface compatibility", "tokenizer compatibility" ], "definition": "Ensuring that a model version, its tokenizer and feature transformations, the schemas it reads, and the consumers of its outputs remain compatible as each is versioned and released independently.", "background": false, "home": "P5.1", "home_id": "release-management", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 20 }, "book_passages": [] }, { "id": "mlsys.secrets-management", "label": "Workload identity and secrets management", "aliases": [ "secrets management", "workload identity", "credential rotation" ], "definition": "Giving training jobs, model servers, and pipelines verifiable identities and provisioning, rotating, and revoking the keys and credentials they use, without embedding secrets in code or images.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.deletion-propagation", "label": "Data deletion propagation", "aliases": [ "right to erasure", "deletion requests" ], "definition": "Carrying a deletion request through every copy and derivative of the data, including datasets, features, caches, replicas, backups, and models trained on it.", "background": false, "home": "P5.7", "home_id": "retirement", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.memory-fragmentation", "label": "Runtime memory fragmentation", "aliases": [ "allocator fragmentation", "reserved versus allocated memory" ], "definition": "Free device memory split into blocks too small to satisfy new allocations, so a job runs out of memory while reserved memory exceeds what live tensors use.", "background": false, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 43 }, "book_passages": [] }, { "id": "mlsys.perception-latency-budget", "label": "Perception latency budget", "aliases": [ "sense-to-act latency budget" ], "definition": "The share of a robot's sense-to-act deadline allotted to sensing and perception, which bounds model size, input resolution, and frame rate on the onboard computer.", "background": false, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 19 }, "book_passages": [] }, { "id": "mlsys.onboard-compute-partitioning", "label": "Onboard compute partitioning", "aliases": [ "onboard workload placement", "robot compute offloading" ], "definition": "Assigning perception, planning, and control workloads to the CPUs, accelerators, and microcontrollers of a robot, or offloading them, so each meets its deadline and criticality level.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.2-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 41 }, "book_passages": [] }, { "id": "mlsys.health-checks", "label": "Health checks and readiness probes", "aliases": [ "readiness probe", "liveness probe", "health endpoint" ], "definition": "Periodic liveness and readiness checks that let an orchestrator or load balancer stop routing traffic to, or restart, a replica that is unhealthy or still loading its model.", "background": false, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.python-gil", "label": "Python GIL and host-language concurrency", "aliases": [ "GIL", "global interpreter lock" ], "definition": "In GIL-enabled CPython, the global interpreter lock serializes Python bytecode across threads, so ML host code achieves concurrency through native kernels that release the lock, multiple processes, or asynchronous I/O.", "background": true, "home": "S2.1", "home_id": "programming-models", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.flash-endurance", "label": "Flash write endurance", "aliases": [ "flash wear-out", "wear leveling", "program/erase cycles" ], "definition": "Flash memory cells wear out after a limited number of program and erase cycles, which bounds how often embedded and edge devices can write logs, checkpoints, or model updates; wear leveling spreads writes to extend life.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 30 }, "book_passages": [] }, { "id": "mlsys.intermittent-computing", "label": "Energy-harvesting intermittent computing", "aliases": [ "energy harvesting", "batteryless inference", "intermittent inference" ], "definition": "Running inference on devices powered by harvested energy that lose power unpredictably, which requires checkpointing progress to nonvolatile memory so computation resumes rather than restarts.", "background": false, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.mmap-weights", "label": "Memory-mapped model weights", "aliases": [ "mmap weights", "shared weight mapping", "zero-copy model loading" ], "definition": "Mapping a weight file into virtual memory so pages load on demand and read-only weights are shared across processes, cutting load time and duplicate memory at the cost of page faults.", "background": false, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 62 }, "book_passages": [] }, { "id": "mlsys.os-memory-pressure", "label": "OS memory pressure and paging", "aliases": [ "page thrashing", "low-memory killer", "memory reclamation" ], "definition": "How an operating system reclaims memory from a model process under pressure, by evicting file-backed pages, swapping or compressing memory, or killing the app, which shows up as page thrashing, latency spikes, or abrupt termination.", "background": true, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 70 }, "book_passages": [] }, { "id": "mlsys.managed-runtime-interop", "label": "Managed-runtime interop and garbage collection", "aliases": [ "JNI overhead", "GC pauses", "native interop" ], "definition": "Crossing between a managed language runtime (Java, Kotlin, Swift, C#) and native inference code, where buffer copies, object pinning, and garbage-collection pauses add latency and memory.", "background": true, "home": "S2.5", "home_id": "runtime-dispatch", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 8 }, "book_passages": [] }, { "id": "mlsys.zero-copy-dma", "label": "Zero-copy sensor input via DMA", "aliases": [ "DMA input buffers", "zero-copy capture" ], "definition": "Having sensors or peripherals write input directly into model input buffers by direct memory access, avoiding CPU copies between capture and inference.", "background": false, "home": "S2.6", "home_id": "data-loading", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 138 }, "book_passages": [] }, { "id": "mlsys.sensor-interference", "label": "Active sensor interference", "aliases": [ "sensor crosstalk", "lidar interference" ], "definition": "Crosstalk between active sensors, such as lidars, radars, or structured-light cameras, that emit into the same environment and corrupt each other's measurements.", "background": false, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.network-propagation-delay", "label": "Network propagation delay", "aliases": [ "speed-of-light latency" ], "definition": "The minimum latency set by the distance a signal travels through the medium, which sets a floor under round-trip time between regions regardless of bandwidth.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 17 }, "book_passages": [] }, { "id": "mlsys.cap-tradeoff", "label": "Consistency-availability trade-off", "aliases": [ "CAP theorem", "partition tolerance" ], "definition": "Under a network partition, a replicated system must choose between answering with possibly stale state and refusing requests until replicas agree.", "background": true, "home": "S3.4", "home_id": "coordination", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.gossip-averaging", "label": "Decentralized gossip averaging", "aliases": [ "gossip SGD", "decentralized SGD" ], "definition": "Training in which each worker averages its model with a few neighbors instead of a central server or global all-reduce, so updates diffuse through the communication graph.", "background": false, "home": "S3.5", "home_id": "federated-learning", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.interrupt-latency", "label": "Interrupt latency", "aliases": [ "ISR latency" ], "definition": "The delay from an interrupt request to the start of its handler, which contributes to event-response time alongside handler execution and subsequent processing.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 23 }, "book_passages": [] }, { "id": "mlsys.transformer-variants", "label": "Encoder-only, decoder-only, and encoder-decoder Transformers", "aliases": [ "encoder-only model", "decoder-only model", "encoder-decoder model", "seq2seq Transformer" ], "definition": "The three arrangements of Transformer stacks, which differ in whether inputs are processed bidirectionally, outputs are generated autoregressively, or both, and therefore in how they are trained, cached, and served.", "background": false, "home": "F2.1", "home_id": "model-families", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 17 }, "book_passages": [] }, { "id": "mlsys.data-transfer-cost", "label": "Data transfer cost", "aliases": [ "egress cost", "egress fees", "network transfer cost" ], "definition": "The money charged for moving data across network, zone, region, or provider boundaries, such as cloud egress fees, which can dominate the cost of distributing data, checkpoints, or responses.", "background": false, "home": "Q6.3", "home_id": "cost-economics", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 30 }, "book_passages": [] }, { "id": "mlsys.duty-cycling", "label": "Duty cycling and sleep states", "aliases": [ "duty cycle", "sleep modes", "active and sleep power states", "low-power modes" ], "definition": "Keeping a device in a low-power sleep state most of the time and waking it briefly to sense and infer, so average power is set by the active fraction and the cost of each wake-up.", "background": false, "home": "S1.4", "home_id": "chip-power-thermal", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 17, "organizational": 0, "user_proxy": 73 }, "book_passages": [] }, { "id": "mlsys.priority-inversion", "label": "Priority inversion", "aliases": [], "definition": "A high-priority task blocked on a resource held by a lower-priority task while a medium-priority task preempts the holder, so the high-priority task's blocking time is unbounded and can exceed its deadline unless the scheduler bounds it.", "background": false, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.randomized-smoothing", "label": "Certified robustness by randomized smoothing", "aliases": [ "randomized smoothing", "certified defense" ], "definition": "Classifying many noise-perturbed copies of an input and taking the majority vote, which yields a provable radius within which no adversarial perturbation can change the prediction, at the cost of many inferences per input.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.sensor-interconnect", "label": "Sensor interconnect bandwidth", "aliases": [ "MIPI CSI-2", "I2S", "SPI bandwidth", "camera link bandwidth" ], "definition": "The data rate of the links that carry sensor data into an embedded processor (MIPI CSI-2 for cameras, I2S for audio, SPI and I2C for low-rate sensors), which bounds the resolution, frame rate, and number of sensors a device can ingest.", "background": false, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 16 }, "book_passages": [] }, { "id": "mlsys.finite-buffer", "label": "Finite buffers and overflow", "aliases": [ "buffer overflow", "bounded queue", "finite-capacity queue", "backlog" ], "definition": "A queue with bounded capacity drops or blocks arrivals once full, so its size trades memory and added delay against the loss rate under bursty input.", "background": false, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 15, "organizational": 0, "user_proxy": 11 }, "book_passages": [] }, { "id": "mlsys.ui-thread-inference", "label": "Asynchronous on-device inference", "aliases": [ "background inference thread", "off-main-thread inference" ], "definition": "Running inference off the application's main thread, so a model call does not block rendering and input handling, with results delivered back by callback or message.", "background": false, "home": "P4.9", "home_id": "on-device-deployment", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 9 }, "book_passages": [] }, { "id": "mlsys.multi-head-latent-attention", "label": "Multi-head latent attention", "aliases": [ "MLA", "latent KV attention" ], "definition": "Attention that caches one low-rank latent vector per token in place of per-head keys and values, reconstructing them or absorbing the up-projections into the query and output weights, which shrinks the KV cache and decode memory traffic.", "background": false, "home": "F2.4", "home_id": "efficient-sequence-models", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.pii-deidentification", "label": "PII detection and de-identification", "aliases": [ "PII scrubbing", "data anonymization", "redaction" ], "definition": "Finding personal identifiers in training corpora, logs, and prompts and removing, masking, or pseudonymizing them before the data is stored, trained on, or shared.", "background": false, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.2-draft", "warrant": { "literary": 0, "corpus_chapters": 8, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.system-identification", "label": "System identification", "aliases": [], "definition": "Estimating the parameters of a dynamics model, such as masses, friction, delays, and actuator gains, from measured input-output data of the physical system.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.off-policy-evaluation", "label": "Off-policy evaluation", "aliases": [ "OPE", "counterfactual policy evaluation" ], "definition": "Estimating a proposed policy's expected outcomes from interactions collected under another policy, by propensity reweighting, outcome modeling, or doubly robust estimators, subject to assumptions about action coverage and confounding.", "background": true, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.pipeline-schedule", "label": "Pipeline schedule", "aliases": [ "GPipe schedule", "1F1B", "interleaved 1F1B", "zero-bubble schedule" ], "definition": "The order in which pipeline stages run the forward and backward passes of microbatches (fill-drain, one-forward-one-backward, interleaved), which sets the bubble fraction and how many microbatches' activations each stage holds.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.3-draft", "warrant": { "literary": 1, "corpus_chapters": 4, "organizational": 2, "user_proxy": 0 }, "book_passages": [ { "volume": "vol2", "chapter": "Distributed Training", "url": "https://mlsysbook.ai/vol2/06_distributed_training/06_distributed_training.html#sec-distributed-training-systems-systems-pipeline-parallelism-8748" } ] }, { "id": "mlsys.sequence-packing", "label": "Sequence packing", "aliases": [ "sample packing", "packed-sequence training" ], "definition": "Combining variable-length training examples into packed token buffers with explicit example boundaries and attention semantics to reduce padding work without unintended cross-example attention.", "background": false, "home": "S2.6", "home_id": "data-loading", "added_in": "0.3-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.document-chunking", "label": "Document chunking", "aliases": [ "retrieval chunking", "passage segmentation", "semantic chunking" ], "definition": "Dividing source documents into retrievable units with chosen boundaries, overlap, and source identifiers so retrieval granularity fits embedding and generation budgets.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.resource-exhaustion-attack", "label": "Resource-exhaustion attack", "aliases": [ "sponge examples", "denial of wallet", "model denial of service", "unbounded consumption", "reasoning slowdown attack" ], "definition": "Crafting inputs that maximize a model service's compute, energy, output length, or tool calls, degrading service for other users or running up the operator's cost.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.generated-content-watermarking", "label": "Generated-content watermarking", "aliases": [ "output watermarking", "text watermarking" ], "definition": "Embedding a statistically detectable signal into generated text, images, or audio at generation time, such as by biasing token sampling, so that content can later be identified as model-generated.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.model-merging", "label": "Model merging", "aliases": [ "model soup", "task arithmetic", "TIES-merging", "weight merging" ], "definition": "Combining compatible model parameters or parameter updates, by averaging or arithmetic on weight differences, into one model that consolidates capabilities without further training or serving the sources separately.", "background": false, "home": "P2.5", "home_id": "fine-tuning", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.optical-circuit-switching", "label": "Optical circuit switching", "aliases": [ "OCS", "reconfigurable optical fabric" ], "definition": "Reconfiguring optical connections between endpoints to match communication demand, trading switching delay and topology coordination against delivered bandwidth and network power.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.3-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.co-packaged-optics", "label": "Co-packaged optics", "aliases": [ "CPO", "co-packaged optical interconnect" ], "definition": "Placing optical transceivers inside the switch or accelerator package beside the silicon, so electrical paths shrink and energy per bit falls, at the cost of field serviceability and added package heat.", "background": false, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.advanced-accelerator-packaging", "label": "Advanced accelerator packaging", "aliases": [ "chiplet integration", "multi-die package", "heterogeneous die integration", "2.5D interposer", "die-to-die interconnect" ], "definition": "Integrating compute and memory dies through chiplets, interposers, or three-dimensional bonding to trade local bandwidth and manufacturing flexibility against power-delivery and heat-removal constraints.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 13 }, "book_passages": [] }, { "id": "mlsys.motion-forecasting", "label": "Motion forecasting", "aliases": [ "behavior prediction", "trajectory prediction", "agent motion prediction" ], "definition": "Predicting the future trajectories of other moving agents, usually as several weighted hypotheses, from their tracked history and the map, so a planner can avoid where they will be rather than where they are.", "background": false, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.representational-bias-evaluation", "label": "Representational bias evaluation", "aliases": [ "generative bias evaluation", "social bias benchmark", "stereotype evaluation" ], "definition": "Measuring stereotyped, demeaning, or unequal portrayals of groups in generated outputs, revealing representational harms that aggregate quality or decision-rate metrics can miss.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.group-relative-policy-optimization", "label": "Group relative policy optimization", "aliases": [ "GRPO" ], "definition": "A policy-gradient method that scores each sampled response against the mean reward of a group of responses to the same prompt, removing the learned critic at the cost of generating several rollouts per prompt.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 2, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.inter-token-latency", "label": "Inter-token latency", "aliases": [ "ITL" ], "definition": "The time between consecutive output tokens or chunks delivered to a streaming client, whose distribution, not only its mean, exposes decode stalls that an average time per output token hides.", "background": false, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.hybrid-retrieval", "label": "Hybrid retrieval", "aliases": [ "hybrid search", "lexical-dense retrieval", "reciprocal rank fusion" ], "definition": "Combining lexical and dense retrieval results through score or rank fusion to recover complementary matches while managing additional search cost and ranking calibration.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.fpga-acceleration", "label": "FPGA-based ML acceleration", "aliases": [ "FPGA inference", "reconfigurable ML acceleration" ], "definition": "Mapping ML operations to reconfigurable logic, arithmetic blocks, and local memories so datapaths can match a workload at the cost of hardware compilation and resource constraints.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.3-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.device-initiated-communication", "label": "Device-initiated communication", "aliases": [ "GPU-initiated communication", "accelerator-initiated communication" ], "definition": "Communication in which kernels on the accelerator issue and synchronize inter-device transfers themselves, without a host call per exchange, allowing fine-grained fusion of compute and communication at the cost of explicit ordering and completion handling.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.3-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-collapse", "label": "Model collapse", "aliases": [ "recursive synthetic-data collapse" ], "definition": "Loss of coverage of the original data distribution across generations trained on model-generated data, making provenance and real-versus-synthetic mixture policies part of quality control.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.action-tokenization", "label": "Action tokenization", "aliases": [ "robot action tokenization", "action-sequence tokenization" ], "definition": "Encoding continuous robot actions or action sequences as discrete model tokens, making reconstruction error and token count determinants of policy fidelity and execution cost.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.wafer-scale-acceleration", "label": "Wafer-scale acceleration", "aliases": [ "wafer-scale accelerator", "wafer-scale engine" ], "definition": "Executing ML on a compute-and-memory array spanning most of a semiconductor wafer, making spatial placement and communication between tiles central to performance.", "background": false, "home": "S1.1", "home_id": "accelerator-architecture", "added_in": "0.3-draft", "warrant": { "literary": 1, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.embedding-migration", "label": "Embedding model migration", "aliases": [ "re-embedding", "vector index backfill", "embedding version upgrade" ], "definition": "Changing the encoder behind a vector index while keeping queries and stored vectors comparable, by re-encoding the corpus, validated mappings, or backward-compatible training, with a controlled serving cutover.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.register-pressure", "label": "Register pressure", "aliases": [ "register spilling", "register file pressure" ], "definition": "Demand for registers from simultaneously live kernel values, which can reduce resident threads or force values to spill into slower memory.", "background": true, "home": "S2.3", "home_id": "gpu-programming", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.software-pipelining", "label": "Software pipelining", "aliases": [ "multi-stage pipelining", "kernel double buffering" ], "definition": "Scheduling independent operations from different loop iterations to overlap, such as loading the next tile while computing the current one, hiding latency at the cost of extra live state and buffering.", "background": true, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.3-draft", "warrant": { "literary": 2, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.ecc-memory", "label": "Error-correcting code memory", "aliases": [ "ECC", "SECDED" ], "definition": "Memory protected by redundant check bits that correct supported error patterns and report detectable uncorrectable errors, reducing the risk of silent corruption.", "background": true, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.integer-only-inference", "label": "Integer-only inference", "aliases": [ "integer-only quantization", "fixed-point inference", "dyadic rescaling" ], "definition": "Executing a quantized model with integer arithmetic throughout its operator graph, including accumulation and rescaling, so no floating-point operations are needed at inference time.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.execute-in-place", "label": "Execute-in-place weights", "aliases": [ "XIP", "flash-resident weights" ], "definition": "Reading immutable model weights from memory-mapped flash during inference without first copying them into SRAM, reducing working memory while exposing flash-access latency.", "background": false, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.confused-deputy", "label": "Confused deputy", "aliases": [ "confused deputy attack", "tool privilege misuse" ], "definition": "A security failure in which a component uses its legitimate authority for an untrusted party's purpose because it does not distinguish the requester's authority from its own.", "background": true, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 8, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.judge-bias", "label": "Model-grader bias", "aliases": [ "LLM-as-judge bias", "position bias", "verbosity bias", "self-preference bias" ], "definition": "Systematic scoring errors in model-based graders, such as preferences for response position, length, or style, that distort comparisons independently of task quality.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.over-refusal", "label": "Over-refusal", "aliases": [ "false refusal", "exaggerated safety" ], "definition": "Refusing benign requests that resemble harmful ones, the false-positive side of a safety filter or refusal policy, which is traded against missed harmful requests when setting the operating point.", "background": false, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.coordinate-frames", "label": "Coordinate frames and transforms", "aliases": [ "TF tree", "rigid-body transforms", "frame transforms" ], "definition": "Reference frames and the spatial transforms between them that express sensor observations, robot geometry, and actions in consistent coordinates.", "background": true, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.sensor-degradation", "label": "Sensor degradation", "aliases": [ "sensor occlusion", "adverse-weather degradation", "perceptual degradation" ], "definition": "Loss of sensor measurement fidelity under environmental or operating conditions, such as glare, rain, fog, dust, or a dirty lens, that degrades perception even when the sensing hardware has not failed.", "background": false, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.data-curriculum", "label": "Data curriculum and annealing", "aliases": [ "curriculum learning", "data annealing", "sequence-length warmup" ], "definition": "Changing the training data distribution over a run, such as raising the share of high-quality or long-sequence data during the final learning-rate decay, to improve final capability per unit of compute.", "background": false, "home": "P1.2", "home_id": "data-curation", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.streaming-dataset-format", "label": "Sharded streaming dataset format", "aliases": [ "shard streaming", "WebDataset-style shards" ], "definition": "Organizing training examples into sequentially readable shard files so workers can stream records from remote storage with few small-object reads and bounded local staging.", "background": false, "home": "P1.4", "home_id": "data-pipelines", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.collective-algorithm-selection", "label": "Collective algorithm selection", "aliases": [ "collective protocol selection", "ring versus tree crossover" ], "definition": "Choosing, per call, among collective algorithms (ring, tree, switch-offloaded) and wire protocols by message size, participant count, and link hierarchy, to minimize latency for small messages and maximize bandwidth for large ones.", "background": false, "home": "S3.3", "home_id": "collectives", "added_in": "0.3-draft", "warrant": { "literary": 1, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.always-on-cascade", "label": "Always-on sensing cascade", "aliases": [ "wake-word cascade", "sensor-hub gating", "event-triggered inference", "trigger-based inference", "motion-triggered inference" ], "definition": "Using a low-power sensing or inference stage to wake a more expensive stage only when needed, cutting continuous-monitoring energy at the risk of missed triggers and wake latency.", "background": false, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.fault-tree-analysis", "label": "Fault tree analysis", "aliases": [ "FTA" ], "definition": "Top-down hazard analysis that decomposes an undesired top event into Boolean combinations of contributing faults, so minimal cut sets and their probabilities can be computed.", "background": true, "home": "Q4.1", "home_id": "hazards-operating-domain", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.metastable-failure", "label": "Metastable failure", "aliases": [ "metastability", "self-sustaining overload" ], "definition": "A state in which a transient trigger, such as a load spike or slowdown, starts a self-sustaining loop of retries, cache misses, or timeouts that keeps a system overloaded after the trigger is gone.", "background": true, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.fan-out-tail-amplification", "label": "Fan-out tail amplification", "aliases": [ "tail at scale" ], "definition": "The growth in a request's latency tail when it waits on many parallel subrequests, because the chance that at least one subrequest is slow rises with the fan-out.", "background": true, "home": "P4.3", "home_id": "latency-goodput", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.proxy-discrimination", "label": "Proxy discrimination", "aliases": [ "redundant encoding", "proxy variables" ], "definition": "Decisions driven by features that encode protected-group membership, so group disparities persist even when the protected attribute itself is omitted.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.cross-embodiment-training", "label": "Cross-embodiment training", "aliases": [ "cross-embodiment pretraining", "multi-robot data pooling" ], "definition": "Training one policy across datasets from several robot embodiments, aligning observations and action interfaces so learning is shared despite differences in sensors and control.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.dual-rate-hierarchical-policy", "label": "Dual-rate hierarchical policy", "aliases": [ "slow-fast policy", "high-level and low-level policy", "System 1 / System 2 control" ], "definition": "Splitting a robot policy into a slow high-level model that sets goals or latent commands and a fast low-level controller that tracks them at control rate, so large-model latency does not set the control loop's rate.", "background": false, "home": "T2.5", "home_id": "robot-learning", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.robot-kinematics", "label": "Robot kinematics", "aliases": [ "forward kinematics", "inverse kinematics", "manipulator Jacobian" ], "definition": "The geometric mapping between joint positions and velocities and the pose and velocity of a robot's links and end effector, including inverse solutions and singular configurations where motion authority is lost.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 10, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.quantization-granularity", "label": "Quantization granularity", "aliases": [ "per-tensor quantization", "per-channel quantization", "group-wise quantization" ], "definition": "The grouping of values that share quantization parameters, such as a scale and an optional zero point, from per tensor to per channel to small groups, trading approximation error against metadata and kernel cost.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.agent-training-environment", "label": "Agent training environment", "aliases": [ "agent rollout environment", "agentic RL environment" ], "definition": "A task system that supplies observations, accepts actions, and manages episode state for training tool-using policies, whose throughput and reset fidelity limit rollout collection.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-directed-control-flow", "label": "Model-directed control flow", "aliases": [ "agent autonomy level", "ReAct loop", "reason-act loop" ], "definition": "A control structure in which model outputs select the next step, tool call, or stopping decision within runtime-enforced limits, increasing flexibility while making cost and behavior harder to predict and test.", "background": false, "home": "T1.3", "home_id": "agent-harness", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.misuse-monitoring", "label": "Misuse monitoring", "aliases": [ "abuse detection", "usage-policy enforcement", "trust and safety monitoring" ], "definition": "Detecting policy-violating use of a deployed model across requests, sessions, and accounts, such as coordinated abuse or capability probing, and responding with throttling, review, or account action.", "background": false, "home": "P5.2", "home_id": "monitoring", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.masked-diffusion-decoding", "label": "Masked diffusion decoding", "aliases": [ "masked diffusion language decoding", "parallel masked-token decoding" ], "definition": "Generating text by repeatedly predicting and committing masked positions in parallel, trading refinement passes for token parallelism while weakening the attention-state reuse that autoregressive decoding relies on.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.3-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.residual-connection", "label": "Residual connection", "aliases": [ "additive skip connection", "skip connection" ], "definition": "A connection that adds a block's input, or a projection of it, to the block's output, so information and gradients can bypass the block and deep stacks stay trainable.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.multi-task-learning", "label": "Multi-task learning", "aliases": [ "multitask learning", "shared-backbone multi-task model" ], "definition": "Jointly learning several tasks through shared parameters, often a shared backbone with task-specific heads, so their training signals can help or interfere with one another.", "background": false, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.recovery-objectives", "label": "Recovery time and point objectives", "aliases": [ "RTO", "RPO", "recovery time objective", "recovery point objective", "recovery time accounting" ], "definition": "Stated limits on how long restoring service may take (RTO) and how far recovered state may lag the state before the failure (RPO), which constrain checkpointing, replication, and failover design.", "background": true, "home": "F1.1", "home_id": "requirements-slos", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.incremental-checkpointing", "label": "Incremental checkpointing", "aliases": [ "delta checkpoint" ], "definition": "Persisting only the state that changed since the preceding checkpoint, with periodic full checkpoints as bases, which cuts write volume when updates are sparse at the cost of reconstructing state from a chain on restore.", "background": false, "home": "Q2.2", "home_id": "checkpoint-recovery", "added_in": "0.3-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.multiple-comparisons", "label": "Multiple comparisons correction", "aliases": [ "multiple hypothesis testing", "multiplicity correction", "Bonferroni correction", "Benjamini-Hochberg procedure" ], "definition": "Adjusting significance thresholds when many hypotheses, metrics, or slices are tested together, so the chance of a false declared effect, by family-wise error or false discovery rate, stays controlled.", "background": true, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.model-assisted-labeling", "label": "Model-assisted labeling", "aliases": [ "pre-labeling", "model-in-the-loop annotation" ], "definition": "Using a model's predictions as draft labels that annotators accept or correct, which raises labeling throughput but can anchor annotators to the model's errors.", "background": false, "home": "P1.1", "home_id": "data-collection", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.forward-error-correction", "label": "Forward error correction", "aliases": [ "FEC", "link-layer FEC" ], "definition": "Adding redundant bits to transmitted data so the receiver corrects supported error patterns without retransmission, at the cost of coding overhead and decoding delay.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.erasure-coding", "label": "Erasure coding", "aliases": [ "Reed-Solomon coding", "erasure-coded storage" ], "definition": "Encoding stored data into data and parity fragments so it can be reconstructed after a bounded number of fragment losses, trading storage overhead against encoding and recovery work.", "background": true, "home": "S4.3", "home_id": "storage-systems", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.bandwidth-delay-product", "label": "Bandwidth-delay product", "aliases": [ "BDP" ], "definition": "A path's bottleneck bandwidth times its round-trip time, the amount of outstanding data a window-controlled flow must sustain to keep the path full.", "background": true, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.backfill-scheduling", "label": "Backfill scheduling", "aliases": [ "backfilling", "EASY backfilling" ], "definition": "Running lower-priority jobs in idle resource-time gaps without delaying the reserved start of a higher-priority job, which raises cluster utilization.", "background": true, "home": "S4.4", "home_id": "cluster-scheduling", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.traffic-classes", "label": "Network traffic classes and QoS", "aliases": [ "quality of service", "traffic class isolation" ], "definition": "Classifying traffic on a shared network into separately scheduled queues with configured priorities or bandwidth shares, limiting interference between collective, storage, and control traffic.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.telemetry-sampling", "label": "Telemetry sampling and cardinality control", "aliases": [ "trace sampling", "tail-based sampling", "metric cardinality management" ], "definition": "Choosing which traces, events, and metric label combinations to keep, by sampling and by bounding or aggregating label values, so fleet telemetry stays affordable at the risk of losing rare failures and fine-grained detail.", "background": true, "home": "P5.2", "home_id": "monitoring", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 7 }, "book_passages": [] }, { "id": "mlsys.dma-cache-maintenance", "label": "DMA cache maintenance", "aliases": [ "cache clean and invalidate for DMA", "DMA buffer synchronization" ], "definition": "Cleaning and invalidating processor caches around DMA transfers on hardware without coherent DMA, with buffers aligned to cache lines, so a device and the CPU do not read stale copies of a shared buffer; each maintenance step adds latency.", "background": true, "home": "S1.5", "home_id": "edge-embedded-hardware", "added_in": "0.3-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.expert-offloading", "label": "Expert offloading", "aliases": [ "MoE expert offloading", "expert-weight offloading" ], "definition": "Keeping some mixture-of-experts weights in host memory or storage and fetching the experts a batch needs into accelerator memory, often prefetching predicted ones, which lets a model larger than accelerator memory run at the cost of transfers.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.3-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 6 }, "book_passages": [] }, { "id": "mlsys.variable-length-parallelism", "label": "Variable-length parallelism planning", "aliases": [ "dynamic context parallelism", "workload-balanced sequence parallelism" ], "definition": "Choosing, for each training step, sequence or context parallel degrees and how packed documents are sharded across ranks from that step's sequence lengths and attention patterns, so ranks receive balanced attention work.", "background": false, "home": "S3.2", "home_id": "parallelism-planning", "added_in": "0.3-draft", "warrant": { "literary": 4, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.fine-grained-accelerator-sharing", "label": "Fine-grained accelerator sharing", "aliases": [ "sub-kernel GPU sharing", "kernel-slice preemption" ], "definition": "Scheduling co-located workloads on one accelerator in small units of a kernel, so latency-sensitive work can preempt background work quickly and interference stays bounded.", "background": false, "home": "S4.5", "home_id": "capacity-provisioning", "added_in": "0.3-draft", "warrant": { "literary": 3, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.intra-device-overlap", "label": "Intra-device overlap of bottlenecks", "aliases": [ "nano-batching", "intra-device operation overlap" ], "definition": "Scheduling generative-inference operations with complementary compute, memory, or communication demands concurrently on one accelerator, using small batches or fused kernels so several hardware resources stay busy together.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.3-draft", "warrant": { "literary": 4, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.graph-pipeline-parallelism", "label": "Graph pipeline parallelism", "aliases": [ "branch-parallel pipelining", "multi-branch pipeline scheduling" ], "definition": "Pipeline parallelism that partitions a model whose layers form a branching graph into stages that follow the graph, so independent branches, such as separate modality encoders, run concurrently in parallel stages instead of being serialized into one chain.", "background": false, "home": "S3.1", "home_id": "parallelism", "added_in": "0.3-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.rag-pipeline-scheduling", "label": "RAG pipeline scheduling", "aliases": [ "retrieval-generation co-scheduling" ], "definition": "Coordinating the retrieval and generation work of retrieval-augmented requests across requests and processors, overlapping and batching work across the stage boundary and optionally adapting retrieval settings per query, to meet latency and answer-quality targets.", "background": false, "home": "P4.8", "home_id": "retrieval-rag", "added_in": "0.3-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.cross-stack-profiling", "label": "Cross-stack profiling", "aliases": [ "framework-to-kernel attribution", "host-device trace correlation" ], "definition": "Profiling that correlates timestamped events and calling contexts from model code, frameworks, native libraries, and accelerator kernels, so that device time and stalls can be attributed to the operations that launched them.", "background": false, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.3-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 1, "user_proxy": 0 }, "book_passages": [ { "volume": "vol1", "chapter": "Model Serving", "url": "https://mlsysbook.ai/vol1/13_model_serving/13_model_serving.html#sec-model-serving-serving-tax-bill-dc6c" } ] }, { "id": "mlsys.model-swapping", "label": "Model swapping and residency", "aliases": [ "weight swapping", "model residency management" ], "definition": "Keeping some models' weights resident in accelerator memory and loading others from host memory or storage on demand, so a shared accelerator serves more models than fit at once, at the cost of transfer latency.", "background": false, "home": "P4.1", "home_id": "serving-architecture", "added_in": "0.3-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.kv-memory-allocation", "label": "KV-cache memory allocation", "aliases": [ "KV allocator", "KV-cache memory layout" ], "definition": "Laying out and allocating accelerator memory for cached attention state, which trades fragmentation and sharing against kernel compatibility and allocation cost.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.3-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.feed-forward-block", "label": "Feed-forward block", "aliases": [ "FFN", "MLP block", "position-wise feed-forward network", "feed-forward sublayer" ], "definition": "The per-position subnetwork of learned projections and a nonlinearity paired with attention in each Transformer block, which in most architectures holds the largest share of a dense model's parameters, executes as large matrix multiplications, and is the sublayer that mixture-of-experts routing replaces.", "background": false, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 10, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.service-level-agreement", "label": "Service-level agreement", "aliases": [ "SLA" ], "definition": "A contract binding a provider to stated service targets, with declared consequences such as credits or penalties when they are missed, whose targets are usually set looser than the internal objectives that trigger engineering response.", "background": true, "home": "F1.1", "home_id": "requirements-slos", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 12, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.head-of-line-blocking", "label": "Head-of-line blocking", "aliases": [ "HOL blocking" ], "definition": "Delay imposed on short or ready work because a queue serves in arrival order and a long or stalled item ahead holds the service point, so an item's wait depends on what is in front of it rather than on its own cost.", "background": true, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.open-closed-loop-load", "label": "Open-loop vs. closed-loop load", "aliases": [ "open-loop load generator", "closed-loop load generator", "open versus closed system model" ], "definition": "A distinction between arrivals scheduled independently of responses and arrivals triggered by prior completions, which changes offered-load feedback and the queueing behavior a load test exposes.", "background": true, "home": "F3.3", "home_id": "queueing-models", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.all-reduce", "label": "All-reduce", "aliases": [ "AllReduce" ], "definition": "A collective in which every participant ends with the same reduction, usually the sum, of all participants' equal-shaped tensors, the operation that synchronizes gradients across replicas in data-parallel training.", "background": true, "home": "S3.3", "home_id": "collectives", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 30, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.reduce-scatter-all-gather", "label": "Reduce-scatter and all-gather", "aliases": [ "ReduceScatter", "AllGather", "reduce-scatter", "all-gather" ], "definition": "The complementary collectives in which each participant receives one reduced shard of the group's tensors (reduce-scatter) or the concatenation of every participant's shard (all-gather), the pair that moves sharded parameters and gradients in fully sharded training and composes into a bandwidth-optimal all-reduce.", "background": true, "home": "S3.3", "home_id": "collectives", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 12, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.torus-topology", "label": "Torus topology", "aliases": [ "torus", "2D torus", "3D torus" ], "definition": "A fabric in which accelerator nodes form a multidimensional grid with wraparound links to fixed neighbors rather than tiers of switches, so neighbor and ring traffic can run at link rate while global exchanges are routed hop by hop with lower bisection bandwidth that depends on dimensions, routing, and contention.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.4-draft", "warrant": { "literary": 1, "corpus_chapters": 5, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.network-oversubscription", "label": "Network oversubscription", "aliases": [ "oversubscription ratio", "downlink-to-uplink ratio", "switch-tier oversubscription" ], "definition": "Provisioning less uplink capacity at a switch tier than the downlink capacity it serves, written as a downlink-to-uplink ratio, which cuts fabric cost but makes cross-tier bandwidth contend when many endpoints communicate at once, as synchronized collectives do.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.host-memory", "label": "Host memory tier", "aliases": [ "host DRAM", "system memory", "DDR memory", "CPU memory" ], "definition": "Memory directly managed by the host processor and used for input staging or offloaded model state, whose capacity and accelerator access costs depend on the platform's memory and interconnect design.", "background": true, "home": "S1.2", "home_id": "memory-systems", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 26, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.schedule-separation", "label": "Algorithm-schedule separation", "aliases": [ "scheduling language", "schedule primitives", "compute-schedule split", "Halide-style scheduling" ], "definition": "A compiler design in which what an operator computes is written once and how it runs, meaning loop order, tiling, vectorization, unrolling, memory placement, and parallel mapping, is a separate composable schedule, so the same algorithm can be retargeted and tuned without changing its semantics.", "background": true, "home": "S2.4", "home_id": "ml-compilers", "added_in": "0.4-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.kernel-synchronization", "label": "Kernel synchronization and atomics", "aliases": [ "thread-block barrier", "__syncthreads", "atomic operations", "atomic reduction", "warp synchronization" ], "definition": "The barriers, memory fences, and atomic read-modify-write operations through which threads of a kernel coordinate access to shared data, where barrier stalls and contended atomics serialize work that the thread hierarchy otherwise runs in parallel.", "background": true, "home": "S2.3", "home_id": "gpu-programming", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.link-reach", "label": "Link reach limit (copper versus optics)", "aliases": [ "copper reach", "electrical reach", "optical reach", "SerDes reach" ], "definition": "The maximum distance over which a link meets its signal-quality requirements for a specified medium and signaling scheme, constraining cabling, retiming, and the geometry of connected accelerator domains.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 5 }, "book_passages": [] }, { "id": "mlsys.storage-access-pattern", "label": "Sequential versus random storage access", "aliases": [ "storage access pattern", "sequential storage access", "random storage access" ], "definition": "Whether a workload reads storage in large contiguous ranges or in many small scattered requests, which decides whether sequential throughput or per-request latency and IOPS bound it, and so which media, layout, and shard sizes suit it.", "background": true, "home": "S4.3", "home_id": "storage-systems", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.dataset-shuffling", "label": "Dataset shuffling", "aliases": [ "shuffle buffer", "global shuffle", "epoch shuffle", "shard shuffle" ], "definition": "Randomizing the order in which training examples are visited, by a full permutation or an approximate buffered shuffle over sharded files, trading the randomness that stochastic optimization assumes against the sequential reads that storage delivers cheaply.", "background": false, "home": "S2.6", "home_id": "data-loading", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.overfitting", "label": "Overfitting", "aliases": [ "overfit" ], "definition": "Fitting training-specific variation that does not persist in new data, causing training performance to overstate generalization.", "background": true, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 16, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.predictive-uncertainty", "label": "Predictive uncertainty estimation", "aliases": [ "epistemic uncertainty", "aleatoric uncertainty", "deep ensembles", "Monte Carlo dropout", "uncertainty quantification" ], "definition": "An estimate of uncertainty in an individual prediction, sometimes decomposed into data noise and model uncertainty, used to support abstention, deferral, or data acquisition.", "background": true, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.statistical-power", "label": "Statistical power", "aliases": [ "power analysis", "sample size calculation", "sample size" ], "definition": "The probability that an experiment detects an effect of a stated size, fixed by sample size, outcome variance, and the significance level, which decides how much traffic or how many evaluation items a comparison needs before its answer can be trusted.", "background": true, "home": "P3.1", "home_id": "evaluation-protocols", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 29, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.retrieval-metrics", "label": "Retrieval quality metrics", "aliases": [ "recall@k", "precision@k", "nDCG", "mean reciprocal rank", "MRR", "retrieval hit rate" ], "definition": "Measures of a retriever's ranked results against relevance judgments, such as recall at k, mean reciprocal rank, and normalized discounted cumulative gain, which show the retrieval depth and ranking quality a downstream reranker or generator can rely on.", "background": false, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.rejection-sampling-fine-tuning", "label": "Rejection-sampling fine-tuning", "aliases": [ "STaR", "ReST", "filtered self-training" ], "definition": "Generating candidate responses, filtering them with an acceptance criterion, and fine-tuning on accepted responses to turn a selection signal into supervised training data, with repeated filtering potentially reducing diversity.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 86, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.early-stopping", "label": "Early stopping", "aliases": [ "validation-based stopping" ], "definition": "Ending training according to a held-out validation criterion and selecting a checkpoint under that criterion, limiting additional compute and potentially reducing overfitting.", "background": true, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.group-dro", "label": "Group distributionally robust optimization", "aliases": [ "group DRO", "worst-group loss minimization" ], "definition": "Training to minimize the highest expected loss among predefined groups so that poor subgroup performance is not hidden by average loss.", "background": true, "home": "P2.2", "home_id": "training-optimization", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.chat-template", "label": "Chat template", "aliases": [ "conversation formatting", "role markers", "prompt format", "turn delimiters" ], "definition": "The fixed scheme of role markers and special tokens that serializes system instructions, user and assistant turns, and tool calls into one token sequence, which must match between fine-tuning and serving or the model misreads turn boundaries.", "background": false, "home": "F2.5", "home_id": "model-artifacts", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.generation-stopping", "label": "Stop conditions and output-length limits", "aliases": [ "stop sequences", "max new tokens", "end-of-sequence handling" ], "definition": "Rules that terminate autoregressive generation when a stopping criterion is met, including an end token, a stop sequence, or an output-token limit.", "background": false, "home": "P4.6", "home_id": "structured-generation", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.runbook", "label": "Runbook", "aliases": [ "playbook", "on-call procedure", "operational procedure" ], "definition": "A written or automated procedure attached to an alert that says how to confirm, mitigate, and escalate a known failure, so that response speed does not depend on who is on call.", "background": true, "home": "P5.4", "home_id": "incident-response", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 11, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.rl-training-colocation", "label": "Colocated versus disaggregated RL post-training", "aliases": [ "actor-learner colocation", "hybrid RL engine", "disaggregated rollout and training" ], "definition": "Whether rollout generation and policy updates share the same accelerators, alternating phases and re-sharding weights in place, or run on separate pools that exchange weights and samples, trading idle time within each phase against transfer cost and policy staleness.", "background": false, "home": "P2.6", "home_id": "post-training", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.sequential-revision", "label": "Sequential revision", "aliases": [ "self-refine", "self-correction", "reflection", "iterative refinement" ], "definition": "Refining an answer through successive model calls conditioned on earlier drafts and internal or external feedback, spending additional inference compute on dependent revisions.", "background": false, "home": "P4.7", "home_id": "test-time-compute", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 21, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.semi-supervised-learning", "label": "Semi-supervised learning", "aliases": [ "pseudo-labeling", "self-training", "consistency regularization" ], "definition": "Learning from a small labeled set together with a large unlabeled one, typically by training on confident model predictions for unlabeled examples or by enforcing consistent predictions under perturbation, which lowers labeling cost but can reinforce the model's own early mistakes.", "background": true, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.contrastive-learning", "label": "Contrastive learning", "aliases": [ "InfoNCE", "SimCLR-style training", "CLIP-style training", "in-batch negatives" ], "definition": "A self-supervised or paired-data objective that pulls representations of matched views or pairs together and pushes mismatched ones apart, whose quality depends on how many negatives each step sees, which makes large batches and their memory a systems cost.", "background": true, "home": "P2.1", "home_id": "learning-paradigms", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.training-memory-estimate", "label": "Training memory estimate", "aliases": [ "bytes per parameter", "mixed-precision memory rule", "activation memory estimate" ], "definition": "Approximating a training step's accelerator memory as a per-parameter cost for weights, gradients, and optimizer state in their chosen precisions, plus activation memory that grows with batch size, sequence length, and depth, to decide what fits before a run is launched.", "background": false, "home": "P2.3", "home_id": "training-memory", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 30, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.privacy-budget", "label": "Privacy budget and accounting", "aliases": [ "privacy accountant", "epsilon-delta budget", "privacy loss composition", "moments accountant", "RDP accountant" ], "definition": "A declared bound on cumulative privacy loss across releases involving overlapping individuals, tracked through composition accounting to constrain further releases.", "background": true, "home": "Q3.3", "home_id": "privacy-dp", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.bias-sources", "label": "Sources of bias in the pipeline", "aliases": [ "historical bias", "representation bias", "measurement bias", "aggregation bias", "evaluation bias", "deployment bias" ], "definition": "The distinct points at which disparity enters an ML system, from the world the data records, who is sampled, how labels and features are measured, pooling heterogeneous groups into one model, an unrepresentative evaluation set, and use outside the intended context, each of which calls for a different remedy.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 4, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.carbon-accounting-method", "label": "Location-based vs. market-based carbon accounting", "aliases": [ "location-based Scope 2 accounting", "market-based Scope 2 accounting" ], "definition": "Accounting for purchased-electricity emissions using grid-average emission factors or qualifying contractual instruments, which changes the emissions attributed to the same electricity consumption.", "background": true, "home": "Q6.1", "home_id": "energy-carbon", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.request-timeout", "label": "Timeouts and deadline propagation", "aliases": [ "request deadline", "deadline propagation", "timeout budget" ], "definition": "A caller's time limit propagated to dependencies so they can stop work after the remaining budget expires, bounding caller wait and reducing abandoned work when cancellation is enforced.", "background": true, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.data-protection-regulation", "label": "Data protection regulation", "aliases": [ "GDPR", "CCPA", "privacy law" ], "definition": "Law governing the collection, use, retention, and automated use of personal data, which imposes obligations such as consent, purpose limitation, deletion, impact assessment, and rights over automated decisions that an ML system must implement in its data pipeline, models, and decision flow.", "background": true, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 8, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.weight-exfiltration", "label": "Model weight theft", "aliases": [ "weight exfiltration" ], "definition": "Theft of a model's parameter files from training, storage, or serving infrastructure by insiders or intruders, as distinct from reconstructing the model through its query interface, which the owner counters with access control, encryption at rest, attested execution, and egress limits.", "background": false, "home": "Q3.2", "home_id": "model-supply-chain", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 4 }, "book_passages": [] }, { "id": "mlsys.ai-bill-of-materials", "label": "AI bill of materials", "aliases": [ "AIBOM", "ML-BOM", "SBOM for ML", "model provenance manifest" ], "definition": "An inventory of the datasets, base weights, libraries, and build steps that produced a model, with versions and hashes, so consumers can check every component against vulnerability and trust lists before depending on the model.", "background": false, "home": "Q3.2", "home_id": "model-supply-chain", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.ai-incident-reporting", "label": "AI incident reporting", "aliases": [ "serious incident reporting", "AI incident database", "near-miss reporting" ], "definition": "Recording and disclosing harms and near-misses caused by a deployed AI system to an internal register or an external authority or database, as regulation or policy requires, so that patterns across deployments feed back into risk management.", "background": false, "home": "Q5.4", "home_id": "governance-regulation", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.mechanistic-interpretability", "label": "Mechanistic interpretability", "aliases": [ "probing classifiers", "activation patching", "sparse autoencoder features", "circuit analysis" ], "definition": "Investigating how a network's internal representations and computations produce behavior through analysis and interventions, to test explanations at the level of internal components.", "background": false, "home": "Q5.2", "home_id": "transparency-explainability", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.safety-certification", "label": "Safety certification", "aliases": [ "independent safety assessment", "type approval", "third-party conformity assessment" ], "definition": "An independent assessment that a system meets specified safety requirements for a stated use, providing evidence for deployment approval and review of relevant changes.", "background": true, "home": "Q4.4", "home_id": "safety-cases", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.tolerable-risk", "label": "Risk acceptance criteria", "aliases": [], "definition": "Criteria used to judge whether residual risk is acceptable or tolerable in a stated application and whether additional risk reduction is required.", "background": true, "home": "Q4.1", "home_id": "hazards-operating-domain", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.fleet-hardware-screening", "label": "Fleet hardware screening tests", "aliases": [ "proactive SDC testing", "fleet burn-in", "accelerator diagnostics", "known-answer tests" ], "definition": "Periodically running test workloads with known outputs on fleet accelerators and processors, at intake and during production, to find defective or silently miscomputing units before they corrupt jobs, at the cost of fleet time spent on tests.", "background": false, "home": "Q2.3", "home_id": "hardware-faults", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.hardware-performance-counters", "label": "Hardware performance counters", "aliases": [ "performance counters", "PMU events", "profiler metrics" ], "definition": "On-chip event counters, such as cycles, issued instructions, cache misses, bytes moved, and active compute units, that profilers sample to compute achieved bandwidth and arithmetic rate and to attribute stalls to their cause rather than to elapsed time alone.", "background": true, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.trajectory-log", "label": "Trajectory log", "aliases": [ "event-sourced trajectory log", "trajectory record", "agent event log" ], "definition": "An append-only harness-owned record of intents before dispatch and outcomes after execution, together with approvals and budget state, that supports recovery, replay, evaluation, and audit.", "background": false, "home": "T1.4", "home_id": "durable-execution", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 21, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.odometry", "label": "Odometry and dead reckoning", "aliases": [ "odometry", "dead reckoning", "wheel odometry", "visual odometry", "visual-inertial odometry", "inertial integration" ], "definition": "Estimating a platform's pose change by integrating its own motion measurements, from wheel encoders, inertial units, or feature motion in camera and lidar frames, whose error grows without bound with distance traveled unless corrected by loop closure or an external reference.", "background": true, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 11, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.capture-epoch", "label": "Capture timing and rolling shutter", "aliases": [ "rolling shutter", "global shutter", "capture timestamp", "capture epoch" ], "definition": "The time or interval over which a measurement is acquired, whose spread across rows or beams can require motion compensation before observations are fused.", "background": true, "home": "T2.1", "home_id": "perception-sensing", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 13, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.timing-jitter", "label": "Timing jitter", "aliases": [ "sampling jitter", "control-loop jitter", "period jitter", "input-output jitter" ], "definition": "Variation in the period or completion time of a periodic sensing, inference, or control task, which degrades a controller designed for a fixed sample interval even when every deadline is met, so real-time budgets bound the spread of completion times and not only their worst case.", "background": true, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 10, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.deterministic-field-network", "label": "Real-time fieldbus", "aliases": [ "EtherCAT", "time-sensitive networking", "TSN", "time-triggered bus", "scheduled CAN", "deterministic field network" ], "definition": "A device network engineered to provide bounded delivery under stated traffic and scheduling assumptions, whose cycle time and occupancy constrain sensing and control deadlines.", "background": true, "home": "T2.4", "home_id": "real-time-systems", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 12, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.loop-stability-margin", "label": "Loop stability margin", "aliases": [ "phase margin", "gain margin", "stability margin" ], "definition": "The allowed gain or phase variation before a specified feedback loop becomes unstable, used to budget delay and model uncertainty around an operating point.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 11, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.impedance-force-control", "label": "Impedance and force control", "aliases": [ "impedance control", "admittance control", "force control", "compliant control", "hybrid position-force control" ], "definition": "Feedback control that regulates the relationship between motion and contact force, by commanding a virtual stiffness and damping or by tracking a force setpoint, so a robot can touch, push, and insert without the force spikes that pure position control produces at contact.", "background": true, "home": "T2.3", "home_id": "planning-control", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 9, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.in-doubt-action", "label": "In-doubt action", "aliases": [ "unknown outcome", "in-doubt operation" ], "definition": "An external action whose outcome is unknown after dispatch, requiring reconciliation or retry-safe endpoint semantics before recovery chooses whether to repeat or skip it.", "background": true, "home": "Q2.1", "home_id": "failure-semantics", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 6, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.model-bandwidth-utilization", "label": "Model bandwidth utilization", "aliases": [ "MBU" ], "definition": "Model-required bytes of weights and KV cache per decoding step divided by step time and by the aggregate peak memory bandwidth of the participating devices, the memory-bound counterpart of model FLOPs utilization, stated for a given batch and cache configuration.", "background": false, "home": "Q1.1", "home_id": "profiling-utilization", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 2, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.hot-spare-capacity", "label": "Hot-spare capacity", "aliases": [ "spare nodes", "standby accelerators" ], "definition": "Idle, pre-provisioned accelerators or nodes reserved to replace failed workers, reducing the delay of acquiring replacement resources at the cost of reserved capacity.", "background": true, "home": "Q2.2", "home_id": "checkpoint-recovery", "added_in": "0.4-draft", "warrant": { "literary": 1, "corpus_chapters": 2, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.geo-distributed-training", "label": "Geo-distributed training", "aliases": [ "cross-datacenter training", "multi-datacenter training", "cross-region training", "WAN training" ], "definition": "Planning one training job across accelerators in several datacenters by placing the low-communication parallelism dimension (pipeline stages, local-SGD replicas, or asynchronous hierarchies) on the long, narrow inter-site links and the bandwidth-hungry dimensions inside each site, and scheduling the WAN transfers against its latency and bandwidth.", "background": false, "home": "S3.2", "home_id": "parallelism-planning", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.quarantined-model-pattern", "label": "Quarantined-model pattern", "aliases": [ "dual-LLM pattern", "privileged and quarantined LLM", "CaMeL", "quarantined reader" ], "definition": "Separating privileged planning from an unprivileged model that parses untrusted content, so that parsed values reach the planner only as references and a trusted runtime, often with provenance tracking, checks derived values before tool execution.", "background": false, "home": "Q3.4", "home_id": "access-isolation", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.sycophancy-evaluation", "label": "Sycophancy evaluation", "aliases": [ "sycophancy benchmark", "persuasion robustness evaluation", "rebuttal flip rate", "agreement-seeking evaluation" ], "definition": "Measuring how often a model abandons a correct answer or endorses a user's mistaken claim under pushback, authority cues, or social pressure, usually across several turns, so that agreement-seeking is measured separately from task accuracy.", "background": false, "home": "P3.4", "home_id": "generative-agent-eval", "added_in": "0.4-draft", "warrant": { "literary": 1, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.communication-intensity", "label": "Communication-to-computation ratio", "aliases": [ "bytes communicated per FLOP", "communication intensity" ], "definition": "The bytes a parallel plan must exchange per unit of local arithmetic, which, set against link bandwidth and the accelerator's FLOP rate, says whether a stage is communication-bound in the same way arithmetic intensity says whether a kernel is memory-bound.", "background": false, "home": "F3.1", "home_id": "roofline-locality", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.technology-scaling", "label": "Technology scaling", "aliases": [ "Moore's law", "Dennard scaling", "end of Dennard scaling", "transistor scaling" ], "definition": "The historical trends by which transistor density doubled on a steady cadence while, for a long period, power density stayed constant as transistors shrank; the end of constant power density is why added transistors can no longer all be clocked at once, pushing designs toward parallelism and specialization.", "background": true, "home": "S1.4", "home_id": "chip-power-thermal", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 8, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.compression-composition", "label": "Compression technique composition", "aliases": [ "compression pipeline", "pruning-quantization-distillation ordering", "combined compression" ], "definition": "Applying several compression methods in a chosen sequence, whose interactions can reinforce or diminish their individual savings and must be evaluated jointly for quality and execution cost.", "background": false, "home": "Q1.3", "home_id": "model-compression", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 1, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.fidelity-adaptive-serving", "label": "Fidelity-adaptive serving", "aliases": [ "accuracy scaling under load", "model switching under load", "quality-elastic serving", "runtime precision switching" ], "definition": "Meeting a latency target under rising load by switching requests to a cheaper variant of the same model, such as a lower-precision copy, a pruned subnetwork, or fewer layers, so capacity is bought with controlled quality loss instead of queueing or rejection.", "background": false, "home": "P4.4", "home_id": "request-scheduling", "added_in": "0.4-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.compound-workflow-pipelining", "label": "Compound workflow pipelining", "aliases": [ "cross-stage pipelining", "inter-agent pipelining", "streaming prefill overlap", "LLM application DAG scheduling" ], "definition": "Overlapping stages of a compound ML application by streaming intermediate results to downstream stages as their dependencies become available, reducing critical-path latency while preserving dependency order.", "background": false, "home": "P4.2", "home_id": "compound-ai", "added_in": "0.4-draft", "warrant": { "literary": 2, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.reachability-analysis", "label": "Reachability analysis", "aliases": [ "forward reachable set", "backward reachable set", "Hamilton-Jacobi reachability", "reachable occupancy" ], "definition": "Computing the set of states a system, or another agent in its workspace, can reach within a time horizon under its dynamics and bounded inputs or disturbances, used to prove a safe set invariant, to size clearance margins, and to screen proposed motions against where other agents could be.", "background": true, "home": "Q4.2", "home_id": "runtime-enforcement", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.fairness-monitoring", "label": "Fairness monitoring", "aliases": [ "production fairness monitoring", "subgroup performance monitoring", "post-market bias monitoring" ], "definition": "Tracking group fairness metrics on production traffic over time, with the sampling, delayed-label, and protected-attribute availability problems that make subgroup estimates noisier than aggregate ones, so that disparities introduced by drift or retraining are caught before they accumulate.", "background": false, "home": "Q5.1", "home_id": "fairness-bias", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.link-signal-integrity", "label": "Link signal degradation and flaps", "aliases": [ "link flap", "link retraining", "marginal link", "signal-integrity failure" ], "definition": "Physical-link degradation that produces excessive bit errors, retraining, or intermittent disconnections, causing retransmission and communication stalls when protection mechanisms cannot absorb it.", "background": true, "home": "S4.2", "home_id": "network-fabrics", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 1, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.activation-function", "label": "Activation function", "aliases": [ "nonlinearity", "ReLU", "GELU", "SwiGLU", "gated linear unit" ], "definition": "The elementwise or gated nonlinearity applied between a network's linear layers, whose choice sets whether activations contain exact zeros, whether a block needs an extra gating projection, and how cheaply the operation fuses into the preceding matrix multiply.", "background": true, "home": "F2.2", "home_id": "architecture-blocks", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 18, "organizational": 0, "user_proxy": 1 }, "book_passages": [] }, { "id": "mlsys.attention-ffn-disaggregation", "label": "Attention-FFN disaggregation", "aliases": [ "AFD", "attention-expert disaggregation", "ping-pong pipelined decoding", "disaggregated expert parallelism" ], "definition": "Running attention layers and the feed-forward or expert layers on separate accelerator pools during decode, so memory-bound attention over the KV cache and compute-bound expert matmuls are each batched and provisioned on hardware suited to them, at the cost of streaming activations between pools at every layer.", "background": false, "home": "P4.5", "home_id": "llm-inference", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.label-shift-estimation", "label": "Label-shift estimation and prior correction", "aliases": [ "BBSE", "black-box shift estimation", "prior shift correction", "EM prior adjustment", "class-prior re-estimation", "class-prior adaptation" ], "definition": "Estimating deployment class priors from unlabeled predictions under stable class-conditional input distributions and adjusting weights or probabilities to reduce errors from the changed priors.", "background": false, "home": "P5.3", "home_id": "drift-retraining", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.map-representation", "label": "Map representation", "aliases": [ "occupancy grid", "signed distance field", "costmap", "semantic map", "metric map", "TSDF", "ESDF", "landmark map", "neural scene map" ], "definition": "The data structure a robot keeps of its environment, such as an occupancy grid, signed-distance field, point cloud, mesh, or semantic layer, chosen by the queries that planning and collision checking must answer at their rates and by the memory and update cost the onboard computer can afford.", "background": true, "home": "T2.2", "home_id": "state-estimation", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 7, "organizational": 1, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.compute-cost-estimate", "label": "Compute cost estimate", "aliases": [ "training run cost", "training run cost estimation", "compute-run cost estimation", "accelerator-hour estimation" ], "definition": "Estimating accelerator time from workload demand and effective hardware performance, then applying resource prices to estimate monetary compute cost under a stated deployment plan.", "background": false, "home": "Q6.3", "home_id": "cost-economics", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 3, "organizational": 0, "user_proxy": 3 }, "book_passages": [] }, { "id": "mlsys.precision-recall", "label": "Precision, recall, and ROC analysis", "aliases": [ "F1 score", "ROC curve", "AUC", "precision-recall curve" ], "definition": "Classifier metrics that separate the two error types, false positives and false negatives, and the curves (ROC, precision-recall) that show how a decision threshold trades them, which matter whenever classes are imbalanced or the two errors cost differently.", "background": true, "home": "P3.2", "home_id": "metrics-uncertainty", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 5, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.instruction-hierarchy", "label": "Instruction privilege levels", "aliases": [ "instruction hierarchy", "system-prompt precedence" ], "definition": "Giving instructions from the system, the user, and retrieved or tool-returned content distinct privilege levels, through prompt formatting, delimiting, or training, so that lower-privilege content is less able to override higher-privilege instructions.", "background": false, "home": "Q3.1", "home_id": "ml-attacks", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 0 }, "book_passages": [] }, { "id": "mlsys.restart-free-recovery", "label": "Restart-free failure recovery", "aliases": [ "recovery without restart" ], "definition": "Reconfiguring distributed training around surviving replicas of model and optimizer state so that the job resumes after supported worker failures without a full-job restart or durable-checkpoint reload, reducing recovery delay at the cost of redundancy and reconfiguration.", "background": false, "home": "Q2.2", "home_id": "checkpoint-recovery", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] }, { "id": "mlsys.shared-memory-bank-conflict", "label": "Shared-memory bank conflict", "aliases": [], "definition": "Serialization of a shared-memory request when threads access different addresses in the same bank, reducing effective shared-memory bandwidth.", "background": false, "home": "S2.3", "home_id": "gpu-programming", "added_in": "0.4-draft", "warrant": { "literary": 0, "corpus_chapters": 0, "organizational": 0, "user_proxy": 2 }, "book_passages": [] } ], "gaps": [ { "id": "gap-emerging-compute-substrates", "topic": "Photonic compute", "subject": "S1.1", "status": "candidate", "reason": "Reviewed 2026-11: photonic arithmetic still has no production ML lineage and no 2025 or 2026 systems-venue paper reports one; compute-in-memory, wafer-scale execution, FPGAs, and advanced packaging cover the admitted substrates." }, { "id": "gap-verifiable-ml-computation", "topic": "Verifiable ML computation (zero-knowledge proofs of training, inference, or unlearning)", "subject": "Q3.2", "status": "candidate", "reason": "Zero-knowledge proofs of ML training, inference, or unlearning have two venue lineages (ZK-APEX, MLSys 2026; zkLLM, CCS 2024) but are expected by no course, standard, or role yet; a proof that a computation ran is also not a proof of an unlearning guarantee." }, { "id": "gap-decoded-video-reuse", "topic": "Decoded-video view reuse", "subject": "S2.6", "status": "candidate", "reason": "Reviewed 2026-11: one venue lineage (SAND, SOSP 2025) and two 2026 preprint production reports describe caching decoded frames across epochs and jobs; the admission test asks for two independent venue, course, or standard sources, so it stays a candidate until a second one appears." }, { "id": "gap-reasoning-trace-monitoring", "topic": "Reasoning-trace monitoring", "subject": "Q4.2", "status": "candidate", "reason": "Reviewed 2026-11: monitoring a model's intermediate reasoning for intent to subvert a task is described by frontier-lab reports and a 2025 multi-lab position paper, but no course, standard, or role yet expects it; it stays a candidate until that evidence exists." } ], "relations": [ { "id": "specializes", "forward": "is a kind of", "inverse": "has kinds such as", "kind": "composition", "meaning": "A is a narrower or more specific case of B." }, { "id": "part-of", "forward": "is part of", "inverse": "includes", "kind": "composition", "meaning": "A is a component or stage within B." }, { "id": "alternative-to", "forward": "is an alternative to", "inverse": "is an alternative to", "kind": "tradeoff", "meaning": "A and B are competing options for the same job (symmetric)." }, { "id": "requires", "forward": "requires", "inverse": "is required by", "kind": "dependency", "meaning": "A cannot work, or cannot be correct, without B." }, { "id": "uses", "forward": "uses", "inverse": "is used by", "kind": "dependency", "meaning": "A commonly relies on B as a building block, without strict necessity." }, { "id": "implements", "forward": "implements", "inverse": "is implemented by", "kind": "mechanism", "meaning": "A is a concrete mechanism or technique that realizes B." }, { "id": "enables", "forward": "enables", "inverse": "is enabled by", "kind": "mechanism", "meaning": "A makes B possible or practical." }, { "id": "produces", "forward": "produces", "inverse": "is produced by", "kind": "mechanism", "meaning": "A creates or emits B as an output or artifact." }, { "id": "reduces", "forward": "reduces", "inverse": "is reduced by", "kind": "mechanism", "meaning": "A lowers the amount, cost, or magnitude of B." }, { "id": "increases", "forward": "increases", "inverse": "is increased by", "kind": "mechanism", "meaning": "A raises the amount, cost, or magnitude of B." }, { "id": "determines", "forward": "determines", "inverse": "is determined by", "kind": "mechanism", "meaning": "A sets or shapes B (direction not simply up/down)." }, { "id": "limits", "forward": "limits", "inverse": "is limited by", "kind": "constraint", "meaning": "A places a limit or ceiling on B." }, { "id": "trades-off", "forward": "trades off against", "inverse": "trades off against", "kind": "tradeoff", "meaning": "Improving A tends to worsen B, or vice versa (symmetric)." }, { "id": "measures", "forward": "measures", "inverse": "is measured by", "kind": "evaluation", "meaning": "A quantifies, evaluates, or tests B." }, { "id": "detects", "forward": "detects", "inverse": "is detected by", "kind": "evaluation", "meaning": "A reveals or identifies occurrences of B." }, { "id": "causes", "forward": "can cause", "inverse": "can be caused by", "kind": "failure", "meaning": "A can lead to the failure, harm, or degradation B." }, { "id": "mitigates", "forward": "mitigates", "inverse": "is mitigated by", "kind": "failure", "meaning": "A protects against, prevents, or recovers from B." } ], "relationships": [ { "source": "mlsys.requirements", "relation": "determines", "target": "mlsys.metrics", "statement": "sets acceptance thresholds for", "conditions": "The operational outcome, evaluation population, and threshold must be declared." }, { "source": "mlsys.requirements", "relation": "determines", "target": "mlsys.slo", "statement": "sets targets for", "conditions": "Service targets are part of the stated deployment requirements." }, { "source": "mlsys.baseline", "relation": "part-of", "target": "mlsys.benchmark", "statement": "is the reference point within", "conditions": "The baseline must be measured using the same protocol." }, { "source": "mlsys.proxy-objective", "relation": "measures", "target": "mlsys.requirements", "statement": "stands in for, and can diverge from,", "conditions": "Improving a measurable proxy need not improve the intended outcome." }, { "source": "mlsys.workload", "relation": "determines", "target": "mlsys.capacity-planning", "statement": "determines resource demand for", "conditions": "Sizing must include input and request distributions, not just a representative average." }, { "source": "mlsys.placement", "relation": "determines", "target": "mlsys.tail-latency", "statement": "changes attainable", "conditions": "Communication paths, locality, and available service resources depend on placement." }, { "source": "mlsys.data-quality", "relation": "limits", "target": "mlsys.generalization", "statement": "bounds achievable", "conditions": "The effect depends on task, coverage, errors, and the learning procedure." }, { "source": "mlsys.labels", "relation": "determines", "target": "mlsys.loss", "statement": "determines the reliability of the supervision in", "conditions": "For supervised learning; label errors change the optimization signal." }, { "source": "mlsys.missing-data", "relation": "determines", "target": "mlsys.features", "statement": "forces imputation and indicator decisions in", "conditions": "Treatment must preserve the meaning and availability of inputs." }, { "source": "mlsys.features", "relation": "requires", "target": "mlsys.point-in-time", "statement": "must preserve", "conditions": "For predictions whose inputs become available over time." }, { "source": "mlsys.leakage", "relation": "determines", "target": "mlsys.metrics", "statement": "can invalidate", "conditions": "When evaluation uses information unavailable at the intended prediction boundary." }, { "source": "mlsys.point-in-time", "relation": "mitigates", "target": "mlsys.leakage", "statement": "prevents", "conditions": "Only when source availability and temporal joins are modeled correctly." }, { "source": "mlsys.artifact-version", "relation": "enables", "target": "mlsys.lineage", "statement": "records the dependencies behind", "conditions": "Provenance must identify versions and transformations sufficient for reconstruction." }, { "source": "mlsys.train-serve-skew", "relation": "reduces", "target": "mlsys.generalization", "statement": "can degrade deployed", "conditions": "A deployed pipeline may compute different inputs from those used in development." }, { "source": "mlsys.deduplication", "relation": "part-of", "target": "mlsys.data-selection", "statement": "is a curation stage of", "conditions": "Removing duplicates changes sample weights and must preserve required coverage." }, { "source": "mlsys.data-selection", "relation": "determines", "target": "mlsys.generalization", "statement": "shapes the data supporting", "conditions": "Selection can reduce cost while changing coverage and distribution." }, { "source": "mlsys.tokenization", "relation": "determines", "target": "mlsys.attention", "statement": "sets the sequence length processed by", "conditions": "For a fixed text, changing vocabulary and segmentation can change sequence length." }, { "source": "mlsys.tokenization", "relation": "determines", "target": "mlsys.embeddings", "statement": "defines the lookup identities of", "conditions": "Embedding rows must correspond to the model's tokenizer and vocabulary." }, { "source": "mlsys.distribution-shift", "relation": "reduces", "target": "mlsys.calibration", "statement": "can degrade", "conditions": "Calibration measured on one distribution need not hold after a shift." }, { "source": "mlsys.retraining", "relation": "mitigates", "target": "mlsys.distribution-shift", "statement": "mitigates, when triggered by,", "conditions": "A response requires diagnosis and evidence; not every detected shift warrants retraining." }, { "source": "mlsys.linear-model", "relation": "uses", "target": "mlsys.features", "statement": "operates on", "conditions": "A linear score can use engineered or learned input features." }, { "source": "mlsys.nearest-neighbor", "relation": "uses", "target": "mlsys.ann", "statement": "can use", "conditions": "Approximation trades exact recovery for search cost in compatible applications." }, { "source": "mlsys.neural-network", "relation": "uses", "target": "mlsys.autodiff", "statement": "is trained using", "conditions": "For gradient-based learning with supported differentiable objectives and derivative rules." }, { "source": "mlsys.inductive-bias", "relation": "determines", "target": "mlsys.generalization", "statement": "changes data needed for", "conditions": "Architecture and task alignment affect which patterns can be learned efficiently." }, { "source": "mlsys.embeddings", "relation": "requires", "target": "mlsys.memory-capacity", "statement": "must be resident within", "conditions": "Residency depends on table dimensions, format, and placement or sharding." }, { "source": "mlsys.attention", "relation": "uses", "target": "mlsys.position-encoding", "statement": "gains sequence-order sensitivity from", "conditions": "Attention without another position-sensitive mechanism is permutation equivariant." }, { "source": "mlsys.attention", "relation": "uses", "target": "mlsys.kv-cache", "statement": "reuses history through", "conditions": "For compatible autoregressive attention architectures." }, { "source": "mlsys.optimizer", "relation": "requires", "target": "mlsys.loss", "statement": "requires an objective from", "conditions": "Optimizing the training objective alone does not certify task success." }, { "source": "mlsys.autodiff", "relation": "enables", "target": "mlsys.optimizer", "statement": "produces the gradients consumed by", "conditions": "Optimizer state and update rules are separate from differentiation." }, { "source": "mlsys.regularization", "relation": "determines", "target": "mlsys.generalization", "statement": "shapes", "conditions": "Effects depend on the data, model, objective, and optimization regime." }, { "source": "mlsys.batch-size", "relation": "determines", "target": "mlsys.optimizer", "statement": "sets gradient noise and step count for", "conditions": "Changing the batch can alter gradient noise and statistical efficiency." }, { "source": "mlsys.memory-capacity", "relation": "limits", "target": "mlsys.batch-size", "statement": "limits the per-device", "conditions": "Activation and workspace demand depend on execution and accumulation choices." }, { "source": "mlsys.optimizer", "relation": "increases", "target": "mlsys.training-state", "statement": "adds optimizer variables to", "conditions": "Different update rules retain different auxiliary state." }, { "source": "mlsys.memory-capacity", "relation": "limits", "target": "mlsys.training-state", "statement": "limits the per-device", "conditions": "Count resident parameters, gradients, optimizer state, and other retained data separately." }, { "source": "mlsys.hyperparameter-search", "relation": "causes", "target": "mlsys.selection-bias", "statement": "can induce", "conditions": "Selecting among noisy validation scores can produce an optimistic winner." }, { "source": "mlsys.peft", "relation": "implements", "target": "mlsys.transfer-learning", "statement": "implements", "conditions": "Parameter-efficient adaptation is one option, not all transfer learning." }, { "source": "mlsys.peft", "relation": "reduces", "target": "mlsys.training-state", "statement": "shrinks trainable and optimizer state in", "conditions": "The method may select existing parameters or introduce new parameterizations; frozen parameters and activations can still remain resident." }, { "source": "mlsys.tensor-layout", "relation": "determines", "target": "mlsys.matrix-multiply", "statement": "sets the data-movement cost of", "conditions": "Views and physical copies have different costs; kernels support particular layouts." }, { "source": "mlsys.precision", "relation": "determines", "target": "mlsys.training-state", "statement": "sets storage width within", "conditions": "Different state components can use different formats." }, { "source": "mlsys.precision", "relation": "determines", "target": "mlsys.numerical-stability", "statement": "affects", "conditions": "Range, rounding, accumulation, and intermediate representation all matter." }, { "source": "mlsys.mixed-precision", "relation": "requires", "target": "mlsys.numerical-stability", "statement": "requires managing", "conditions": "Loss scaling or wider accumulation is useful only for the relevant numerical regime." }, { "source": "mlsys.quantization", "relation": "reduces", "target": "mlsys.precision", "statement": "lowers the numerical precision of", "conditions": "The encoding, scale, and rounding rule determine the approximation." }, { "source": "mlsys.quantization", "relation": "requires", "target": "mlsys.metrics", "statement": "requires re-evaluation of", "conditions": "Quality and calibration must be evaluated for the resulting deployed system." }, { "source": "mlsys.distillation", "relation": "produces", "target": "mlsys.neural-network", "statement": "produces a smaller student", "conditions": "The student architecture and task evidence determine the actual trade-off." }, { "source": "mlsys.compilation", "relation": "enables", "target": "mlsys.fusion", "statement": "can apply", "conditions": "Only when semantics and execution dependencies permit combination." }, { "source": "mlsys.fusion", "relation": "reduces", "target": "mlsys.dispatch", "statement": "can reduce", "conditions": "Fewer launches can reduce host work, depending on the original execution boundary." }, { "source": "mlsys.memory-bandwidth", "relation": "determines", "target": "mlsys.roofline", "statement": "sets the sloped ceiling of", "conditions": "The bandwidth must correspond to the memory boundary used for arithmetic intensity." }, { "source": "mlsys.compute-throughput", "relation": "determines", "target": "mlsys.roofline", "statement": "sets the flat ceiling of", "conditions": "The ceiling must match the precision and operation mix of the workload." }, { "source": "mlsys.arithmetic-intensity", "relation": "determines", "target": "mlsys.roofline", "statement": "selects the active bound in", "conditions": "The compute and bandwidth ceilings describe the declared operating model." }, { "source": "mlsys.tiling", "relation": "increases", "target": "mlsys.locality", "statement": "improves", "conditions": "Tile size, storage capacity, and execution support constrain useful reuse." }, { "source": "mlsys.locality", "relation": "increases", "target": "mlsys.arithmetic-intensity", "statement": "can increase", "conditions": "When reuse reduces transfers across the measured memory boundary." }, { "source": "mlsys.profiling", "relation": "measures", "target": "mlsys.utilization", "statement": "estimates realized", "conditions": "Instrumentation and measurement boundaries affect interpretation." }, { "source": "mlsys.data-parallel", "relation": "uses", "target": "mlsys.collectives", "statement": "uses", "conditions": "Synchronous replicas commonly aggregate gradients; other learning semantics differ." }, { "source": "mlsys.tensor-parallel", "relation": "requires", "target": "mlsys.collectives", "statement": "requires per-layer", "conditions": "Partitioned operations require data exchange according to their decomposition." }, { "source": "mlsys.stragglers", "relation": "limits", "target": "mlsys.pipeline-parallel", "statement": "stall", "conditions": "Dependencies and pipeline bubbles expose unequal stage or worker times." }, { "source": "mlsys.state-sharding", "relation": "reduces", "target": "mlsys.training-state", "statement": "reduces per-worker copies of", "conditions": "Replication, communication, and reconstruction depend on the sharding stage." }, { "source": "mlsys.communication-overlap", "relation": "reduces", "target": "mlsys.collectives", "statement": "hides some exposed cost of", "conditions": "Only independent work overlaps; contention or dependencies can expose residual communication." }, { "source": "mlsys.network-topology", "relation": "limits", "target": "mlsys.collectives", "statement": "constrains achievable rates of", "conditions": "Available paths and contention determine achievable exchange rates." }, { "source": "mlsys.training-state", "relation": "part-of", "target": "mlsys.checkpoint", "statement": "is persisted, in selected form, in", "conditions": "Resumption semantics determine which state must be saved." }, { "source": "mlsys.checkpoint", "relation": "requires", "target": "mlsys.consistent-cut", "statement": "requires, for distributed jobs, a", "conditions": "The saved worker states and dependencies must form a recoverable state." }, { "source": "mlsys.consensus", "relation": "enables", "target": "mlsys.fencing", "statement": "maintains the epoch authority used by", "conditions": "Resource-enforced epochs remain necessary; a consensus log alone does not fence effects." }, { "source": "mlsys.federated-learning", "relation": "requires", "target": "mlsys.privacy", "statement": "must address", "conditions": "Keeping raw records local alone does not establish privacy of updates or outputs." }, { "source": "mlsys.service-capacity", "relation": "limits", "target": "mlsys.arrival-rate", "statement": "bounds the sustainable", "conditions": "The request mix and service-time distribution determine demand." }, { "source": "mlsys.utilization", "relation": "increases", "target": "mlsys.queueing", "statement": "increases, near saturation,", "conditions": "With fixed service conditions, waits typically rise near saturation; this is not a universal scalar model." }, { "source": "mlsys.queueing", "relation": "increases", "target": "mlsys.tail-latency", "statement": "contributes to", "conditions": "Service, batching, network, and other delays also contribute." }, { "source": "mlsys.batching", "relation": "increases", "target": "mlsys.service-capacity", "statement": "can increase", "conditions": "Workload, memory, and kernel efficiency bound the gain." }, { "source": "mlsys.batching", "relation": "increases", "target": "mlsys.tail-latency", "statement": "can increase", "conditions": "Waiting to form a batch can add delay; the net effect depends on queueing and service changes." }, { "source": "mlsys.admission", "relation": "requires", "target": "mlsys.service-capacity", "statement": "admits work against a known", "conditions": "Limits must be expressed in resources relevant to the workload." }, { "source": "mlsys.autoscaling", "relation": "determines", "target": "mlsys.service-capacity", "statement": "adjusts provisioned", "conditions": "Provisioning delay, state transfer, and warm-up affect when capacity becomes useful." }, { "source": "mlsys.memory-capacity", "relation": "limits", "target": "mlsys.kv-cache", "statement": "limits the resident", "conditions": "Size depends on architecture, sequence length, format, and concurrency." }, { "source": "mlsys.paged-kv", "relation": "implements", "target": "mlsys.kv-cache", "statement": "stores, in fixed-size blocks mapped by a block table, the", "conditions": "When a block table maps each request's positions to retained key-value blocks that attention can retrieve." }, { "source": "mlsys.prefix-cache", "relation": "uses", "target": "mlsys.kv-cache", "statement": "reuses compatible", "conditions": "Token prefix, model state, position conventions, and access boundaries must match." }, { "source": "mlsys.memory-bandwidth", "relation": "limits", "target": "mlsys.prefill-decode", "statement": "bounds the decode phase of", "conditions": "The dominant resource depends on architecture, batch, context length, and implementation." }, { "source": "mlsys.speculative-decoding", "relation": "increases", "target": "mlsys.service-capacity", "statement": "can raise effective decode throughput of", "conditions": "Acceptance rate, proposer cost, verification, and workload determine the net gain." }, { "source": "mlsys.slo", "relation": "determines", "target": "mlsys.slo-goodput", "statement": "defines the counting rule for", "conditions": "The counting rule must specify latency, quality, failures, and the workload population." }, { "source": "mlsys.evaluation-split", "relation": "measures", "target": "mlsys.generalization", "statement": "supports unbiased estimates of", "conditions": "The partition and sampling unit must match the intended deployment question." }, { "source": "mlsys.selection-bias", "relation": "increases", "target": "mlsys.metrics", "statement": "inflates reported", "conditions": "Independent final evaluation is needed after selection on noisy scores." }, { "source": "mlsys.calibration", "relation": "enables", "target": "mlsys.selective-prediction", "statement": "supports", "conditions": "Abstention thresholds still require risk and coverage evaluation under the deployment distribution." }, { "source": "mlsys.statistical-uncertainty", "relation": "limits", "target": "mlsys.metrics", "statement": "bounds the precision of conclusions from", "conditions": "Dependence, sample selection, and repeated comparison matter." }, { "source": "mlsys.benchmark", "relation": "measures", "target": "mlsys.tail-latency", "statement": "specifies how to measure", "conditions": "Measurement needs a declared load model, warm-up, boundary, and quantile method." }, { "source": "mlsys.online-experiment", "relation": "measures", "target": "mlsys.metrics", "statement": "estimates causal effects on", "conditions": "Causal interpretation depends on assignment, interference, observation, and stopping assumptions." }, { "source": "mlsys.sliced-evaluation", "relation": "measures", "target": "mlsys.fairness", "statement": "tests specified", "conditions": "An aggregate score cannot establish performance for every subgroup." }, { "source": "mlsys.rollout", "relation": "requires", "target": "mlsys.artifact-version", "statement": "requires pinned", "conditions": "Evidence and compatibility must refer to the exact candidate being deployed." }, { "source": "mlsys.monitoring", "relation": "detects", "target": "mlsys.distribution-shift", "statement": "observes signs of", "conditions": "Signals require diagnosis and do not establish harmfulness by themselves." }, { "source": "mlsys.rollout", "relation": "uses", "target": "mlsys.rollback", "statement": "can use", "conditions": "A rollout can use rollback when acceptance criteria fail; state compatibility and external effects may limit recovery." }, { "source": "mlsys.feedback-loop", "relation": "reduces", "target": "mlsys.data-quality", "statement": "can degrade", "conditions": "Deployed decisions can alter which labels and examples are subsequently observed." }, { "source": "mlsys.observability", "relation": "enables", "target": "mlsys.incident-response", "statement": "supports", "conditions": "Useful diagnosis requires recorded context and trustworthy instrumentation." }, { "source": "mlsys.threat-model", "relation": "determines", "target": "mlsys.poisoning", "statement": "scopes the attack surfaces considered for", "conditions": "The adversary's ability to modify training inputs or signals must be specified." }, { "source": "mlsys.threat-model", "relation": "determines", "target": "mlsys.evasion", "statement": "scopes the attack surfaces considered for", "conditions": "The adversary's knowledge, perturbations, and target behavior must be specified." }, { "source": "mlsys.differential-privacy", "relation": "implements", "target": "mlsys.privacy", "statement": "provides a formal guarantee for", "conditions": "The guarantee bounds output-distribution change under a declared neighboring relation and accounting parameters; it does not establish every privacy requirement." }, { "source": "mlsys.human-oversight", "relation": "requires", "target": "mlsys.observability", "statement": "requires actionable", "conditions": "Recorded signals must be interpretable and available in time to support intervention." }, { "source": "mlsys.energy", "relation": "increases", "target": "mlsys.carbon", "statement": "contributes operational emissions to", "conditions": "Conversion depends on operational accounting boundaries and energy-source emissions; embodied emissions are separate." }, { "source": "mlsys.unit-cost", "relation": "requires", "target": "mlsys.metrics", "statement": "requires an acceptance criterion from", "conditions": "Counting attempted work as successful outcomes understates the unit cost." }, { "source": "mlsys.capacity-planning", "relation": "requires", "target": "mlsys.slo", "statement": "is sized against", "conditions": "Provisioning decisions must account for demand variation and resource constraints." }, { "source": "mlsys.context", "relation": "limits", "target": "mlsys.trajectory", "statement": "bounds the history carried across an", "conditions": "Only selected observations and records fit in a model invocation." }, { "source": "mlsys.planning", "relation": "determines", "target": "mlsys.trajectory", "statement": "revises the actions within", "conditions": "Execution observations and changing constraints can invalidate a prospective plan." }, { "source": "mlsys.context", "relation": "limits", "target": "mlsys.retrieval", "statement": "bounds the passage volume injected by", "conditions": "Relevance, authorization, provenance, and token budget constrain what may be supplied." }, { "source": "mlsys.retrieval", "relation": "uses", "target": "mlsys.ann", "statement": "can use", "conditions": "Approximation and index freshness affect retrieval quality and cost." }, { "source": "mlsys.trajectory", "relation": "uses", "target": "mlsys.compensation", "statement": "can use, to remedy earlier steps,", "conditions": "A remedy may not restore the original state or undo all consequences." }, { "source": "mlsys.authorization", "relation": "mitigates", "target": "mlsys.prompt-injection", "statement": "blocks unauthorized actions triggered by", "conditions": "Untrusted content must not silently acquire authority through model interpretation." }, { "source": "mlsys.agent-evaluation", "relation": "measures", "target": "mlsys.trajectory", "statement": "measures accepted outcomes of", "conditions": "Evaluate external effects and constraint compliance, not only final text." }, { "source": "mlsys.proxy-objective", "relation": "causes", "target": "mlsys.reward-hacking", "statement": "can invite", "conditions": "A reward is not itself evidence that the intended task was performed." }, { "source": "mlsys.dynamics", "relation": "limits", "target": "mlsys.trajectory-feasibility", "statement": "constrains", "conditions": "Physical states, inputs, disturbances, and constraints determine admissible motion." }, { "source": "mlsys.sensing", "relation": "enables", "target": "mlsys.state-estimation", "statement": "supplies the measurement model and frame alignment for", "conditions": "When the estimator's measurement models and frame and time alignment are established by sensor calibration." }, { "source": "mlsys.state-observability", "relation": "limits", "target": "mlsys.state-estimation", "statement": "limits", "conditions": "A low reported covariance does not by itself establish that the state is distinguishable." }, { "source": "mlsys.observation-age", "relation": "determines", "target": "mlsys.state-estimation", "statement": "inflates the uncertainty of", "conditions": "State can evolve between capture and use; propagation depends on the plant and disturbances." }, { "source": "mlsys.feedback-delay", "relation": "limits", "target": "mlsys.feedback-control", "statement": "limits the stability margin of", "conditions": "Delay can reduce stability margins; the effect depends on sampling and dynamics." }, { "source": "mlsys.wcet", "relation": "limits", "target": "mlsys.feedback-delay", "statement": "contributes to bounding", "conditions": "Execution-time bounds alone do not include all queueing, scheduling, sensing, and transport delays." }, { "source": "mlsys.contact-compliance", "relation": "determines", "target": "mlsys.dynamics", "statement": "changes", "conditions": "Contact modes, stiffness, damping, and impact alter physical response." }, { "source": "mlsys.sim-to-real", "relation": "requires", "target": "mlsys.dynamics", "statement": "requires assessing simulator mismatch in", "conditions": "Simulation evidence must be evaluated against the actual deployment conditions." }, { "source": "mlsys.runtime-assurance", "relation": "uses", "target": "mlsys.trajectory-feasibility", "statement": "checks proposed actions against", "conditions": "When the monitor's feasibility model, timing, and implementation assumptions hold; the guarantee is only as good as that model." }, { "source": "mlsys.feedback-delay", "relation": "reduces", "target": "mlsys.stopping-envelope", "statement": "shrinks", "conditions": "Delay adds motion before braking; braking authority and uncertainty also matter." }, { "source": "mlsys.operating-domain", "relation": "limits", "target": "mlsys.safety-case", "statement": "bounds the claims in", "conditions": "Evidence must match the intended operating conditions and configuration." }, { "source": "mlsys.safety-case", "relation": "uses", "target": "mlsys.runtime-assurance", "statement": "argues over the premises and fallbacks of", "conditions": "When runtime monitoring and fallback are part of the safety argument rather than an unargued numerical filter." }, { "source": "mlsys.isolation", "relation": "implements", "target": "mlsys.authorization", "statement": "enforces the access boundaries of", "conditions": "Isolation can restrict resource access, but it does not choose the authorization policy or prove external-effect correctness." }, { "source": "mlsys.data-acquisition", "relation": "determines", "target": "mlsys.evaluation-split", "statement": "determines the observed population for", "conditions": "Selection and missing outcomes can change the population to which an estimate applies." }, { "source": "mlsys.data-contract", "relation": "determines", "target": "mlsys.data-quality", "statement": "specifies the validity criteria for", "conditions": "A schema check alone cannot establish semantic correctness or task fitness." }, { "source": "mlsys.delayed-outcomes", "relation": "limits", "target": "mlsys.monitoring", "statement": "limits timely conclusions from", "conditions": "Outcome maturation and selective observation require explicit evaluation windows." }, { "source": "mlsys.decision-policy", "relation": "uses", "target": "mlsys.calibration", "statement": "uses outputs assessed by", "conditions": "A probability threshold needs task costs and deployment-distribution evidence as well as calibration." }, { "source": "mlsys.supervised-learning", "relation": "requires", "target": "mlsys.labels", "statement": "uses targets assessed by", "conditions": "The target semantics and observation process must match the intended task." }, { "source": "mlsys.self-supervised-learning", "relation": "determines", "target": "mlsys.loss", "statement": "constructs targets for its own", "conditions": "Targets are derived from the data under a specified objective and corruption or prediction scheme." }, { "source": "mlsys.unsupervised-learning", "relation": "uses", "target": "mlsys.features", "statement": "takes its inputs from", "conditions": "When an unsupervised pipeline explicitly constructs or transforms its input features before fitting the learner." }, { "source": "mlsys.reinforcement-learning", "relation": "uses", "target": "mlsys.proxy-objective", "statement": "optimizes behavior against a reward acting as", "conditions": "The reward and interaction model may differ from the desired real-world outcome." }, { "source": "mlsys.imitation-learning", "relation": "causes", "target": "mlsys.distribution-shift", "statement": "induces compounding state-distribution shift", "conditions": "Policy actions change the states visited relative to demonstration data." }, { "source": "mlsys.model-artifact", "relation": "requires", "target": "mlsys.memory-capacity", "statement": "must be resident within", "conditions": "Residency can include parameters, tree structure, support vectors, or stored examples, depending on the model." }, { "source": "mlsys.linear-model", "relation": "determines", "target": "mlsys.model-artifact", "statement": "determines the contents of", "conditions": "Feature transforms and output rules are part of the execution contract." }, { "source": "mlsys.tree-ensemble", "relation": "determines", "target": "mlsys.model-artifact", "statement": "determines the contents of", "conditions": "Tree structures, thresholds, leaf values, and aggregation rules require storage and compatible execution." }, { "source": "mlsys.kernel-method", "relation": "determines", "target": "mlsys.model-artifact", "statement": "determines the size of", "conditions": "Support vectors or an explicit approximation affect the stored state and evaluation cost." }, { "source": "mlsys.nearest-neighbor", "relation": "determines", "target": "mlsys.model-artifact", "statement": "determines the contents of", "conditions": "Stored vectors, metadata, and optional indexes contribute capacity and search cost." }, { "source": "mlsys.failure-model", "relation": "limits", "target": "mlsys.consensus", "statement": "bounds the guarantees of", "conditions": "Fault classes, membership, quorum availability, and timing assumptions bound the agreement claim." }, { "source": "mlsys.failure-model", "relation": "determines", "target": "mlsys.checkpoint", "statement": "determines the recovery requirements for", "conditions": "Recovery semantics must cover the faults and state dependencies actually assumed." }, { "source": "mlsys.backpressure", "relation": "limits", "target": "mlsys.arrival-rate", "statement": "limits offered", "conditions": "The signal must reach producers in time and avoid shifting overload into unbounded buffers." }, { "source": "mlsys.privacy", "relation": "limits", "target": "mlsys.privacy-leakage", "statement": "bounds acceptable", "conditions": "Risk depends on the information, recipient, purpose, consent, and declared protections." }, { "source": "mlsys.roofline", "relation": "limits", "target": "mlsys.execution-time", "statement": "implies a lower bound on", "conditions": "For fixed arithmetic work, a throughput ceiling implies a lower execution-time bound. Additional overheads can lengthen actual runtime." }, { "source": "mlsys.execution-time", "relation": "increases", "target": "mlsys.tail-latency", "statement": "contributes to", "conditions": "Execution contributes to each request latency; component quantiles cannot generally be added to obtain an end-to-end quantile." }, { "source": "mlsys.memory-capacity", "relation": "limits", "target": "mlsys.batching", "statement": "limits feasible", "conditions": "Model state, request state, intermediates, and allocator overhead must fit for the selected execution strategy." }, { "source": "mlsys.metrics", "relation": "measures", "target": "mlsys.generalization", "statement": "measure", "conditions": "The metric, sampling protocol, and target distribution determine what the estimate means." }, { "source": "mlsys.metrics", "relation": "determines", "target": "mlsys.rollout", "statement": "gate advancement during", "conditions": "Release criteria and populations must be specified before interpreting candidate results." }, { "source": "mlsys.monitoring", "relation": "measures", "target": "mlsys.metrics", "statement": "tracks deployed", "conditions": "Observed task metrics can be delayed, biased, or unavailable; system health alone does not establish task quality." }, { "source": "mlsys.data-parallel", "relation": "increases", "target": "mlsys.training-state", "statement": "replicates, multiplying the total footprint of,", "conditions": "Conventional replicated data parallelism keeps model and update state at each worker; ZeRO and hybrid strategies shard selected components." }, { "source": "mlsys.tensor-parallel", "relation": "uses", "target": "mlsys.matrix-multiply", "statement": "partitions", "conditions": "For tensor decompositions whose operand and reduction partitions require compatible communication and layout." }, { "source": "mlsys.profiling", "relation": "measures", "target": "mlsys.execution-time", "statement": "measures observed", "conditions": "Measurement boundaries, instrumentation, and workload determine the observation; a sample is not a universal time bound." }, { "source": "mlsys.fairness", "relation": "limits", "target": "mlsys.decision-policy", "statement": "constrains acceptable", "conditions": "The protected populations, outcome criterion, operating conditions, and trade-offs must be stated; one criterion does not establish all fairness goals." }, { "source": "mlsys.threat-model", "relation": "determines", "target": "mlsys.robustness-evaluation", "statement": "selects attack cases for", "conditions": "Tests exercise a declared adversary and interface; passing selected attacks does not establish universal robustness." }, { "source": "mlsys.failure-model", "relation": "limits", "target": "mlsys.incident-response", "statement": "bounds planned recovery in", "conditions": "A design's recovery guarantees are limited to its declared faults and assumptions; operations may also encounter unmodeled incidents." }, { "source": "mlsys.feedback-control", "relation": "uses", "target": "mlsys.state-estimation", "statement": "uses estimated state from", "conditions": "For controllers using estimated state, uncertainty and observation timing matter; other output-feedback designs need not reconstruct full state." }, { "source": "mlsys.dynamics", "relation": "determines", "target": "mlsys.feedback-control", "statement": "governs the plant response to", "conditions": "The closed-loop response depends on the plant, disturbances, constraints, sensing, and controller implementation." }, { "source": "mlsys.runtime-assurance", "relation": "uses", "target": "mlsys.state-estimation", "statement": "relies on qualified", "conditions": "State-dependent checks need valid observation timing and error bounds; the guarantee remains conditional on those premises." }, { "source": "mlsys.operating-domain", "relation": "determines", "target": "mlsys.robustness-evaluation", "statement": "scopes the relevant conditions for", "conditions": "The intended domain identifies relevant stresses and coupled corners, but finite selected tests do not exhaust continuous operating conditions." }, { "source": "mlsys.service-capacity", "relation": "limits", "target": "mlsys.slo-goodput", "statement": "bounds sustainable", "conditions": "Only for matching work units, counting boundary, workload, operating state, and horizon. A measured throughput is not a universal capacity ceiling." }, { "source": "mlsys.energy", "relation": "increases", "target": "mlsys.unit-cost", "statement": "adds monetary cost to", "conditions": "Convert energy through the applicable tariff or cost model; do not add it again when already included in cloud or facility charges." }, { "source": "mlsys.irreversible-effects", "relation": "limits", "target": "mlsys.rollback", "statement": "can prevent full reversal by", "conditions": "Restoring software or configuration cannot necessarily undo physical, informational, financial, or human consequences already caused." }, { "source": "mlsys.irreversible-effects", "relation": "limits", "target": "mlsys.compensation", "statement": "can remain after", "conditions": "A remedy may address an obligation or consequence without restoring the original world state." }, { "source": "mlsys.runtime-assurance", "relation": "mitigates", "target": "mlsys.irreversible-effects", "statement": "can gate prospective", "conditions": "Only 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." }, { "source": "mlsys.privacy-leakage", "relation": "causes", "target": "mlsys.irreversible-effects", "statement": "can create", "conditions": "Disclosed information may be copied or retained outside the system's control, so subsequent deletion cannot guarantee full retraction." }, { "source": "mlsys.human-oversight", "relation": "limits", "target": "mlsys.authorization", "statement": "can constrain", "conditions": "People need usable evidence, sufficient time, and effective authority; a nominal approval step alone does not establish meaningful control." }, { "source": "mlsys.neural-network", "relation": "determines", "target": "mlsys.model-artifact", "statement": "determines the contents of", "conditions": "Execution requires compatible parameters, operators, and configuration. Training state can include additional data not needed for inference." }, { "source": "mlsys.metrics", "relation": "measures", "target": "mlsys.hallucination", "statement": "can assess support for", "conditions": "For tasks requiring factual or source-grounded assertions, evaluate support in addition to fluency or schema validity; not every task has the same truth criterion." }, { "source": "mlsys.dataset-storage-layout", "relation": "determines", "target": "mlsys.data-ingestion", "statement": "sets the I/O required by", "conditions": "Effects depend on access pattern, projection/filtering, encoding, storage medium, and caching; a layout does not guarantee higher throughput." }, { "source": "mlsys.schema-evolution", "relation": "requires", "target": "mlsys.data-contract", "statement": "requires compatibility rules in", "conditions": "Contracts must identify producers/consumers, field meaning, allowed changes, defaults, and migration expectations; format-supported changes need not preserve application semantics." }, { "source": "mlsys.data-ingestion", "relation": "uses", "target": "mlsys.backpressure", "statement": "applies, under overload,", "conditions": "When record arrival exceeds sustainable processing capacity and the source permits slowing producers; sources that cannot be slowed need buffering or shedding instead." }, { "source": "mlsys.point-in-time", "relation": "requires", "target": "mlsys.data-ingestion", "statement": "requires event-time-preserving", "conditions": "When downstream features or evaluation depend on event-time correctness, ingestion/availability timestamps and late-data policies must be retained; arrival order alone is insufficient." }, { "source": "mlsys.integration-testing", "relation": "requires", "target": "mlsys.test-oracle", "statement": "requires expectations from a", "conditions": "The oracle may be partial, probabilistic, or human-assisted and only covers declared properties and tested cases." }, { "source": "mlsys.integration-testing", "relation": "detects", "target": "mlsys.train-serve-skew", "statement": "can expose", "conditions": "Tests compare relevant preprocessing and model interfaces across representative paths; untested distributions or deployment differences can remain." }, { "source": "mlsys.integration-testing", "relation": "measures", "target": "mlsys.dependency-environment", "statement": "tests compatibility within", "conditions": "Test environments must represent the target packages, binaries, resources, and configuration; passing one environment does not validate all others." }, { "source": "mlsys.failure-model", "relation": "determines", "target": "mlsys.failure-injection", "statement": "selects candidate scenarios for", "conditions": "Injected faults instantiate selected assumptions; realistic timing, combinations, observation criteria, and containment must be declared." }, { "source": "mlsys.failure-injection", "relation": "measures", "target": "mlsys.incident-response", "statement": "exercises", "conditions": "A controlled scenario can test detection and response paths; success does not demonstrate preparedness for every incident." }, { "source": "mlsys.failure-injection", "relation": "requires", "target": "mlsys.test-oracle", "statement": "requires observable acceptance criteria from", "conditions": "Criteria may evaluate latency, error rates, recovery behavior, or invariants; injecting a fault without assessing outcomes is not sufficient evidence." }, { "source": "mlsys.experiment-reproducibility", "relation": "requires", "target": "mlsys.dependency-environment", "statement": "requires a pinned", "conditions": "Package identity alone is insufficient: hardware, data, seeds, nondeterministic operations, external services, and acceptance criteria may also matter." }, { "source": "mlsys.artifact-version", "relation": "enables", "target": "mlsys.experiment-reproducibility", "statement": "identifies the exact inputs for", "conditions": "Reproduction requires resolvable versions and the execution procedure/environment; version labels alone do not reproduce results." }, { "source": "mlsys.workflow-orchestration", "relation": "produces", "target": "mlsys.lineage", "statement": "can record execution evidence as", "conditions": "Only when actual input/output artifacts, parameters, steps, and execution identities are captured; orchestration alone does not establish complete provenance." }, { "source": "mlsys.workflow-orchestration", "relation": "requires", "target": "mlsys.idempotency", "statement": "requires retry-safe steps through", "conditions": "When steps can be retried after ambiguous completion, repeated external effects need explicit handling; idempotency is one option rather than a universal orchestration guarantee." }, { "source": "mlsys.benchmark", "relation": "requires", "target": "mlsys.experiment-reproducibility", "statement": "requires", "conditions": "The benchmark protocol must state tolerances, repetitions, environment, and workload; reproducibility does not establish external validity." }, { "source": "mlsys.memory-hierarchy", "relation": "enables", "target": "mlsys.locality", "statement": "provides storage levels for", "conditions": "The algorithm, placement, and implementation determine which levels actually serve an access; effective rates depend on contention and access pattern." }, { "source": "mlsys.host-device-interconnect", "relation": "increases", "target": "mlsys.execution-time", "statement": "can add transfer delay to", "conditions": "When 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." }, { "source": "mlsys.accelerator-execution-architecture", "relation": "determines", "target": "mlsys.matrix-multiply", "statement": "constrains supported mappings of", "conditions": "Operations, data types, shapes, and implementation must match the execution resources; peak throughput is not achieved throughput." }, { "source": "mlsys.power-delivery", "relation": "limits", "target": "mlsys.compute-throughput", "statement": "constrains feasible operating points for", "conditions": "Supply limits, voltage integrity, and platform power management can constrain sustained execution; power is distinct from energy." }, { "source": "mlsys.thermal-dynamics", "relation": "increases", "target": "mlsys.execution-time", "statement": "can increase", "conditions": "When thermal controls reduce frequency or service rate or suspend operation, a cold unloaded timing measurement need not characterize sustained or shared operation." }, { "source": "mlsys.thermal-dynamics", "relation": "increases", "target": "mlsys.energy", "statement": "adds cooling overhead to", "conditions": "The system boundary and cooling overhead must be stated; carbon conclusions additionally require energy mix and lifecycle assumptions." }, { "source": "mlsys.sli", "relation": "part-of", "target": "mlsys.slo", "statement": "is the measured quantity in", "conditions": "The indicator's population, window, and aggregation must match the objective's." }, { "source": "mlsys.slo", "relation": "determines", "target": "mlsys.error-budget", "statement": "sets the size of", "conditions": "The objective is stated as a fraction over a defined window." }, { "source": "mlsys.error-budget", "relation": "determines", "target": "mlsys.rollout", "statement": "paces", "conditions": "Teams agree to slow or halt changes when the remaining budget is exhausted." }, { "source": "mlsys.latency-budget", "relation": "limits", "target": "mlsys.tail-latency", "statement": "allocates permitted", "conditions": "Stage latencies compose along a request path whose end-to-end target is fixed." }, { "source": "mlsys.flop-accounting", "relation": "measures", "target": "mlsys.arithmetic-intensity", "statement": "supplies the operation and byte counts for", "conditions": "Operation and byte counts are taken at the same memory boundary." }, { "source": "mlsys.memory-capacity", "relation": "limits", "target": "mlsys.memory-footprint", "statement": "caps", "conditions": "All resident state is counted at its peak, including runtime buffers." }, { "source": "mlsys.capacity-planning", "relation": "uses", "target": "mlsys.workload-trace", "statement": "grounds demand estimates in", "conditions": "The trace is representative of the arrival, size, and mix distribution being provisioned for." }, { "source": "mlsys.flop-accounting", "relation": "measures", "target": "mlsys.workload", "statement": "estimates the operations and bytes of", "conditions": "Model shapes and input sizes are known." }, { "source": "mlsys.cost-sensitive-decision", "relation": "determines", "target": "mlsys.decision-policy", "statement": "sets thresholds in", "conditions": "Error costs and class priors can be stated for the deployment context." }, { "source": "mlsys.pareto-frontier", "relation": "limits", "target": "mlsys.requirements", "statement": "bounds the non-dominated trade-offs available to", "conditions": "Two or more objectives are measured on the same workload and cannot all be optimized at once." }, { "source": "mlsys.transformer", "relation": "uses", "target": "mlsys.attention", "statement": "is built from", "conditions": "The architecture uses attention layers interleaved with feed-forward blocks." }, { "source": "mlsys.recommendation-model", "relation": "uses", "target": "mlsys.embeddings", "statement": "is dominated by lookups in", "conditions": "Categorical features are encoded through large embedding tables." }, { "source": "mlsys.diffusion-model", "relation": "increases", "target": "mlsys.execution-time", "statement": "multiplies by sampling step count", "conditions": "Generation runs the denoising network once per sampling step." }, { "source": "mlsys.convolution", "relation": "increases", "target": "mlsys.arithmetic-intensity", "statement": "typically raises", "conditions": "Filters are reused across many spatial positions within one kernel." }, { "source": "mlsys.recurrence", "relation": "increases", "target": "mlsys.serial-fraction", "statement": "adds sequential dependence to", "conditions": "Each step depends on the previous hidden state, preventing parallelism across time steps." }, { "source": "mlsys.normalization-layer", "relation": "increases", "target": "mlsys.numerical-stability", "statement": "improves", "conditions": "Activation statistics would otherwise drift with depth or batch composition." }, { "source": "mlsys.mixture-of-experts", "relation": "reduces", "target": "mlsys.active-parameters", "statement": "reduces relative to total parameters", "conditions": "Each token is routed to only a small subset of experts." }, { "source": "mlsys.mixture-of-experts", "relation": "requires", "target": "mlsys.memory-capacity", "statement": "still requires storing all experts within", "conditions": "Every expert must be resident somewhere in the system even if inactive for a token." }, { "source": "mlsys.expert-routing", "relation": "determines", "target": "mlsys.all-to-all", "statement": "determines token traffic in", "conditions": "Experts are placed on different devices from the tokens routed to them." }, { "source": "mlsys.expert-capacity", "relation": "limits", "target": "mlsys.expert-routing", "statement": "bounds assignments made by", "conditions": "A fixed per-expert buffer is allocated per batch." }, { "source": "mlsys.load-balancing-loss", "relation": "determines", "target": "mlsys.expert-routing", "statement": "regularizes toward balance", "conditions": "The auxiliary loss weight is large enough to counter router collapse without dominating the task loss." }, { "source": "mlsys.active-parameters", "relation": "determines", "target": "mlsys.flop-accounting", "statement": "determines per-token compute in", "conditions": "Compute is counted for the experts actually activated rather than all stored parameters." }, { "source": "mlsys.early-exit", "relation": "reduces", "target": "mlsys.execution-time", "statement": "reduces average, and makes input-dependent,", "conditions": "Exit criteria differ across inputs, so latency varies with input difficulty." }, { "source": "mlsys.context", "relation": "increases", "target": "mlsys.kv-cache", "statement": "scales the size of", "conditions": "Autoregressive decoding retains keys and values for every prior position." }, { "source": "mlsys.sliding-window-attention", "relation": "enables", "target": "mlsys.context", "statement": "makes longer contexts more affordable by reducing attention cost over", "conditions": "The attention pattern reduces attended query-key pairs relative to dense attention; linear cost requires bounded attended positions per query, and quality on long-range tasks may change." }, { "source": "mlsys.linear-attention", "relation": "reduces", "target": "mlsys.kv-cache", "statement": "replaces with a fixed-size state", "conditions": "The attention kernel is approximated by a feature map that permits a recurrent formulation." }, { "source": "mlsys.state-space-model", "relation": "uses", "target": "mlsys.recurrence", "statement": "updates its state through", "conditions": "When the state update runs step by step at inference." }, { "source": "mlsys.grouped-query-attention", "relation": "reduces", "target": "mlsys.kv-cache", "statement": "reduces the size of", "conditions": "Fewer key-value heads are stored than query heads." }, { "source": "mlsys.context-extension", "relation": "uses", "target": "mlsys.position-encoding", "statement": "rescales or interpolates", "conditions": "The model uses a position encoding, such as rotary embeddings, that can be rescaled or interpolated." }, { "source": "mlsys.weight-serialization", "relation": "part-of", "target": "mlsys.model-artifact", "statement": "persists the parameters of", "conditions": "The artifact's weights are persisted for later loading." }, { "source": "mlsys.model-exchange-format", "relation": "enables", "target": "mlsys.compilation", "statement": "provides input to", "conditions": "The target runtime or compiler supports the exchanged operator set." }, { "source": "mlsys.model-export", "relation": "produces", "target": "mlsys.model-exchange-format", "statement": "can emit", "conditions": "The program's operators and control flow are expressible in the format." }, { "source": "mlsys.integration-testing", "relation": "measures", "target": "mlsys.model-signature", "statement": "checks callers against", "conditions": "Callers and the artifact agree on input and output names, types, and shapes." }, { "source": "mlsys.ridge-point", "relation": "determines", "target": "mlsys.bound-regime", "statement": "separates", "conditions": "Throughput and bandwidth ceilings are measured for the same device and memory level." }, { "source": "mlsys.prefill-decode", "relation": "determines", "target": "mlsys.bound-regime", "statement": "classifies prefill as compute-bound and decode as memory-bound in", "conditions": "Prefill processes many tokens per weight read while decode processes few." }, { "source": "mlsys.data-movement-energy", "relation": "increases", "target": "mlsys.energy", "statement": "often dominates", "conditions": "Operands are fetched from off-chip memory rather than reused locally." }, { "source": "mlsys.locality", "relation": "reduces", "target": "mlsys.data-movement-energy", "statement": "reduces", "conditions": "Reused data is served from nearer, lower-energy storage levels." }, { "source": "mlsys.amdahls-law", "relation": "uses", "target": "mlsys.serial-fraction", "statement": "parameterizes", "conditions": "The unaccelerated fraction is measured on the same workload as the speedup." }, { "source": "mlsys.amdahls-law", "relation": "limits", "target": "mlsys.strong-scaling", "statement": "bounds speedup under", "conditions": "Problem size is fixed while resources grow." }, { "source": "mlsys.stragglers", "relation": "reduces", "target": "mlsys.parallel-efficiency", "statement": "reduce", "conditions": "Workers synchronize so the slowest determines step time." }, { "source": "mlsys.collectives", "relation": "reduces", "target": "mlsys.parallel-efficiency", "statement": "can reduce", "conditions": "Communication time is exposed rather than overlapped with computation." }, { "source": "mlsys.littles-law", "relation": "uses", "target": "mlsys.arrival-rate", "statement": "relates system occupancy to", "conditions": "The system is stable and averages are taken over a long window." }, { "source": "mlsys.latency-throughput-curve", "relation": "determines", "target": "mlsys.slo-goodput", "statement": "locates the maximum load meeting", "conditions": "Latency is measured at the percentile the objective specifies." }, { "source": "mlsys.arrival-burstiness", "relation": "increases", "target": "mlsys.queueing", "statement": "increases", "conditions": "Mean utilization is held fixed while inter-arrival variability grows." }, { "source": "mlsys.compute-optimal-allocation", "relation": "requires", "target": "mlsys.scaling-laws", "statement": "is derived from", "conditions": "The fitted law is used within the range of model and data sizes it was fit on." }, { "source": "mlsys.compute-optimal-allocation", "relation": "uses", "target": "mlsys.training-compute-estimate", "statement": "uses", "conditions": "Dense models are trained where the six-operations-per-parameter-per-token approximation holds." }, { "source": "mlsys.compute-optimal-allocation", "relation": "trades-off", "target": "mlsys.unit-cost", "statement": "can trade off against lifecycle", "conditions": "Large inference volume can favor smaller models trained beyond the compute-optimal point." }, { "source": "mlsys.training-compute-estimate", "relation": "specializes", "target": "mlsys.flop-accounting", "statement": "is a special case of", "conditions": "The model is a dense transformer and attention cost is small relative to parameter compute." }, { "source": "mlsys.systolic-array", "relation": "implements", "target": "mlsys.matrix-multiply", "statement": "executes", "conditions": "Operand shapes are tiled to the array dimensions." }, { "source": "mlsys.matrix-unit", "relation": "requires", "target": "mlsys.precision", "statement": "reaches peak throughput only at supported", "conditions": "Operations use input and accumulation formats the unit implements." }, { "source": "mlsys.simt", "relation": "determines", "target": "mlsys.compute-throughput", "statement": "sets the achieved fraction of, through lane masking under divergence,", "conditions": "When threads in a warp take different control-flow paths, so masked lanes idle while each path executes in turn." }, { "source": "mlsys.dataflow", "relation": "determines", "target": "mlsys.data-movement-energy", "statement": "determines", "conditions": "The accelerator's local storage cannot hold all operands at once." }, { "source": "mlsys.domain-specific-architecture", "relation": "increases", "target": "mlsys.compute-throughput", "statement": "raises, on its target operations,", "conditions": "The target operations dominate the workload the hardware runs." }, { "source": "mlsys.hbm", "relation": "increases", "target": "mlsys.memory-bandwidth", "statement": "provides high", "conditions": "Data resides in device memory rather than host memory." }, { "source": "mlsys.on-chip-sram", "relation": "limits", "target": "mlsys.tiling", "statement": "bounds tile size for", "conditions": "Tiles must fit in on-chip storage for reuse to occur." }, { "source": "mlsys.memory-wall", "relation": "determines", "target": "mlsys.bound-regime", "statement": "shifts workloads toward the memory-bound side of", "conditions": "Peak arithmetic throughput grows faster than memory bandwidth across hardware generations." }, { "source": "mlsys.scale-up-interconnect", "relation": "limits", "target": "mlsys.tensor-parallel", "statement": "bounds the efficient group size of", "conditions": "Tensor-parallel collectives occur every layer and need the highest-bandwidth links." }, { "source": "mlsys.coherent-interconnect", "relation": "increases", "target": "mlsys.memory-capacity", "statement": "can extend", "conditions": "Pooled or attached memory is acceptable at its higher latency." }, { "source": "mlsys.peer-to-peer-transfer", "relation": "uses", "target": "mlsys.host-device-interconnect", "statement": "bypasses host staging on", "conditions": "Devices, drivers, and topology support direct transfers." }, { "source": "mlsys.tdp", "relation": "limits", "target": "mlsys.thermal-dynamics", "statement": "sets the design envelope for", "conditions": "Sustained workloads, rather than brief bursts, determine heat removal needs." }, { "source": "mlsys.dvfs", "relation": "reduces", "target": "mlsys.energy", "statement": "can reduce, at some cost in speed,", "conditions": "Workload phases tolerate lower clocks, such as memory-bound phases." }, { "source": "mlsys.dark-silicon", "relation": "enables", "target": "mlsys.domain-specific-architecture", "statement": "motivates", "conditions": "Power, not transistor count, limits how much of a chip can be active." }, { "source": "mlsys.power-capping", "relation": "reduces", "target": "mlsys.compute-throughput", "statement": "can lower", "conditions": "The cap binds below the power the workload would otherwise draw." }, { "source": "mlsys.microcontroller", "relation": "determines", "target": "mlsys.on-device-memory-budget", "statement": "imposes", "conditions": "Weights and activations must fit in on-chip flash and SRAM." }, { "source": "mlsys.edge-npu", "relation": "requires", "target": "mlsys.precision", "statement": "requires supported low", "conditions": "The accelerator implements only specific integer or reduced-precision formats." }, { "source": "mlsys.heterogeneous-soc", "relation": "requires", "target": "mlsys.placement", "statement": "requires operator-level", "conditions": "Different blocks support different operators, precisions, and power states." }, { "source": "mlsys.energy-per-inference", "relation": "specializes", "target": "mlsys.energy", "statement": "is a per-input case of", "conditions": "The boundary is one inference on one device, including memory and idle overhead." }, { "source": "mlsys.tensor-layout", "relation": "part-of", "target": "mlsys.tensor", "statement": "is metadata carried by", "conditions": "The framework records strides or format metadata with each tensor." }, { "source": "mlsys.autodiff", "relation": "uses", "target": "mlsys.computation-graph", "statement": "traverses", "conditions": "Derivatives are composed in reverse over recorded operations." }, { "source": "mlsys.eager-execution", "relation": "determines", "target": "mlsys.dispatch", "statement": "incurs per-operation", "conditions": "Each operation is launched separately from the host program." }, { "source": "mlsys.tracing-jit", "relation": "enables", "target": "mlsys.compilation", "statement": "enables", "conditions": "A compiler accepts the traced intermediate graph and its assumptions about inputs and control flow hold." }, { "source": "mlsys.io-aware-attention", "relation": "requires", "target": "mlsys.on-chip-sram", "statement": "depends on the capacity of", "conditions": "Tile sizes are chosen to fit on-chip memory." }, { "source": "mlsys.convolution-lowering", "relation": "uses", "target": "mlsys.matrix-multiply", "statement": "maps convolution onto", "conditions": "The im2col-style transformation is used, at the cost of duplicated input data." }, { "source": "mlsys.convolution-lowering", "relation": "implements", "target": "mlsys.convolution", "statement": "implements", "conditions": "A lowering algorithm is chosen per layer shape." }, { "source": "mlsys.kernel-library", "relation": "reduces", "target": "mlsys.fusion", "statement": "can limit opportunities for", "conditions": "Library calls are opaque to the compiler and cannot be fused with neighbors." }, { "source": "mlsys.intermediate-representation", "relation": "enables", "target": "mlsys.fusion", "statement": "enables", "conditions": "The IR exposes producer-consumer structure across operators." }, { "source": "mlsys.autotuning", "relation": "determines", "target": "mlsys.tiling", "statement": "selects parameters for", "conditions": "The schedule space is large and performance is hard to predict analytically." }, { "source": "mlsys.kernel-dsl", "relation": "uses", "target": "mlsys.tiling", "statement": "is written in terms of", "conditions": "Kernels are written in terms of blocks that the compiler maps to threads and memory." }, { "source": "mlsys.memory-planning", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "reduces peak", "conditions": "Tensor lifetimes are known at compile time." }, { "source": "mlsys.device-streams", "relation": "enables", "target": "mlsys.communication-overlap", "statement": "enables", "conditions": "Communication and computation are issued on independent streams with correct event dependencies." }, { "source": "mlsys.graph-replay", "relation": "reduces", "target": "mlsys.dispatch", "statement": "reduces", "conditions": "The launch sequence and tensor addresses are static across iterations." }, { "source": "mlsys.loss-scaling", "relation": "part-of", "target": "mlsys.mixed-precision", "statement": "is a component of", "conditions": "Gradients would otherwise underflow in the reduced-precision format." }, { "source": "mlsys.block-scaled-format", "relation": "determines", "target": "mlsys.precision", "statement": "extends the representable range, at fixed element bits, of", "conditions": "When values within each block have similar magnitude so one shared scale serves them all." }, { "source": "mlsys.accumulation-precision", "relation": "determines", "target": "mlsys.numerical-stability", "statement": "affects", "conditions": "Long reductions accumulate rounding error in low precision." }, { "source": "mlsys.expert-parallel", "relation": "requires", "target": "mlsys.all-to-all", "statement": "requires", "conditions": "Tokens are dispatched to and gathered from experts on other devices." }, { "source": "mlsys.expert-parallel", "relation": "enables", "target": "mlsys.mixture-of-experts", "statement": "distributes the experts of", "conditions": "Experts do not fit, or run too slowly, on one device." }, { "source": "mlsys.sequence-parallel", "relation": "increases", "target": "mlsys.context", "statement": "extends the feasible", "conditions": "Activation or attention memory for one sequence exceeds one device." }, { "source": "mlsys.network-topology", "relation": "determines", "target": "mlsys.hybrid-parallel", "statement": "shapes the group mapping of", "conditions": "Link bandwidth differs between levels of the hardware hierarchy." }, { "source": "mlsys.pipeline-parallel", "relation": "produces", "target": "mlsys.pipeline-bubble", "statement": "incurs", "conditions": "Microbatches per step are few relative to pipeline stages." }, { "source": "mlsys.ring-allreduce", "relation": "implements", "target": "mlsys.collectives", "statement": "is an algorithm for", "conditions": "Participants can be arranged in a ring with comparable link bandwidth." }, { "source": "mlsys.alpha-beta-model", "relation": "measures", "target": "mlsys.ring-allreduce", "statement": "predicts the cost of", "conditions": "Link latency and bandwidth are roughly uniform and contention is ignored." }, { "source": "mlsys.tree-allreduce", "relation": "alternative-to", "target": "mlsys.ring-allreduce", "statement": "is an alternative to", "conditions": "Messages are small or participants many, so latency terms dominate." }, { "source": "mlsys.bisection-bandwidth", "relation": "limits", "target": "mlsys.all-to-all", "statement": "bounds throughput of", "conditions": "Traffic crosses the network bisection rather than staying within local groups." }, { "source": "mlsys.data-parallel", "relation": "uses", "target": "mlsys.gradient-compression", "statement": "cuts gradient-synchronization volume with", "conditions": "Convergence is preserved, for example through error feedback, at the chosen compression ratio." }, { "source": "mlsys.lease", "relation": "requires", "target": "mlsys.fencing", "statement": "needs epoch fencing to be safe, as in", "conditions": "Clocks are bounded in drift so expired holders can be rejected safely." }, { "source": "mlsys.leader-election", "relation": "uses", "target": "mlsys.consensus", "statement": "commonly relies on", "conditions": "Replicas must agree on a single leader under partial failure." }, { "source": "mlsys.bounded-staleness", "relation": "mitigates", "target": "mlsys.stragglers", "statement": "reduces waiting on", "conditions": "Convergence tolerates updates computed on stale parameters within the bound." }, { "source": "mlsys.federated-averaging", "relation": "implements", "target": "mlsys.federated-learning", "statement": "is a baseline algorithm for", "conditions": "Clients can run local optimization steps between communication rounds." }, { "source": "mlsys.non-iid-data", "relation": "limits", "target": "mlsys.federated-averaging", "statement": "slows convergence of", "conditions": "Client data distributions differ and clients take many local steps." }, { "source": "mlsys.client-sampling", "relation": "determines", "target": "mlsys.non-iid-data", "statement": "sets which clients' distributions enter each round, shaping the effective", "conditions": "When participation correlates with client characteristics such as connectivity, device class, or time zone." }, { "source": "mlsys.federation-setting", "relation": "determines", "target": "mlsys.client-sampling", "statement": "determines constraints on", "conditions": "Device availability and reliability differ between cross-device and cross-silo settings." }, { "source": "mlsys.pue", "relation": "determines", "target": "mlsys.energy", "statement": "scales IT energy to facility energy in", "conditions": "The accounting boundary includes facility overheads." }, { "source": "mlsys.liquid-cooling", "relation": "increases", "target": "mlsys.rack-power-density", "statement": "raises the feasible", "conditions": "When air cooling cannot remove the heat generated at the target density." }, { "source": "mlsys.rack-power-density", "relation": "limits", "target": "mlsys.scale-up-interconnect", "statement": "constrains the physical extent of", "conditions": "Short copper links require accelerators to be packed into few racks." }, { "source": "mlsys.power-oversubscription", "relation": "requires", "target": "mlsys.power-capping", "statement": "relies on", "conditions": "Simultaneous peaks across equipment are rare but must be bounded when they occur." }, { "source": "mlsys.rdma", "relation": "reduces", "target": "mlsys.collectives", "statement": "lowers per-message overhead of", "conditions": "Transfers bypass host network stacks and remote CPUs." }, { "source": "mlsys.lossless-fabric", "relation": "enables", "target": "mlsys.rdma", "statement": "provides drop-free transport for", "conditions": "The RDMA transport performs poorly under packet loss, as with RoCE." }, { "source": "mlsys.congestion-control", "relation": "mitigates", "target": "mlsys.stragglers", "statement": "can reduce network-induced", "conditions": "Synchronized collective traffic or incast creates switch queue buildup." }, { "source": "mlsys.fat-tree", "relation": "increases", "target": "mlsys.bisection-bandwidth", "statement": "can provide full", "conditions": "Switch tiers are not oversubscribed." }, { "source": "mlsys.rail-optimized", "relation": "determines", "target": "mlsys.hybrid-parallel", "statement": "shapes the group-to-rail placement of", "conditions": "When parallel groups are placed so their peers share a rail, keeping most of their collective traffic within one leaf switch." }, { "source": "mlsys.in-network-reduction", "relation": "implements", "target": "mlsys.tree-allreduce", "statement": "offloads reduction steps of", "conditions": "Switches support aggregation for the data types and operations used." }, { "source": "mlsys.bisection-bandwidth", "relation": "measures", "target": "mlsys.network-topology", "statement": "summarizes the capacity of", "conditions": "Traffic is uniform enough that the worst-case cut is relevant." }, { "source": "mlsys.parallel-file-system", "relation": "enables", "target": "mlsys.input-pipeline", "statement": "supplies aggregate read bandwidth for", "conditions": "Many workers read training data concurrently from shared storage." }, { "source": "mlsys.object-storage", "relation": "part-of", "target": "mlsys.storage-tiering", "statement": "commonly forms the capacity tier in", "conditions": "Hot data is cached closer to compute." }, { "source": "mlsys.input-pipeline", "relation": "uses", "target": "mlsys.storage-tiering", "statement": "reads through the nearer tiers of", "conditions": "Working sets fit in nearer tiers or are prefetched." }, { "source": "mlsys.checkpoint-io", "relation": "determines", "target": "mlsys.checkpoint", "statement": "sets the cost of taking", "conditions": "Training state is large relative to available storage bandwidth." }, { "source": "mlsys.gang-scheduling", "relation": "reduces", "target": "mlsys.utilization", "statement": "can reduce cluster", "conditions": "Resources sit idle while waiting for a full allocation, unless backfilled." }, { "source": "mlsys.job-preemption", "relation": "requires", "target": "mlsys.checkpoint", "statement": "depends on", "conditions": "Preempted work is resumed rather than restarted from scratch." }, { "source": "mlsys.dominant-resource-fairness", "relation": "implements", "target": "mlsys.multi-tenancy", "statement": "allocates shares among tenants in", "conditions": "Jobs demand several resource types in different proportions." }, { "source": "mlsys.elastic-training", "relation": "mitigates", "target": "mlsys.job-preemption", "statement": "reduces lost work from", "conditions": "The job can reshard state and continue with fewer workers." }, { "source": "mlsys.topology-aware-placement", "relation": "reduces", "target": "mlsys.collectives", "statement": "reduces the cost of", "conditions": "Communicating workers are placed on shared fast links or switches." }, { "source": "mlsys.topology-aware-placement", "relation": "specializes", "target": "mlsys.placement", "statement": "refines", "conditions": "Placement decisions are made at node and device granularity." }, { "source": "mlsys.bin-packing", "relation": "trades-off", "target": "mlsys.topology-aware-placement", "statement": "trades off against", "conditions": "Dense packing fragments the contiguous allocations that locality requires." }, { "source": "mlsys.multi-tenancy", "relation": "requires", "target": "mlsys.isolation", "statement": "requires", "conditions": "Tenants must not observe or disrupt one another's work." }, { "source": "mlsys.resource-quota", "relation": "part-of", "target": "mlsys.capacity-planning", "statement": "partitions shared capacity within", "conditions": "Shared capacity is divided among tenants with stated limits." }, { "source": "mlsys.gpu-partitioning", "relation": "increases", "target": "mlsys.utilization", "statement": "can raise", "conditions": "Individual workloads cannot saturate a full accelerator." }, { "source": "mlsys.performance-interference", "relation": "increases", "target": "mlsys.tail-latency", "statement": "can inflate", "conditions": "Co-located workloads contend for unpartitioned shared resources." }, { "source": "mlsys.spot-capacity", "relation": "causes", "target": "mlsys.job-preemption", "statement": "exposes workloads to", "conditions": "The provider reclaims capacity on short notice." }, { "source": "mlsys.spot-capacity", "relation": "reduces", "target": "mlsys.unit-cost", "statement": "can lower", "conditions": "Interruption overhead is smaller than the price discount." }, { "source": "mlsys.weak-supervision", "relation": "determines", "target": "mlsys.labels", "statement": "produces higher-volume, noisier", "conditions": "Holds when labeling sources are noisy and their errors must be modeled or denoised before training." }, { "source": "mlsys.active-learning", "relation": "determines", "target": "mlsys.data-acquisition", "statement": "biases sampling in", "conditions": "The selected examples depend on the current model, so the labeled set is not a random sample of the deployment population." }, { "source": "mlsys.annotator-agreement", "relation": "measures", "target": "mlsys.labels", "statement": "measures reliability of", "conditions": "Holds when multiple independent annotations of the same items are collected." }, { "source": "mlsys.quality-filtering", "relation": "implements", "target": "mlsys.data-selection", "statement": "implements", "conditions": "The quality score must correlate with downstream value; filtering on a narrow proxy can remove useful diversity." }, { "source": "mlsys.data-mixture", "relation": "determines", "target": "mlsys.generalization", "statement": "shifts across domains", "conditions": "Holds under a fixed training budget, where raising one domain's share reduces exposure to others." }, { "source": "mlsys.synthetic-data", "relation": "causes", "target": "mlsys.distribution-shift", "statement": "can introduce", "conditions": "Holds when the generator's output distribution differs from the deployment distribution or repeats the generator's errors." }, { "source": "mlsys.data-augmentation", "relation": "increases", "target": "mlsys.generalization", "statement": "can improve", "conditions": "The transformations must preserve labels and reflect invariances present at deployment." }, { "source": "mlsys.data-validation", "relation": "implements", "target": "mlsys.data-contract", "statement": "enforces", "conditions": "The contract's schema and validity rules must be machine-checkable." }, { "source": "mlsys.feature-store", "relation": "mitigates", "target": "mlsys.train-serve-skew", "statement": "reduces", "conditions": "Holds when training and serving read features from the same definitions; it does not prevent skew from logic outside the store." }, { "source": "mlsys.feature-store", "relation": "enables", "target": "mlsys.point-in-time", "statement": "supports point-in-time joins for", "conditions": "Requires timestamped feature values and as-of joins when building training sets." }, { "source": "mlsys.stream-processing", "relation": "limits", "target": "mlsys.feature-freshness", "statement": "bounds achievable", "conditions": "Processing latency and windowing determine how recent a served feature value can be." }, { "source": "mlsys.feature-backfill", "relation": "requires", "target": "mlsys.point-in-time", "statement": "must preserve", "conditions": "Backfilled values must be computed only from data available at each historical timestamp." }, { "source": "mlsys.experiment-reproducibility", "relation": "requires", "target": "mlsys.data-versioning", "statement": "requires", "conditions": "Holds whenever training or evaluation data can change after a run." }, { "source": "mlsys.data-dependency", "relation": "increases", "target": "mlsys.entanglement", "statement": "amplifies", "conditions": "Holds when unstable or unused inputs remain in a model whose behavior depends jointly on all inputs." }, { "source": "mlsys.feature-freshness", "relation": "causes", "target": "mlsys.train-serve-skew", "statement": "stale values can cause", "conditions": "Holds when serving-time feature age differs from the age assumed when training examples were built." }, { "source": "mlsys.learning-rate-schedule", "relation": "mitigates", "target": "mlsys.training-instability", "statement": "mitigates", "conditions": "Warmup and decay reduce early and late divergence for many optimizers but do not remove instabilities from data or numerics." }, { "source": "mlsys.critical-batch-size", "relation": "limits", "target": "mlsys.data-parallel", "statement": "limits useful scaling of", "conditions": "Holds when adding replicas increases the global batch beyond the critical batch size." }, { "source": "mlsys.checkpoint", "relation": "mitigates", "target": "mlsys.training-instability", "statement": "enables restarts that recover from", "conditions": "Holds when loss spikes are handled by rolling back to an earlier checkpoint and changing data order or hyperparameters." }, { "source": "mlsys.gradient-accumulation", "relation": "enables", "target": "mlsys.batch-size", "statement": "enables large effective batch sizes within limited memory for", "conditions": "Holds for losses that sum over examples; per-batch statistics such as batch normalization still see micro-batches." }, { "source": "mlsys.memory-offloading", "relation": "reduces", "target": "mlsys.training-state", "statement": "reduces the accelerator-resident footprint of", "conditions": "Benefit depends on host-device bandwidth and how well transfers overlap with compute." }, { "source": "mlsys.experiment-tracking", "relation": "enables", "target": "mlsys.experiment-reproducibility", "statement": "records evidence for", "conditions": "Tracking must capture code, data, configuration, and environment versions, not only metrics." }, { "source": "mlsys.hyperparameter-transfer", "relation": "reduces", "target": "mlsys.experiment-budget", "statement": "reduces large-scale spend of", "conditions": "Holds only under a parameterization in which optimal hyperparameters remain stable as width or depth grows." }, { "source": "mlsys.lora", "relation": "specializes", "target": "mlsys.peft", "statement": "is an instance of", "conditions": "Always; LoRA restricts updates to low-rank added matrices." }, { "source": "mlsys.supervised-fine-tuning", "relation": "causes", "target": "mlsys.catastrophic-forgetting", "statement": "can cause", "conditions": "Holds when fine-tuning data are narrow relative to pretraining and no replay or regularization is used." }, { "source": "mlsys.reward-model", "relation": "part-of", "target": "mlsys.rlhf", "statement": "supplies the training signal in", "conditions": "Holds for RLHF pipelines that learn a reward model from human comparisons." }, { "source": "mlsys.rlhf", "relation": "causes", "target": "mlsys.reward-hacking", "statement": "is exposed to", "conditions": "Holds 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." }, { "source": "mlsys.rlvr", "relation": "requires", "target": "mlsys.verifier", "statement": "requires", "conditions": "Rewards are only as reliable as the programmatic checker; exploitable checkers produce reward hacking." }, { "source": "mlsys.weight-sync", "relation": "determines", "target": "mlsys.policy-staleness", "statement": "frequency determines", "conditions": "Holds when rollout workers continue generating between synchronizations." }, { "source": "mlsys.policy-staleness", "relation": "causes", "target": "mlsys.training-instability", "statement": "can induce", "conditions": "Holds when the update rule assumes on-policy data and lag is not corrected by importance weighting or clipping." }, { "source": "mlsys.online-learning", "relation": "determines", "target": "mlsys.feedback-loop", "statement": "shortens the cycle of", "conditions": "Holds when the model's own predictions influence the data it next learns from." }, { "source": "mlsys.incremental-training", "relation": "implements", "target": "mlsys.retraining", "statement": "is one way to carry out", "conditions": "When warm-started training produces the replacement for a deployed model." }, { "source": "mlsys.experience-replay", "relation": "mitigates", "target": "mlsys.catastrophic-forgetting", "statement": "mitigates", "conditions": "Effectiveness depends on buffer size and how representative stored examples are of earlier tasks." }, { "source": "mlsys.memory-capacity", "relation": "limits", "target": "mlsys.on-device-training", "statement": "constrains", "conditions": "Holds on devices whose memory cannot hold full activations and optimizer state, forcing sparse or partial updates." }, { "source": "mlsys.cross-validation", "relation": "reduces", "target": "mlsys.statistical-uncertainty", "statement": "reduces variance of estimates under", "conditions": "Holds for i.i.d. data; temporal or grouped data need splits that respect dependence." }, { "source": "mlsys.sequential-testing", "relation": "enables", "target": "mlsys.online-experiment", "statement": "enables valid early stopping of", "conditions": "Holds when the analysis uses a sequential procedure; repeated peeking at fixed-horizon tests inflates false positives." }, { "source": "mlsys.experiment-interference", "relation": "limits", "target": "mlsys.online-experiment", "statement": "can bias estimates from", "conditions": "Holds when units share resources, markets, or networks, so treatment of one unit affects others." }, { "source": "mlsys.conformal-prediction", "relation": "enables", "target": "mlsys.selective-prediction", "statement": "provides coverage-calibrated sets for", "conditions": "Prediction sets inform an abstention rule; ordinary marginal coverage under exchangeability does not automatically guarantee coverage under distribution shift or conditional accuracy among accepted predictions." }, { "source": "mlsys.slice-discovery", "relation": "enables", "target": "mlsys.sliced-evaluation", "statement": "proposes slices for", "conditions": "Discovered slices require human review for coherence and multiple-comparison effects." }, { "source": "mlsys.llm-judge", "relation": "requires", "target": "mlsys.human-evaluation", "statement": "must be calibrated against", "conditions": "Holds whenever model-graded scores are used as evidence; judges show position, length, and self-preference biases." }, { "source": "mlsys.pass-at-k", "relation": "limits", "target": "mlsys.best-of-n", "statement": "upper-bounds accuracy of", "conditions": "Best-of-n reaches pass@n only with a perfect selector; real verifiers lose part of the gap." }, { "source": "mlsys.statistical-uncertainty", "relation": "limits", "target": "mlsys.pairwise-preference-eval", "statement": "bounds the ranking resolution of", "conditions": "Holds because rankings from finite pairwise votes carry sampling error that can reorder close models." }, { "source": "mlsys.benchmark-contamination", "relation": "specializes", "target": "mlsys.leakage", "statement": "is a form of", "conditions": "Holds when evaluation items or near-duplicates appear in training corpora; detection needs overlap checks against training data." }, { "source": "mlsys.deduplication", "relation": "mitigates", "target": "mlsys.benchmark-contamination", "statement": "reduces", "conditions": "Holds when deduplication is run against evaluation sets, not only within the training corpus." }, { "source": "mlsys.time-to-train", "relation": "part-of", "target": "mlsys.benchmark", "statement": "is the headline metric of a training", "conditions": "Holds for training benchmarks that fix model, dataset, and quality target under run rules." }, { "source": "mlsys.inference-load-scenario", "relation": "limits", "target": "mlsys.tail-latency", "statement": "sets the latency-percentile constraint on", "conditions": "Holds for server-style scenarios that report throughput only at a stated latency percentile." }, { "source": "mlsys.metamorphic-testing", "relation": "implements", "target": "mlsys.test-oracle", "statement": "substitutes output relations for a missing", "conditions": "Holds when a valid metamorphic relation between inputs and outputs is known." }, { "source": "mlsys.production-readiness", "relation": "determines", "target": "mlsys.release-gate", "statement": "supplies criteria for", "conditions": "Holds when rubric items are converted into automated or reviewed checks." }, { "source": "mlsys.tail-latency", "relation": "measures", "target": "mlsys.online-inference", "statement": "is the judging metric of", "conditions": "Holds for user-facing request paths with latency targets." }, { "source": "mlsys.model-server", "relation": "uses", "target": "mlsys.batching", "statement": "hosts", "conditions": "Holds for servers that queue concurrent requests to batch them on shared accelerators." }, { "source": "mlsys.multi-model-serving", "relation": "increases", "target": "mlsys.cold-start", "statement": "increases exposure to load-time", "conditions": "Holds when more models than fit in memory share resources, so infrequently used models must be loaded on demand." }, { "source": "mlsys.model-cascade", "relation": "reduces", "target": "mlsys.unit-cost", "statement": "can reduce", "conditions": "Holds when most requests are resolved by cheap models and the escalation criterion is reliable." }, { "source": "mlsys.model-cascade", "relation": "requires", "target": "mlsys.selective-prediction", "statement": "relies on", "conditions": "Each stage must abstain and defer when confidence is low; poorly calibrated scores misroute requests." }, { "source": "mlsys.slo", "relation": "determines", "target": "mlsys.latency-budget", "statement": "sets the total allocated by", "conditions": "Holds when the end-to-end SLO is decomposed into per-stage targets that sum within it." }, { "source": "mlsys.prefill-decode", "relation": "determines", "target": "mlsys.ttft", "statement": "prefill phase dominates", "conditions": "Holds when queueing delay is small relative to prompt processing time." }, { "source": "mlsys.memory-bandwidth", "relation": "limits", "target": "mlsys.tpot", "statement": "bounds", "conditions": "Holds in decode at small batch, where each step reads weights and KV cache with low arithmetic intensity." }, { "source": "mlsys.hedged-requests", "relation": "reduces", "target": "mlsys.tail-latency", "statement": "reduces", "conditions": "Holds when slow responses come from replica-specific variability and the extra load from duplicates is small." }, { "source": "mlsys.continuous-batching", "relation": "specializes", "target": "mlsys.batching", "statement": "specializes", "conditions": "Applies to iterative generation, where requests finish at different steps." }, { "source": "mlsys.continuous-batching", "relation": "increases", "target": "mlsys.slo-goodput", "statement": "increases", "conditions": "Holds when KV memory admits enough concurrent sequences; interference between prefill and decode can raise latency." }, { "source": "mlsys.load-balancing", "relation": "determines", "target": "mlsys.prefix-cache", "statement": "determines hit rate of", "conditions": "Holds when cached prefixes are replica-local, so cache-aware routing raises reuse but can unbalance load." }, { "source": "mlsys.cold-start", "relation": "limits", "target": "mlsys.autoscaling", "statement": "bounds the reaction time of", "conditions": "When new replicas must load large weights before serving, so scaling must anticipate demand rather than follow it." }, { "source": "mlsys.chunked-prefill", "relation": "reduces", "target": "mlsys.tpot", "statement": "reduces decode stalls in", "conditions": "Holds when long prefills would otherwise stall decode iterations of co-batched requests; it can raise TTFT for the long prompt." }, { "source": "mlsys.pd-disaggregation", "relation": "determines", "target": "mlsys.prefill-decode", "statement": "provisions separate resources for each phase of", "conditions": "Pays off when interference between phases costs more than transferring KV state between pools." }, { "source": "mlsys.pd-disaggregation", "relation": "increases", "target": "mlsys.slo-goodput", "statement": "can increase", "conditions": "Holds when TTFT and TPOT targets would otherwise force over-provisioning of a colocated deployment." }, { "source": "mlsys.constrained-decoding", "relation": "implements", "target": "mlsys.grammar", "statement": "enforces", "conditions": "The constraint representation is supported, enforcement is compatible with the tokenizer, and generation completes successfully." }, { "source": "mlsys.output-validation", "relation": "alternative-to", "target": "mlsys.constrained-decoding", "statement": "is a post-generation alternative for structural checks to", "conditions": "For the same expressible structural constraint, checking after generation can substitute for enforcing it during generation; other checks may still follow constrained decoding." }, { "source": "mlsys.self-consistency", "relation": "requires", "target": "mlsys.sampling-controls", "statement": "requires sampling diversity from", "conditions": "For the sampled-reasoning-path procedure, diversity comes from stochastic decoding; repeated deterministic greedy generation with the same model and prompt produces the same path." }, { "source": "mlsys.reasoning-budget", "relation": "increases", "target": "mlsys.tail-latency", "statement": "larger budgets increase", "conditions": "Holds because generated reasoning length varies widely across requests and adds decode steps." }, { "source": "mlsys.verifier", "relation": "limits", "target": "mlsys.best-of-n", "statement": "precision bounds the gains of", "conditions": "Holds as n grows, since selection pressure increasingly exploits verifier errors." }, { "source": "mlsys.process-reward-model", "relation": "specializes", "target": "mlsys.verifier", "statement": "is a step-level form of", "conditions": "Requires step-level labels or estimates of step correctness." }, { "source": "mlsys.inference-time-search", "relation": "uses", "target": "mlsys.process-reward-model", "statement": "uses for pruning", "conditions": "Holds for searches that score partial solutions before completion." }, { "source": "mlsys.test-time-scaling", "relation": "specializes", "target": "mlsys.scaling-laws", "statement": "extends to inference compute", "conditions": "Holds within task families where extra samples or reasoning steps improve accuracy; gains saturate and vary by difficulty." }, { "source": "mlsys.test-time-scaling", "relation": "determines", "target": "mlsys.unit-cost", "statement": "sets the quality gained per unit of", "conditions": "Holds whenever additional per-query compute is billed or consumes shared serving capacity." }, { "source": "mlsys.rag", "relation": "mitigates", "target": "mlsys.hallucination", "statement": "can reduce", "conditions": "Holds when retrieved passages are relevant and the model grounds its claims in them; irrelevant retrievals can mislead." }, { "source": "mlsys.rag", "relation": "requires", "target": "mlsys.retrieval", "statement": "depends on", "conditions": "Generation quality is bounded by retrieval recall for the facts the answer needs." }, { "source": "mlsys.dense-retrieval", "relation": "uses", "target": "mlsys.ann", "statement": "uses at scale", "conditions": "Holds when the collection is too large for exact similarity search within latency limits." }, { "source": "mlsys.lexical-retrieval", "relation": "alternative-to", "target": "mlsys.dense-retrieval", "statement": "is a complementary alternative to", "conditions": "Lexical matching handles rare terms and exact identifiers that embeddings may miss; hybrid systems combine both scores." }, { "source": "mlsys.reranking", "relation": "part-of", "target": "mlsys.retrieval", "statement": "is the precision stage of", "conditions": "Holds when first-stage recall is adequate and reranker cost fits the latency budget." }, { "source": "mlsys.model-conversion", "relation": "uses", "target": "mlsys.quantization", "statement": "often applies", "conditions": "Holds for integer-only targets; converted models must be re-evaluated because numerics change." }, { "source": "mlsys.memory-capacity", "relation": "limits", "target": "mlsys.on-device-runtime", "statement": "constrains", "conditions": "Holds on microcontrollers and mobile devices where weights, activations, and runtime must fit in limited RAM and flash." }, { "source": "mlsys.ota-model-update", "relation": "uses", "target": "mlsys.rollout", "statement": "applies staged exposure from", "conditions": "Holds for device fleets where updates are staged by cohort and devices may be offline." }, { "source": "mlsys.device-cloud-partitioning", "relation": "specializes", "target": "mlsys.placement", "statement": "is a form of", "conditions": "Holds when part of the computation can run on either side of a network link." }, { "source": "mlsys.device-heterogeneity", "relation": "increases", "target": "mlsys.model-conversion", "statement": "multiplies targets and effort for", "conditions": "Holds when devices differ in supported operators, precision, or accelerators." }, { "source": "mlsys.model-registry", "relation": "implements", "target": "mlsys.artifact-version", "statement": "stores and governs", "conditions": "Holds when every deployable model is registered with its lineage and stage." }, { "source": "mlsys.shadow-deployment", "relation": "part-of", "target": "mlsys.rollout", "statement": "is the zero-exposure stage before", "conditions": "Holds when candidate outputs can be compared without user exposure; shadowing cannot measure effects of users reacting to the new model." }, { "source": "mlsys.ml-ci-cd", "relation": "implements", "target": "mlsys.release-gate", "statement": "automates", "conditions": "Holds when gate criteria are machine-checkable; some require human review." }, { "source": "mlsys.release-gate", "relation": "determines", "target": "mlsys.rollout", "statement": "controls advancement of", "conditions": "Holds when each rollout stage has predeclared pass criteria." }, { "source": "mlsys.drift-detection", "relation": "detects", "target": "mlsys.distribution-shift", "statement": "detects", "conditions": "Holds for shifts visible in monitored statistics; detected drift need not degrade quality, and some harmful shifts are not detected." }, { "source": "mlsys.inference-logging", "relation": "causes", "target": "mlsys.privacy-leakage", "statement": "can increase exposure to", "conditions": "Holds when logs retain personal inputs or outputs without redaction, access control, or retention limits." }, { "source": "mlsys.alerting", "relation": "part-of", "target": "mlsys.incident-response", "statement": "initiates", "conditions": "Holds when alerts are actionable; excessive alerting degrades response through fatigue." }, { "source": "mlsys.retraining", "relation": "mitigates", "target": "mlsys.concept-drift", "statement": "mitigates, given fresh labels,", "conditions": "Holds because input monitoring alone cannot reveal changed input-target relations; updated outcomes are needed." }, { "source": "mlsys.continuous-training", "relation": "implements", "target": "mlsys.retraining", "statement": "automates", "conditions": "Holds when data validation and release gates are automated so retrained models can be promoted without manual steps." }, { "source": "mlsys.postmortem", "relation": "part-of", "target": "mlsys.incident-response", "statement": "closes the loop on", "conditions": "Holds when corrective actions are tracked to completion." }, { "source": "mlsys.failure-attribution", "relation": "requires", "target": "mlsys.observability", "statement": "depends on", "conditions": "Distinguishing data, model, and infrastructure causes requires versioned, correlated signals from each layer." }, { "source": "mlsys.configuration-management", "relation": "reduces", "target": "mlsys.technical-debt", "statement": "reduces configuration", "conditions": "Holds when configuration is reviewed, tested, and versioned like code." }, { "source": "mlsys.infrastructure-as-code", "relation": "determines", "target": "mlsys.dependency-environment", "statement": "declares and reproduces", "conditions": "Holds when declarations pin versions and are applied without manual drift." }, { "source": "mlsys.ml-platform", "relation": "reduces", "target": "mlsys.technical-debt", "statement": "can amortize", "conditions": "Holds when many teams adopt shared components; a poorly maintained platform becomes debt itself." }, { "source": "mlsys.undeclared-consumers", "relation": "causes", "target": "mlsys.feedback-loop", "statement": "can create hidden", "conditions": "Holds when unknown consumers act on model outputs in ways that change future training data." }, { "source": "mlsys.entanglement", "relation": "increases", "target": "mlsys.technical-debt", "statement": "is a source of", "conditions": "Holds for models where inputs interact, so isolated changes have non-local effects." }, { "source": "mlsys.continuous-training", "relation": "part-of", "target": "mlsys.mlops-maturity", "statement": "is a defining capability of higher levels of", "conditions": "Holds for maturity models that define levels by degree of pipeline automation." }, { "source": "mlsys.ml-roles", "relation": "determines", "target": "mlsys.ml-platform", "statement": "assigns ownership of", "conditions": "Holds in organizations large enough to separate platform engineering from model development." }, { "source": "mlsys.undeclared-consumers", "relation": "limits", "target": "mlsys.model-deprecation", "statement": "can block or break", "conditions": "Holds when consumers cannot be identified and migrated before the model is removed." }, { "source": "mlsys.artifact-archival", "relation": "enables", "target": "mlsys.experiment-reproducibility", "statement": "preserves", "conditions": "Holds when archives include data references, environments, and configuration as well as weights." }, { "source": "mlsys.data-retention", "relation": "implements", "target": "mlsys.privacy", "statement": "implements requirements of", "conditions": "Holds when retention limits apply to raw data, logs, and derived artifacts such as features and models." }, { "source": "mlsys.data-retention", "relation": "trades-off", "target": "mlsys.artifact-archival", "statement": "trades off against", "conditions": "Holds when audit or reproduction needs conflict with deletion obligations." }, { "source": "mlsys.model-flops-utilization", "relation": "specializes", "target": "mlsys.utilization", "statement": "specializes", "conditions": "Holds when utilization is defined as model-required FLOPs per second over aggregate peak FLOP rate, excluding recomputed work." }, { "source": "mlsys.roofline", "relation": "limits", "target": "mlsys.model-flops-utilization", "statement": "bounds attainable", "conditions": "For a given kernel mix, attainable MFU cannot exceed the roofline bound implied by each kernel's arithmetic intensity." }, { "source": "mlsys.profiling", "relation": "enables", "target": "mlsys.bottleneck-analysis", "statement": "supplies evidence for", "conditions": "When the profile covers the same workload, configuration, and execution boundary as the bottleneck claim." }, { "source": "mlsys.bottleneck-analysis", "relation": "uses", "target": "mlsys.roofline", "statement": "uses", "conditions": "When the candidate bottleneck is a kernel's compute or memory-bandwidth bound rather than I/O, scheduling, or communication." }, { "source": "mlsys.pruning", "relation": "produces", "target": "mlsys.sparsity", "statement": "produces", "conditions": "When pruning removes individual weights, leaving an irregular zero pattern whose speedup depends on sparse-execution support." }, { "source": "mlsys.structured-pruning", "relation": "specializes", "target": "mlsys.pruning", "statement": "specializes", "conditions": "When the pruned unit is a whole channel, filter, head, or layer rather than an individual weight." }, { "source": "mlsys.low-rank-factorization", "relation": "reduces", "target": "mlsys.matrix-multiply", "statement": "reduces the cost of", "conditions": "When weight matrices are well approximated at a rank much smaller than their dimensions and the smaller products still run efficiently." }, { "source": "mlsys.quantization-aware-training", "relation": "uses", "target": "mlsys.quantization", "statement": "simulates during training", "conditions": "When training models the rounding of the format the model will deploy in." }, { "source": "mlsys.neural-architecture-search", "relation": "specializes", "target": "mlsys.hyperparameter-search", "statement": "is a structured form of", "conditions": "When architectural choices are treated as a structured search space evaluated under an experiment budget." }, { "source": "mlsys.hardware-aware-nas", "relation": "specializes", "target": "mlsys.neural-architecture-search", "statement": "specializes", "conditions": "When the search objective includes a cost measured or predicted on the target platform." }, { "source": "mlsys.hardware-aware-nas", "relation": "reduces", "target": "mlsys.execution-time", "statement": "optimizes measured", "conditions": "When latency on the target device is measured or predicted directly, since FLOP counts can rank architectures differently from real latency." }, { "source": "mlsys.depthwise-separable-convolution", "relation": "reduces", "target": "mlsys.arithmetic-intensity", "statement": "lowers", "conditions": "Fewer operations per output come with less data reuse, so memory-bound depthwise layers may not speed up in proportion to the FLOP reduction." }, { "source": "mlsys.efficiency-frontier", "relation": "requires", "target": "mlsys.metrics", "statement": "is defined only for stated", "conditions": "A frontier is defined only for a stated quality metric, cost measure, and target platform." }, { "source": "mlsys.model-hardware-co-design", "relation": "determines", "target": "mlsys.efficiency-frontier", "statement": "can shift outward", "conditions": "When model structure and hardware or kernels are designed together rather than one being fixed while the other is tuned." }, { "source": "mlsys.model-hardware-co-design", "relation": "determines", "target": "mlsys.accelerator-execution-architecture", "statement": "jointly specifies", "conditions": "When the accelerator's dataflow, precision support, or memory hierarchy is a design variable alongside the model." }, { "source": "mlsys.retry-backoff", "relation": "requires", "target": "mlsys.idempotency", "statement": "requires", "conditions": "Retrying an operation with external effects is safe only when the operation is idempotent or deduplicated." }, { "source": "mlsys.circuit-breaker", "relation": "limits", "target": "mlsys.retry-backoff", "statement": "bounds", "conditions": "When the breaker's error threshold trips before retries saturate the failing dependency." }, { "source": "mlsys.replication-failover", "relation": "increases", "target": "mlsys.availability", "statement": "raises", "conditions": "When replica failures are independent and failover completes faster than the outage it masks." }, { "source": "mlsys.failure-domain", "relation": "limits", "target": "mlsys.replication-failover", "statement": "limits the protection of", "conditions": "When replicas share a failure domain, a single cause can take out all of them together." }, { "source": "mlsys.mtbf", "relation": "determines", "target": "mlsys.checkpoint-interval", "statement": "determines", "conditions": "Under approximately exponential failure times, where the optimal interval scales with the square root of checkpoint cost times MTBF." }, { "source": "mlsys.mtbf", "relation": "increases", "target": "mlsys.availability", "statement": "raises", "conditions": "For a repairable component in steady state, availability is approximately MTBF divided by MTBF plus mean time to repair." }, { "source": "mlsys.checkpoint-interval", "relation": "determines", "target": "mlsys.training-goodput", "statement": "maximizes, when well chosen,", "conditions": "Under a stated MTBF and checkpoint write cost, the interval trades checkpoint overhead against expected lost work." }, { "source": "mlsys.async-checkpointing", "relation": "reduces", "target": "mlsys.checkpoint-interval", "statement": "shortens", "conditions": "When the blocking snapshot to host memory costs much less than the durable write, lowering the effective checkpoint cost." }, { "source": "mlsys.elastic-training", "relation": "uses", "target": "mlsys.checkpoint", "statement": "resumes from", "conditions": "When the job restarts on a changed worker set and state must be resharded from the latest consistent checkpoint." }, { "source": "mlsys.elastic-training", "relation": "increases", "target": "mlsys.training-goodput", "statement": "can raise", "conditions": "When replacing failed nodes takes longer than continuing at reduced width and the optimizer tolerates the batch-size change." }, { "source": "mlsys.abft", "relation": "detects", "target": "mlsys.silent-data-corruption", "statement": "detects", "conditions": "For linear-algebra operations whose input checksums predict output checksums, within the floating-point tolerance used for comparison." }, { "source": "mlsys.gray-failure", "relation": "causes", "target": "mlsys.stragglers", "statement": "can manifest as", "conditions": "In synchronous distributed jobs, where one degraded but live node slows every step." }, { "source": "mlsys.chaos-engineering", "relation": "uses", "target": "mlsys.failure-injection", "statement": "uses", "conditions": "When injected faults are framed as experiments against a stated steady-state hypothesis." }, { "source": "mlsys.blast-radius", "relation": "limits", "target": "mlsys.chaos-engineering", "statement": "bounds", "conditions": "When experiments run against production traffic and must limit the users and components they can affect." }, { "source": "mlsys.failure-domain", "relation": "limits", "target": "mlsys.blast-radius", "statement": "bounds", "conditions": "When isolation between failure domains holds, so a fault does not propagate across shared dependencies." }, { "source": "mlsys.game-day", "relation": "measures", "target": "mlsys.incident-response", "statement": "rehearses", "conditions": "When the exercise uses the same detection, paging, and runbook paths as real incidents." }, { "source": "mlsys.hardware-fault-injection", "relation": "measures", "target": "mlsys.silent-data-corruption", "statement": "estimates model sensitivity to", "conditions": "When the injected fault model, location, and rate approximate the hardware's actual failure modes." }, { "source": "mlsys.backdoor", "relation": "uses", "target": "mlsys.poisoning", "statement": "is commonly implanted through", "conditions": "When the trigger enters through manipulated training data rather than direct weight modification." }, { "source": "mlsys.supply-chain-compromise", "relation": "causes", "target": "mlsys.backdoor", "statement": "can implant a", "conditions": "When pretrained weights or fine-tuning data come from parties outside the defender's trust boundary." }, { "source": "mlsys.jailbreak", "relation": "specializes", "target": "mlsys.evasion", "statement": "specializes", "conditions": "When 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." }, { "source": "mlsys.adversarial-training", "relation": "mitigates", "target": "mlsys.evasion", "statement": "reduces success of", "conditions": "Only within the perturbation set and threat model used to generate training adversaries." }, { "source": "mlsys.adversarial-training", "relation": "trades-off", "target": "mlsys.generalization", "statement": "trades off against clean", "conditions": "Robust training typically costs clean accuracy and multiplies training compute by the cost of generating adversaries." }, { "source": "mlsys.model-extraction", "relation": "produces", "target": "mlsys.model-artifact", "statement": "produces a functional copy of", "conditions": "When the attacker has sufficient query access and output detail, such as confidence scores, relative to model complexity." }, { "source": "mlsys.model-extraction", "relation": "enables", "target": "mlsys.evasion", "statement": "enables transfer attacks for", "conditions": "When adversarial inputs crafted against the extracted surrogate transfer to the deployed model." }, { "source": "mlsys.model-watermarking", "relation": "detects", "target": "mlsys.model-extraction", "statement": "provides ownership evidence against", "conditions": "When the watermark survives the adversary's fine-tuning, pruning, or distillation of the copied model." }, { "source": "mlsys.artifact-signing", "relation": "measures", "target": "mlsys.model-artifact", "statement": "verifies integrity of", "conditions": "When consumers check signatures against trusted identities before loading and the signing key is not compromised." }, { "source": "mlsys.artifact-signing", "relation": "uses", "target": "mlsys.lineage", "statement": "binds provenance from", "conditions": "When the signature covers provenance metadata such as training data and build inputs, not only the weight bytes." }, { "source": "mlsys.unsafe-deserialization", "relation": "enables", "target": "mlsys.supply-chain-compromise", "statement": "enables", "conditions": "When model files in executable serialization formats are loaded from untrusted sources without a data-only format such as safetensors." }, { "source": "mlsys.dependency-environment", "relation": "causes", "target": "mlsys.supply-chain-compromise", "statement": "can introduce, when unpinned or unverified,", "conditions": "When dependencies are resolved from mutable or unverified sources at build or load time." }, { "source": "mlsys.membership-inference", "relation": "measures", "target": "mlsys.privacy-leakage", "statement": "measures", "conditions": "When attack advantage over a baseline is used as an empirical estimate of record-level leakage." }, { "source": "mlsys.differential-privacy", "relation": "limits", "target": "mlsys.membership-inference", "statement": "bounds the advantage of", "conditions": "At the unit defined by the neighboring-dataset relation; the bound weakens as epsilon and delta grow." }, { "source": "mlsys.model-inversion", "relation": "causes", "target": "mlsys.privacy-leakage", "statement": "causes", "conditions": "When outputs, confidence scores, or shared gradients carry enough information about sensitive attributes or inputs." }, { "source": "mlsys.dp-sgd", "relation": "implements", "target": "mlsys.differential-privacy", "statement": "implements", "conditions": "When clipping norm, noise multiplier, sampling rate, and step count are tracked by a privacy accountant." }, { "source": "mlsys.machine-unlearning", "relation": "alternative-to", "target": "mlsys.retraining", "statement": "is a cheaper, approximate alternative to full", "conditions": "Exact unlearning equals retraining without the removed records; approximate methods trade the strength of the guarantee for lower cost." }, { "source": "mlsys.machine-unlearning", "relation": "requires", "target": "mlsys.lineage", "statement": "requires", "conditions": "To locate every model, shard, and derived artifact that was trained on the records to be removed." }, { "source": "mlsys.federated-learning", "relation": "uses", "target": "mlsys.secure-aggregation", "statement": "can use, to hide individual client updates,", "conditions": "When enough clients participate per round that the aggregate does not reveal any single client's update." }, { "source": "mlsys.secure-aggregation", "relation": "increases", "target": "mlsys.poisoning", "statement": "hinders detection of, increasing exposure to", "conditions": "Because the server sees only aggregated updates, it cannot inspect individual client updates for anomalies." }, { "source": "mlsys.least-privilege", "relation": "limits", "target": "mlsys.authorization", "statement": "constrains grants made by", "conditions": "When permissions are scoped to the task and expire or are revoked when the task ends." }, { "source": "mlsys.least-privilege", "relation": "mitigates", "target": "mlsys.prompt-injection", "statement": "limits the consequences of", "conditions": "When a model-driven component processes untrusted content and can act only through narrowly scoped credentials and tools." }, { "source": "mlsys.confidential-computing", "relation": "specializes", "target": "mlsys.isolation", "statement": "is a hardware-attested form of", "conditions": "When the trusted execution environment's hardware and firmware are trusted and the host operator is not." }, { "source": "mlsys.remote-attestation", "relation": "enables", "target": "mlsys.confidential-computing", "statement": "establishes trust in", "conditions": "When the verifier checks attestation evidence against reference values before releasing keys, data, or model weights." }, { "source": "mlsys.side-channel", "relation": "limits", "target": "mlsys.isolation", "statement": "can bypass", "conditions": "When tenants share accelerators, caches, memory, or interconnect whose timing or contention is observable across the boundary." }, { "source": "mlsys.operational-carbon", "relation": "part-of", "target": "mlsys.carbon", "statement": "is a component of", "conditions": "Lifecycle carbon sums operational and embodied components under one declared accounting boundary." }, { "source": "mlsys.embodied-carbon", "relation": "part-of", "target": "mlsys.carbon", "statement": "is a component of", "conditions": "Lifecycle carbon sums operational and embodied components under one declared accounting boundary." }, { "source": "mlsys.carbon-intensity", "relation": "increases", "target": "mlsys.operational-carbon", "statement": "scales", "conditions": "Operational carbon is energy consumed times the carbon intensity of supply at the time and place of consumption." }, { "source": "mlsys.carbon-aware-scheduling", "relation": "uses", "target": "mlsys.carbon-intensity", "statement": "exploits variation in", "conditions": "When workloads are deferrable or relocatable and intensity varies materially across hours or regions." }, { "source": "mlsys.carbon-aware-scheduling", "relation": "trades-off", "target": "mlsys.slo", "statement": "trades off against", "conditions": "When shifting work in time or place delays completion or adds latency beyond a stated objective." }, { "source": "mlsys.energy-efficiency", "relation": "reduces", "target": "mlsys.energy", "statement": "lowers, for a fixed amount of work,", "conditions": "When useful work and energy are measured over the same system boundary and workload." }, { "source": "mlsys.emission-scopes", "relation": "determines", "target": "mlsys.embodied-carbon", "statement": "classifies as value-chain (Scope 3)", "conditions": "For an operator buying hardware, manufacturing emissions fall in Scope 3 while purchased electricity falls in Scope 2." }, { "source": "mlsys.lifecycle-assessment", "relation": "measures", "target": "mlsys.embodied-carbon", "statement": "estimates", "conditions": "When the assessment's goal, scope, and functional unit are stated and manufacturer data are available or proxied." }, { "source": "mlsys.energy", "relation": "increases", "target": "mlsys.water-footprint", "statement": "increases", "conditions": "Through on-site evaporative cooling and off-site water used in electricity generation, both of which scale with energy use." }, { "source": "mlsys.water-footprint", "relation": "trades-off", "target": "mlsys.carbon-aware-scheduling", "statement": "can conflict with", "conditions": "When low-carbon hours or regions coincide with high cooling water demand, such as hot sunny afternoons." }, { "source": "mlsys.hardware-lifetime", "relation": "reduces", "target": "mlsys.embodied-carbon", "statement": "amortizes, lowering per-unit-work", "conditions": "When embodied emissions are attributed per year or per unit of work over the asset's service life." }, { "source": "mlsys.hardware-lifetime", "relation": "trades-off", "target": "mlsys.energy-efficiency", "statement": "trades off against", "conditions": "Keeping older hardware saves embodied emissions and capital but forgoes newer generations' work per joule." }, { "source": "mlsys.hardware-lifetime", "relation": "reduces", "target": "mlsys.e-waste", "statement": "lowers the annual rate of", "conditions": "For a fleet of fixed size, shorter refresh cycles retire more equipment per year." }, { "source": "mlsys.tco", "relation": "part-of", "target": "mlsys.unit-cost", "statement": "supplies the numerator of", "conditions": "When the TCO window and the count of accepted outcomes cover the same period and system boundary." }, { "source": "mlsys.hardware-lifetime", "relation": "determines", "target": "mlsys.tco", "statement": "sets the amortization period in", "conditions": "Capital cost enters TCO through amortization over the assumed service life." }, { "source": "mlsys.cost-per-token", "relation": "measures", "target": "mlsys.unit-cost", "statement": "approximates", "conditions": "Only when tokens per accepted outcome are stable; longer reasoning, retries, or context resends make cheap tokens an expensive outcome." }, { "source": "mlsys.batching", "relation": "reduces", "target": "mlsys.cost-per-token", "statement": "reduces", "conditions": "When larger or continuous batches raise tokens per accelerator-second without violating latency objectives." }, { "source": "mlsys.model-routing", "relation": "requires", "target": "mlsys.calibration", "statement": "depends on", "conditions": "When routers or cascade gates escalate on predicted quality or confidence, which must be calibrated to the request distribution." }, { "source": "mlsys.cost-attribution", "relation": "enables", "target": "mlsys.unit-cost", "statement": "enables", "conditions": "When shared spend is allocated to the workloads that produce the counted outcomes by a declared allocation rule." }, { "source": "mlsys.safety-case", "relation": "requires", "target": "mlsys.hazard-analysis", "statement": "argues only over hazards from", "conditions": "A safety case is complete only relative to the hazards identified for its stated operating domain." }, { "source": "mlsys.stpa", "relation": "specializes", "target": "mlsys.hazard-analysis", "statement": "is a method of", "conditions": "STPA addresses hazards from unsafe interactions and control flaws that component-failure methods can miss." }, { "source": "mlsys.fmea", "relation": "specializes", "target": "mlsys.hazard-analysis", "statement": "is a method of", "conditions": "FMEA is bottom-up over component failures and does not by itself capture hazards without a failed component." }, { "source": "mlsys.control-barrier-function", "relation": "enables", "target": "mlsys.safety-filter", "statement": "provides the certificate for", "conditions": "Guarantees depend on an accurate dynamics model, bounded disturbance, and a feasible filtering problem at each tick." }, { "source": "mlsys.safety-filter", "relation": "implements", "target": "mlsys.runtime-assurance", "statement": "implements", "conditions": "The filter must run independently of, and with authority over, the component it constrains." }, { "source": "mlsys.simplex-architecture", "relation": "implements", "target": "mlsys.runtime-assurance", "statement": "implements", "conditions": "Switching must occur while the verified controller can still recover the state." }, { "source": "mlsys.simplex-architecture", "relation": "uses", "target": "mlsys.minimal-risk-condition", "statement": "falls back toward", "conditions": "Holds when the safety controller's goal is to reach a defined minimal-risk state rather than continue the mission." }, { "source": "mlsys.stopping-envelope", "relation": "limits", "target": "mlsys.minimal-risk-condition", "statement": "bounds reachability of", "conditions": "A stop-type fallback is feasible only while the state remains inside its stopping envelope." }, { "source": "mlsys.guardrails", "relation": "detects", "target": "mlsys.prompt-injection", "statement": "partially detects", "conditions": "Classifier guardrails are probabilistic and can be evaded, so they do not replace authority boundaries." }, { "source": "mlsys.guardrails", "relation": "specializes", "target": "mlsys.runtime-assurance", "statement": "is a content-level form of", "conditions": "Applies when the safe set is a content or action policy rather than a physical state region." }, { "source": "mlsys.falsification", "relation": "uses", "target": "mlsys.scenario-based-testing", "statement": "searches the parameter space of", "conditions": "Found counterexamples are only as realistic as the scenario model and simulator." }, { "source": "mlsys.safety-case", "relation": "uses", "target": "mlsys.rare-event-validation", "statement": "draws statistical evidence from", "conditions": "Bounds hold for the sampled operating distribution and configuration, not for unmodeled conditions." }, { "source": "mlsys.neural-network-verification", "relation": "specializes", "target": "mlsys.formal-verification", "statement": "specializes", "conditions": "Exact methods scale poorly with network size; bound-propagation methods trade completeness for scale." }, { "source": "mlsys.neural-network-verification", "relation": "measures", "target": "mlsys.evasion", "statement": "certifies absence, within a bounded region, of", "conditions": "Holds only for the specified perturbation set, typically a norm ball around given inputs." }, { "source": "mlsys.xil-testing", "relation": "measures", "target": "mlsys.sim-to-real", "statement": "progressively tests", "conditions": "Each rung adds real software, timing, or hardware but still relies on simulated or recorded environments." }, { "source": "mlsys.formal-verification", "relation": "measures", "target": "mlsys.control-barrier-function", "statement": "can certify", "conditions": "Certification covers the barrier condition under the modeled dynamics, not the fidelity of the model." }, { "source": "mlsys.goal-structuring-notation", "relation": "implements", "target": "mlsys.safety-case", "statement": "expresses", "conditions": "Notation makes the argument structure reviewable; it does not make the argument sound." }, { "source": "mlsys.safety-performance-indicator", "relation": "measures", "target": "mlsys.safety-case", "statement": "monitors validity of", "conditions": "Each indicator must trace to a specific claim and have a threshold that triggers review." }, { "source": "mlsys.safety-performance-indicator", "relation": "requires", "target": "mlsys.monitoring", "statement": "is collected through", "conditions": "Holds when field telemetry captures the events the indicator counts." }, { "source": "mlsys.hazard-analysis", "relation": "determines", "target": "mlsys.safety-integrity-level", "statement": "assigns", "conditions": "Integrity levels derive from severity, exposure, and controllability of analyzed hazards." }, { "source": "mlsys.group-fairness-metric", "relation": "measures", "target": "mlsys.fairness", "statement": "measures", "conditions": "A metric measures one criterion on a chosen population and grouping; other criteria remain unmeasured." }, { "source": "mlsys.fairness-incompatibility", "relation": "limits", "target": "mlsys.group-fairness-metric", "statement": "forces a choice among", "conditions": "Holds when base rates differ across groups and the predictor is imperfect." }, { "source": "mlsys.individual-fairness", "relation": "trades-off", "target": "mlsys.group-fairness-metric", "statement": "can conflict with", "conditions": "Holds when similarity under the task metric does not align with group membership." }, { "source": "mlsys.bias-mitigation", "relation": "trades-off", "target": "mlsys.metrics", "statement": "trades off against aggregate", "conditions": "Reducing a disparity often lowers aggregate accuracy or shifts error to other groups or metrics." }, { "source": "mlsys.feature-attribution", "relation": "increases", "target": "mlsys.tail-latency", "statement": "adds to", "conditions": "Holds when explanations are computed on the request path; perturbation-based methods need many extra model evaluations." }, { "source": "mlsys.counterfactual-explanation", "relation": "enables", "target": "mlsys.contestability", "statement": "supports", "conditions": "Useful for recourse only when the suggested changes are feasible and actionable for the person." }, { "source": "mlsys.interpretable-model", "relation": "alternative-to", "target": "mlsys.feature-attribution", "statement": "removes the need for post-hoc", "conditions": "Holds in high-stakes settings where an interpretable model reaches comparable accuracy." }, { "source": "mlsys.model-card", "relation": "uses", "target": "mlsys.sliced-evaluation", "statement": "reports results of", "conditions": "A model card is informative only for the conditions and groups actually evaluated." }, { "source": "mlsys.datasheet", "relation": "uses", "target": "mlsys.lineage", "statement": "summarizes, in human-readable form, provenance from", "conditions": "A datasheet summarizes provenance; machine-checkable lineage is still needed for reproducibility." }, { "source": "mlsys.automation-bias", "relation": "reduces", "target": "mlsys.human-oversight", "statement": "undermines", "conditions": "Holds when automation is usually right, error cues are weak, and reviewers face time pressure." }, { "source": "mlsys.feedback-delay", "relation": "limits", "target": "mlsys.authority-handover", "statement": "bounds, through human reaction time,", "conditions": "Takeover is safe only if human reaction and re-orientation time fits within the remaining safety margin." }, { "source": "mlsys.contestability", "relation": "part-of", "target": "mlsys.human-oversight", "statement": "extends to affected people", "conditions": "Requires a channel, decision records, and authority to reverse outcomes." }, { "source": "mlsys.impact-assessment", "relation": "part-of", "target": "mlsys.ai-risk-management", "statement": "is part of", "conditions": "Mapping risks to affected parties precedes measuring and managing them." }, { "source": "mlsys.algorithmic-audit", "relation": "uses", "target": "mlsys.model-card", "statement": "draws evidence from", "conditions": "Audits also need access to data, code, and logs; documentation alone is insufficient evidence." }, { "source": "mlsys.risk-tiered-regulation", "relation": "requires", "target": "mlsys.human-oversight", "statement": "mandates for high-risk uses", "conditions": "Applies to systems classified high-risk under the governing regulation." }, { "source": "mlsys.risk-tiered-regulation", "relation": "requires", "target": "mlsys.algorithmic-audit", "statement": "requires conformity assessment by", "conditions": "Applies to high-risk categories; obligations differ by role (provider, deployer)." }, { "source": "mlsys.frontier-safety-framework", "relation": "implements", "target": "mlsys.ai-risk-management", "statement": "is a developer's instrument for", "conditions": "When a developer's published capability thresholds and commitments govern its model releases." }, { "source": "mlsys.context-compaction", "relation": "trades-off", "target": "mlsys.prefix-cache", "statement": "trades saved tokens against the prefix reuse lost in", "conditions": "When rewriting earlier context invalidates the cached prefix, so compaction pays off only when the saved execution work over the remaining session exceeds the summarization and prefix-recompute cost." }, { "source": "mlsys.context-degradation", "relation": "limits", "target": "mlsys.context", "statement": "limits useful size of", "conditions": "Holds even when the context fits the nominal window; strength varies by model and task." }, { "source": "mlsys.agent-memory", "relation": "uses", "target": "mlsys.retrieval", "statement": "is read through", "conditions": "Memory recall quality is bounded by the retrieval contract and index freshness." }, { "source": "mlsys.function-calling", "relation": "requires", "target": "mlsys.tool-contract", "statement": "is validated against", "conditions": "The harness must validate calls against the contract; schema conformance does not imply authority." }, { "source": "mlsys.model-context-protocol", "relation": "enables", "target": "mlsys.function-calling", "statement": "standardizes tool discovery for", "conditions": "Discovering a tool through a protocol does not grant authority to use it." }, { "source": "mlsys.computer-use", "relation": "increases", "target": "mlsys.irreversible-effects", "statement": "raises exposure to", "conditions": "GUI actions often lack typed preconditions, idempotency keys, or undo, unlike designed tool APIs." }, { "source": "mlsys.computer-use", "relation": "increases", "target": "mlsys.prompt-injection", "statement": "widens the surface for", "conditions": "Holds when on-screen content from untrusted sources is read as part of the agent's context." }, { "source": "mlsys.agent-harness", "relation": "produces", "target": "mlsys.trajectory", "statement": "produces and records", "conditions": "The harness, not the model, owns continuation and the authoritative trajectory record." }, { "source": "mlsys.agent-budget", "relation": "determines", "target": "mlsys.unit-cost", "statement": "caps per-trajectory", "conditions": "Budgets cap spend per trajectory; cost per accepted task also depends on success rate." }, { "source": "mlsys.approval-gate", "relation": "implements", "target": "mlsys.human-oversight", "statement": "implements", "conditions": "Effective only if approvals bind to the exact action and reviewers can judge it; see automation bias." }, { "source": "mlsys.approval-gate", "relation": "mitigates", "target": "mlsys.irreversible-effects", "statement": "guards against", "conditions": "Holds when the gate is placed before dispatch of actions classified irreversible." }, { "source": "mlsys.durable-execution", "relation": "requires", "target": "mlsys.idempotency", "statement": "relies on", "conditions": "Resuming after a crash between dispatch and record can re-issue a step, which is safe only for idempotent or deduplicated effects." }, { "source": "mlsys.deterministic-replay", "relation": "uses", "target": "mlsys.durable-execution", "statement": "reuses the log of", "conditions": "Replay is exact only if every nondeterministic input, including model samples and tool results, was recorded." }, { "source": "mlsys.suspend-resume", "relation": "enables", "target": "mlsys.approval-gate", "statement": "enables long waits for", "conditions": "Resumption must revalidate state and authority that may have changed during the wait." }, { "source": "mlsys.suspend-resume", "relation": "requires", "target": "mlsys.fencing", "statement": "requires", "conditions": "Holds when a resumed worker could race a stale one that still believes it owns the trajectory." }, { "source": "mlsys.multi-agent-orchestration", "relation": "causes", "target": "mlsys.coordination-failure", "statement": "can introduce", "conditions": "Each delegation and handoff adds a point where specification or context can be lost." }, { "source": "mlsys.multi-agent-orchestration", "relation": "uses", "target": "mlsys.baseline", "statement": "is justified by comparison against a single-agent", "conditions": "When the orchestration's added cost is evaluated against a single agent at matched total compute or token budget." }, { "source": "mlsys.shared-agent-state", "relation": "uses", "target": "mlsys.fencing", "statement": "uses", "conditions": "Holds when leases can expire while a slow agent still attempts writes." }, { "source": "mlsys.correlated-agent-errors", "relation": "limits", "target": "mlsys.multi-agent-orchestration", "statement": "limits the benefit of", "conditions": "Holds when agents share a base model, prompts, or retrieved evidence." }, { "source": "mlsys.sensor-fusion", "relation": "requires", "target": "mlsys.time-synchronization", "statement": "requires", "conditions": "Timestamp error translates to spatial error proportional to relative motion." }, { "source": "mlsys.sensor-fusion", "relation": "requires", "target": "mlsys.sensing", "statement": "requires extrinsic calibration from", "conditions": "Fused estimates inherit calibration errors between sensor frames." }, { "source": "mlsys.robot-perception", "relation": "increases", "target": "mlsys.observation-age", "statement": "adds inference latency to", "conditions": "Inference time adds directly to the age of the state used by planning and control." }, { "source": "mlsys.recursive-filtering", "relation": "implements", "target": "mlsys.state-estimation", "statement": "implements", "conditions": "Kalman-family filters assume approximately Gaussian, well-modeled noise; violations make reported covariance overconfident." }, { "source": "mlsys.slam", "relation": "specializes", "target": "mlsys.state-estimation", "statement": "specializes", "conditions": "Drift accumulates without loop closure or external references." }, { "source": "mlsys.model-predictive-control", "relation": "uses", "target": "mlsys.world-model", "statement": "can use as its prediction model a", "conditions": "Plans exploit model errors, so performance depends on accuracy in the regions the optimizer visits." }, { "source": "mlsys.motion-planning", "relation": "requires", "target": "mlsys.trajectory-feasibility", "statement": "must satisfy", "conditions": "Geometric planners need a later step to enforce dynamic and actuation limits." }, { "source": "mlsys.model-predictive-control", "relation": "requires", "target": "mlsys.wcet", "statement": "requires a solver WCET within its control period, i.e. a", "conditions": "Holds when each optimization must finish within the control period; warm starts and horizon length trade quality for time." }, { "source": "mlsys.model-predictive-control", "relation": "specializes", "target": "mlsys.feedback-control", "statement": "is a form of", "conditions": "Re-solving from each new state estimate closes the loop." }, { "source": "mlsys.real-time-scheduling", "relation": "requires", "target": "mlsys.wcet", "statement": "takes as input", "conditions": "Schedulability results are valid only if WCET bounds hold on the deployed platform." }, { "source": "mlsys.mixed-criticality", "relation": "enables", "target": "mlsys.runtime-assurance", "statement": "protects the timing of", "conditions": "Holds when the safety monitor shares hardware with learned components that can overrun or contend for memory." }, { "source": "mlsys.mixed-criticality", "relation": "requires", "target": "mlsys.isolation", "statement": "requires temporal and spatial", "conditions": "Shared caches, memory controllers, and buses can break isolation that address-space separation provides." }, { "source": "mlsys.vla-model", "relation": "increases", "target": "mlsys.observation-age", "statement": "increases", "conditions": "Holds when billion-parameter inference on robot compute runs slower than the control loop." }, { "source": "mlsys.vla-model", "relation": "uses", "target": "mlsys.action-chunking", "statement": "uses, to decouple control rate from inference rate,", "conditions": "Chunks executed open-loop delay reaction to new observations until the next inference." }, { "source": "mlsys.teleoperation-data", "relation": "enables", "target": "mlsys.imitation-learning", "statement": "supplies demonstrations for", "conditions": "Policy quality is bounded by demonstration coverage, consistency, and action-label timing." }, { "source": "mlsys.domain-randomization", "relation": "enables", "target": "mlsys.sim-to-real", "statement": "improves", "conditions": "Holds when the randomized ranges cover the real parameters without making the task unlearnable." }, { "source": "mlsys.thread-hierarchy", "relation": "uses", "target": "mlsys.simt", "statement": "is executed through", "conditions": "Warps of a thread block are scheduled together on one streaming multiprocessor." }, { "source": "mlsys.memory-coalescing", "relation": "determines", "target": "mlsys.memory-bandwidth", "statement": "helps determine achieved", "conditions": "For global-memory accesses by a warp, coalescing reduces wasted transactions; achieved bandwidth also depends on the workload and hardware." }, { "source": "mlsys.async-data-movement", "relation": "implements", "target": "mlsys.communication-overlap", "statement": "applies within a device the idea of", "conditions": "Loads into on-chip memory proceed while matrix units compute on the previous tile." }, { "source": "mlsys.matrix-engine-programming", "relation": "enables", "target": "mlsys.compute-throughput", "statement": "is required to reach peak", "conditions": "Peak matrix throughput is available only to instructions with supported shapes, layouts, and precisions." }, { "source": "mlsys.thread-hierarchy", "relation": "implements", "target": "mlsys.tiling", "statement": "maps output tiles onto thread blocks in", "conditions": "Each thread block typically computes one output tile staged through shared memory." }, { "source": "mlsys.embedding-table-sharding", "relation": "enables", "target": "mlsys.recommendation-model", "statement": "makes trainable", "conditions": "Embedding tables exceed one device's memory and must be partitioned across devices." }, { "source": "mlsys.embedding-table-sharding", "relation": "requires", "target": "mlsys.collectives", "statement": "requires all-to-all", "conditions": "Lookups for a batch touch rows held on other devices, so activations are exchanged." }, { "source": "mlsys.parameter-server", "relation": "uses", "target": "mlsys.bounded-staleness", "statement": "commonly operates under", "conditions": "Workers may read parameters that lag the latest updates by a bounded number of steps." }, { "source": "mlsys.latency-budget", "relation": "limits", "target": "mlsys.ranking-funnel", "statement": "bounds the per-stage cost of", "conditions": "Each stage receives a share of the end-to-end budget and a candidate count it can afford to score." }, { "source": "mlsys.ranking-funnel", "relation": "uses", "target": "mlsys.ann", "statement": "uses for candidate generation", "conditions": "The first stage retrieves candidates by approximate similarity before learned ranking." }, { "source": "mlsys.multimodal-model", "relation": "uses", "target": "mlsys.modality-encoder", "statement": "is composed with", "conditions": "Non-text inputs are encoded into tokens the shared backbone attends over." }, { "source": "mlsys.encoder-disaggregation", "relation": "reduces", "target": "mlsys.tail-latency", "statement": "reduces", "conditions": "Large media inputs would otherwise block token generation for requests sharing the batch." }, { "source": "mlsys.auto-parallelization", "relation": "produces", "target": "mlsys.hybrid-parallel", "statement": "derives", "conditions": "The compiler's cost model must reflect the target device mesh's compute and link bandwidths for the chosen plan to be competitive." }, { "source": "mlsys.sharding-propagation", "relation": "implements", "target": "mlsys.auto-parallelization", "statement": "implements", "conditions": "Holds when users supply seed annotations and the compiler propagates layouts through operators with known sharding rules." }, { "source": "mlsys.sharding-propagation", "relation": "produces", "target": "mlsys.collectives", "statement": "inserts", "conditions": "Collectives are inserted wherever an operator's input layouts differ from what its sharded implementation requires." }, { "source": "mlsys.sharding-propagation", "relation": "requires", "target": "mlsys.device-mesh", "statement": "is defined over the axes of", "conditions": "Tensor dimensions are sharded over named mesh axes, so mesh shape bounds which layouts are expressible." }, { "source": "mlsys.function-transformations", "relation": "implements", "target": "mlsys.autodiff", "statement": "exposes as a transform", "conditions": "Holds when programs are pure functions so differentiation composes with vectorization and compilation." }, { "source": "mlsys.function-transformations", "relation": "requires", "target": "mlsys.tracing-jit", "statement": "relies on", "conditions": "For transformations implemented over a traced intermediate representation, such as JAX transformations; this is not a requirement for every transformation implementation." }, { "source": "mlsys.task-actor-runtime", "relation": "enables", "target": "mlsys.rl-rollout", "statement": "orchestrates", "conditions": "Applies when rollout workers, trainers, and reward models run as separate stateful actors exchanging data through the runtime." }, { "source": "mlsys.single-vs-multi-controller", "relation": "determines", "target": "mlsys.task-actor-runtime", "statement": "is a design axis of", "conditions": "A single controller simplifies heterogeneous pipelines but must keep per-step dispatch off the critical path to scale." }, { "source": "mlsys.analytical-performance-model", "relation": "uses", "target": "mlsys.roofline", "statement": "composes", "conditions": "Per-operator roofline bounds are summed or overlapped with communication and queueing terms under stated overlap assumptions." }, { "source": "mlsys.analytical-performance-model", "relation": "uses", "target": "mlsys.alpha-beta-model", "statement": "composes", "conditions": "Communication terms use per-collective latency and bandwidth parameters measured or specified for the target fabric." }, { "source": "mlsys.model-fidelity", "relation": "measures", "target": "mlsys.analytical-performance-model", "statement": "bounds trust in", "conditions": "Predictions are credible only within the configurations where validation error against measurement has been reported." }, { "source": "mlsys.performance-simulation", "relation": "alternative-to", "target": "mlsys.analytical-performance-model", "statement": "trades speed for detail against", "conditions": "Simulation captures contention and scheduling effects closed forms omit, at much higher evaluation cost per configuration." }, { "source": "mlsys.post-training-quantization", "relation": "specializes", "target": "mlsys.quantization", "statement": "applies", "conditions": "Holds when the calibration set is representative of deployment inputs; otherwise chosen ranges misfit and accuracy drops." }, { "source": "mlsys.weight-only-quantization", "relation": "uses", "target": "mlsys.post-training-quantization", "statement": "is commonly produced by", "conditions": "For GPTQ- and AWQ-class methods applied after training." }, { "source": "mlsys.activation-outliers", "relation": "limits", "target": "mlsys.quantization", "statement": "can degrade the accuracy of", "conditions": "Applies to per-tensor activation quantization of large transformers; per-channel handling or scale migration mitigates it." }, { "source": "mlsys.kv-cache-quantization", "relation": "reduces", "target": "mlsys.kv-cache", "statement": "shrinks", "conditions": "Capacity gains translate into throughput only when serving is KV-memory-bound and attention kernels read the quantized format directly." }, { "source": "mlsys.nm-sparsity", "relation": "requires", "target": "mlsys.matrix-unit", "statement": "requires support in", "conditions": "Speedup materializes only on matrix engines with sparse-operand instructions for the specific N:M ratio." }, { "source": "mlsys.nm-sparsity", "relation": "specializes", "target": "mlsys.sparsity", "statement": "specializes", "conditions": "Constrains sparsity to a fixed per-group pattern so hardware can index nonzeros with small metadata." }, { "source": "mlsys.kv-cache-offload", "relation": "enables", "target": "mlsys.prefix-cache", "statement": "increases the capacity and hit rate of", "conditions": "Pays off when reload time from host or SSD is shorter than recomputing the prefix's prefill." }, { "source": "mlsys.pd-disaggregation", "relation": "requires", "target": "mlsys.kv-cache-transfer", "statement": "requires", "conditions": "Disaggregation is profitable only when KV transfer between pools is faster than the decode-latency gain it buys." }, { "source": "mlsys.compound-ai-system", "relation": "uses", "target": "mlsys.model-routing", "statement": "composes", "conditions": "Holds when the program chooses among models per step, so routing decisions are optimized against end-to-end quality and cost." }, { "source": "mlsys.rag", "relation": "part-of", "target": "mlsys.compound-ai-system", "statement": "is a common stage of", "conditions": "When retrieval and generation run as separately served and tuned components." }, { "source": "mlsys.lm-program-optimization", "relation": "determines", "target": "mlsys.compound-ai-system", "statement": "tunes the prompts and weights of", "conditions": "Requires a declared program structure and an end-to-end metric evaluable on a training set of inputs." }, { "source": "mlsys.multi-adapter-serving", "relation": "uses", "target": "mlsys.lora", "statement": "serves many instances of", "conditions": "Holds when adapters share one frozen base model so base weights are loaded once and only low-rank factors differ per request." }, { "source": "mlsys.multi-adapter-serving", "relation": "specializes", "target": "mlsys.multi-model-serving", "statement": "specializes", "conditions": "Applies when hosted variants differ only by adapters, letting requests for different variants share one batch." }, { "source": "mlsys.graph-vector-index", "relation": "implements", "target": "mlsys.ann", "statement": "implements", "conditions": "Best recall-latency trade-off when the full graph and vectors fit in memory." }, { "source": "mlsys.quantized-vector-index", "relation": "implements", "target": "mlsys.ann", "statement": "implements", "conditions": "Preferred when memory per vector is the binding constraint, accepting lower recall or a reranking pass on exact vectors." }, { "source": "mlsys.quantized-vector-index", "relation": "alternative-to", "target": "mlsys.graph-vector-index", "statement": "trades memory against recall versus", "conditions": "Compared at equal query latency on the same corpus and recall target." }, { "source": "mlsys.filtered-vector-search", "relation": "uses", "target": "mlsys.ann", "statement": "commonly runs on", "conditions": "When the filter leaves too many candidates for exact search." }, { "source": "mlsys.graph-vector-index", "relation": "limits", "target": "mlsys.index-freshness", "statement": "update and delete cost bounds attainable", "conditions": "Deletes and in-place updates degrade graph connectivity unless repaired or periodically rebuilt." }, { "source": "mlsys.numerical-nondeterminism", "relation": "reduces", "target": "mlsys.experiment-reproducibility", "statement": "limits", "conditions": "Bitwise reproducibility requires fixing kernels, reduction order, parallel layout, and hardware, not only random seeds." }, { "source": "mlsys.accumulation-precision", "relation": "determines", "target": "mlsys.numerical-nondeterminism", "statement": "sets the magnitude of", "conditions": "Order-dependent rounding differences grow as accumulation precision falls and reduction length grows." }, { "source": "mlsys.batch-invariance", "relation": "mitigates", "target": "mlsys.numerical-nondeterminism", "statement": "removes the batch-composition source of", "conditions": "Achieved by fixing reduction strategies independent of batch size, typically at some kernel throughput cost." }, { "source": "mlsys.batch-invariance", "relation": "enables", "target": "mlsys.rl-rollout", "statement": "aligns inference numerics with training for", "conditions": "Matters when rollout log-probabilities from the serving engine are reused in on-policy training objectives." }, { "source": "mlsys.elastic-training", "relation": "requires", "target": "mlsys.checkpoint-resharding", "statement": "requires", "conditions": "Holds whenever a restart changes world size or parallel layout relative to the saved checkpoint." }, { "source": "mlsys.checkpoint-resharding", "relation": "uses", "target": "mlsys.state-sharding", "statement": "remaps between layouts of", "conditions": "Requires checkpoints that record each shard's global tensor coordinates rather than rank-local layouts." }, { "source": "mlsys.batch-data-processing", "relation": "enables", "target": "mlsys.deduplication", "statement": "executes at corpus scale", "conditions": "Applies when the corpus exceeds one machine, so near-duplicate detection is partitioned and shuffled across workers." }, { "source": "mlsys.batch-data-processing", "relation": "enables", "target": "mlsys.feature-backfill", "statement": "executes", "conditions": "Backfills over long histories run as bounded batch jobs whose logic must match the serving feature definition." }, { "source": "mlsys.distributed-tracing", "relation": "part-of", "target": "mlsys.observability", "statement": "provides request-level evidence within", "conditions": "Requires trace context to be propagated across every service and asynchronous hop the request traverses." }, { "source": "mlsys.llm-call-tracing", "relation": "specializes", "target": "mlsys.distributed-tracing", "statement": "specializes", "conditions": "Adds model-specific span attributes such as token counts, model version, and cost under a shared semantic convention." }, { "source": "mlsys.llm-call-tracing", "relation": "measures", "target": "mlsys.agent-harness", "statement": "instruments", "conditions": "Each loop iteration, model call, and tool call is emitted as a child span of the task's root span." }, { "source": "mlsys.evaluation-harness", "relation": "implements", "target": "mlsys.benchmark", "statement": "implements", "conditions": "Comparability holds only when prompt templates, decoding settings, and scorer versions are pinned with the results." }, { "source": "mlsys.evaluation-harness", "relation": "uses", "target": "mlsys.experiment-tracking", "statement": "records results into", "conditions": "Holds when each run logs checkpoint, suite version, and configuration alongside scores." }, { "source": "mlsys.energy-measurement", "relation": "measures", "target": "mlsys.energy", "statement": "determines the reported value of", "conditions": "Reported energy is comparable only across measurements with the same boundary, sensor class, and sampling rate." }, { "source": "mlsys.idle-dynamic-power", "relation": "determines", "target": "mlsys.energy-measurement", "statement": "sets attribution rules for", "conditions": "Matters when hardware is shared or partly idle, so idle power must be charged by an explicit allocation rule." }, { "source": "mlsys.he-inference", "relation": "alternative-to", "target": "mlsys.confidential-computing", "statement": "trades compute overhead for hardware trust versus", "conditions": "Homomorphic encryption avoids trusting server hardware but costs orders of magnitude more compute than enclave execution." }, { "source": "mlsys.secure-computation-inference", "relation": "alternative-to", "target": "mlsys.he-inference", "statement": "trades communication for computation versus", "conditions": "Hybrid protocols use homomorphic encryption for linear layers and secret sharing or garbled circuits for nonlinear ones." }, { "source": "mlsys.secure-computation-inference", "relation": "mitigates", "target": "mlsys.privacy-leakage", "statement": "prevents input disclosure in", "conditions": "Protects inputs during computation but not leakage from the revealed outputs themselves." }, { "source": "mlsys.red-teaming", "relation": "detects", "target": "mlsys.jailbreak", "statement": "discovers", "conditions": "Coverage depends on attacker diversity and budget, so finding no failures does not establish their absence." }, { "source": "mlsys.red-teaming", "relation": "requires", "target": "mlsys.threat-model", "statement": "must be scoped by", "conditions": "The adversary capabilities, interfaces, and harms probed must be declared for findings to be interpretable." }, { "source": "mlsys.dangerous-capability-eval", "relation": "part-of", "target": "mlsys.frontier-safety-framework", "statement": "supplies the threshold tests that trigger safeguards in", "conditions": "Holds when the framework ties stated capability thresholds to required mitigations before further training or deployment." }, { "source": "mlsys.dangerous-capability-eval", "relation": "uses", "target": "mlsys.red-teaming", "statement": "uses elicitation from", "conditions": "Capability estimates are lower bounds unless elicitation, including expert red teaming, is strong." }, { "source": "mlsys.efficiency-frontier", "relation": "specializes", "target": "mlsys.pareto-frontier", "statement": "specializes", "conditions": "The objectives are model quality and a stated cost measure on a stated platform." }, { "source": "mlsys.energy-per-inference", "relation": "measures", "target": "mlsys.energy-efficiency", "statement": "is the per-input reciprocal of", "conditions": "Both are measured within the same declared system boundary." }, { "source": "mlsys.requirements", "relation": "determines", "target": "mlsys.placement", "statement": "determine feasible", "conditions": "Latency, privacy, connectivity, and power requirements rule out some device, edge, or datacenter placements." }, { "source": "mlsys.release-gate", "relation": "requires", "target": "mlsys.baseline", "statement": "compares candidates against", "conditions": "A candidate advances only if it beats or matches the incumbent baseline on the gated metrics under the same protocol." }, { "source": "mlsys.baseline", "relation": "part-of", "target": "mlsys.online-experiment", "statement": "serves as the control arm of", "conditions": "The incumbent system or model is served to the control group during the experiment." }, { "source": "mlsys.benchmark", "relation": "requires", "target": "mlsys.workload", "statement": "is only meaningful for a stated", "conditions": "Benchmark results transfer only to deployments whose input, request, and model distributions match the benchmark workload." }, { "source": "mlsys.workload", "relation": "uses", "target": "mlsys.workload-trace", "statement": "is characterized empirically using", "conditions": "The trace window covers the arrival, size, and mix variation that matters for the decision being evaluated." }, { "source": "mlsys.reward-model", "relation": "causes", "target": "mlsys.reward-hacking", "statement": "can invite", "conditions": "Policies optimized hard against a learned reward model can exploit its errors, scoring higher while true quality falls." }, { "source": "mlsys.monitoring", "relation": "produces", "target": "mlsys.sli", "statement": "computes", "conditions": "Monitoring records the events and counts over the population and window the indicator definition specifies." }, { "source": "mlsys.sli", "relation": "measures", "target": "mlsys.tail-latency", "statement": "commonly measures, as a percentile,", "conditions": "Latency indicators are computed as a high percentile over a stated request population and window." }, { "source": "mlsys.sli", "relation": "measures", "target": "mlsys.availability", "statement": "commonly measures", "conditions": "Availability indicators count successful requests or uptime over a defined population and window." }, { "source": "mlsys.error-budget", "relation": "determines", "target": "mlsys.release-gate", "statement": "can freeze releases through", "conditions": "Teams agree that an exhausted budget blocks or slows promotion of new versions." }, { "source": "mlsys.error-budget", "relation": "determines", "target": "mlsys.alerting", "statement": "sets burn-rate thresholds for", "conditions": "Alerts fire when the budget is being consumed faster than the window allows." }, { "source": "mlsys.latency-budget", "relation": "limits", "target": "mlsys.reasoning-budget", "statement": "caps the intermediate tokens allowed by", "conditions": "Interactive requests must answer within a fixed end-to-end latency target, so extra reasoning tokens consume the same budget." }, { "source": "mlsys.activation-recompute", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "reduces the peak", "conditions": "Retained activations dominate training memory and spare compute is available to recompute them." }, { "source": "mlsys.kv-cache", "relation": "increases", "target": "mlsys.memory-footprint", "statement": "adds per-request state to", "conditions": "Autoregressive serving retains keys and values for every in-flight token; the term grows with batch size and context length." }, { "source": "mlsys.workload-trace", "relation": "measures", "target": "mlsys.arrival-burstiness", "statement": "reveals the", "conditions": "Inter-arrival times are recorded at fine enough resolution to show bursts rather than only averages." }, { "source": "mlsys.performance-simulation", "relation": "uses", "target": "mlsys.workload-trace", "statement": "replays", "conditions": "Trace-driven simulation is used to compare scheduling, capacity, or serving policies on realistic demand." }, { "source": "mlsys.cost-sensitive-decision", "relation": "requires", "target": "mlsys.calibration", "statement": "requires calibrated probabilities from", "conditions": "Expected-cost thresholds are computed from predicted probabilities, which must match observed frequencies on the deployment distribution." }, { "source": "mlsys.cost-sensitive-decision", "relation": "determines", "target": "mlsys.selective-prediction", "statement": "sets the abstention threshold in", "conditions": "Deferring to a human or fallback has a known cost that can be compared with the expected cost of each error type." }, { "source": "mlsys.tokenization", "relation": "determines", "target": "mlsys.context", "statement": "sets how much text fits in", "conditions": "For a fixed window, a tokenizer that produces more tokens per word admits less source text." }, { "source": "mlsys.linear-model", "relation": "alternative-to", "target": "mlsys.tree-ensemble", "statement": "is an alternative to", "conditions": "Both are standard choices for tabular supervised prediction; trees capture interactions linear scores need engineered features for." }, { "source": "mlsys.tree-ensemble", "relation": "uses", "target": "mlsys.features", "statement": "splits on", "conditions": "Trees operate on tabular features without requiring scaling; feature engineering still determines what interactions are available." }, { "source": "mlsys.kernel-method", "relation": "increases", "target": "mlsys.memory-footprint", "statement": "has a quadratic training", "conditions": "Exact training forms or stores an n-by-n kernel matrix, so memory grows with the square of the training set size unless approximated." }, { "source": "mlsys.kernel-method", "relation": "determines", "target": "mlsys.inductive-bias", "statement": "encodes its inductive bias through the choice of kernel in", "conditions": "The kernel function fixes which similarity structure the model can exploit." }, { "source": "mlsys.nearest-neighbor", "relation": "uses", "target": "mlsys.embeddings", "statement": "searches over", "conditions": "Neighbors are found by distance in a learned or engineered vector space whose quality determines result quality." }, { "source": "mlsys.neural-network", "relation": "uses", "target": "mlsys.matrix-multiply", "statement": "spends most compute in", "conditions": "Dense and attention layers dominate; small or irregular layers can be bound by other operations." }, { "source": "mlsys.convolution", "relation": "implements", "target": "mlsys.inductive-bias", "statement": "encodes translation equivariance and locality as an", "conditions": "Weights are shared across positions, so the model assumes local patterns recur across the input." }, { "source": "mlsys.position-encoding", "relation": "limits", "target": "mlsys.context", "statement": "bounds the usable length of", "conditions": "Without extension, sequences longer than the positions seen in training degrade quality." }, { "source": "mlsys.transformer", "relation": "uses", "target": "mlsys.normalization-layer", "statement": "stabilizes residual blocks with", "conditions": "Each block applies layer or RMS normalization around attention and feed-forward sublayers." }, { "source": "mlsys.transformer", "relation": "uses", "target": "mlsys.position-encoding", "statement": "injects order with", "conditions": "Attention alone is permutation equivariant, so the architecture adds absolute, relative, or rotary positions." }, { "source": "mlsys.transformer", "relation": "specializes", "target": "mlsys.neural-network", "statement": "is a kind of", "conditions": "The architecture is a differentiable composition trained by gradient descent." }, { "source": "mlsys.ranking-funnel", "relation": "uses", "target": "mlsys.recommendation-model", "statement": "runs in its ranking stages a", "conditions": "Candidate generation narrows the set before costly deep ranking models score it within the latency budget." }, { "source": "mlsys.depthwise-separable-convolution", "relation": "specializes", "target": "mlsys.convolution", "statement": "is a factorized kind of", "conditions": "A standard convolution is split into depthwise and pointwise steps." }, { "source": "mlsys.depthwise-separable-convolution", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "cuts parameter storage in", "conditions": "Parameter memory is a significant share of the deployed footprint, as on mobile and embedded targets." }, { "source": "mlsys.recurrence", "relation": "alternative-to", "target": "mlsys.attention", "statement": "is an alternative to", "conditions": "Both mix information across sequence positions; recurrence keeps fixed state but serializes time, attention parallelizes but grows cost with length." }, { "source": "mlsys.normalization-layer", "relation": "mitigates", "target": "mlsys.training-instability", "statement": "mitigates", "conditions": "Deep networks whose activation scales would otherwise drift with depth or batch composition." }, { "source": "mlsys.expert-routing", "relation": "part-of", "target": "mlsys.mixture-of-experts", "statement": "is the gating component of", "conditions": "The layer routes each token to a small subset of experts." }, { "source": "mlsys.expert-capacity", "relation": "part-of", "target": "mlsys.mixture-of-experts", "statement": "is a buffer-size parameter of", "conditions": "Experts process fixed-size per-batch buffers so computation shapes stay static." }, { "source": "mlsys.expert-capacity", "relation": "increases", "target": "mlsys.all-to-all", "statement": "increases padded token volume in", "conditions": "A higher capacity factor allocates larger per-expert buffers that are exchanged whether or not they are full." }, { "source": "mlsys.load-balancing-loss", "relation": "mitigates", "target": "mlsys.stragglers", "statement": "reduces hot-expert imbalance behind", "conditions": "Under expert parallelism, an overloaded expert's device finishes last and holds back the step." }, { "source": "mlsys.load-balancing-loss", "relation": "part-of", "target": "mlsys.mixture-of-experts", "statement": "is an auxiliary objective in training of", "conditions": "Routing is learned and can otherwise collapse onto a few experts." }, { "source": "mlsys.active-parameters", "relation": "determines", "target": "mlsys.training-compute-estimate", "statement": "sets the parameter count used in", "conditions": "For conditionally computed models, the six-operations-per-parameter-per-token estimate uses active rather than total parameters." }, { "source": "mlsys.early-exit", "relation": "requires", "target": "mlsys.calibration", "statement": "relies on confidence quality assessed by", "conditions": "Exit decisions threshold an intermediate confidence; overconfident exits return wrong answers early." }, { "source": "mlsys.early-exit", "relation": "reduces", "target": "mlsys.energy-per-inference", "statement": "reduces average", "conditions": "Many inputs are easy enough to exit at an early classifier and the exit overhead is small." }, { "source": "mlsys.sliding-window-attention", "relation": "specializes", "target": "mlsys.attention", "statement": "is a restricted kind of", "conditions": "Each position attends only within a window or fixed sparse pattern." }, { "source": "mlsys.sliding-window-attention", "relation": "reduces", "target": "mlsys.kv-cache", "statement": "caps the size of", "conditions": "During decoding only the most recent window of keys and values must be retained." }, { "source": "mlsys.linear-attention", "relation": "specializes", "target": "mlsys.attention", "statement": "is a kernelized variant of", "conditions": "Softmax similarity is replaced by a feature-map inner product, changing model quality." }, { "source": "mlsys.linear-attention", "relation": "alternative-to", "target": "mlsys.sliding-window-attention", "statement": "is an alternative to", "conditions": "Both reduce attention cost growth for long sequences, trading different quality losses." }, { "source": "mlsys.state-space-model", "relation": "alternative-to", "target": "mlsys.attention", "statement": "is an alternative to", "conditions": "Both serve as the sequence-mixing layer; state-space layers trade exact long-range lookup for linear cost and fixed state." }, { "source": "mlsys.state-space-model", "relation": "reduces", "target": "mlsys.kv-cache", "statement": "eliminates the growing", "conditions": "Inference carries a constant-size state per layer instead of keys and values for every prior token." }, { "source": "mlsys.grouped-query-attention", "relation": "specializes", "target": "mlsys.attention", "statement": "is a kind of", "conditions": "Multiple query heads share each key-value head." }, { "source": "mlsys.grouped-query-attention", "relation": "reduces", "target": "mlsys.tpot", "statement": "reduces", "conditions": "Decode is bound by reading the KV cache, so fewer key-value heads cut bytes read per token." }, { "source": "mlsys.context-extension", "relation": "increases", "target": "mlsys.context", "statement": "extends the usable length of", "conditions": "The model is fine-tuned briefly on longer sequences after rescaling positions; quality at the new length must be checked." }, { "source": "mlsys.weight-serialization", "relation": "determines", "target": "mlsys.unsafe-deserialization", "statement": "determines exposure to", "conditions": "Executable formats such as pickle run code on load; data-only formats such as safetensors do not." }, { "source": "mlsys.weight-serialization", "relation": "determines", "target": "mlsys.cold-start", "statement": "sets weight-loading time in", "conditions": "Memory-mappable, zero-copy formats load faster than formats requiring full deserialization." }, { "source": "mlsys.model-conversion", "relation": "uses", "target": "mlsys.model-exchange-format", "statement": "commonly passes through a", "conditions": "Training and target runtimes differ and both support the exchange format's operator set." }, { "source": "mlsys.model-export", "relation": "produces", "target": "mlsys.model-artifact", "statement": "produces a deployable", "conditions": "The artifact represents supported operators, parameters, and input constraints for its target runtime." }, { "source": "mlsys.model-export", "relation": "uses", "target": "mlsys.tracing-jit", "statement": "can obtain its graph through", "conditions": "In export workflows that obtain a graph by tracing, such as JAX export; tracing and lowering may precede execution on another machine." }, { "source": "mlsys.model-signature", "relation": "part-of", "target": "mlsys.model-artifact", "statement": "is the interface declared in", "conditions": "The artifact records input and output names, types, and shapes." }, { "source": "mlsys.model-server", "relation": "requires", "target": "mlsys.model-signature", "statement": "validates requests against", "conditions": "The server decodes and checks requests using the artifact's declared inputs and outputs." }, { "source": "mlsys.vla-model", "relation": "specializes", "target": "mlsys.multimodal-model", "statement": "is a kind of", "conditions": "The model consumes images and language and emits actions through a shared backbone." }, { "source": "mlsys.multimodal-model", "relation": "uses", "target": "mlsys.transformer", "statement": "commonly uses as its shared backbone a", "conditions": "Encoded modality tokens are processed together with text tokens by a transformer." }, { "source": "mlsys.encoder-disaggregation", "relation": "uses", "target": "mlsys.modality-encoder", "statement": "schedules separately the", "conditions": "Image or audio encoding is heavy enough to stall token generation if co-scheduled." }, { "source": "mlsys.context-compaction", "relation": "mitigates", "target": "mlsys.context-degradation", "statement": "mitigates", "conditions": "Older material is summarized so relevant information stays short and near the end of the window." }, { "source": "mlsys.reranking", "relation": "mitigates", "target": "mlsys.context-degradation", "statement": "mitigates position effects behind", "conditions": "Reranked evidence is placed where models use it best, near the start or end of the context." }, { "source": "mlsys.autodiff", "relation": "increases", "target": "mlsys.memory-footprint", "statement": "retains forward activations, increasing", "conditions": "Reverse-mode differentiation stores intermediates from the forward pass for use in the backward pass." }, { "source": "mlsys.tensor-layout", "relation": "determines", "target": "mlsys.memory-coalescing", "statement": "determines whether accesses achieve", "conditions": "Strides decide whether adjacent threads touch adjacent addresses." }, { "source": "mlsys.mixed-precision", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "reduces activation and gradient bytes in", "conditions": "Activations and gradients are stored in 16-bit formats while a master copy of weights stays in higher precision." }, { "source": "mlsys.mixed-precision", "relation": "uses", "target": "mlsys.matrix-unit", "statement": "gains throughput from", "conditions": "The matrix engine supports the reduced input formats with wider accumulation." }, { "source": "mlsys.computation-graph", "relation": "uses", "target": "mlsys.tensor", "statement": "passes between operations a", "conditions": "Edges carry tensors whose shapes and types the graph records." }, { "source": "mlsys.tensor", "relation": "uses", "target": "mlsys.precision", "statement": "carries an element type drawn from", "conditions": "Each tensor records a numerical format that kernels must support." }, { "source": "mlsys.compilation", "relation": "uses", "target": "mlsys.computation-graph", "statement": "takes as input a", "conditions": "The compiler needs a captured graph rather than one operation at a time." }, { "source": "mlsys.eager-execution", "relation": "determines", "target": "mlsys.fusion", "statement": "forecloses cross-operation", "conditions": "Each operation is launched as the host issues it, so no compiler sees neighboring operations." }, { "source": "mlsys.io-aware-attention", "relation": "implements", "target": "mlsys.attention", "statement": "exactly computes", "conditions": "The result matches standard softmax attention up to floating-point reordering." }, { "source": "mlsys.io-aware-attention", "relation": "uses", "target": "mlsys.tiling", "statement": "is built on", "conditions": "Query, key, and value blocks are staged through on-chip memory with an online softmax." }, { "source": "mlsys.convolution-lowering", "relation": "increases", "target": "mlsys.memory-footprint", "statement": "can increase, through duplicated patches,", "conditions": "im2col-style lowering copies overlapping input patches into a larger matrix." }, { "source": "mlsys.kernel-library", "relation": "implements", "target": "mlsys.matrix-multiply", "statement": "provides tuned", "conditions": "Vendor libraries supply GEMM and convolution kernels selected per shape and precision." }, { "source": "mlsys.kernel-library", "relation": "alternative-to", "target": "mlsys.kernel-dsl", "statement": "is an alternative to", "conditions": "Both supply operator kernels; libraries cover common shapes, DSLs let engineers write fused or novel kernels." }, { "source": "mlsys.compilation", "relation": "uses", "target": "mlsys.intermediate-representation", "statement": "lowers through", "conditions": "Graph, tensor, and loop-level representations carry the program between passes." }, { "source": "mlsys.sharding-propagation", "relation": "uses", "target": "mlsys.intermediate-representation", "statement": "propagates layouts through", "conditions": "Sharding annotations are attached to operations in the compiler IR." }, { "source": "mlsys.autotuning", "relation": "uses", "target": "mlsys.profiling", "statement": "measures candidates with", "conditions": "Candidate schedules are compiled and timed on the target, or a cost model is trained from such measurements." }, { "source": "mlsys.autotuning", "relation": "part-of", "target": "mlsys.compilation", "statement": "is a search stage within", "conditions": "The compiler exposes a schedule space whose performance is hard to predict analytically." }, { "source": "mlsys.kernel-dsl", "relation": "uses", "target": "mlsys.on-chip-sram", "statement": "lets the compiler manage", "conditions": "Programs are written over blocks while the compiler allocates shared memory and registers." }, { "source": "mlsys.kernel-dsl", "relation": "enables", "target": "mlsys.fusion", "statement": "enables custom", "conditions": "Engineers can write a single kernel that combines operations no library provides fused." }, { "source": "mlsys.memory-planning", "relation": "part-of", "target": "mlsys.compilation", "statement": "is a pass within", "conditions": "Static shapes let the compiler compute tensor lifetimes ahead of execution." }, { "source": "mlsys.memory-planning", "relation": "alternative-to", "target": "mlsys.caching-allocator", "statement": "is an alternative to", "conditions": "Both manage device buffers; static planning needs known lifetimes, caching allocation handles dynamic shapes." }, { "source": "mlsys.device-streams", "relation": "reduces", "target": "mlsys.dispatch", "statement": "hides, through asynchronous enqueue, exposed", "conditions": "The host can run ahead of the device, so launch work overlaps device execution until queues drain." }, { "source": "mlsys.graph-replay", "relation": "uses", "target": "mlsys.device-streams", "statement": "is captured from and replayed on", "conditions": "Work enqueued on a stream is recorded as a graph and replayed as one launch." }, { "source": "mlsys.caching-allocator", "relation": "reduces", "target": "mlsys.dispatch", "statement": "avoids synchronizing allocation calls in", "conditions": "Freed blocks are reused instead of calling the device allocator, which can synchronize the host." }, { "source": "mlsys.graph-replay", "relation": "reduces", "target": "mlsys.tpot", "statement": "reduces per-step launch overhead in", "conditions": "Decode steps launch many small kernels with static shapes, so host overhead is a large share of step time." }, { "source": "mlsys.loss-scaling", "relation": "increases", "target": "mlsys.numerical-stability", "statement": "keeps small gradients representable, improving", "conditions": "Reduced-precision gradient formats would otherwise underflow; overflow is detected and the step skipped." }, { "source": "mlsys.block-scaled-format", "relation": "mitigates", "target": "mlsys.activation-outliers", "statement": "limits the range damage from", "conditions": "A shared scale per small block confines an outlier's effect to its own block." }, { "source": "mlsys.quantization", "relation": "uses", "target": "mlsys.block-scaled-format", "statement": "can encode values in", "conditions": "Hardware or kernels support the block-scaled element and scale formats." }, { "source": "mlsys.matrix-unit", "relation": "determines", "target": "mlsys.accumulation-precision", "statement": "fixes the available", "conditions": "The matrix engine implements specific accumulation formats for each input format." }, { "source": "mlsys.function-transformations", "relation": "uses", "target": "mlsys.compilation", "statement": "exposes as a transform", "conditions": "Pure array functions are traced and handed to an XLA-style compiler." }, { "source": "mlsys.single-vs-multi-controller", "relation": "determines", "target": "mlsys.dispatch", "statement": "sets where per-step dispatch cost falls in", "conditions": "One controller issuing work to many devices must keep dispatch asynchronous and off the critical path." }, { "source": "mlsys.thread-hierarchy", "relation": "uses", "target": "mlsys.on-chip-sram", "statement": "shares within a thread block", "conditions": "Threads of a block cooperate through shared memory sized per block." }, { "source": "mlsys.thread-hierarchy", "relation": "determines", "target": "mlsys.memory-coalescing", "statement": "determines, through thread-to-address mapping,", "conditions": "Adjacent threads of a warp should access adjacent addresses." }, { "source": "mlsys.on-chip-sram", "relation": "limits", "target": "mlsys.occupancy", "statement": "bounds", "conditions": "Shared memory and registers per block limit how many warps can be resident." }, { "source": "mlsys.thread-hierarchy", "relation": "determines", "target": "mlsys.occupancy", "statement": "sets, through block size and per-thread resources,", "conditions": "Block shape and register or shared-memory use determine resident warps per multiprocessor." }, { "source": "mlsys.async-data-movement", "relation": "uses", "target": "mlsys.on-chip-sram", "statement": "stages operands into", "conditions": "Bulk copies land in shared memory while compute proceeds on the previous buffer." }, { "source": "mlsys.async-data-movement", "relation": "increases", "target": "mlsys.utilization", "statement": "raises matrix-engine", "conditions": "The kernel would otherwise stall its matrix units waiting on loads." }, { "source": "mlsys.matrix-engine-programming", "relation": "uses", "target": "mlsys.matrix-unit", "statement": "issues instructions to", "conditions": "Fragments are loaded in the shapes and layouts the unit accepts." }, { "source": "mlsys.matrix-engine-programming", "relation": "requires", "target": "mlsys.tensor-layout", "statement": "requires specific operand", "conditions": "Matrix instructions accept only particular fragment layouts, so data may need rearrangement." }, { "source": "mlsys.batch-invariance", "relation": "enables", "target": "mlsys.experiment-reproducibility", "statement": "enables bitwise-reproducible serving within", "conditions": "Kernels use fixed reduction strategies independent of batch size and co-batched requests." }, { "source": "mlsys.systolic-array", "relation": "specializes", "target": "mlsys.accelerator-execution-architecture", "statement": "is a kind of", "conditions": "Processing elements pass operands to neighbors under a fixed schedule." }, { "source": "mlsys.simt", "relation": "specializes", "target": "mlsys.accelerator-execution-architecture", "statement": "is a kind of", "conditions": "Thread groups issue the same instruction and divergent paths serialize." }, { "source": "mlsys.matrix-unit", "relation": "part-of", "target": "mlsys.accelerator-execution-architecture", "statement": "is a functional unit within", "conditions": "The accelerator pairs matrix engines with vector or scalar units." }, { "source": "mlsys.hbm", "relation": "part-of", "target": "mlsys.memory-hierarchy", "statement": "is the device-memory level of", "conditions": "The accelerator uses stacked DRAM as off-chip device memory." }, { "source": "mlsys.on-chip-sram", "relation": "part-of", "target": "mlsys.memory-hierarchy", "statement": "is the fastest managed level of", "conditions": "The platform exposes caches, shared memory, or scratchpads near compute." }, { "source": "mlsys.host-device-interconnect", "relation": "limits", "target": "mlsys.memory-offloading", "statement": "bounds the throughput of", "conditions": "Offloaded optimizer state, parameters, or activations cross the host link every step." }, { "source": "mlsys.host-device-interconnect", "relation": "limits", "target": "mlsys.kv-cache-offload", "statement": "bounds reload rate in", "conditions": "Offloaded KV blocks return to device memory over the host link on reuse." }, { "source": "mlsys.power-delivery", "relation": "limits", "target": "mlsys.rack-power-density", "statement": "bounds", "conditions": "Facility and rack feeds, conversion stages, and breakers limit sustained draw per rack." }, { "source": "mlsys.power-delivery", "relation": "causes", "target": "mlsys.silent-data-corruption", "statement": "can cause, through insufficient voltage margin,", "conditions": "Droop or marginal supply under heavy load leaves circuits below the voltage they need for correct timing." }, { "source": "mlsys.systolic-array", "relation": "reduces", "target": "mlsys.data-movement-energy", "statement": "reduces", "conditions": "Operands are reused across processing elements instead of being refetched from register files or memory." }, { "source": "mlsys.dataflow", "relation": "determines", "target": "mlsys.locality", "statement": "determines", "conditions": "Local storage cannot hold every operand, so the schedule chooses which data is reused in place." }, { "source": "mlsys.systolic-array", "relation": "uses", "target": "mlsys.dataflow", "statement": "operates under a chosen", "conditions": "The array keeps one operand stationary, such as weights, while others stream." }, { "source": "mlsys.domain-specific-architecture", "relation": "increases", "target": "mlsys.energy-efficiency", "statement": "raises", "conditions": "Target operations dominate the workload and are supported by the specialized hardware." }, { "source": "mlsys.hbm", "relation": "determines", "target": "mlsys.memory-capacity", "statement": "sets device", "conditions": "Stack count and stack height fix the capacity attached to each accelerator." }, { "source": "mlsys.hbm", "relation": "mitigates", "target": "mlsys.memory-wall", "statement": "narrows, without closing,", "conditions": "Stacking and wide interfaces raise bandwidth, though compute still grows faster across generations." }, { "source": "mlsys.memory-wall", "relation": "reduces", "target": "mlsys.model-flops-utilization", "statement": "depresses", "conditions": "Memory-bound phases such as decode cannot feed the arithmetic units at peak rate." }, { "source": "mlsys.scale-up-interconnect", "relation": "part-of", "target": "mlsys.network-topology", "statement": "forms the fastest tier of", "conditions": "The cluster topology is hierarchical, with a node or rack domain joined by a slower scale-out fabric." }, { "source": "mlsys.coherent-interconnect", "relation": "specializes", "target": "mlsys.host-device-interconnect", "statement": "is a kind of", "conditions": "The link adds hardware coherence and load-store memory semantics over a host-device physical layer." }, { "source": "mlsys.coherent-interconnect", "relation": "enables", "target": "mlsys.kv-cache-offload", "statement": "can provide a pooled tier for", "conditions": "Attached or pooled memory is cheaper than device memory and its higher latency can be overlapped with compute." }, { "source": "mlsys.kv-cache-transfer", "relation": "uses", "target": "mlsys.peer-to-peer-transfer", "statement": "moves state with", "conditions": "Direct device or NIC transfers avoid staging KV blocks through host memory." }, { "source": "mlsys.rdma", "relation": "enables", "target": "mlsys.peer-to-peer-transfer", "statement": "can read and write device memory through", "conditions": "The NIC and driver support direct access to accelerator memory." }, { "source": "mlsys.tdp", "relation": "determines", "target": "mlsys.rack-power-density", "statement": "sets, with accelerators per rack,", "conditions": "Racks are provisioned for the sustained power of the chips they hold." }, { "source": "mlsys.tdp", "relation": "determines", "target": "mlsys.dark-silicon", "statement": "sets the fraction of", "conditions": "Transistor counts grow while the power envelope does not." }, { "source": "mlsys.power-capping", "relation": "uses", "target": "mlsys.dvfs", "statement": "enforces caps through", "conditions": "The device lowers clocks and voltage when measured power exceeds the cap." }, { "source": "mlsys.thermal-dynamics", "relation": "uses", "target": "mlsys.dvfs", "statement": "throttles through", "conditions": "Thermal controls lower frequency and voltage when temperature limits are approached." }, { "source": "mlsys.microcontroller", "relation": "uses", "target": "mlsys.on-device-runtime", "statement": "runs models through a lightweight", "conditions": "No operating system or virtual memory is available, so the runtime is statically linked and memory-planned." }, { "source": "mlsys.edge-npu", "relation": "part-of", "target": "mlsys.heterogeneous-soc", "statement": "is one block of", "conditions": "The SoC exposes the NPU alongside CPU, GPU, and DSP under shared memory and power." }, { "source": "mlsys.on-device-runtime", "relation": "uses", "target": "mlsys.heterogeneous-soc", "statement": "delegates operators across", "conditions": "The runtime partitions supported operators onto the NPU, GPU, or DSP and falls back to CPU." }, { "source": "mlsys.edge-npu", "relation": "reduces", "target": "mlsys.energy-per-inference", "statement": "reduces", "conditions": "The model's operators and precisions are supported, so inference avoids CPU fallback." }, { "source": "mlsys.hardware-aware-nas", "relation": "uses", "target": "mlsys.on-device-memory-budget", "statement": "searches within", "conditions": "The search treats flash and SRAM limits as hard constraints on candidate models." }, { "source": "mlsys.on-device-memory-budget", "relation": "limits", "target": "mlsys.on-device-training", "statement": "bounds", "conditions": "Training must fit weights, activations, and update state in on-chip memory." }, { "source": "mlsys.recommendation-model", "relation": "uses", "target": "mlsys.embedding-cache", "statement": "serves hot rows from an", "conditions": "Lookups are skewed so a small set of rows receives most accesses." }, { "source": "mlsys.embedding-cache", "relation": "specializes", "target": "mlsys.storage-tiering", "statement": "is a kind of", "conditions": "Rows are placed across accelerator memory, host memory, and SSD by access frequency." }, { "source": "mlsys.pipeline-parallel", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "divides per-device parameter", "conditions": "Each stage holds only its layers' parameters and their state." }, { "source": "mlsys.pipeline-parallel", "relation": "part-of", "target": "mlsys.hybrid-parallel", "statement": "is one dimension of", "conditions": "Pipeline stages are combined with data and tensor parallelism across nodes." }, { "source": "mlsys.state-sharding", "relation": "uses", "target": "mlsys.collectives", "statement": "gathers and scatters state through", "conditions": "Sharded parameters are all-gathered before use and gradients reduce-scattered after." }, { "source": "mlsys.state-sharding", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "divides per-worker", "conditions": "Parameters, gradients, or optimizer state are partitioned instead of replicated." }, { "source": "mlsys.fat-tree", "relation": "specializes", "target": "mlsys.network-topology", "statement": "is a kind of", "conditions": "The fabric is built from tiers of commodity switches." }, { "source": "mlsys.rail-optimized", "relation": "specializes", "target": "mlsys.network-topology", "statement": "is a kind of", "conditions": "Same-index accelerators across nodes share a leaf switch." }, { "source": "mlsys.expert-parallel", "relation": "part-of", "target": "mlsys.hybrid-parallel", "statement": "is one dimension of", "conditions": "MoE layers place experts across devices alongside data and tensor parallelism." }, { "source": "mlsys.sequence-parallel", "relation": "part-of", "target": "mlsys.hybrid-parallel", "statement": "is one dimension of", "conditions": "Long-sequence training adds a sequence axis to the parallel mapping." }, { "source": "mlsys.sequence-parallel", "relation": "uses", "target": "mlsys.collectives", "statement": "exchanges keys and values through", "conditions": "Blocks of keys and values circulate among devices while attention is computed." }, { "source": "mlsys.gradient-accumulation", "relation": "reduces", "target": "mlsys.pipeline-bubble", "statement": "shrinks, through more microbatches,", "conditions": "More microbatches per optimizer step fill the pipeline relative to the number of stages." }, { "source": "mlsys.pipeline-bubble", "relation": "reduces", "target": "mlsys.model-flops-utilization", "statement": "lowers", "conditions": "Devices idle during fill and drain while their peak FLOPs still count in the denominator." }, { "source": "mlsys.tree-allreduce", "relation": "implements", "target": "mlsys.collectives", "statement": "is an algorithm for", "conditions": "Participants are arranged in trees or hierarchical groups." }, { "source": "mlsys.all-to-all", "relation": "specializes", "target": "mlsys.collectives", "statement": "is a kind of", "conditions": "Every participant sends distinct data to every other." }, { "source": "mlsys.alpha-beta-model", "relation": "measures", "target": "mlsys.tree-allreduce", "statement": "predicts the cost of", "conditions": "Link latency and bandwidth are roughly uniform and contention is ignored." }, { "source": "mlsys.gradient-compression", "relation": "reduces", "target": "mlsys.time-to-train", "statement": "can reduce", "conditions": "Training is communication-bound and compression preserves convergence at the chosen ratio." }, { "source": "mlsys.gradient-compression", "relation": "uses", "target": "mlsys.sparsity", "statement": "can transmit only top-magnitude gradients, using", "conditions": "Residuals are accumulated locally so dropped values are sent later." }, { "source": "mlsys.leader-election", "relation": "uses", "target": "mlsys.lease", "statement": "holds leadership through a", "conditions": "The leader renews a time-bounded grant so a failed leader's authority lapses." }, { "source": "mlsys.lease", "relation": "enables", "target": "mlsys.replication-failover", "statement": "enables timed", "conditions": "A standby can take over once the failed holder's lease has expired." }, { "source": "mlsys.leader-election", "relation": "requires", "target": "mlsys.fencing", "statement": "requires", "conditions": "A deposed leader may still act; resources must reject its stale epoch." }, { "source": "mlsys.federated-averaging", "relation": "uses", "target": "mlsys.secure-aggregation", "statement": "can combine updates through", "conditions": "The server must not see individual client updates." }, { "source": "mlsys.client-sampling", "relation": "part-of", "target": "mlsys.federated-averaging", "statement": "is the per-round selection step of", "conditions": "Each round samples a subset of available clients." }, { "source": "mlsys.federation-setting", "relation": "determines", "target": "mlsys.device-heterogeneity", "statement": "determines exposure to", "conditions": "Cross-device federations span varied hardware and runtimes; cross-silo ones are few and uniform." }, { "source": "mlsys.federation-setting", "relation": "determines", "target": "mlsys.failure-model", "statement": "determines the", "conditions": "Cross-device clients drop out routinely, while cross-silo participants are expected to stay available." }, { "source": "mlsys.embedding-table-sharding", "relation": "part-of", "target": "mlsys.hybrid-parallel", "statement": "is the model-parallel part of", "conditions": "Dense layers are data-parallel while embedding tables are partitioned." }, { "source": "mlsys.parameter-server", "relation": "alternative-to", "target": "mlsys.ring-allreduce", "statement": "is an alternative to", "conditions": "Both aggregate gradients for data-parallel training; servers suit sparse or asynchronous updates, all-reduce suits dense synchronous ones." }, { "source": "mlsys.recommendation-model", "relation": "uses", "target": "mlsys.parameter-server", "statement": "commonly trains sparse tables through a", "conditions": "Updates touch only the rows looked up in each batch." }, { "source": "mlsys.auto-parallelization", "relation": "uses", "target": "mlsys.analytical-performance-model", "statement": "ranks plans with", "conditions": "The search evaluates candidate partitionings with a cost model of compute and communication." }, { "source": "mlsys.network-topology", "relation": "determines", "target": "mlsys.device-mesh", "statement": "determines the axis alignment of", "conditions": "The busiest parallel axis is mapped onto the fastest links." }, { "source": "mlsys.hybrid-parallel", "relation": "uses", "target": "mlsys.device-mesh", "statement": "assigns parallel dimensions to the axes of", "conditions": "Each parallelism dimension corresponds to a named mesh axis." }, { "source": "mlsys.pue", "relation": "measures", "target": "mlsys.warehouse-scale-computer", "statement": "measures facility overhead of", "conditions": "IT and total facility energy are metered over the same period." }, { "source": "mlsys.liquid-cooling", "relation": "reduces", "target": "mlsys.pue", "statement": "can reduce", "conditions": "Cooling energy is a large share of facility overhead and liquid loops need less fan and chiller work." }, { "source": "mlsys.liquid-cooling", "relation": "mitigates", "target": "mlsys.thermal-dynamics", "statement": "mitigates throttling in", "conditions": "Air cooling cannot hold chips below their thermal limits at the target power." }, { "source": "mlsys.power-oversubscription", "relation": "reduces", "target": "mlsys.tco", "statement": "reduces", "conditions": "More servers share provisioned power capacity because their peaks rarely coincide." }, { "source": "mlsys.warehouse-scale-computer", "relation": "uses", "target": "mlsys.power-oversubscription", "statement": "provisions power through", "conditions": "Fleet-wide power measurements show aggregate peaks below summed nameplate." }, { "source": "mlsys.warehouse-scale-computer", "relation": "uses", "target": "mlsys.tco", "statement": "is designed against the cost model of", "conditions": "When facility, hardware, and software choices are evaluated as one cost model over the fleet's lifetime." }, { "source": "mlsys.lossless-fabric", "relation": "requires", "target": "mlsys.congestion-control", "statement": "requires", "conditions": "Priority flow control alone spreads congestion and pauses; end-to-end rate control keeps queues short." }, { "source": "mlsys.lossless-fabric", "relation": "increases", "target": "mlsys.tail-latency", "statement": "can inflate", "conditions": "Pause frames cause head-of-line blocking that delays unrelated flows under congestion." }, { "source": "mlsys.congestion-control", "relation": "mitigates", "target": "mlsys.performance-interference", "statement": "mitigates network", "conditions": "Jobs share switch queues on a common fabric." }, { "source": "mlsys.rail-optimized", "relation": "uses", "target": "mlsys.scale-up-interconnect", "statement": "relies for cross-rail traffic on", "conditions": "Traffic between different rails is forwarded through the node's scale-up domain." }, { "source": "mlsys.in-network-reduction", "relation": "reduces", "target": "mlsys.collectives", "statement": "reduces link traffic and latency of", "conditions": "Switches support aggregation for the data types and reduction operations used." }, { "source": "mlsys.in-network-reduction", "relation": "alternative-to", "target": "mlsys.ring-allreduce", "statement": "is an alternative to", "conditions": "Both perform all-reduce; switch aggregation shortens paths while rings need only end-host support." }, { "source": "mlsys.parallel-file-system", "relation": "alternative-to", "target": "mlsys.object-storage", "statement": "is an alternative to", "conditions": "Both serve shared training data and checkpoints; file systems offer lower latency, object stores cheaper scalable capacity." }, { "source": "mlsys.checkpoint-io", "relation": "uses", "target": "mlsys.parallel-file-system", "statement": "commonly writes to a", "conditions": "Sharded checkpoints are written concurrently by many ranks." }, { "source": "mlsys.object-storage", "relation": "determines", "target": "mlsys.dataset-storage-layout", "statement": "favors large sequential shards in", "conditions": "Per-request latency is high, so many small reads are slow." }, { "source": "mlsys.async-checkpointing", "relation": "reduces", "target": "mlsys.checkpoint-io", "statement": "hides the training stall from", "conditions": "A fast snapshot to host memory lets persistence proceed in the background." }, { "source": "mlsys.checkpoint-io", "relation": "determines", "target": "mlsys.checkpoint-interval", "statement": "sets the write-cost input to", "conditions": "The optimal interval scales with the square root of checkpoint cost times MTBF." }, { "source": "mlsys.data-parallel", "relation": "requires", "target": "mlsys.gang-scheduling", "statement": "requires", "conditions": "Synchronous replicas cannot start or step until every worker holds its allocation." }, { "source": "mlsys.gang-scheduling", "relation": "increases", "target": "mlsys.queueing", "statement": "can increase job", "conditions": "Large jobs wait until enough resources are free at once." }, { "source": "mlsys.bin-packing", "relation": "increases", "target": "mlsys.utilization", "statement": "raises", "conditions": "Jobs can share nodes without violating resource or isolation constraints." }, { "source": "mlsys.bin-packing", "relation": "increases", "target": "mlsys.performance-interference", "statement": "can increase", "conditions": "Dense co-location puts jobs on shared caches, memory bandwidth, and links." }, { "source": "mlsys.multi-tenancy", "relation": "uses", "target": "mlsys.gpu-partitioning", "statement": "can share devices through", "conditions": "Tenant workloads are too small to saturate whole accelerators." }, { "source": "mlsys.resource-quota", "relation": "part-of", "target": "mlsys.multi-tenancy", "statement": "is a policy within", "conditions": "Tenants share a cluster under stated limits and guarantees." }, { "source": "mlsys.resource-quota", "relation": "alternative-to", "target": "mlsys.dominant-resource-fairness", "statement": "is an alternative to", "conditions": "Both divide shared capacity among tenants; quotas are static limits, DRF adapts shares to demand." }, { "source": "mlsys.gpu-partitioning", "relation": "mitigates", "target": "mlsys.performance-interference", "statement": "mitigates, with hardware partitions,", "conditions": "Partitions give each workload dedicated memory and compute slices; time-slicing does not." }, { "source": "mlsys.elastic-training", "relation": "enables", "target": "mlsys.spot-capacity", "statement": "makes usable", "conditions": "Training continues while preempted workers leave and replacements join." }, { "source": "mlsys.ridge-point", "relation": "part-of", "target": "mlsys.roofline", "statement": "is the corner of", "conditions": "Both ceilings refer to the same device and memory level." }, { "source": "mlsys.memory-wall", "relation": "increases", "target": "mlsys.ridge-point", "statement": "moves to higher intensity the", "conditions": "Peak compute grows faster than bandwidth across hardware generations." }, { "source": "mlsys.parallel-efficiency", "relation": "measures", "target": "mlsys.strong-scaling", "statement": "measures", "conditions": "Problem size is fixed while resources grow." }, { "source": "mlsys.parallel-efficiency", "relation": "measures", "target": "mlsys.weak-scaling", "statement": "measures", "conditions": "Work per resource is fixed while resources and problem size grow." }, { "source": "mlsys.weak-scaling", "relation": "alternative-to", "target": "mlsys.strong-scaling", "statement": "is an alternative to", "conditions": "Both describe how to use added resources; one grows the problem, the other fixes it." }, { "source": "mlsys.critical-batch-size", "relation": "limits", "target": "mlsys.weak-scaling", "statement": "bounds useful data-parallel", "conditions": "Growing the global batch with worker count stops reducing optimization steps beyond the critical batch size." }, { "source": "mlsys.capacity-planning", "relation": "uses", "target": "mlsys.littles-law", "statement": "sizes concurrency with", "conditions": "Arrival rate and time in system are measured for a stable operating point." }, { "source": "mlsys.batching", "relation": "determines", "target": "mlsys.latency-throughput-curve", "statement": "moves the operating point along", "conditions": "Larger batches raise throughput while adding batch-formation and execution delay." }, { "source": "mlsys.queueing", "relation": "determines", "target": "mlsys.latency-throughput-curve", "statement": "bends upward the", "conditions": "Utilization approaches service capacity, so waiting time grows sharply." }, { "source": "mlsys.arrival-burstiness", "relation": "determines", "target": "mlsys.capacity-planning", "statement": "raises headroom required by", "conditions": "Capacity is provisioned for bursts at the target latency percentile, not mean load." }, { "source": "mlsys.model-fidelity", "relation": "measures", "target": "mlsys.performance-simulation", "statement": "bounds trust in", "conditions": "Simulator predictions are validated against measurement on reference workloads." }, { "source": "mlsys.model-fidelity", "relation": "requires", "target": "mlsys.profiling", "statement": "requires measurements from", "conditions": "Validation compares predictions with profiled or benchmarked execution." }, { "source": "mlsys.amdahls-law", "relation": "limits", "target": "mlsys.parallel-efficiency", "statement": "bounds", "conditions": "The unparallelized fraction is fixed while workers are added to a fixed-size problem." }, { "source": "mlsys.dataset-storage-layout", "relation": "determines", "target": "mlsys.input-pipeline", "statement": "sets the read and decode cost of", "conditions": "Record grouping, compression, and columnar layout decide how many bytes each batch must fetch and decode." }, { "source": "mlsys.dark-silicon", "relation": "limits", "target": "mlsys.compute-throughput", "statement": "bounds sustained all-unit", "conditions": "Power and thermal limits prevent every unit from running at full frequency at once." }, { "source": "mlsys.fat-tree", "relation": "reduces", "target": "mlsys.tco", "statement": "lowers network cost within", "conditions": "Identical commodity switches replace large high-radix core switches at equal bisection bandwidth." }, { "source": "mlsys.lineage", "relation": "detects", "target": "mlsys.data-dependency", "statement": "reveals undeclared and underused", "conditions": "Lineage must be captured at the feature or column level and compared against what the model actually consumes." }, { "source": "mlsys.data-contract", "relation": "mitigates", "target": "mlsys.data-dependency", "statement": "makes explicit and versions", "conditions": "Holds when every model input is consumed through a declared, owned contract rather than an ad hoc signal." }, { "source": "mlsys.sliced-evaluation", "relation": "detects", "target": "mlsys.entanglement", "statement": "exposes subset regressions caused by", "conditions": "Applies when a change to one input or hyperparameter is evaluated on the same fixed slices before and after." }, { "source": "mlsys.configuration-management", "relation": "enables", "target": "mlsys.experiment-reproducibility", "statement": "pins the settings required for", "conditions": "Holds when every run records the exact reviewed configuration version it executed." }, { "source": "mlsys.configuration-management", "relation": "enables", "target": "mlsys.rollback", "statement": "makes configuration changes revertible through", "conditions": "Applies when configuration is versioned and deployed through the same pipeline as code." }, { "source": "mlsys.feature-store", "relation": "part-of", "target": "mlsys.ml-platform", "statement": "is a shared component of", "conditions": "Typical when an organization centralizes feature definitions for many teams." }, { "source": "mlsys.model-registry", "relation": "part-of", "target": "mlsys.ml-platform", "statement": "is a shared component of", "conditions": "Typical when promotion and deployment are standardized across teams." }, { "source": "mlsys.infrastructure-as-code", "relation": "enables", "target": "mlsys.ml-platform", "statement": "provisions repeatable environments for", "conditions": "Holds when platform compute, storage, and services are declared in versioned files." }, { "source": "mlsys.ml-roles", "relation": "determines", "target": "mlsys.incident-response", "statement": "assigns on-call ownership in", "conditions": "Applies when data, model, and platform owners are named for each production model." }, { "source": "mlsys.undeclared-consumers", "relation": "increases", "target": "mlsys.technical-debt", "statement": "add hidden coupling to", "conditions": "Holds whenever model outputs are consumed without access control or registration." }, { "source": "mlsys.authorization", "relation": "mitigates", "target": "mlsys.undeclared-consumers", "statement": "prevents unregistered", "conditions": "Applies when prediction endpoints and output tables require granted access per consumer." }, { "source": "mlsys.production-readiness", "relation": "measures", "target": "mlsys.technical-debt", "statement": "scores accumulated", "conditions": "The rubric must be applied repeatedly to the same system to track change over time." }, { "source": "mlsys.production-readiness", "relation": "requires", "target": "mlsys.integration-testing", "statement": "requires evidence from", "conditions": "Holds for the infrastructure and pipeline sections of the rubric." }, { "source": "mlsys.supervised-fine-tuning", "relation": "specializes", "target": "mlsys.supervised-learning", "statement": "is a kind of", "conditions": "Holds when training uses input-output demonstrations as targets." }, { "source": "mlsys.supervised-fine-tuning", "relation": "implements", "target": "mlsys.transfer-learning", "statement": "implements", "conditions": "Applies when a pretrained model is adapted to a new task or behavior." }, { "source": "mlsys.rlhf", "relation": "requires", "target": "mlsys.supervised-fine-tuning", "statement": "starts from a policy produced by", "conditions": "Holds for the standard InstructGPT-style pipeline; some recipes skip or shorten the SFT stage." }, { "source": "mlsys.synthetic-data", "relation": "enables", "target": "mlsys.supervised-fine-tuning", "statement": "supplies instruction data for", "conditions": "Applies when generated demonstrations are filtered for quality and diversity." }, { "source": "mlsys.incremental-training", "relation": "causes", "target": "mlsys.catastrophic-forgetting", "statement": "can cause", "conditions": "Holds when new data differs from the original distribution and old data is not replayed." }, { "source": "mlsys.incremental-training", "relation": "uses", "target": "mlsys.experience-replay", "statement": "commonly mixes in", "conditions": "Applies when earlier data is retained and permitted for reuse." }, { "source": "mlsys.data-retention", "relation": "limits", "target": "mlsys.experience-replay", "statement": "limits which past examples are available to", "conditions": "Holds when retention or deletion policies remove raw training data after a fixed period." }, { "source": "mlsys.process-reward-model", "relation": "implements", "target": "mlsys.reward-model", "statement": "can supply the training signal of", "conditions": "When step scores train the policy rather than only guide inference-time search." }, { "source": "mlsys.rlvr", "relation": "specializes", "target": "mlsys.reinforcement-learning", "statement": "is a kind of", "conditions": "Holds when rewards come from programmatic outcome checks." }, { "source": "mlsys.rlvr", "relation": "requires", "target": "mlsys.rl-rollout", "statement": "requires many sampled responses from", "conditions": "Holds for policy-gradient training where rewards are computed on generated outputs." }, { "source": "mlsys.verifier", "relation": "causes", "target": "mlsys.reward-hacking", "statement": "imperfect checks can cause", "conditions": "Holds when a policy or search is optimized against a verifier with exploitable false positives." }, { "source": "mlsys.best-of-n", "relation": "increases", "target": "mlsys.unit-cost", "statement": "multiplies generation cost in", "conditions": "Generation cost scales roughly linearly with n unless prefixes are shared." }, { "source": "mlsys.best-of-n", "relation": "alternative-to", "target": "mlsys.self-consistency", "statement": "is an alternative to", "conditions": "Both spend parallel samples; best-of-n needs a scorer while self-consistency needs answers that can be voted on." }, { "source": "mlsys.best-of-n", "relation": "requires", "target": "mlsys.sampling-controls", "statement": "requires candidate diversity from", "conditions": "For repeated generation with the same model and prompt when stochastic token selection supplies candidate diversity; deterministic greedy decoding with unchanged inputs repeats the same output." }, { "source": "mlsys.pass-at-k", "relation": "specializes", "target": "mlsys.metrics", "statement": "is a kind of", "conditions": "Applies when outputs have an automatic correctness check, as in code generation." }, { "source": "mlsys.sampling-controls", "relation": "determines", "target": "mlsys.pass-at-k", "statement": "temperature shapes", "conditions": "For a fixed model, task set, sampling procedure, and correctness criterion; the preferred temperature depends on the evaluation setting and sample budget." }, { "source": "mlsys.inference-time-search", "relation": "increases", "target": "mlsys.unit-cost", "statement": "multiplies per-query compute in", "conditions": "Holds in proportion to the number of expanded and scored nodes." }, { "source": "mlsys.inference-time-search", "relation": "uses", "target": "mlsys.prefix-cache", "statement": "shares partial-solution state through", "conditions": "Applies when branches share prompt and partial-solution prefixes on one server." }, { "source": "mlsys.self-consistency", "relation": "increases", "target": "mlsys.unit-cost", "statement": "multiplies generation cost in", "conditions": "Cost scales with the number of sampled reasoning paths." }, { "source": "mlsys.test-time-scaling", "relation": "uses", "target": "mlsys.verifier", "statement": "uses search guided by", "conditions": "Applies to the search-based way of spending inference compute, as opposed to sequential revision." }, { "source": "mlsys.pairwise-preference-eval", "relation": "uses", "target": "mlsys.llm-judge", "statement": "can use as the judge", "conditions": "Applies when pairwise comparisons are graded by a model rather than people." }, { "source": "mlsys.pairwise-preference-eval", "relation": "uses", "target": "mlsys.human-evaluation", "statement": "aggregates votes from", "conditions": "Holds for crowd-sourced arenas where people choose between two outputs." }, { "source": "mlsys.annotator-agreement", "relation": "measures", "target": "mlsys.human-evaluation", "statement": "measures reliability of", "conditions": "Requires several raters to judge overlapping items." }, { "source": "mlsys.evaluation-harness", "relation": "uses", "target": "mlsys.llm-judge", "statement": "can score open-ended outputs with", "conditions": "Applies when tasks lack exact-match references." }, { "source": "mlsys.llm-judge", "relation": "detects", "target": "mlsys.hallucination", "statement": "can flag", "conditions": "Holds when the judge sees the source evidence and its error rate has been measured against people." }, { "source": "mlsys.missing-data", "relation": "causes", "target": "mlsys.train-serve-skew", "statement": "can cause", "conditions": "Holds when missingness patterns or imputation differ between offline data and live requests." }, { "source": "mlsys.data-validation", "relation": "detects", "target": "mlsys.missing-data", "statement": "flags unexpected", "conditions": "Requires declared expectations on completeness per feature." }, { "source": "mlsys.data-validation", "relation": "measures", "target": "mlsys.data-quality", "statement": "checks", "conditions": "Covers only the properties the checks encode, not semantic correctness." }, { "source": "mlsys.data-validation", "relation": "detects", "target": "mlsys.schema-evolution", "statement": "flags undeclared", "conditions": "Holds when incoming data is compared against the expected schema version." }, { "source": "mlsys.schema-evolution", "relation": "causes", "target": "mlsys.train-serve-skew", "statement": "uncoordinated changes can cause", "conditions": "Holds when training and serving read different schema versions or interpret a changed field differently." }, { "source": "mlsys.evaluation-split", "relation": "mitigates", "target": "mlsys.leakage", "statement": "guards against", "conditions": "Holds when splits respect time, entity, and duplicate boundaries." }, { "source": "mlsys.evaluation-split", "relation": "mitigates", "target": "mlsys.selection-bias", "statement": "keeps a final test partition free of", "conditions": "Holds only if the final test partition is not used for any selection decision." }, { "source": "mlsys.data-mixture", "relation": "part-of", "target": "mlsys.data-selection", "statement": "records the sampling proportions chosen in", "conditions": "When learning effort is allocated across data sources by weighted sampling." }, { "source": "mlsys.active-learning", "relation": "specializes", "target": "mlsys.data-selection", "statement": "is label-acquisition", "conditions": "Applies when the cost being allocated is annotation effort." }, { "source": "mlsys.weak-supervision", "relation": "alternative-to", "target": "mlsys.active-learning", "statement": "is an alternative to", "conditions": "Both reduce manual labeling cost; one writes labeling heuristics, the other chooses what humans label." }, { "source": "mlsys.quality-filtering", "relation": "increases", "target": "mlsys.data-quality", "statement": "raises average", "conditions": "Holds when the filter's notion of quality matches the target task; filters can also remove useful diversity." }, { "source": "mlsys.quality-filtering", "relation": "uses", "target": "mlsys.batch-data-processing", "statement": "runs at corpus scale on", "conditions": "Applies to web-scale pretraining corpora." }, { "source": "mlsys.data-augmentation", "relation": "specializes", "target": "mlsys.regularization", "statement": "acts as a form of", "conditions": "Holds when transformations are label-preserving." }, { "source": "mlsys.data-augmentation", "relation": "mitigates", "target": "mlsys.distribution-shift", "statement": "can reduce sensitivity to", "conditions": "Holds only for shifts that the chosen transformations actually model." }, { "source": "mlsys.synthetic-data", "relation": "alternative-to", "target": "mlsys.data-augmentation", "statement": "is an alternative to", "conditions": "Both enlarge training data; augmentation transforms real examples while synthetic data generates new ones." }, { "source": "mlsys.domain-randomization", "relation": "uses", "target": "mlsys.synthetic-data", "statement": "generates varied", "conditions": "Applies to simulator-generated training data for transfer to the real world." }, { "source": "mlsys.data-acquisition", "relation": "causes", "target": "mlsys.distribution-shift", "statement": "sampling bias can cause", "conditions": "Holds when the collected population differs from the deployment population." }, { "source": "mlsys.stream-processing", "relation": "alternative-to", "target": "mlsys.batch-data-processing", "statement": "is an alternative to", "conditions": "Choose by whether the data is unbounded and how fresh outputs must be." }, { "source": "mlsys.online-learning", "relation": "uses", "target": "mlsys.stream-processing", "statement": "consumes examples from", "conditions": "Applies when updates happen as events arrive." }, { "source": "mlsys.online-learning", "relation": "mitigates", "target": "mlsys.concept-drift", "statement": "adapts to", "conditions": "Holds when labels arrive quickly enough to track the change." }, { "source": "mlsys.feature-backfill", "relation": "mitigates", "target": "mlsys.train-serve-skew", "statement": "realigns history to reduce", "conditions": "Holds when the backfill uses the same definition that serving computes." }, { "source": "mlsys.data-versioning", "relation": "part-of", "target": "mlsys.artifact-version", "statement": "is part of", "conditions": "Holds when a release records the exact dataset versions it was trained and evaluated on." }, { "source": "mlsys.lineage", "relation": "requires", "target": "mlsys.data-versioning", "statement": "references immutable", "conditions": "Lineage is only replayable if the referenced datasets are versioned." }, { "source": "mlsys.batch-data-processing", "relation": "uses", "target": "mlsys.dataset-storage-layout", "statement": "prunes reads through", "conditions": "Holds for columnar, partitioned layouts with statistics." }, { "source": "mlsys.backpressure", "relation": "limits", "target": "mlsys.queueing", "statement": "bounds queue growth and", "conditions": "Holds when upstream producers honor the signal rather than buffering unboundedly." }, { "source": "mlsys.regularization", "relation": "mitigates", "target": "mlsys.membership-inference", "statement": "reduces the overfitting exploited by", "conditions": "Reduces but does not bound leakage; formal bounds need differential privacy." }, { "source": "mlsys.critical-batch-size", "relation": "limits", "target": "mlsys.batch-size", "statement": "bounds the useful", "conditions": "Holds for a given target loss and training phase; it grows as loss falls." }, { "source": "mlsys.activation-recompute", "relation": "increases", "target": "mlsys.execution-time", "statement": "adds about one extra forward pass to", "conditions": "Holds for full recomputation; selective schemes recompute less." }, { "source": "mlsys.activation-recompute", "relation": "alternative-to", "target": "mlsys.memory-offloading", "statement": "is an alternative to", "conditions": "Both relieve accelerator memory; one spends compute, the other interconnect bandwidth." }, { "source": "mlsys.self-supervised-learning", "relation": "enables", "target": "mlsys.transfer-learning", "statement": "pretrains representations for", "conditions": "Holds when the pretraining data covers the target domain." }, { "source": "mlsys.self-supervised-learning", "relation": "produces", "target": "mlsys.embeddings", "statement": "produces", "conditions": "Applies to contrastive and masked-prediction objectives." }, { "source": "mlsys.slice-discovery", "relation": "uses", "target": "mlsys.unsupervised-learning", "statement": "clusters errors with", "conditions": "Applies to methods that group examples without predefined slice labels." }, { "source": "mlsys.slice-discovery", "relation": "uses", "target": "mlsys.embeddings", "statement": "searches for coherent slices in", "conditions": "Holds for embedding-based methods such as Domino." }, { "source": "mlsys.rlhf", "relation": "specializes", "target": "mlsys.reinforcement-learning", "statement": "is a kind of", "conditions": "Holds for PPO-style optimization against a learned reward." }, { "source": "mlsys.preference-optimization", "relation": "alternative-to", "target": "mlsys.rlhf", "statement": "is an alternative to", "conditions": "Both align to preference data; DPO skips the separate reward model and RL loop." }, { "source": "mlsys.preference-optimization", "relation": "reduces", "target": "mlsys.training-state", "statement": "avoids reward and value models in", "conditions": "Holds relative to PPO-based RLHF; a frozen reference model is still needed." }, { "source": "mlsys.reward-model", "relation": "requires", "target": "mlsys.labels", "statement": "requires reliable preference", "conditions": "Holds because reward accuracy is bounded by comparison label agreement." }, { "source": "mlsys.vla-model", "relation": "uses", "target": "mlsys.imitation-learning", "statement": "is commonly trained by", "conditions": "Applies when robot demonstrations supply the action targets." }, { "source": "mlsys.batch-size", "relation": "determines", "target": "mlsys.learning-rate-schedule", "statement": "shifts the best peak rate of", "conditions": "Holds below the critical batch size, where learning rate is typically scaled with batch." }, { "source": "mlsys.hyperparameter-transfer", "relation": "determines", "target": "mlsys.learning-rate-schedule", "statement": "carries tuned settings into", "conditions": "Holds under a parameterization such as muP that keeps optima stable across width." }, { "source": "mlsys.hyperparameter-transfer", "relation": "uses", "target": "mlsys.hyperparameter-search", "statement": "runs on small proxies", "conditions": "Holds when proxy models share the target's parameterization." }, { "source": "mlsys.hyperparameter-search", "relation": "uses", "target": "mlsys.experiment-tracking", "statement": "compares trials through", "conditions": "Applies whenever more than a handful of configurations are evaluated." }, { "source": "mlsys.gradient-accumulation", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "reduces per-step activation", "conditions": "Holds because activation memory scales with micro-batch, not effective batch." }, { "source": "mlsys.gradient-accumulation", "relation": "reduces", "target": "mlsys.collectives", "statement": "reduces the frequency of", "conditions": "Holds when gradient all-reduce is deferred to the last micro-batch." }, { "source": "mlsys.lora", "relation": "uses", "target": "mlsys.low-rank-factorization", "statement": "parameterizes updates by", "conditions": "Holds when the update rank is much smaller than the frozen weight dimensions." }, { "source": "mlsys.weight-sync", "relation": "uses", "target": "mlsys.collectives", "statement": "broadcasts weights through", "conditions": "Applies when training and rollout workers sit on separate device groups." }, { "source": "mlsys.weight-sync", "relation": "reduces", "target": "mlsys.training-goodput", "statement": "sync time reduces", "conditions": "Holds when transfers are not overlapped with rollout or training." }, { "source": "mlsys.rl-rollout", "relation": "increases", "target": "mlsys.policy-staleness", "statement": "asynchronous generation increases", "conditions": "Holds when rollouts overlap with policy updates." }, { "source": "mlsys.on-device-training", "relation": "uses", "target": "mlsys.peft", "statement": "updates sparse parameter subsets via", "conditions": "Applies under tight device memory where full backpropagation does not fit." }, { "source": "mlsys.federated-learning", "relation": "requires", "target": "mlsys.on-device-training", "statement": "requires", "conditions": "Holds for cross-device federation; cross-silo clients train on servers." }, { "source": "mlsys.cross-validation", "relation": "specializes", "target": "mlsys.evaluation-split", "statement": "is a rotating", "conditions": "Applies when data is too scarce for a single fixed validation split and examples are exchangeable." }, { "source": "mlsys.conformal-prediction", "relation": "requires", "target": "mlsys.evaluation-split", "statement": "requires a held-out calibration", "conditions": "For split conformal methods, the calibration data are separated from score-function fitting, with the stated exchangeability assumptions required for the coverage guarantee." }, { "source": "mlsys.sequential-testing", "relation": "enables", "target": "mlsys.rollout", "statement": "enables early stopping of a harmful stage in", "conditions": "Holds when the test's error guarantees allow continuous monitoring." }, { "source": "mlsys.feedback-loop", "relation": "causes", "target": "mlsys.experiment-interference", "statement": "shared retraining data can cause", "conditions": "Holds when control and treatment models learn from logs that both arms influenced." }, { "source": "mlsys.benchmark-contamination", "relation": "increases", "target": "mlsys.metrics", "statement": "inflates", "conditions": "Holds when the evaluated items or close variants were seen in training." }, { "source": "mlsys.behavioral-testing", "relation": "uses", "target": "mlsys.metamorphic-testing", "statement": "builds invariance and directional tests from", "conditions": "Applies to CheckList-style INV and DIR tests." }, { "source": "mlsys.release-gate", "relation": "uses", "target": "mlsys.behavioral-testing", "statement": "can block on failures of", "conditions": "Applies when capability tests are part of the predeclared checks." }, { "source": "mlsys.release-gate", "relation": "uses", "target": "mlsys.dangerous-capability-eval", "statement": "can block on thresholds crossed in", "conditions": "Applies to frontier-model releases governed by a scaling policy." }, { "source": "mlsys.admission", "relation": "limits", "target": "mlsys.queueing", "statement": "bounds by shedding excess load", "conditions": "Holds when rejected work is not retried immediately into the same queue." }, { "source": "mlsys.admission", "relation": "implements", "target": "mlsys.graceful-degradation", "statement": "implements load shedding for", "conditions": "Holds when shedding low-priority work keeps the remaining requests within their SLO." }, { "source": "mlsys.autoscaling", "relation": "increases", "target": "mlsys.utilization", "statement": "raises average", "conditions": "Holds when demand varies and scale-down is permitted." }, { "source": "mlsys.paged-kv", "relation": "enables", "target": "mlsys.prefix-cache", "statement": "enables block sharing for", "conditions": "Holds when the allocator supports copy-on-write sharing of identical prefix blocks." }, { "source": "mlsys.paged-kv", "relation": "increases", "target": "mlsys.service-capacity", "statement": "raises batch size and so", "conditions": "Holds when KV memory fragmentation was limiting concurrency." }, { "source": "mlsys.speculative-decoding", "relation": "reduces", "target": "mlsys.tpot", "statement": "reduces", "conditions": "Holds when the draft acceptance rate is high and decode is memory-bound." }, { "source": "mlsys.speculative-decoding", "relation": "increases", "target": "mlsys.arithmetic-intensity", "statement": "verifies several tokens per weight read, raising", "conditions": "Holds at small batch sizes where decode is memory-bound." }, { "source": "mlsys.prefix-cache", "relation": "reduces", "target": "mlsys.ttft", "statement": "skips recomputed prefill to reduce", "conditions": "Holds when requests share long identical prefixes and the cache hits." }, { "source": "mlsys.queueing", "relation": "increases", "target": "mlsys.ttft", "statement": "increases", "conditions": "Holds under load when requests wait for a prefill slot." }, { "source": "mlsys.rag", "relation": "increases", "target": "mlsys.ttft", "statement": "lengthens prompts and so", "conditions": "Holds because retrieved passages add prefill tokens and retrieval adds a serial step." }, { "source": "mlsys.online-inference", "relation": "alternative-to", "target": "mlsys.batch-inference", "statement": "is an alternative to", "conditions": "Choose by required freshness, latency, and cost per prediction." }, { "source": "mlsys.online-inference", "relation": "uses", "target": "mlsys.model-server", "statement": "uses", "conditions": "Applies when predictions are served per request behind a network interface." }, { "source": "mlsys.batch-inference", "relation": "reduces", "target": "mlsys.unit-cost", "statement": "reduces", "conditions": "Holds when predictions can be precomputed for a known input set." }, { "source": "mlsys.model-server", "relation": "requires", "target": "mlsys.model-artifact", "statement": "loads and executes", "conditions": "Holds for any server that loads serialized weights and graph at startup or on version change." }, { "source": "mlsys.multi-model-serving", "relation": "increases", "target": "mlsys.utilization", "statement": "increases", "conditions": "Holds when individual models have low or bursty traffic." }, { "source": "mlsys.model-cascade", "relation": "alternative-to", "target": "mlsys.model-routing", "statement": "is an alternative to", "conditions": "Routing decides before calling a model; a cascade escalates after a cheap model's answer." }, { "source": "mlsys.hedged-requests", "relation": "increases", "target": "mlsys.arrival-rate", "statement": "adds duplicate load to", "conditions": "Extra load stays small when hedges fire only after a high-percentile delay." }, { "source": "mlsys.hedged-requests", "relation": "mitigates", "target": "mlsys.stragglers", "statement": "mitigates", "conditions": "Holds when slowness is replica-specific rather than shared." }, { "source": "mlsys.continuous-batching", "relation": "uses", "target": "mlsys.paged-kv", "statement": "relies on dynamic allocation from", "conditions": "Holds for vLLM-style engines where sequences join and leave each iteration." }, { "source": "mlsys.load-balancing", "relation": "reduces", "target": "mlsys.queueing", "statement": "reduces", "conditions": "Holds when load signals are fresh and replicas are interchangeable." }, { "source": "mlsys.load-balancing", "relation": "reduces", "target": "mlsys.tail-latency", "statement": "reduces", "conditions": "Holds when imbalance, not total capacity, causes the tail." }, { "source": "mlsys.chunked-prefill", "relation": "alternative-to", "target": "mlsys.pd-disaggregation", "statement": "is an alternative to", "conditions": "Both limit prefill-decode interference; one colocates with chunking, the other separates pools." }, { "source": "mlsys.chunked-prefill", "relation": "uses", "target": "mlsys.continuous-batching", "statement": "piggybacks prefill chunks on", "conditions": "Holds when the scheduler admits work at each iteration under a per-iteration token budget." }, { "source": "mlsys.constrained-decoding", "relation": "enables", "target": "mlsys.function-calling", "statement": "guarantees well-formed arguments for", "conditions": "The supported argument constraints are enforced during decoding and the output completes successfully; this does not establish argument truth or tool authorization." }, { "source": "mlsys.function-calling", "relation": "uses", "target": "mlsys.grammar", "statement": "constrains arguments with", "conditions": "Applies when tool arguments must match a declared schema." }, { "source": "mlsys.output-validation", "relation": "implements", "target": "mlsys.grammar", "statement": "checks after generation", "conditions": "Outputs are checked against the declared schema after generation, whether or not constrained decoding was used." }, { "source": "mlsys.output-validation", "relation": "increases", "target": "mlsys.tail-latency", "statement": "retries increase", "conditions": "Validation failure triggers retries on the response’s critical path, compared with an otherwise comparable run without those retries." }, { "source": "mlsys.reasoning-budget", "relation": "limits", "target": "mlsys.unit-cost", "statement": "caps the reasoning share of", "conditions": "Holds when reasoning tokens are billed or provisioned like output tokens." }, { "source": "mlsys.reasoning-budget", "relation": "limits", "target": "mlsys.kv-cache", "statement": "bounds growth of", "conditions": "Holds because reasoning tokens stay in context until the answer." }, { "source": "mlsys.dense-retrieval", "relation": "uses", "target": "mlsys.embeddings", "statement": "uses", "conditions": "Holds for bi-encoder retrieval where queries and records are embedded separately." }, { "source": "mlsys.dense-retrieval", "relation": "specializes", "target": "mlsys.retrieval", "statement": "is a kind of", "conditions": "Applies when relevance is scored by similarity of learned vectors." }, { "source": "mlsys.lexical-retrieval", "relation": "specializes", "target": "mlsys.retrieval", "statement": "is a kind of", "conditions": "Applies when relevance is scored by weighted term overlap over an inverted index." }, { "source": "mlsys.reranking", "relation": "part-of", "target": "mlsys.ranking-funnel", "statement": "is a late stage of", "conditions": "Holds when a cheap candidate stage precedes a costlier scoring model under one latency budget." }, { "source": "mlsys.reranking", "relation": "uses", "target": "mlsys.transformer", "statement": "commonly uses cross-encoder", "conditions": "Applies to neural rerankers that jointly encode query and candidate." }, { "source": "mlsys.on-device-runtime", "relation": "requires", "target": "mlsys.model-conversion", "statement": "requires", "conditions": "Holds because runtimes accept only their own formats and operator sets." }, { "source": "mlsys.on-device-runtime", "relation": "uses", "target": "mlsys.edge-npu", "statement": "delegates supported operators to", "conditions": "Applies when the device has an accelerator and a delegate for it." }, { "source": "mlsys.ota-model-update", "relation": "requires", "target": "mlsys.artifact-signing", "statement": "must verify", "conditions": "Holds for any fleet where tampered models could be pushed." }, { "source": "mlsys.ota-model-update", "relation": "uses", "target": "mlsys.rollback", "statement": "relies on on-device", "conditions": "Requires the previous model to stay on the device." }, { "source": "mlsys.device-cloud-partitioning", "relation": "reduces", "target": "mlsys.energy-per-inference", "statement": "can reduce on-device", "conditions": "Holds when uplink cost is lower than the compute it moves off device." }, { "source": "mlsys.device-heterogeneity", "relation": "determines", "target": "mlsys.device-cloud-partitioning", "statement": "shifts the best split in", "conditions": "Holds because device speed and network differ widely across the fleet." }, { "source": "mlsys.model-registry", "relation": "enables", "target": "mlsys.rollback", "statement": "keeps prior versions available for", "conditions": "Holds when prior versions and their serving configuration remain registered and deployable." }, { "source": "mlsys.shadow-deployment", "relation": "detects", "target": "mlsys.train-serve-skew", "statement": "can expose", "conditions": "Holds when shadow outputs are compared with offline predictions on the same inputs." }, { "source": "mlsys.shadow-deployment", "relation": "increases", "target": "mlsys.unit-cost", "statement": "duplicates inference cost in", "conditions": "Holds for the shadow period because every mirrored request runs on both models." }, { "source": "mlsys.ml-ci-cd", "relation": "uses", "target": "mlsys.integration-testing", "statement": "runs", "conditions": "Applies on each change to code, data checks, or model artifacts." }, { "source": "mlsys.ml-ci-cd", "relation": "part-of", "target": "mlsys.mlops-maturity", "statement": "marks higher levels of", "conditions": "Holds for maturity models in which automated pipeline delivery defines the top level." }, { "source": "mlsys.drift-detection", "relation": "part-of", "target": "mlsys.monitoring", "statement": "is part of", "conditions": "Applies when distribution checks run alongside service and quality signals in production." }, { "source": "mlsys.drift-detection", "relation": "requires", "target": "mlsys.inference-logging", "statement": "compares distributions from", "conditions": "Holds because production distributions are reconstructed from logged inputs and predictions." }, { "source": "mlsys.delayed-outcomes", "relation": "limits", "target": "mlsys.drift-detection", "statement": "limits outcome-based", "conditions": "Holds when labels arrive late; input-distribution checks remain available." }, { "source": "mlsys.alerting", "relation": "uses", "target": "mlsys.sli", "statement": "fires on thresholds over", "conditions": "Applies when alerts are defined over user-facing indicators rather than raw resource metrics." }, { "source": "mlsys.concept-drift", "relation": "specializes", "target": "mlsys.distribution-shift", "statement": "is the input-target form of", "conditions": "Holds when P(y|x) changes even if the input distribution does not." }, { "source": "mlsys.workflow-orchestration", "relation": "implements", "target": "mlsys.continuous-training", "statement": "runs the pipelines of", "conditions": "Holds when retraining, validation, and registration steps run as an orchestrated pipeline." }, { "source": "mlsys.postmortem", "relation": "uses", "target": "mlsys.failure-attribution", "statement": "uses", "conditions": "Applies to ML incidents whose cause could lie in data, model, code, or infrastructure." }, { "source": "mlsys.failure-attribution", "relation": "uses", "target": "mlsys.lineage", "statement": "traces data and artifact versions through", "conditions": "Holds when the failing prediction can be tied to the data and artifact versions that produced it." }, { "source": "mlsys.distributed-tracing", "relation": "enables", "target": "mlsys.failure-attribution", "statement": "localizes the component behind", "conditions": "Holds when every component a request traverses propagates the trace context." }, { "source": "mlsys.llm-call-tracing", "relation": "enables", "target": "mlsys.cost-attribution", "statement": "attributes token spend per step for", "conditions": "Holds when spans record token counts and model identity per call." }, { "source": "mlsys.model-deprecation", "relation": "requires", "target": "mlsys.lineage", "statement": "finds downstream consumers through", "conditions": "Holds when lineage records which downstream jobs and services read the model's outputs." }, { "source": "mlsys.artifact-archival", "relation": "part-of", "target": "mlsys.model-deprecation", "statement": "is the final stage of", "conditions": "Applies when a retired model must stay reproducible or auditable after removal from service." }, { "source": "mlsys.artifact-archival", "relation": "enables", "target": "mlsys.algorithmic-audit", "statement": "preserves evidence for", "conditions": "Holds when evaluation records and data references are kept for the audit horizon." }, { "source": "mlsys.kv-cache-offload", "relation": "specializes", "target": "mlsys.storage-tiering", "statement": "is a kind of", "conditions": "Applies when KV blocks move between accelerator memory, host memory, and SSD by reuse likelihood." }, { "source": "mlsys.kv-cache-transfer", "relation": "uses", "target": "mlsys.rdma", "statement": "uses", "conditions": "Applies in disaggregated clusters with RDMA fabrics." }, { "source": "mlsys.kv-cache-transfer", "relation": "uses", "target": "mlsys.communication-overlap", "statement": "hides latency with", "conditions": "Holds when layer-wise KV transfer can proceed while later layers still compute." }, { "source": "mlsys.lm-program-optimization", "relation": "requires", "target": "mlsys.metrics", "statement": "optimizes against an end-to-end", "conditions": "Holds because the optimizer can only improve what the declared metric scores." }, { "source": "mlsys.multi-adapter-serving", "relation": "reduces", "target": "mlsys.unit-cost", "statement": "reduces", "conditions": "Holds relative to one dedicated deployment per fine-tuned model." }, { "source": "mlsys.denoising-step-reduction", "relation": "uses", "target": "mlsys.distillation", "statement": "often uses step", "conditions": "Applies to progressive and consistency distillation, not to training-free fast samplers." }, { "source": "mlsys.denoising-step-reduction", "relation": "reduces", "target": "mlsys.execution-time", "statement": "reduces per-sample", "conditions": "Holds roughly in proportion to the number of network evaluations removed." }, { "source": "mlsys.diffusion-feature-caching", "relation": "increases", "target": "mlsys.memory-footprint", "statement": "increases cached-activation", "conditions": "Holds because cached features from earlier steps must stay resident." }, { "source": "mlsys.diffusion-feature-caching", "relation": "reduces", "target": "mlsys.execution-time", "statement": "reduces", "conditions": "Holds when adjacent-step features change little, so skipped blocks lose little quality." }, { "source": "mlsys.quantized-vector-index", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "compresses vectors to reduce", "conditions": "Vectors are stored as shorter quantized codes rather than full-precision coordinates; recall depends on quantization, partition probing, reranking, and the data." }, { "source": "mlsys.filtered-vector-search", "relation": "uses", "target": "mlsys.graph-vector-index", "statement": "extends traversal of", "conditions": "Applies to predicate-agnostic methods that densify a proximity graph to keep filtered recall." }, { "source": "mlsys.index-freshness", "relation": "limits", "target": "mlsys.retrieval", "statement": "bounds recall on recently changed records in", "conditions": "Holds when inserts and deletes reach the index later than the source of truth." }, { "source": "mlsys.agent-evaluation", "relation": "uses", "target": "mlsys.llm-judge", "statement": "can grade trajectories with", "conditions": "Applies where task success lacks a programmatic check." }, { "source": "mlsys.deterministic-replay", "relation": "enables", "target": "mlsys.agent-evaluation", "statement": "makes repeatable", "conditions": "Holds when tool responses and model outputs are recorded so a run can be re-executed exactly." }, { "source": "mlsys.critical-batch-size", "relation": "limits", "target": "mlsys.time-to-train", "statement": "bounds the data-parallel speedup of", "conditions": "Holds because beyond it more examples per step no longer cut the steps needed to reach the target loss." }, { "source": "mlsys.distillation", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "reduces the serving memory footprint by producing a smaller student", "conditions": "When the student architecture is smaller than the teacher and reaches acceptable quality on the target metric." }, { "source": "mlsys.distillation", "relation": "enables", "target": "mlsys.speculative-decoding", "statement": "produces well-aligned draft models for", "conditions": "When the draft is distilled from the target model so its token distribution matches closely enough to raise acceptance rates." }, { "source": "mlsys.pruning", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "reduces", "conditions": "When the pruned weights are stored in a compressed or smaller dense form rather than as explicit zeros." }, { "source": "mlsys.structured-pruning", "relation": "reduces", "target": "mlsys.execution-time", "statement": "reduces dense-hardware", "conditions": "When whole channels, filters, or heads are removed so the remaining layers are smaller dense operations." }, { "source": "mlsys.low-rank-factorization", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "reduces", "conditions": "When the chosen rank is much smaller than the matrix dimensions and the factored model keeps acceptable accuracy." }, { "source": "mlsys.quantization-aware-training", "relation": "alternative-to", "target": "mlsys.post-training-quantization", "statement": "is an alternative to", "conditions": "QAT recovers more accuracy at low bit widths but needs a training run; PTQ needs only a calibration set." }, { "source": "mlsys.neural-architecture-search", "relation": "determines", "target": "mlsys.efficiency-frontier", "statement": "searches for points on", "conditions": "When the search objective combines quality and a stated cost measure on the target platform." }, { "source": "mlsys.hardware-aware-nas", "relation": "uses", "target": "mlsys.analytical-performance-model", "statement": "uses latency predictors from", "conditions": "When on-device measurement of every candidate is too slow, so a calibrated cost model ranks candidates." }, { "source": "mlsys.model-hardware-co-design", "relation": "enables", "target": "mlsys.nm-sparsity", "statement": "enables", "conditions": "When the sparsity pattern is chosen to match a sparse matrix-engine instruction, as with 2:4 sparsity on sparse tensor cores." }, { "source": "mlsys.weight-only-quantization", "relation": "reduces", "target": "mlsys.tpot", "statement": "reduces", "conditions": "When small-batch decoding is memory-bandwidth-bound and dequantization is fused into the matrix multiply." }, { "source": "mlsys.weight-only-quantization", "relation": "mitigates", "target": "mlsys.activation-outliers", "statement": "sidesteps", "conditions": "Because activations stay in higher precision, outlier channels never enter a low-bit activation range." }, { "source": "mlsys.transformer", "relation": "produces", "target": "mlsys.activation-outliers", "statement": "develops, at scale,", "conditions": "In large transformers, roughly beyond a few billion parameters, where a few hidden dimensions carry systematic large-magnitude activations." }, { "source": "mlsys.kv-cache-quantization", "relation": "reduces", "target": "mlsys.tpot", "statement": "reduces", "conditions": "When long-context decoding time is dominated by reading the KV cache from memory." }, { "source": "mlsys.kv-cache-quantization", "relation": "enables", "target": "mlsys.batching", "statement": "raises the feasible batch size of", "conditions": "When serving is KV-memory-bound, so freed capacity admits more concurrent sequences." }, { "source": "mlsys.training-goodput", "relation": "reduces", "target": "mlsys.time-to-train", "statement": "reduces", "conditions": "For a fixed model, data, and quality target, since lost and redone work lengthens wall-clock time." }, { "source": "mlsys.inference-load-scenario", "relation": "determines", "target": "mlsys.batching", "statement": "determines the admissible", "conditions": "Offline scenarios permit large batches; server and single-stream scenarios limit batching by their latency constraint." }, { "source": "mlsys.inference-load-scenario", "relation": "uses", "target": "mlsys.arrival-rate", "statement": "specifies the query", "conditions": "In server scenarios, where queries arrive by a stated random process at a target rate." }, { "source": "mlsys.activation-recompute", "relation": "reduces", "target": "mlsys.model-flops-utilization", "statement": "lowers measured", "conditions": "Because recomputed FLOPs are excluded from MFU's model-required count, so MFU falls while hardware FLOPs utilization rises." }, { "source": "mlsys.bottleneck-analysis", "relation": "detects", "target": "mlsys.bound-regime", "statement": "identifies the operating", "conditions": "When per-kernel time, FLOPs, and bytes are measured, so each kernel can be placed relative to the ridge point." }, { "source": "mlsys.job-preemption", "relation": "reduces", "target": "mlsys.training-goodput", "statement": "reduces", "conditions": "When preempted work since the last checkpoint is lost and restart costs are paid on resumption." }, { "source": "mlsys.benchmark", "relation": "uses", "target": "mlsys.energy-measurement", "statement": "uses", "conditions": "For energy or power benchmarks, which must declare sensor class, sampling, and system boundary to be comparable." }, { "source": "mlsys.consistent-cut", "relation": "enables", "target": "mlsys.checkpoint-resharding", "statement": "enables", "conditions": "A sharded checkpoint can be remapped to a new layout only when all shards were saved from the same training step." }, { "source": "mlsys.async-checkpointing", "relation": "requires", "target": "mlsys.consistent-cut", "statement": "requires", "conditions": "When every rank snapshots its shard at the same step boundary before background persistence begins." }, { "source": "mlsys.async-checkpointing", "relation": "increases", "target": "mlsys.training-goodput", "statement": "increases", "conditions": "When the blocking snapshot is much cheaper than the durable write, so training stalls less per checkpoint." }, { "source": "mlsys.circuit-breaker", "relation": "enables", "target": "mlsys.graceful-degradation", "statement": "enables", "conditions": "When an open breaker routes requests to a fallback response rather than returning errors." }, { "source": "mlsys.circuit-breaker", "relation": "reduces", "target": "mlsys.arrival-rate", "statement": "cuts the request rate reaching a failing dependency, lowering its", "conditions": "While the breaker is open, which gives the overloaded dependency time to recover." }, { "source": "mlsys.replication-failover", "relation": "requires", "target": "mlsys.fencing", "statement": "requires", "conditions": "When a failed-over primary may still be alive, so stale writers must be fenced to avoid split brain." }, { "source": "mlsys.replication-failover", "relation": "increases", "target": "mlsys.tco", "statement": "increases", "conditions": "Redundant replicas add capital and operating cost in proportion to the replication factor." }, { "source": "mlsys.mtbf", "relation": "limits", "target": "mlsys.training-goodput", "statement": "bounds", "conditions": "For large synchronous jobs, where job MTBF shrinks with component count and each failure costs rework plus restart." }, { "source": "mlsys.distributed-tracing", "relation": "detects", "target": "mlsys.gray-failure", "statement": "detects", "conditions": "When traces record the latency and errors that callers actually experience rather than health-check results." }, { "source": "mlsys.network-topology", "relation": "determines", "target": "mlsys.failure-domain", "statement": "helps determine network-related", "conditions": "Components depend on shared network elements; power, software, and other dependencies may define additional or overlapping failure domains." }, { "source": "mlsys.abft", "relation": "uses", "target": "mlsys.matrix-multiply", "statement": "carries checksums through", "conditions": "For linear-algebra kernels whose input checksums predict output checksums." }, { "source": "mlsys.abft", "relation": "increases", "target": "mlsys.execution-time", "statement": "adds checksum overhead to", "conditions": "The overhead is small relative to large matrix multiplications but nonzero." }, { "source": "mlsys.game-day", "relation": "measures", "target": "mlsys.alerting", "statement": "tests whether alerts fire and reach responders through", "conditions": "When the exercise uses real paging and runbook paths." }, { "source": "mlsys.hardware-fault-injection", "relation": "specializes", "target": "mlsys.failure-injection", "statement": "is a kind of", "conditions": "When the injected faults are bit flips or stuck-at errors in model state rather than service-level faults." }, { "source": "mlsys.rollout", "relation": "reduces", "target": "mlsys.blast-radius", "statement": "reduces", "conditions": "When each stage exposes a small fraction of traffic and gates promotion on health signals." }, { "source": "mlsys.least-privilege", "relation": "reduces", "target": "mlsys.blast-radius", "statement": "reduces", "conditions": "When a compromised or faulty component holds only the permissions its task requires." }, { "source": "mlsys.transfer-learning", "relation": "increases", "target": "mlsys.backdoor", "statement": "increases exposure to", "conditions": "When pretrained weights come from untrusted parties, since backdoors can survive fine-tuning on clean data." }, { "source": "mlsys.guardrails", "relation": "mitigates", "target": "mlsys.jailbreak", "statement": "partially mitigates", "conditions": "Input and output classifiers catch known patterns but can be evaded by novel or obfuscated prompts." }, { "source": "mlsys.rlhf", "relation": "mitigates", "target": "mlsys.jailbreak", "statement": "partially mitigates", "conditions": "Safety training reduces disallowed outputs but fails under competing objectives and mismatched generalization." }, { "source": "mlsys.adversarial-training", "relation": "increases", "target": "mlsys.time-to-train", "statement": "increases", "conditions": "Each training step runs an inner attack, multiplying compute by roughly the number of attack steps." }, { "source": "mlsys.distillation", "relation": "limits", "target": "mlsys.model-watermarking", "statement": "can strip parameter-embedded", "conditions": "For watermarks embedded in weights rather than in behavior on trigger inputs, since a distilled student has new weights." }, { "source": "mlsys.artifact-signing", "relation": "mitigates", "target": "mlsys.supply-chain-compromise", "statement": "mitigates", "conditions": "When consumers verify signatures and provenance against trusted identities before loading artifacts." }, { "source": "mlsys.model-registry", "relation": "uses", "target": "mlsys.artifact-signing", "statement": "uses", "conditions": "When the registry signs artifacts on promotion and verifies them before deployment." }, { "source": "mlsys.isolation", "relation": "mitigates", "target": "mlsys.unsafe-deserialization", "statement": "contains the effects of", "conditions": "When untrusted model files are loaded inside a sandbox with no credentials or network access." }, { "source": "mlsys.generalization", "relation": "reduces", "target": "mlsys.membership-inference", "statement": "lowers the advantage of", "conditions": "Attack advantage grows with the train-test generalization gap, so better-generalizing models leak less membership signal." }, { "source": "mlsys.dp-sgd", "relation": "mitigates", "target": "mlsys.model-inversion", "statement": "mitigates", "conditions": "For reconstruction of individual training records; inference of population-level attributes is not bounded by DP." }, { "source": "mlsys.secure-aggregation", "relation": "mitigates", "target": "mlsys.model-inversion", "statement": "blocks gradient-based", "conditions": "When the server could otherwise reconstruct a client's data from its individual update." }, { "source": "mlsys.data-retention", "relation": "requires", "target": "mlsys.machine-unlearning", "statement": "requires", "conditions": "When deletion obligations extend to models trained on the deleted records." }, { "source": "mlsys.remote-attestation", "relation": "detects", "target": "mlsys.supply-chain-compromise", "statement": "detects", "conditions": "For platform firmware or software whose measured state differs from trusted reference values." }, { "source": "mlsys.multi-tenancy", "relation": "increases", "target": "mlsys.side-channel", "statement": "increases exposure to", "conditions": "When tenants share accelerators, caches, memory, or interconnect whose contention is observable." }, { "source": "mlsys.side-channel", "relation": "causes", "target": "mlsys.privacy-leakage", "statement": "can cause", "conditions": "When observed timing, power, or memory behavior correlates with secret inputs or model parameters." }, { "source": "mlsys.he-inference", "relation": "increases", "target": "mlsys.execution-time", "statement": "multiplies", "conditions": "Homomorphic evaluation costs orders of magnitude more compute than plaintext inference, especially for nonlinear layers." }, { "source": "mlsys.secure-computation-inference", "relation": "increases", "target": "mlsys.execution-time", "statement": "increases", "conditions": "Latency is dominated by communication rounds and garbled-circuit evaluation of nonlinear layers." }, { "source": "mlsys.pue", "relation": "increases", "target": "mlsys.operational-carbon", "statement": "multiplies IT energy into", "conditions": "Facility overhead for cooling and power conversion scales IT energy by PUE before carbon intensity applies." }, { "source": "mlsys.emission-scopes", "relation": "determines", "target": "mlsys.operational-carbon", "statement": "classifies purchased-electricity emissions (Scope 2) in", "conditions": "For an operator buying grid electricity under the GHG Protocol." }, { "source": "mlsys.lifecycle-assessment", "relation": "measures", "target": "mlsys.water-footprint", "statement": "measures", "conditions": "When the assessment's inventory includes on-site cooling and off-site generation water." }, { "source": "mlsys.lifecycle-assessment", "relation": "measures", "target": "mlsys.e-waste", "statement": "accounts end-of-life impacts of", "conditions": "When the system boundary includes disposal and recycling stages." }, { "source": "mlsys.spot-capacity", "relation": "reduces", "target": "mlsys.tco", "statement": "reduces", "conditions": "When workloads tolerate preemption through checkpointing or elasticity so the discount is not lost to rework." }, { "source": "mlsys.model-routing", "relation": "reduces", "target": "mlsys.cost-per-token", "statement": "lowers the average", "conditions": "When a router sends requests a cheaper model can handle to that model while meeting the quality target." }, { "source": "mlsys.multi-tenancy", "relation": "requires", "target": "mlsys.cost-attribution", "statement": "requires", "conditions": "When shared infrastructure must be charged back to the teams or workloads that consume it." }, { "source": "mlsys.cost-attribution", "relation": "uses", "target": "mlsys.idle-dynamic-power", "statement": "allocates shared idle cost using", "conditions": "When provisioned but idle capacity must be charged to owners by an explicit rule." }, { "source": "mlsys.idle-dynamic-power", "relation": "determines", "target": "mlsys.energy-efficiency", "statement": "determines", "conditions": "At low load, idle power dominates, so poorly energy-proportional hardware loses efficiency." }, { "source": "mlsys.stpa", "relation": "uses", "target": "mlsys.feedback-control", "statement": "models the system as", "conditions": "STPA represents the system as controllers, control actions, and feedback paths." }, { "source": "mlsys.fmea", "relation": "uses", "target": "mlsys.failure-model", "statement": "enumerates component failure modes from", "conditions": "FMEA is bottom-up over declared component failure modes." }, { "source": "mlsys.functional-safety", "relation": "requires", "target": "mlsys.hazard-analysis", "statement": "requires", "conditions": "ISO 26262 requires hazard analysis and risk assessment before safety goals are set." }, { "source": "mlsys.functional-safety", "relation": "uses", "target": "mlsys.safety-integrity-level", "statement": "uses", "conditions": "Safety goals are assigned an integrity level, such as an ASIL, that sets development rigor." }, { "source": "mlsys.functional-safety", "relation": "uses", "target": "mlsys.fmea", "statement": "uses", "conditions": "When component failure modes must be traced to effects on safety goals." }, { "source": "mlsys.safety-filter", "relation": "uses", "target": "mlsys.model-predictive-control", "statement": "uses", "conditions": "For predictive safety filters that solve a constrained finite-horizon problem to find the nearest safe action." }, { "source": "mlsys.control-barrier-function", "relation": "requires", "target": "mlsys.dynamics", "statement": "requires a model of", "conditions": "The barrier condition holds only under the modeled dynamics and bounded disturbance." }, { "source": "mlsys.simplex-architecture", "relation": "requires", "target": "mlsys.mixed-criticality", "statement": "requires", "conditions": "The verified safety controller and decision module must keep their timing even when the complex controller overruns." }, { "source": "mlsys.minimal-risk-condition", "relation": "part-of", "target": "mlsys.graceful-degradation", "statement": "can be the final stage of", "conditions": "When degradation in a physical system proceeds to a safe stop." }, { "source": "mlsys.minimal-risk-condition", "relation": "alternative-to", "target": "mlsys.authority-handover", "statement": "is an alternative to", "conditions": "When the system can no longer operate safely, it can either hand control to a human or reach a minimal-risk state itself." }, { "source": "mlsys.operating-domain", "relation": "determines", "target": "mlsys.scenario-based-testing", "statement": "scopes the scenario catalog of", "conditions": "Scenarios and parameter ranges are drawn from the declared operational design domain." }, { "source": "mlsys.safety-case", "relation": "uses", "target": "mlsys.scenario-based-testing", "statement": "draws coverage evidence from", "conditions": "When scenario coverage is traced to hazards and operating-domain conditions." }, { "source": "mlsys.falsification", "relation": "enables", "target": "mlsys.rare-event-validation", "statement": "enables", "conditions": "Failure regions found by falsification can seed importance sampling for failure-probability estimates." }, { "source": "mlsys.falsification", "relation": "alternative-to", "target": "mlsys.formal-verification", "statement": "is an alternative to", "conditions": "For systems too complex for formal proof, where searching for counterexamples replaces proving their absence." }, { "source": "mlsys.neural-network-verification", "relation": "alternative-to", "target": "mlsys.robustness-evaluation", "statement": "is a certified alternative to empirical", "conditions": "For properties over input regions small enough for the verifier to scale." }, { "source": "mlsys.statistical-uncertainty", "relation": "limits", "target": "mlsys.rare-event-validation", "statement": "limits the confidence achievable in", "conditions": "When the number of observed failures is small, confidence bounds dominate the estimate." }, { "source": "mlsys.xil-testing", "relation": "measures", "target": "mlsys.feedback-delay", "statement": "measures", "conditions": "When the real compute and hardware run in the loop, exposing timing that pure simulation omits." }, { "source": "mlsys.xil-testing", "relation": "specializes", "target": "mlsys.integration-testing", "statement": "is a kind of", "conditions": "Each rung tests the assembled software or hardware against a simulated or recorded environment." }, { "source": "mlsys.goal-structuring-notation", "relation": "enables", "target": "mlsys.algorithmic-audit", "statement": "enables", "conditions": "When auditors review the assurance argument, not only its evidence artifacts." }, { "source": "mlsys.compensation", "relation": "alternative-to", "target": "mlsys.rollback", "statement": "is an alternative to", "conditions": "Rollback restores software or configuration state; compensation remedies effects when such a reversal is unavailable." }, { "source": "mlsys.runtime-assurance", "relation": "uses", "target": "mlsys.stopping-envelope", "statement": "checks the current state against", "conditions": "When the fallback is a stop and the monitor runs fast enough to trigger it before the state leaves the envelope." }, { "source": "mlsys.hardware-fault-injection", "relation": "measures", "target": "mlsys.abft", "statement": "measures the detection coverage of", "conditions": "When injected faults hit the operations the checksums protect, at realistic rates and locations." }, { "source": "mlsys.individual-fairness", "relation": "alternative-to", "target": "mlsys.fairness", "statement": "states fairness over similar individuals rather than across groups as in", "conditions": "When a system must choose which fairness notion to enforce or audit." }, { "source": "mlsys.individual-fairness", "relation": "limits", "target": "mlsys.decision-policy", "statement": "constrains", "conditions": "When a task-specific similarity metric is declared and accepted by stakeholders." }, { "source": "mlsys.fairness-incompatibility", "relation": "limits", "target": "mlsys.calibration", "statement": "limits per-group", "conditions": "When base rates differ and error-rate parity is also required." }, { "source": "mlsys.fairness-incompatibility", "relation": "limits", "target": "mlsys.fairness", "statement": "limits which criteria can jointly hold in", "conditions": "When base rates differ across groups and the predictor is imperfect." }, { "source": "mlsys.bias-mitigation", "relation": "implements", "target": "mlsys.fairness", "statement": "implements", "conditions": "When the intervention targets a declared criterion and is re-evaluated on held-out data." }, { "source": "mlsys.bias-mitigation", "relation": "uses", "target": "mlsys.data-selection", "statement": "uses reweighting and resampling from", "conditions": "For pre-processing methods that change the training distribution across groups." }, { "source": "mlsys.feature-attribution", "relation": "detects", "target": "mlsys.leakage", "statement": "detects", "conditions": "When a feature unavailable at prediction time dominates attributions." }, { "source": "mlsys.counterfactual-explanation", "relation": "alternative-to", "target": "mlsys.feature-attribution", "statement": "is an alternative to", "conditions": "Both explain one decision post hoc; counterfactuals give actionable changes, attributions give per-feature contributions." }, { "source": "mlsys.interpretable-model", "relation": "enables", "target": "mlsys.contestability", "statement": "enables", "conditions": "When affected people or reviewers can read the decision logic that produced their outcome." }, { "source": "mlsys.interpretable-model", "relation": "enables", "target": "mlsys.algorithmic-audit", "statement": "enables", "conditions": "When auditors can inspect the model directly instead of relying on post-hoc approximations." }, { "source": "mlsys.model-registry", "relation": "uses", "target": "mlsys.model-card", "statement": "stores and versions", "conditions": "When documentation is required for promotion alongside the model artifact." }, { "source": "mlsys.algorithmic-audit", "relation": "uses", "target": "mlsys.datasheet", "statement": "draws evidence from", "conditions": "Datasheets document data provenance and composition but do not replace access to the data." }, { "source": "mlsys.automation-bias", "relation": "limits", "target": "mlsys.approval-gate", "statement": "can reduce to rubber stamps", "conditions": "When reviewers see many correct proposals, weak error cues, and time pressure." }, { "source": "mlsys.automation-bias", "relation": "limits", "target": "mlsys.authority-handover", "statement": "can degrade", "conditions": "When operators lose vigilance and manual skill during long automated stretches." }, { "source": "mlsys.ai-risk-management", "relation": "uses", "target": "mlsys.red-teaming", "statement": "uses", "conditions": "When red-team findings feed the measure and manage functions of the risk process." }, { "source": "mlsys.risk-tiered-regulation", "relation": "requires", "target": "mlsys.impact-assessment", "statement": "requires", "conditions": "For deployers of high-risk systems in categories where a fundamental-rights impact assessment is required." }, { "source": "mlsys.impact-assessment", "relation": "determines", "target": "mlsys.fairness", "statement": "identifies the affected groups and criteria for", "conditions": "When the assessment precedes deployment and is revisited after material changes." }, { "source": "mlsys.planning", "relation": "uses", "target": "mlsys.inference-time-search", "statement": "uses", "conditions": "When the agent expands and scores alternative partial plans rather than committing to one chain." }, { "source": "mlsys.agent-evaluation", "relation": "measures", "target": "mlsys.planning", "statement": "measures plan quality through", "conditions": "When intermediate steps, not only final outcomes, are scored across the trajectory." }, { "source": "mlsys.agent-budget", "relation": "limits", "target": "mlsys.planning", "statement": "bounds the search depth of", "conditions": "When turn, token, or time limits cap how many alternatives a plan can explore." }, { "source": "mlsys.tool-contract", "relation": "uses", "target": "mlsys.idempotency", "statement": "declares retry safety in terms of", "conditions": "When the contract states an idempotency key or at-most-once semantics so the harness can decide whether a re-issued call is safe." }, { "source": "mlsys.robot-perception", "relation": "requires", "target": "mlsys.sensing", "statement": "requires calibrated", "conditions": "Detections and depth are only as accurate as the intrinsic, extrinsic, and timing calibration of the sensors." }, { "source": "mlsys.sensor-fusion", "relation": "increases", "target": "mlsys.state-observability", "statement": "increases", "conditions": "When an added modality measures a state the others cannot distinguish, such as scale for monocular vision." }, { "source": "mlsys.state-observability", "relation": "limits", "target": "mlsys.slam", "statement": "bounds drift correction in", "conditions": "Absolute pose and heading are unobservable from relative measurements alone, so drift persists without loop closure or external references." }, { "source": "mlsys.contact-compliance", "relation": "limits", "target": "mlsys.sim-to-real", "statement": "can widen the gap in", "conditions": "Contact forces depend on stiffness, friction, and geometry that simulators approximate poorly." }, { "source": "mlsys.model-context-protocol", "relation": "increases", "target": "mlsys.prompt-injection", "statement": "widens the surface for", "conditions": "When tool descriptions and results from third-party servers enter the model's context as if trusted." }, { "source": "mlsys.model-context-protocol", "relation": "uses", "target": "mlsys.authorization", "statement": "uses", "conditions": "When servers require scoped, delegated credentials before exposing tools or resources." }, { "source": "mlsys.agent-evaluation", "relation": "measures", "target": "mlsys.computer-use", "statement": "measures", "conditions": "When tasks run in real applications and success is checked from the resulting system state." }, { "source": "mlsys.agent-harness", "relation": "uses", "target": "mlsys.function-calling", "statement": "dispatches tool actions through", "conditions": "When the model emits structured calls that the harness validates and executes." }, { "source": "mlsys.agent-harness", "relation": "uses", "target": "mlsys.agent-budget", "statement": "enforces stopping rules from", "conditions": "The harness, not the model, applies turn, token, time, and spend limits." }, { "source": "mlsys.agent-budget", "relation": "limits", "target": "mlsys.trajectory", "statement": "bounds the length of", "conditions": "When limits on turns, tokens, or wall-clock time stop the run before completion." }, { "source": "mlsys.durable-execution", "relation": "specializes", "target": "mlsys.workflow-orchestration", "statement": "is a kind of", "conditions": "When each step's intent and result are logged so recovery reuses recorded results; an effect completed before its result was recorded may need retry, deduplication, or reconciliation." }, { "source": "mlsys.deterministic-replay", "relation": "uses", "target": "mlsys.llm-call-tracing", "statement": "can take recorded model outputs and tool results from", "conditions": "When traces retain full payloads rather than sampled or truncated spans; otherwise replay needs the trajectory log." }, { "source": "mlsys.deterministic-replay", "relation": "enables", "target": "mlsys.failure-attribution", "statement": "enables", "conditions": "When a failing trajectory can be re-run step by step with its original inputs." }, { "source": "mlsys.suspend-resume", "relation": "uses", "target": "mlsys.durable-execution", "statement": "persists waiting state through", "conditions": "When a trajectory must survive process restarts while waiting on a human or external event." }, { "source": "mlsys.shared-agent-state", "relation": "part-of", "target": "mlsys.multi-agent-orchestration", "statement": "is part of", "conditions": "When several agents read and write task state under one orchestration topology." }, { "source": "mlsys.shared-agent-state", "relation": "causes", "target": "mlsys.coordination-failure", "statement": "can cause", "conditions": "When concurrent writes or stale reads go uncontrolled, so agents act on conflicting views of the task." }, { "source": "mlsys.llm-call-tracing", "relation": "detects", "target": "mlsys.coordination-failure", "statement": "detects", "conditions": "When handoffs, delegated instructions, and per-agent outputs are recorded as linked spans." }, { "source": "mlsys.correlated-agent-errors", "relation": "limits", "target": "mlsys.self-consistency", "statement": "limits the benefit of", "conditions": "When the sampled answers or voting agents share a base model and prompt, so their errors are not independent." }, { "source": "mlsys.correlated-agent-errors", "relation": "limits", "target": "mlsys.llm-judge", "statement": "weakens the independence of", "conditions": "When the judge shares a base model with the system it grades." }, { "source": "mlsys.observation-age", "relation": "requires", "target": "mlsys.time-synchronization", "statement": "is measurable across sensors only under", "conditions": "Observation age can be computed only when capture and use timestamps share a synchronized clock." }, { "source": "mlsys.recursive-filtering", "relation": "uses", "target": "mlsys.dynamics", "statement": "predicts state with a model of", "conditions": "The prediction step propagates the estimate through modeled dynamics between measurements." }, { "source": "mlsys.sensor-fusion", "relation": "uses", "target": "mlsys.recursive-filtering", "statement": "uses", "conditions": "When fused measurements arrive asynchronously and are combined through prediction and correction steps." }, { "source": "mlsys.world-model", "relation": "enables", "target": "mlsys.reinforcement-learning", "statement": "enables sample-efficient", "conditions": "When policies are trained on imagined rollouts from the learned model rather than only on environment interaction." }, { "source": "mlsys.motion-planning", "relation": "requires", "target": "mlsys.state-estimation", "statement": "requires", "conditions": "Plans are valid only relative to the estimated pose and map they start from." }, { "source": "mlsys.motion-planning", "relation": "uses", "target": "mlsys.slam", "statement": "plans over maps from", "conditions": "When the environment is not known in advance and the map is built online." }, { "source": "mlsys.mixed-criticality", "relation": "uses", "target": "mlsys.real-time-scheduling", "statement": "uses", "conditions": "Schedulability is analyzed with criticality-dependent WCET estimates for each task." }, { "source": "mlsys.real-time-scheduling", "relation": "limits", "target": "mlsys.utilization", "statement": "bounds", "conditions": "Rate-monotonic scheduling guarantees deadlines only when processor utilization stays below the Liu-Layland bound." }, { "source": "mlsys.vla-model", "relation": "uses", "target": "mlsys.transfer-learning", "statement": "transfers web-scale vision-language knowledge through", "conditions": "When the policy is fine-tuned from a pretrained vision-language backbone." }, { "source": "mlsys.vla-model", "relation": "uses", "target": "mlsys.teleoperation-data", "statement": "is fine-tuned on", "conditions": "When robot action labels come from human-driven demonstrations." }, { "source": "mlsys.synthetic-data", "relation": "alternative-to", "target": "mlsys.teleoperation-data", "statement": "is an alternative to", "conditions": "Simulated demonstrations are cheaper to scale but carry a sim-to-real gap that teleoperated demonstrations avoid." }, { "source": "mlsys.action-chunking", "relation": "increases", "target": "mlsys.observation-age", "statement": "increases, for actions late in a chunk,", "conditions": "When chunks execute open loop, so later actions rest on older observations." }, { "source": "mlsys.imitation-learning", "relation": "uses", "target": "mlsys.action-chunking", "statement": "reduces compounding error with", "conditions": "When the policy predicts short action sequences rather than single steps." }, { "source": "mlsys.domain-randomization", "relation": "alternative-to", "target": "mlsys.data-augmentation", "statement": "generates varied simulated examples instead of transforming recorded ones as in", "conditions": "When training data for a policy or perception model can be simulated rather than only recorded." }, { "source": "mlsys.domain-randomization", "relation": "mitigates", "target": "mlsys.distribution-shift", "statement": "mitigates", "conditions": "When the randomized ranges cover the real-world parameters without making the task unlearnable." }, { "source": "mlsys.safety-filter", "relation": "limits", "target": "mlsys.vla-model", "statement": "constrains the commanded actions of", "conditions": "When the filter runs on every action at control rate with authority over the learned policy." }, { "source": "mlsys.world-model", "relation": "uses", "target": "mlsys.self-supervised-learning", "statement": "is trained to predict future observations by", "conditions": "When the model learns from action-conditioned sequences without task labels." }, { "source": "mlsys.time-synchronization", "relation": "part-of", "target": "mlsys.sensing", "statement": "is part of", "conditions": "Sensing establishes how observations relate to time as well as to physical quantities and coordinates." }, { "source": "mlsys.diffusion-model", "relation": "specializes", "target": "mlsys.neural-network", "statement": "is a kind of", "conditions": "Diffusion models are neural networks trained to reverse a gradual noising process over many denoising steps." }, { "source": "mlsys.compute-in-memory", "relation": "reduces", "target": "mlsys.data-movement-energy", "statement": "lowers by computing where weights are stored", "conditions": "Weight-stationary matrix-vector workloads whose weights fit in the memory arrays." }, { "source": "mlsys.compute-in-memory", "relation": "limits", "target": "mlsys.precision", "statement": "caps the effective bit width of", "conditions": "Analog implementations, where device noise, drift, and converter resolution bound precision." }, { "source": "mlsys.emerging-nvm", "relation": "enables", "target": "mlsys.compute-in-memory", "statement": "provides programmable analog cells for", "conditions": "Device conductance can be programmed to represent weight values." }, { "source": "mlsys.emerging-nvm", "relation": "alternative-to", "target": "mlsys.on-chip-sram", "statement": "competes for on-chip weight storage with", "conditions": "Density and standby power outweigh write endurance and write latency." }, { "source": "mlsys.embedded-dsl", "relation": "produces", "target": "mlsys.computation-graph", "statement": "builds", "conditions": "Staged or traced execution rather than immediate evaluation." }, { "source": "mlsys.embedded-dsl", "relation": "enables", "target": "mlsys.compilation", "statement": "exposes whole programs to", "conditions": "The DSL restricts host-language features that cannot be staged." }, { "source": "mlsys.differentiable-programming", "relation": "requires", "target": "mlsys.autodiff", "statement": "relies on", "conditions": "Derivatives must propagate through control flow and data structures." }, { "source": "mlsys.function-transformations", "relation": "implements", "target": "mlsys.differentiable-programming", "statement": "realizes through a composable grad transformation", "conditions": "Programs written as pure array functions." }, { "source": "mlsys.persistent-kernel", "relation": "reduces", "target": "mlsys.dispatch", "statement": "removes per-operator launch overhead from", "conditions": "Small-batch, latency-bound workloads made of many short kernels." }, { "source": "mlsys.persistent-kernel", "relation": "alternative-to", "target": "mlsys.graph-replay", "statement": "is an alternative way to remove launch overhead to", "conditions": "Persistent kernels also overlap work across operator boundaries inside one launch." }, { "source": "mlsys.persistent-kernel", "relation": "alternative-to", "target": "mlsys.fusion", "statement": "keeps one kernel resident rather than merging operations as in", "conditions": "Both reduce host launches for repeated work." }, { "source": "mlsys.graph-superoptimization", "relation": "part-of", "target": "mlsys.compilation", "statement": "is an optimization stage within", "conditions": "Rewrites are verified equivalent before code generation." }, { "source": "mlsys.graph-superoptimization", "relation": "uses", "target": "mlsys.intermediate-representation", "statement": "searches rewrites expressed in", "conditions": "Graph- or kernel-level IR with defined operator semantics." }, { "source": "mlsys.graph-superoptimization", "relation": "enables", "target": "mlsys.fusion", "statement": "discovers new opportunities for", "conditions": "Search spans kernel and thread-block levels as well as the graph level." }, { "source": "mlsys.llm-training-framework", "relation": "implements", "target": "mlsys.hybrid-parallel", "statement": "packages", "conditions": "Parallelism degrees are set by configuration and mapped to the device topology." }, { "source": "mlsys.llm-training-framework", "relation": "uses", "target": "mlsys.state-sharding", "statement": "commonly applies", "conditions": "Optimizer and parameter state exceed per-device memory." }, { "source": "mlsys.llm-training-framework", "relation": "uses", "target": "mlsys.activation-recompute", "statement": "commonly applies", "conditions": "Activation memory limits sequence length or micro-batch size." }, { "source": "mlsys.cloud-service-model", "relation": "determines", "target": "mlsys.tco", "statement": "shifts the cost structure of", "conditions": "Comparing rented infrastructure, managed platforms, and hosted APIs for the same workload." }, { "source": "mlsys.cloud-service-model", "relation": "uses", "target": "mlsys.multi-tenancy", "statement": "relies on resource pooling through", "conditions": "Providers serve many customers from shared physical resources." }, { "source": "mlsys.lakehouse", "relation": "uses", "target": "mlsys.object-storage", "statement": "stores table files in", "conditions": "Lakehouse designs over open file formats." }, { "source": "mlsys.lakehouse", "relation": "enables", "target": "mlsys.data-versioning", "statement": "provides snapshot time travel for", "conditions": "Table formats that retain prior snapshots." }, { "source": "mlsys.lakehouse", "relation": "uses", "target": "mlsys.dataset-storage-layout", "statement": "stores data in columnar layouts such as", "conditions": "Open columnar formats such as Parquet." }, { "source": "mlsys.training-power-swings", "relation": "limits", "target": "mlsys.power-oversubscription", "statement": "removes headroom assumed by", "conditions": "Swings are coincident across accelerators, defeating non-coincident-peak assumptions." }, { "source": "mlsys.facility-energy-storage", "relation": "mitigates", "target": "mlsys.training-power-swings", "statement": "buffers", "conditions": "Storage sized and controlled to respond at the swing frequency." }, { "source": "mlsys.power-capping", "relation": "mitigates", "target": "mlsys.training-power-swings", "statement": "smooths", "conditions": "Device power caps and floors applied to bound ramp rates." }, { "source": "mlsys.facility-energy-storage", "relation": "increases", "target": "mlsys.availability", "statement": "raises", "conditions": "Carrying load through utility outages until generators take over." }, { "source": "mlsys.grid-interconnection", "relation": "limits", "target": "mlsys.capacity-planning", "statement": "caps the power available to", "conditions": "Sites whose deliverable grid power is below planned IT load." }, { "source": "mlsys.on-site-generation", "relation": "alternative-to", "target": "mlsys.grid-interconnection", "statement": "substitutes for or supplements", "conditions": "Interconnection queues delay or cap grid supply." }, { "source": "mlsys.on-site-generation", "relation": "determines", "target": "mlsys.operational-carbon", "statement": "sets the emissions of supply for", "conditions": "On-site sources supply a material share of load." }, { "source": "mlsys.data-cleaning", "relation": "uses", "target": "mlsys.data-validation", "statement": "repairs violations found by", "conditions": "Validation rules or statistics flag the records to repair." }, { "source": "mlsys.data-cleaning", "relation": "mitigates", "target": "mlsys.missing-data", "statement": "imputes or removes", "conditions": "The missingness mechanism is understood well enough to choose a treatment." }, { "source": "mlsys.data-cleaning", "relation": "increases", "target": "mlsys.data-quality", "statement": "raises", "conditions": "Repairs are validated and do not introduce label leakage." }, { "source": "mlsys.rebalancing", "relation": "mitigates", "target": "mlsys.class-imbalance", "statement": "compensates for", "conditions": "Resampling or weighting is applied to training data only." }, { "source": "mlsys.class-imbalance", "relation": "determines", "target": "mlsys.metrics", "statement": "shapes the choice of", "conditions": "Per-class, precision-recall, or cost-weighted metrics replace accuracy." }, { "source": "mlsys.rebalancing", "relation": "reduces", "target": "mlsys.calibration", "statement": "shifts predicted probabilities away from true base rates, degrading", "conditions": "When posteriors track the altered training prior, deployment retains a different prior under stable class-conditional distributions, and no prior correction is applied." }, { "source": "mlsys.feature-hashing", "relation": "specializes", "target": "mlsys.features", "statement": "is a technique within", "conditions": "Categorical inputs with large or open vocabularies." }, { "source": "mlsys.feature-hashing", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "bounds", "conditions": "Weight or embedding tables indexed by hash bucket." }, { "source": "mlsys.feature-cross", "relation": "specializes", "target": "mlsys.features", "statement": "is a technique within", "conditions": "Interactions among categorical inputs." }, { "source": "mlsys.feature-cross", "relation": "uses", "target": "mlsys.feature-hashing", "statement": "is often bucketed by", "conditions": "Crosses of high-cardinality features whose joint vocabulary would explode." }, { "source": "mlsys.linear-model", "relation": "uses", "target": "mlsys.feature-cross", "statement": "represents interactions through", "conditions": "Interactions not otherwise expressible by an additive model." }, { "source": "mlsys.data-governance", "relation": "determines", "target": "mlsys.data-retention", "statement": "sets policy for", "conditions": "Retention periods follow purpose and legal obligations." }, { "source": "mlsys.data-governance", "relation": "requires", "target": "mlsys.lineage", "statement": "depends on traceability from", "conditions": "Accountability for which data reached which model must be demonstrable." }, { "source": "mlsys.data-governance", "relation": "uses", "target": "mlsys.authorization", "statement": "enforces access decisions through", "conditions": "Access policies are encoded as enforceable permissions." }, { "source": "mlsys.data-consent", "relation": "part-of", "target": "mlsys.data-governance", "statement": "is a policy element of", "conditions": "Personal data is processed." }, { "source": "mlsys.data-consent", "relation": "limits", "target": "mlsys.data-acquisition", "statement": "restricts", "conditions": "Personal data for which a lawful basis or consent is required." }, { "source": "mlsys.exploratory-data-analysis", "relation": "detects", "target": "mlsys.missing-data", "statement": "reveals patterns of", "conditions": "Profiles are computed per feature and per slice." }, { "source": "mlsys.exploratory-data-analysis", "relation": "enables", "target": "mlsys.data-validation", "statement": "informs constraints for", "conditions": "Observed statistics seed a schema or expectation set." }, { "source": "mlsys.model-visualization", "relation": "uses", "target": "mlsys.experiment-tracking", "statement": "plots metrics recorded by", "conditions": "Runs log curves and artifacts in a queryable store." }, { "source": "mlsys.model-visualization", "relation": "enables", "target": "mlsys.error-analysis", "statement": "supports", "conditions": "Errors are displayed alongside inputs and slices." }, { "source": "mlsys.model-selection", "relation": "uses", "target": "mlsys.cross-validation", "statement": "estimates candidate performance with", "conditions": "Data is too small for a dedicated validation split." }, { "source": "mlsys.model-selection", "relation": "causes", "target": "mlsys.selection-bias", "statement": "can inflate estimates through", "conditions": "Selection and final assessment reuse the same validation data." }, { "source": "mlsys.model-selection", "relation": "uses", "target": "mlsys.baseline", "statement": "compares candidates against", "conditions": "The baseline is evaluated under the same protocol." }, { "source": "mlsys.hyperparameter-search", "relation": "part-of", "target": "mlsys.model-selection", "statement": "is a stage within", "conditions": "Hyperparameters are chosen on validation data." }, { "source": "mlsys.pretraining", "relation": "uses", "target": "mlsys.self-supervised-learning", "statement": "optimizes an objective from", "conditions": "Unlabeled corpora supply the training targets." }, { "source": "mlsys.pretraining", "relation": "enables", "target": "mlsys.transfer-learning", "statement": "supplies base models for", "conditions": "Downstream tasks share structure with the pretraining corpus." }, { "source": "mlsys.pretraining", "relation": "uses", "target": "mlsys.data-mixture", "statement": "draws its corpus from", "conditions": "Corpora combine multiple domains or sources." }, { "source": "mlsys.pretraining", "relation": "uses", "target": "mlsys.causal-lm", "statement": "commonly optimizes", "conditions": "Decoder-only language-model pretraining." }, { "source": "mlsys.causal-lm", "relation": "specializes", "target": "mlsys.self-supervised-learning", "statement": "is a text case of", "conditions": "Targets are the next tokens of the input itself." }, { "source": "mlsys.gan", "relation": "alternative-to", "target": "mlsys.diffusion-model", "statement": "competes for sample generation with", "conditions": "Single-step sampling versus many denoising steps, with different stability and diversity." }, { "source": "mlsys.gan", "relation": "causes", "target": "mlsys.mode-collapse", "statement": "is prone to", "conditions": "The generator outpaces the discriminator during adversarial training." }, { "source": "mlsys.mode-collapse", "relation": "specializes", "target": "mlsys.training-instability", "statement": "is a diversity-loss form of", "conditions": "Collapse of output diversity rather than divergence of the loss." }, { "source": "mlsys.perplexity", "relation": "specializes", "target": "mlsys.metrics", "statement": "is a likelihood-based case of", "conditions": "Models that assign token probabilities." }, { "source": "mlsys.tokenization", "relation": "determines", "target": "mlsys.perplexity", "statement": "sets the unit of", "conditions": "Per-token perplexities are comparable only under one tokenizer; bits per byte normalizes." }, { "source": "mlsys.perplexity", "relation": "measures", "target": "mlsys.causal-lm", "statement": "evaluates held-out performance of", "conditions": "Held-out text is free of training contamination." }, { "source": "mlsys.reference-match-scoring", "relation": "specializes", "target": "mlsys.metrics", "statement": "is a reference-based case of", "conditions": "Tasks with canonical reference answers." }, { "source": "mlsys.reference-match-scoring", "relation": "alternative-to", "target": "mlsys.llm-judge", "statement": "competes for scoring open outputs with", "conditions": "Reproducibility and cost weighed against tolerance of paraphrase." }, { "source": "mlsys.error-analysis", "relation": "uses", "target": "mlsys.slice-discovery", "statement": "groups mistakes with", "conditions": "Error categories are not known in advance." }, { "source": "mlsys.training-diagnostics", "relation": "detects", "target": "mlsys.training-instability", "statement": "reveals", "conditions": "Gradient and loss statistics are logged at step granularity." }, { "source": "mlsys.training-diagnostics", "relation": "uses", "target": "mlsys.model-visualization", "statement": "reads plots from", "conditions": "Learning curves and activation statistics are recorded during training." }, { "source": "mlsys.batch-inference", "relation": "uses", "target": "mlsys.keyed-predictions", "statement": "joins outputs to inputs with", "conditions": "Distributed batch jobs return outputs out of order." }, { "source": "mlsys.inference-logging", "relation": "uses", "target": "mlsys.keyed-predictions", "statement": "joins predictions to later outcomes with", "conditions": "Outcomes arrive after the prediction and must be matched." }, { "source": "mlsys.stateless-serving-function", "relation": "enables", "target": "mlsys.autoscaling", "statement": "lets replicas be added or removed for", "conditions": "No request depends on replica-local state." }, { "source": "mlsys.stateless-serving-function", "relation": "uses", "target": "mlsys.model-signature", "statement": "exposes its interface as", "conditions": "Inputs and outputs are declared at export." }, { "source": "mlsys.model-gateway", "relation": "uses", "target": "mlsys.model-routing", "statement": "directs requests by", "conditions": "Several models or providers serve the same application." }, { "source": "mlsys.model-gateway", "relation": "enables", "target": "mlsys.cost-attribution", "statement": "records usage for", "conditions": "All model calls pass through the gateway." }, { "source": "mlsys.model-gateway", "relation": "uses", "target": "mlsys.semantic-cache", "statement": "can serve repeated requests from", "conditions": "Responses are reusable across users and sessions." }, { "source": "mlsys.semantic-cache", "relation": "uses", "target": "mlsys.ann", "statement": "finds similar past requests with", "conditions": "Requests are embedded and indexed." }, { "source": "mlsys.semantic-cache", "relation": "reduces", "target": "mlsys.unit-cost", "statement": "avoids model calls that drive", "conditions": "Workloads with many near-duplicate requests." }, { "source": "mlsys.glue-code", "relation": "causes", "target": "mlsys.technical-debt", "statement": "accumulates, when wired in without a common interface,", "conditions": "When supporting code around a general package grows to fit one system." }, { "source": "mlsys.ml-platform", "relation": "mitigates", "target": "mlsys.glue-code", "statement": "reduces the need for", "conditions": "The platform wraps packages behind common APIs." }, { "source": "mlsys.pipeline-jungle", "relation": "specializes", "target": "mlsys.technical-debt", "statement": "is a form of", "conditions": "Data preparation evolves without holistic redesign." }, { "source": "mlsys.pipeline-jungle", "relation": "limits", "target": "mlsys.integration-testing", "statement": "impedes end-to-end", "conditions": "Failures can only be detected by expensive end-to-end tests." }, { "source": "mlsys.build-vs-buy", "relation": "determines", "target": "mlsys.ml-platform", "statement": "shapes the composition of", "conditions": "Platform components are chosen one capability at a time." }, { "source": "mlsys.build-vs-buy", "relation": "uses", "target": "mlsys.tco", "statement": "compares options by", "conditions": "Engineering, licensing, and operating costs are counted over the same horizon." }, { "source": "mlsys.user-feedback", "relation": "uses", "target": "mlsys.inference-logging", "statement": "attaches signals to records in", "conditions": "Feedback is joined to the prediction that elicited it." }, { "source": "mlsys.user-feedback", "relation": "enables", "target": "mlsys.retraining", "statement": "supplies labels for", "conditions": "Feedback is filtered for noise and position or exposure bias." }, { "source": "mlsys.user-feedback", "relation": "causes", "target": "mlsys.feedback-loop", "statement": "can drive", "conditions": "Feedback is gathered only on outputs the model chose to show." }, { "source": "mlsys.human-ai-interaction", "relation": "mitigates", "target": "mlsys.automation-bias", "statement": "counters", "conditions": "Interfaces convey uncertainty and make errors visible." }, { "source": "mlsys.human-ai-interaction", "relation": "enables", "target": "mlsys.contestability", "statement": "provides interfaces for", "conditions": "Users can see why an output was produced and correct or override it." }, { "source": "mlsys.demand-response", "relation": "uses", "target": "mlsys.power-capping", "statement": "sheds load through", "conditions": "Caps are applied in response to grid signals." }, { "source": "mlsys.demand-response", "relation": "uses", "target": "mlsys.job-preemption", "statement": "defers flexible work through", "conditions": "Interruptible training or batch jobs checkpoint before suspension." }, { "source": "mlsys.prompt-engineering", "relation": "alternative-to", "target": "mlsys.supervised-fine-tuning", "statement": "competes for task adaptation with", "conditions": "Iteration speed and no training cost weighed against reliability, context cost, and control." }, { "source": "mlsys.prompt-engineering", "relation": "uses", "target": "mlsys.in-context-learning", "statement": "elicits behavior through", "conditions": "The model follows instructions or demonstrations in its input." }, { "source": "mlsys.lm-program-optimization", "relation": "implements", "target": "mlsys.prompt-engineering", "statement": "automates", "conditions": "An end-to-end metric and training examples are available." }, { "source": "mlsys.context", "relation": "limits", "target": "mlsys.in-context-learning", "statement": "caps the demonstrations available to", "conditions": "Demonstrations and instructions share one input budget." }, { "source": "mlsys.pretraining", "relation": "enables", "target": "mlsys.in-context-learning", "statement": "gives rise to", "conditions": "Sufficient model and data scale." }, { "source": "mlsys.streaming-asr", "relation": "part-of", "target": "mlsys.voice-pipeline", "statement": "is the input stage of", "conditions": "Cascaded designs that transcribe speech before the language model." }, { "source": "mlsys.streaming-asr", "relation": "uses", "target": "mlsys.on-device-runtime", "statement": "runs in", "conditions": "On-device recognizers for privacy and latency." }, { "source": "mlsys.latency-budget", "relation": "limits", "target": "mlsys.voice-pipeline", "statement": "bounds the stages of", "conditions": "Turn latency must stay within conversational expectations." }, { "source": "mlsys.kv-cache-eviction", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "reduces the KV portion of", "conditions": "Generation is long enough that cached keys and values dominate accelerator memory." }, { "source": "mlsys.kv-cache-eviction", "relation": "alternative-to", "target": "mlsys.kv-cache-quantization", "statement": "competes as a KV-compression option with", "conditions": "Both target KV memory; they can also be combined." }, { "source": "mlsys.kv-cache-eviction", "relation": "uses", "target": "mlsys.attention", "statement": "selects entries using scores from", "conditions": "Importance is estimated from accumulated or recent attention weights." }, { "source": "mlsys.dynamic-sparse-attention", "relation": "specializes", "target": "mlsys.attention", "statement": "is a runtime-sparsified form of", "conditions": "Only predicted-important query-key blocks are computed." }, { "source": "mlsys.dynamic-sparse-attention", "relation": "alternative-to", "target": "mlsys.sliding-window-attention", "statement": "replaces fixed sparsity patterns of", "conditions": "Sparsity is chosen per input rather than fixed in advance." }, { "source": "mlsys.dynamic-sparse-attention", "relation": "reduces", "target": "mlsys.ttft", "statement": "shortens long-prompt", "conditions": "Prefill attention over long contexts dominates first-token latency." }, { "source": "mlsys.host-device-interconnect", "relation": "limits", "target": "mlsys.cpu-gpu-hybrid-inference", "statement": "caps the transfer rate available to", "conditions": "Activations or KV state cross the host link at each step." }, { "source": "mlsys.cpu-gpu-hybrid-inference", "relation": "alternative-to", "target": "mlsys.kv-cache-offload", "statement": "computes on the host instead of moving state back as in", "conditions": "Host memory holds state that exceeds accelerator capacity." }, { "source": "mlsys.heterogeneous-accelerator-execution", "relation": "uses", "target": "mlsys.pipeline-parallel", "statement": "assigns unequal stages through", "conditions": "Stage sizes are matched to each device type's compute and memory." }, { "source": "mlsys.heterogeneous-accelerator-execution", "relation": "uses", "target": "mlsys.auto-parallelization", "statement": "searches placements and plans with", "conditions": "Device types differ in throughput, memory, and interconnect." }, { "source": "mlsys.graph-neural-network", "relation": "specializes", "target": "mlsys.neural-network", "statement": "is a graph-structured form of", "conditions": "Computation follows the input graph's edges." }, { "source": "mlsys.graph-neural-network", "relation": "uses", "target": "mlsys.sparse-tensor-kernels", "statement": "aggregates neighbor features with", "conditions": "Graph adjacency is stored sparsely." }, { "source": "mlsys.graph-neural-network", "relation": "uses", "target": "mlsys.neighbor-sampling", "statement": "trains on large graphs with", "conditions": "Full multi-hop neighborhoods exceed device memory." }, { "source": "mlsys.neighbor-sampling", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "bounds the per-batch", "conditions": "Fanout per hop is capped." }, { "source": "mlsys.neighbor-sampling", "relation": "increases", "target": "mlsys.statistical-uncertainty", "statement": "adds sampling variance to gradient estimates and thus", "conditions": "Sampled neighborhoods approximate full aggregation." }, { "source": "mlsys.dynamic-resharding", "relation": "uses", "target": "mlsys.collectives", "statement": "moves weight shards with", "conditions": "Layouts differ between phases." }, { "source": "mlsys.dynamic-resharding", "relation": "implements", "target": "mlsys.weight-sync", "statement": "realizes", "conditions": "Training and generation share the same devices." }, { "source": "mlsys.dynamic-resharding", "relation": "uses", "target": "mlsys.device-mesh", "statement": "remaps shards across", "conditions": "Each phase uses a different mesh layout." }, { "source": "mlsys.activation-sparsity", "relation": "specializes", "target": "mlsys.sparsity", "statement": "is an input-dependent form of", "conditions": "The zero pattern changes per token." }, { "source": "mlsys.activation-sparsity", "relation": "alternative-to", "target": "mlsys.pruning", "statement": "skips work per input rather than permanently as in", "conditions": "Static pruning removes the same weights for every input." }, { "source": "mlsys.activation-sparsity", "relation": "reduces", "target": "mlsys.tpot", "statement": "shortens", "conditions": "Decode is memory-bound and skipped weights need not be read." }, { "source": "mlsys.non-prefix-kv-reuse", "relation": "alternative-to", "target": "mlsys.prefix-cache", "statement": "generalizes beyond the exact-prefix matching of", "conditions": "Shared chunks appear at different positions across requests." }, { "source": "mlsys.non-prefix-kv-reuse", "relation": "reduces", "target": "mlsys.ttft", "statement": "shortens", "conditions": "Reused chunks such as retrieved documents dominate prompt length." }, { "source": "mlsys.ml-aware-scheduling", "relation": "uses", "target": "mlsys.elastic-training", "statement": "resizes jobs through", "conditions": "Jobs tolerate changes in worker count." }, { "source": "mlsys.ml-aware-scheduling", "relation": "uses", "target": "mlsys.gang-scheduling", "statement": "co-allocates workers with", "conditions": "Distributed jobs need all workers simultaneously." }, { "source": "mlsys.ml-aware-scheduling", "relation": "increases", "target": "mlsys.training-goodput", "statement": "raises cluster-wide", "conditions": "Jobs' scaling behavior is measured or predicted." }, { "source": "mlsys.adaptive-routing", "relation": "reduces", "target": "mlsys.tail-latency", "statement": "shortens collective", "conditions": "Few large synchronized flows collide under static hashing." }, { "source": "mlsys.adaptive-routing", "relation": "uses", "target": "mlsys.fat-tree", "statement": "spreads traffic over the multiple paths of", "conditions": "The fabric offers path diversity between endpoints." }, { "source": "mlsys.all-to-all", "relation": "causes", "target": "mlsys.incast", "statement": "can trigger", "conditions": "Many ranks send to the same receiver simultaneously." }, { "source": "mlsys.congestion-control", "relation": "mitigates", "target": "mlsys.incast", "statement": "throttles senders during", "conditions": "Senders react to congestion signals before buffers overflow." }, { "source": "mlsys.incast", "relation": "increases", "target": "mlsys.tail-latency", "statement": "raises", "conditions": "Buffers fill and packets queue or drop." }, { "source": "mlsys.context", "relation": "determines", "target": "mlsys.context-selection", "statement": "sets the token budget for", "conditions": "Candidate content exceeds the input budget." }, { "source": "mlsys.context-selection", "relation": "uses", "target": "mlsys.retrieval", "statement": "gathers candidates with", "conditions": "Relevant items are fetched on demand rather than preloaded." }, { "source": "mlsys.context-selection", "relation": "mitigates", "target": "mlsys.context-degradation", "statement": "limits distractor content that drives", "conditions": "Irrelevant tokens otherwise dilute attention." }, { "source": "mlsys.agent-memory", "relation": "causes", "target": "mlsys.context-staleness", "statement": "can retain facts that produce", "conditions": "Stored observations are not revalidated after the environment changes." }, { "source": "mlsys.observation-age", "relation": "measures", "target": "mlsys.context-staleness", "statement": "quantifies one dimension of", "conditions": "Observations carry acquisition timestamps." }, { "source": "mlsys.context-staleness", "relation": "causes", "target": "mlsys.hallucination", "statement": "can lead to", "conditions": "The agent reports or acts on state it observed before later changes." }, { "source": "mlsys.multi-hop-retrieval", "relation": "specializes", "target": "mlsys.retrieval", "statement": "is an iterative form of", "conditions": "Later queries depend on earlier results." }, { "source": "mlsys.rag", "relation": "uses", "target": "mlsys.multi-hop-retrieval", "statement": "answers multi-step questions with", "conditions": "No single query retrieves all needed evidence." }, { "source": "mlsys.multi-hop-retrieval", "relation": "increases", "target": "mlsys.execution-time", "statement": "adds retrieval rounds to", "conditions": "Each hop requires another query and model call." }, { "source": "mlsys.capability-based-security", "relation": "specializes", "target": "mlsys.authorization", "statement": "is a reference-based form of", "conditions": "Rights travel with unforgeable tokens." }, { "source": "mlsys.capability-based-security", "relation": "implements", "target": "mlsys.least-privilege", "statement": "realizes", "conditions": "Delegated capabilities are attenuated to the task's needs." }, { "source": "mlsys.ephemeral-sandbox", "relation": "implements", "target": "mlsys.isolation", "statement": "provides per-task", "conditions": "Each task or tenant receives a fresh environment." }, { "source": "mlsys.ephemeral-sandbox", "relation": "mitigates", "target": "mlsys.irreversible-effects", "statement": "confines local damage that would otherwise become", "conditions": "Effects stay inside the sandbox rather than reaching external systems." }, { "source": "mlsys.egress-control", "relation": "implements", "target": "mlsys.least-privilege", "statement": "applies to network access", "conditions": "Only destinations the task needs are allowed." }, { "source": "mlsys.egress-control", "relation": "mitigates", "target": "mlsys.privacy-leakage", "statement": "blocks exfiltration channels for", "conditions": "Leakage would otherwise travel over outbound connections." }, { "source": "mlsys.ephemeral-sandbox", "relation": "uses", "target": "mlsys.egress-control", "statement": "restricts outbound network access with", "conditions": "The sandbox has network access." }, { "source": "mlsys.information-flow-control", "relation": "mitigates", "target": "mlsys.prompt-injection", "statement": "prevents untrusted data from steering actions in", "conditions": "Data from tools and documents carries low-integrity labels." }, { "source": "mlsys.information-flow-control", "relation": "mitigates", "target": "mlsys.privacy-leakage", "statement": "blocks labeled secrets from", "conditions": "Confidential data is tagged at its source." }, { "source": "mlsys.distribution-shift", "relation": "causes", "target": "mlsys.compounding-error", "statement": "feeds, in closed loop, the growth of", "conditions": "When the shift is the policy-induced change in visited states, so each off-distribution state raises the next per-step error." }, { "source": "mlsys.imitation-learning", "relation": "causes", "target": "mlsys.compounding-error", "statement": "can suffer", "conditions": "Training sees only expert-visited states." }, { "source": "mlsys.action-chunking", "relation": "mitigates", "target": "mlsys.compounding-error", "statement": "shortens the effective horizon of", "conditions": "Actions are predicted in multi-step chunks." }, { "source": "mlsys.interactive-imitation", "relation": "specializes", "target": "mlsys.imitation-learning", "statement": "is an interactive form of", "conditions": "An expert can label states the learner visits." }, { "source": "mlsys.interactive-imitation", "relation": "mitigates", "target": "mlsys.compounding-error", "statement": "trains on learner-visited states to counter", "conditions": "Expert corrections are available during data collection." }, { "source": "mlsys.interactive-imitation", "relation": "uses", "target": "mlsys.teleoperation-data", "statement": "collects corrections as", "conditions": "Human operators supply corrective actions." }, { "source": "mlsys.loss-masking", "relation": "part-of", "target": "mlsys.supervised-fine-tuning", "statement": "is a configuration step in", "conditions": "Examples contain non-target tokens such as prompts or tool output." }, { "source": "mlsys.loss-masking", "relation": "determines", "target": "mlsys.loss", "statement": "defines which tokens contribute to", "conditions": "Sequences and turns vary in length." }, { "source": "mlsys.credit-assignment", "relation": "part-of", "target": "mlsys.reinforcement-learning", "statement": "is a core problem within", "conditions": "Rewards are delayed or episode-level." }, { "source": "mlsys.process-reward-model", "relation": "implements", "target": "mlsys.credit-assignment", "statement": "provides step-level signals for", "conditions": "Intermediate steps can be judged." }, { "source": "mlsys.actuator-limits", "relation": "determines", "target": "mlsys.trajectory-feasibility", "statement": "sets which motions satisfy", "conditions": "Planned motions demand force or speed near the actuator's capability." }, { "source": "mlsys.model-predictive-control", "relation": "uses", "target": "mlsys.actuator-limits", "statement": "encodes as input constraints", "conditions": "Limits are known and modeled." }, { "source": "mlsys.watchdog-timer", "relation": "part-of", "target": "mlsys.runtime-assurance", "statement": "is a liveness mechanism within", "conditions": "The monitored component can hang or stall." }, { "source": "mlsys.watchdog-timer", "relation": "uses", "target": "mlsys.minimal-risk-condition", "statement": "triggers on expiry", "conditions": "A safe fallback state is defined." }, { "source": "mlsys.closed-loop-evaluation", "relation": "detects", "target": "mlsys.compounding-error", "statement": "reveals", "conditions": "Errors alter later observations, unlike open-loop replay." }, { "source": "mlsys.xil-testing", "relation": "implements", "target": "mlsys.closed-loop-evaluation", "statement": "provides simulated loops for", "conditions": "The environment model reacts to the system's actions." }, { "source": "mlsys.event-data-recorder", "relation": "enables", "target": "mlsys.failure-attribution", "statement": "supplies reconstruction evidence for", "conditions": "Records cover the interval before and after the event." }, { "source": "mlsys.safety-performance-indicator", "relation": "uses", "target": "mlsys.event-data-recorder", "statement": "is computed from events logged by", "conditions": "Field events are recorded consistently." }, { "source": "mlsys.decision-accountability", "relation": "enables", "target": "mlsys.algorithmic-audit", "statement": "provides traceable decisions for", "conditions": "Records name the approver and evidence." }, { "source": "mlsys.decision-accountability", "relation": "enables", "target": "mlsys.contestability", "statement": "identifies whom to challenge in", "conditions": "Affected people can see who decided." }, { "source": "mlsys.approval-gate", "relation": "produces", "target": "mlsys.decision-accountability", "statement": "records approvals as", "conditions": "Each approval is logged with approver identity." }, { "source": "mlsys.agent-interruption", "relation": "part-of", "target": "mlsys.human-oversight", "statement": "is an intervention mechanism within", "conditions": "Operators can act during execution." }, { "source": "mlsys.agent-interruption", "relation": "uses", "target": "mlsys.compensation", "statement": "undoes partial work with", "conditions": "Cancelled work already produced side effects." }, { "source": "mlsys.agent-loop-detection", "relation": "part-of", "target": "mlsys.agent-harness", "statement": "is a progress monitor within", "conditions": "The harness can inspect the trajectory between steps." }, { "source": "mlsys.agent-loop-detection", "relation": "detects", "target": "mlsys.coordination-failure", "statement": "flags looping instances of", "conditions": "Agents repeat steps or hand work back and forth." }, { "source": "mlsys.agent-loop-detection", "relation": "alternative-to", "target": "mlsys.agent-budget", "statement": "stops stuck runs earlier than", "conditions": "Progress can be judged before a budget is exhausted." }, { "source": "mlsys.task-and-motion-planning", "relation": "uses", "target": "mlsys.motion-planning", "statement": "checks each symbolic step with", "conditions": "Actions have geometric preconditions." }, { "source": "mlsys.task-and-motion-planning", "relation": "requires", "target": "mlsys.trajectory-feasibility", "statement": "accepts plans only with", "conditions": "Symbolic plans can be geometrically infeasible." }, { "source": "mlsys.multi-object-tracking", "relation": "specializes", "target": "mlsys.state-estimation", "statement": "is a multi-target form of", "conditions": "Several objects are observed with unknown correspondence." }, { "source": "mlsys.multi-object-tracking", "relation": "uses", "target": "mlsys.recursive-filtering", "statement": "propagates each track with", "conditions": "Object motion is modeled between frames." }, { "source": "mlsys.change-impact-analysis", "relation": "requires", "target": "mlsys.safety-case", "statement": "traces affected claims through", "conditions": "Claims and evidence are explicitly linked." }, { "source": "mlsys.change-impact-analysis", "relation": "determines", "target": "mlsys.release-gate", "statement": "sets the re-verification scope for", "conditions": "A model, software, data, or ODD change is proposed." }, { "source": "mlsys.pruning", "relation": "uses", "target": "mlsys.sparse-tensor-kernels", "statement": "relies on, to turn zeros into wall-clock speedup,", "conditions": "Sparsity is high enough to overcome irregular access." }, { "source": "mlsys.sparse-tensor-kernels", "relation": "uses", "target": "mlsys.tensor-layout", "statement": "stores nonzeros in compressed formats described by", "conditions": "Nonzeros are kept in formats such as CSR or coordinate lists." }, { "source": "mlsys.local-sgd", "relation": "specializes", "target": "mlsys.data-parallel", "statement": "is a periodically synchronized form of", "conditions": "Workers hold full model replicas." }, { "source": "mlsys.local-sgd", "relation": "alternative-to", "target": "mlsys.gradient-compression", "statement": "reduces communication by frequency rather than size as in", "conditions": "Bandwidth between workers is scarce." }, { "source": "mlsys.federated-averaging", "relation": "specializes", "target": "mlsys.local-sgd", "statement": "is a federated instance of", "conditions": "Clients run several local epochs before averaging." }, { "source": "mlsys.expert-placement", "relation": "part-of", "target": "mlsys.expert-parallel", "statement": "is a placement decision within", "conditions": "Experts are distributed across devices." }, { "source": "mlsys.expert-placement", "relation": "mitigates", "target": "mlsys.stragglers", "statement": "relieves overloaded devices that become", "conditions": "Routing concentrates tokens on a few experts." }, { "source": "mlsys.expert-placement", "relation": "alternative-to", "target": "mlsys.load-balancing-loss", "statement": "balances load in the system rather than the model as in", "conditions": "Routing remains skewed despite training-time balancing." }, { "source": "mlsys.memory-efficient-optimizer", "relation": "specializes", "target": "mlsys.optimizer", "statement": "is a reduced-state form of", "conditions": "Full adaptive state exceeds memory." }, { "source": "mlsys.memory-efficient-optimizer", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "shrinks the optimizer-state share of", "conditions": "Optimizer state dominates training memory." }, { "source": "mlsys.memory-efficient-optimizer", "relation": "alternative-to", "target": "mlsys.state-sharding", "statement": "shrinks state rather than distributing it as in", "conditions": "Few devices are available to shard across." }, { "source": "mlsys.profiling", "relation": "produces", "target": "mlsys.execution-trace", "statement": "records", "conditions": "Operator dependencies are captured along with timings." }, { "source": "mlsys.performance-simulation", "relation": "uses", "target": "mlsys.execution-trace", "statement": "replays workloads from", "conditions": "Traces are in a standardized format." }, { "source": "mlsys.offline-rl", "relation": "specializes", "target": "mlsys.reinforcement-learning", "statement": "is a fixed-dataset form of", "conditions": "No further environment interaction is allowed." }, { "source": "mlsys.offline-rl", "relation": "alternative-to", "target": "mlsys.imitation-learning", "statement": "can improve on demonstrations unlike", "conditions": "Logged data carries reward labels." }, { "source": "mlsys.distribution-shift", "relation": "limits", "target": "mlsys.offline-rl", "statement": "bounds what can be learned in", "conditions": "The learned policy visits states the dataset does not cover." }, { "source": "mlsys.byzantine-robust-aggregation", "relation": "mitigates", "target": "mlsys.poisoning", "statement": "bounds the influence of", "conditions": "Only a minority of participants are malicious." }, { "source": "mlsys.federated-learning", "relation": "uses", "target": "mlsys.byzantine-robust-aggregation", "statement": "combines client updates with", "conditions": "Clients are untrusted." }, { "source": "mlsys.cluster-manager", "relation": "uses", "target": "mlsys.bin-packing", "statement": "places tasks with", "conditions": "Machines have multidimensional resource capacity." }, { "source": "mlsys.cluster-manager", "relation": "uses", "target": "mlsys.resource-quota", "statement": "enforces", "conditions": "Multiple teams share the cluster." }, { "source": "mlsys.cluster-manager", "relation": "uses", "target": "mlsys.job-preemption", "statement": "honors priorities through", "conditions": "Higher-priority work arrives while capacity is full." }, { "source": "mlsys.energy-efficiency", "relation": "causes", "target": "mlsys.rebound-effect", "statement": "can trigger", "conditions": "Demand for the service is price-elastic." }, { "source": "mlsys.rebound-effect", "relation": "increases", "target": "mlsys.energy", "statement": "raises total", "conditions": "Demand growth outpaces per-unit savings." }, { "source": "mlsys.data-validation", "relation": "mitigates", "target": "mlsys.data-cascade", "statement": "catches early defects that start", "conditions": "Checks run at ingestion and between stages." }, { "source": "mlsys.data-dependency", "relation": "causes", "target": "mlsys.data-cascade", "statement": "can propagate", "conditions": "Downstream consumers depend on undeclared upstream signals." }, { "source": "mlsys.data-valuation", "relation": "enables", "target": "mlsys.data-selection", "statement": "ranks candidates for", "conditions": "Value estimates are cheaper than retraining on each subset." }, { "source": "mlsys.data-valuation", "relation": "measures", "target": "mlsys.data-quality", "statement": "estimates the contribution side of", "conditions": "A utility metric and baseline model are fixed." }, { "source": "mlsys.diffusion-policy", "relation": "uses", "target": "mlsys.diffusion-model", "statement": "generates actions with", "conditions": "When the action generator reverses a noising process; flow-matching policies integrate a learned flow instead." }, { "source": "mlsys.diffusion-policy", "relation": "uses", "target": "mlsys.action-chunking", "statement": "predicts action sequences as in", "conditions": "Actions are generated over a short horizon." }, { "source": "mlsys.sensor-spoofing", "relation": "implements", "target": "mlsys.evasion", "statement": "realizes, through counterfeit measurements,", "conditions": "When the counterfeit measurements target a learned perception model." }, { "source": "mlsys.sensor-fusion", "relation": "mitigates", "target": "mlsys.sensor-spoofing", "statement": "cross-checks modalities against", "conditions": "The attacker cannot spoof all modalities consistently." }, { "source": "mlsys.paired-comparison", "relation": "reduces", "target": "mlsys.statistical-uncertainty", "statement": "cancels shared item variance from", "conditions": "Both systems are scored on the same items." }, { "source": "mlsys.paired-comparison", "relation": "uses", "target": "mlsys.baseline", "statement": "tests candidates against", "conditions": "A noninferiority margin is declared in advance." }, { "source": "mlsys.release-gate", "relation": "uses", "target": "mlsys.paired-comparison", "statement": "decides promotion with", "conditions": "Candidate and incumbent are evaluated on the same set." }, { "source": "mlsys.cpu-execution-architecture", "relation": "alternative-to", "target": "mlsys.accelerator-execution-architecture", "statement": "is a competing execution target to", "conditions": "The workload is small, sparse, or latency-bound enough that host execution is viable." }, { "source": "mlsys.cpu-execution-architecture", "relation": "uses", "target": "mlsys.memory-hierarchy", "statement": "relies on caches and DRAM levels of", "conditions": "Performance depends on working-set reuse in the cache levels." }, { "source": "mlsys.cpu-gpu-hybrid-inference", "relation": "uses", "target": "mlsys.cpu-execution-architecture", "statement": "offloads selected operators to", "conditions": "Model or KV state exceeds accelerator memory." }, { "source": "mlsys.cpu-threading-numa", "relation": "uses", "target": "mlsys.cpu-execution-architecture", "statement": "maps operator work onto the cores and caches of", "conditions": "Inference or preprocessing runs on multicore hosts." }, { "source": "mlsys.cpu-threading-numa", "relation": "determines", "target": "mlsys.utilization", "statement": "sets achieved core utilization through", "conditions": "Thread counts and affinity are configured relative to available cores and NUMA domains." }, { "source": "mlsys.dynamic-shapes", "relation": "part-of", "target": "mlsys.compilation", "statement": "is a concern within", "conditions": "Input dimensions vary across calls to a compiled function." }, { "source": "mlsys.dynamic-shapes", "relation": "limits", "target": "mlsys.graph-replay", "statement": "restricts reuse of", "conditions": "Captured graphs are valid only for the shapes recorded at capture time." }, { "source": "mlsys.rule-based-graph-rewriting", "relation": "part-of", "target": "mlsys.compilation", "statement": "is a pass family within", "conditions": "The compiler applies pattern-based passes before lowering." }, { "source": "mlsys.rule-based-graph-rewriting", "relation": "alternative-to", "target": "mlsys.graph-superoptimization", "statement": "is the hand-written counterpart to", "conditions": "Both seek semantics-preserving graph transformations; one uses fixed rules, the other searched rewrites." }, { "source": "mlsys.rule-based-graph-rewriting", "relation": "enables", "target": "mlsys.fusion", "statement": "exposes fusible patterns for", "conditions": "Canonicalization places compatible operators adjacent in the graph." }, { "source": "mlsys.sparse-ragged-tensors", "relation": "specializes", "target": "mlsys.tensor-layout", "statement": "is a compressed form of", "conditions": "Storage records only nonzero entries or actual sequence extents." }, { "source": "mlsys.sparse-ragged-tensors", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "shrinks", "conditions": "Density is low or lengths vary widely, so padding would dominate storage." }, { "source": "mlsys.sparse-tensor-kernels", "relation": "uses", "target": "mlsys.sparse-ragged-tensors", "statement": "operates on operands stored in", "conditions": "Kernels read compressed index structures rather than dense arrays." }, { "source": "mlsys.execution-providers", "relation": "uses", "target": "mlsys.model-exchange-format", "statement": "partitions graphs expressed in", "conditions": "The runtime loads a framework-neutral operator graph." }, { "source": "mlsys.on-device-runtime", "relation": "uses", "target": "mlsys.execution-providers", "statement": "delegates subgraphs through", "conditions": "The target device exposes an NPU, GPU, or DSP backend with partial operator coverage." }, { "source": "mlsys.execution-providers", "relation": "increases", "target": "mlsys.dispatch", "statement": "adds host orchestration through", "conditions": "Operator fallback splits the graph and forces cross-backend copies." }, { "source": "mlsys.large-language-model", "relation": "uses", "target": "mlsys.transformer", "statement": "is usually built on", "conditions": "Most current language models use decoder-only Transformer stacks." }, { "source": "mlsys.pretraining", "relation": "produces", "target": "mlsys.large-language-model", "statement": "yields the base weights of", "conditions": "Training uses broad corpora and a self-supervised language-modeling objective." }, { "source": "mlsys.large-language-model", "relation": "specializes", "target": "mlsys.neural-network", "statement": "is a large-scale instance of", "conditions": "The model is a parameterized composition of differentiable operations." }, { "source": "mlsys.inference-compute-estimate", "relation": "measures", "target": "mlsys.kv-cache", "statement": "sizes", "conditions": "Dense attention with known layers, KV heads, head dimension, context, and precision." }, { "source": "mlsys.capacity-planning", "relation": "uses", "target": "mlsys.inference-compute-estimate", "statement": "sizes accelerators from", "conditions": "Planning precedes measurement on the target system." }, { "source": "mlsys.inference-compute-estimate", "relation": "specializes", "target": "mlsys.flop-accounting", "statement": "is the per-token inference case of", "conditions": "The model is a dense transformer served autoregressively." }, { "source": "mlsys.sync-async-training", "relation": "part-of", "target": "mlsys.data-parallel", "statement": "is a design choice within", "conditions": "Replicas train on different examples and must coordinate updates." }, { "source": "mlsys.parameter-server", "relation": "enables", "target": "mlsys.sync-async-training", "statement": "makes asynchronous updates practical for", "conditions": "Workers push and pull parameters without a global barrier." }, { "source": "mlsys.input-prefetching", "relation": "part-of", "target": "mlsys.input-pipeline", "statement": "is a stage within", "conditions": "Batches are prepared ahead of consumption." }, { "source": "mlsys.input-prefetching", "relation": "increases", "target": "mlsys.utilization", "statement": "raises accelerator utilization by hiding input latency behind", "conditions": "Input preparation would otherwise stall compute." }, { "source": "mlsys.event-time-windowing", "relation": "part-of", "target": "mlsys.stream-processing", "statement": "is a mechanism within", "conditions": "Events carry timestamps and can arrive out of order." }, { "source": "mlsys.event-time-windowing", "relation": "determines", "target": "mlsys.feature-freshness", "statement": "sets the staleness of streaming features through", "conditions": "Window size and allowed lateness bound when an aggregate is emitted." }, { "source": "mlsys.token-streaming", "relation": "uses", "target": "mlsys.backpressure", "statement": "relies on flow control from", "conditions": "Clients consume output more slowly than the server generates it." }, { "source": "mlsys.tpot", "relation": "measures", "target": "mlsys.token-streaming", "statement": "quantifies the delivery cadence of", "conditions": "Output is returned token by token after the first." }, { "source": "mlsys.serverless-inference", "relation": "uses", "target": "mlsys.stateless-serving-function", "statement": "packages models as", "conditions": "Any instance must be able to serve any request." }, { "source": "mlsys.cold-start", "relation": "limits", "target": "mlsys.serverless-inference", "statement": "bounds the latency attainable by", "conditions": "Replicas scale from zero and must load weights before serving." }, { "source": "mlsys.llm-request-scheduling", "relation": "uses", "target": "mlsys.continuous-batching", "statement": "reorders requests inside", "conditions": "Scheduling decisions occur at each generation iteration." }, { "source": "mlsys.llm-request-scheduling", "relation": "determines", "target": "mlsys.tail-latency", "statement": "shapes", "conditions": "Long and short generations share capacity under load." }, { "source": "mlsys.containerization", "relation": "implements", "target": "mlsys.dependency-environment", "statement": "packages", "conditions": "The image pins packages, native libraries, and runtime configuration." }, { "source": "mlsys.containerization", "relation": "enables", "target": "mlsys.experiment-reproducibility", "statement": "makes reruns practical for", "conditions": "Images are versioned and run unchanged across machines." }, { "source": "mlsys.distributed-hang-diagnosis", "relation": "detects", "target": "mlsys.gray-failure", "statement": "identifies", "conditions": "A slow or stuck rank stalls collectives without crashing." }, { "source": "mlsys.distributed-hang-diagnosis", "relation": "uses", "target": "mlsys.watchdog-timer", "statement": "flags stalls with", "conditions": "Collectives and ranks are expected to make progress within a timeout." }, { "source": "mlsys.benchmark-representativeness", "relation": "requires", "target": "mlsys.workload", "statement": "depends on an accurate", "conditions": "The benchmark is used to predict deployed behavior." }, { "source": "mlsys.inference-load-scenario", "relation": "determines", "target": "mlsys.benchmark-representativeness", "statement": "sets how closely measured results match deployment for", "conditions": "Arrival pattern and latency constraint are chosen to mirror the target service." }, { "source": "mlsys.data-model-licensing", "relation": "limits", "target": "mlsys.data-acquisition", "statement": "restricts which sources can enter", "conditions": "Datasets or weights carry use, redistribution, or field-of-use terms." }, { "source": "mlsys.data-governance", "relation": "uses", "target": "mlsys.data-model-licensing", "statement": "enforces terms recorded in", "conditions": "Access and use decisions must respect license obligations." }, { "source": "mlsys.risk-tiered-regulation", "relation": "requires", "target": "mlsys.content-provenance", "statement": "can mandate", "conditions": "Regulation requires machine-readable marking of synthetic content." }, { "source": "mlsys.fleet-data-engine", "relation": "uses", "target": "mlsys.active-learning", "statement": "selects uploads with", "conditions": "Triggers prioritize uncertain or rare events for labeling." }, { "source": "mlsys.fleet-data-engine", "relation": "mitigates", "target": "mlsys.distribution-shift", "statement": "counters", "conditions": "Mined data covers conditions the deployed fleet now encounters." }, { "source": "mlsys.robotics-middleware", "relation": "increases", "target": "mlsys.feedback-delay", "statement": "adds transport, serialization, and executor queuing delay to", "conditions": "When sensing, inference, and control run as separate communicating nodes rather than in one process." }, { "source": "mlsys.sensor-fusion", "relation": "uses", "target": "mlsys.robotics-middleware", "statement": "receives synchronized sensor streams through", "conditions": "Sensors publish on separate topics consumed by a fusion node." }, { "source": "mlsys.reinforcement-learning", "relation": "uses", "target": "mlsys.parallel-simulation", "statement": "gathers experience from", "conditions": "Training runs in simulation before physical deployment." }, { "source": "mlsys.domain-randomization", "relation": "uses", "target": "mlsys.parallel-simulation", "statement": "samples randomized environments from", "conditions": "Simulator parameters can be varied per environment instance." }, { "source": "mlsys.agent-credentials", "relation": "implements", "target": "mlsys.least-privilege", "statement": "realizes", "conditions": "Credentials are scoped to the task and expire after it." }, { "source": "mlsys.agent-credentials", "relation": "uses", "target": "mlsys.capability-based-security", "statement": "issues attenuated rights through", "conditions": "Delegated tokens carry narrower scope than the user holds." }, { "source": "mlsys.graph-neural-network", "relation": "uses", "target": "mlsys.gnn-graph-partitioning", "statement": "scales training across workers with", "conditions": "The graph and its features exceed one device or machine." }, { "source": "mlsys.neighbor-sampling", "relation": "uses", "target": "mlsys.gnn-graph-partitioning", "statement": "draws mini-batches across", "conditions": "Distributed training samples neighborhoods that span partitions." }, { "source": "mlsys.host-data-loader", "relation": "part-of", "target": "mlsys.input-pipeline", "statement": "is the single-machine stage of", "conditions": "Examples are decoded and collated on the training host." }, { "source": "mlsys.host-data-loader", "relation": "determines", "target": "mlsys.utilization", "statement": "bounds accelerator utilization through", "conditions": "Per-batch preprocessing time approaches the accelerator step time." }, { "source": "mlsys.input-prefetching", "relation": "part-of", "target": "mlsys.host-data-loader", "statement": "overlaps loading with compute in", "conditions": "Worker processes prepare batches ahead of the training step." }, { "source": "mlsys.streaming-asr", "relation": "specializes", "target": "mlsys.streaming-inference", "statement": "is streaming inference over audio", "conditions": "Audio arrives in chunks and partial transcripts are emitted before the utterance ends." }, { "source": "mlsys.streaming-inference", "relation": "uses", "target": "mlsys.backpressure", "statement": "relies on", "conditions": "Input or output rates can outpace the other side of the stream." }, { "source": "mlsys.inference-api", "relation": "part-of", "target": "mlsys.model-server", "statement": "is the request interface of", "conditions": "Clients call the model over a network protocol." }, { "source": "mlsys.inference-api", "relation": "increases", "target": "mlsys.tail-latency", "statement": "adds serialization time to", "conditions": "Payloads are large, such as images or long token sequences." }, { "source": "mlsys.request-throughput", "relation": "part-of", "target": "mlsys.latency-throughput-curve", "statement": "is the horizontal axis of", "conditions": "Latency is measured across offered loads." }, { "source": "mlsys.request-throughput", "relation": "limits", "target": "mlsys.slo-goodput", "statement": "is an upper bound on", "conditions": "Goodput counts only completed requests that meet their latency targets." }, { "source": "mlsys.latency-percentiles", "relation": "measures", "target": "mlsys.tail-latency", "statement": "quantifies", "conditions": "High percentiles such as p99 are reported." }, { "source": "mlsys.latency-percentiles", "relation": "implements", "target": "mlsys.sli", "statement": "is a common form of", "conditions": "The SLO is stated as a percentile under a threshold." }, { "source": "mlsys.cache-aware-routing", "relation": "uses", "target": "mlsys.prefix-cache", "statement": "routes requests to replicas holding compatible cached prefixes of", "conditions": "When requests share prompt prefixes or belong to multi-turn sessions and the cached prefix is replica-local." }, { "source": "mlsys.cache-aware-routing", "relation": "trades-off", "target": "mlsys.load-balancing", "statement": "trades cache reuse against", "conditions": "Popular prefixes concentrate traffic on a few replicas." }, { "source": "mlsys.process-group-rendezvous", "relation": "enables", "target": "mlsys.collectives", "statement": "establishes the communicators used by", "conditions": "Every rank must join before the first collective." }, { "source": "mlsys.elastic-training", "relation": "uses", "target": "mlsys.process-group-rendezvous", "statement": "re-runs", "conditions": "Workers join or leave and groups are rebuilt." }, { "source": "mlsys.agentic-workload", "relation": "increases", "target": "mlsys.kv-cache", "statement": "grows the size and reuse of", "conditions": "Each step resends a growing shared context to the same model." }, { "source": "mlsys.workload", "relation": "measures", "target": "mlsys.agentic-workload", "statement": "quantifies the demands of", "conditions": "The workload is a multi-step agent trajectory rather than independent requests." }, { "source": "mlsys.performance-regression-testing", "relation": "uses", "target": "mlsys.benchmark", "statement": "repeatedly runs", "conditions": "Each change is benchmarked against a stored baseline." }, { "source": "mlsys.performance-regression-testing", "relation": "part-of", "target": "mlsys.ml-ci-cd", "statement": "is a gate in", "conditions": "Performance thresholds can block a merge or release." }, { "source": "mlsys.measurement-variability", "relation": "limits", "target": "mlsys.benchmark", "statement": "limits the precision of", "conditions": "Differences between systems are comparable to run-to-run spread." }, { "source": "mlsys.measurement-variability", "relation": "limits", "target": "mlsys.profiling", "statement": "confounds", "conditions": "Asynchronous kernels are timed without synchronization or warm-up." }, { "source": "mlsys.pii-deidentification", "relation": "mitigates", "target": "mlsys.privacy-leakage", "statement": "reduces the risk of", "conditions": "Identifiers are removed before training or logging." }, { "source": "mlsys.data-governance", "relation": "requires", "target": "mlsys.pii-deidentification", "statement": "requires", "conditions": "Data contains personal information subject to regulation." }, { "source": "mlsys.gradient-clipping", "relation": "mitigates", "target": "mlsys.training-instability", "statement": "suppresses", "conditions": "Instability arises from occasional very large gradients." }, { "source": "mlsys.gradient-clipping", "relation": "part-of", "target": "mlsys.optimizer", "statement": "is applied in the update step of", "conditions": "Gradients are clipped before the optimizer step." }, { "source": "mlsys.weight-initialization", "relation": "mitigates", "target": "mlsys.training-instability", "statement": "prevents early", "conditions": "Networks are deep and use rectifier or residual blocks." }, { "source": "mlsys.hyperparameter-transfer", "relation": "determines", "target": "mlsys.weight-initialization", "statement": "prescribes width-dependent scales for", "conditions": "Hyperparameters tuned on a small model are reused at a larger width." }, { "source": "mlsys.delivery-semantics", "relation": "uses", "target": "mlsys.idempotency", "statement": "achieves effectively-once results with", "conditions": "Records may be redelivered after failures." }, { "source": "mlsys.delivery-semantics", "relation": "part-of", "target": "mlsys.stream-processing", "statement": "is a guarantee of", "conditions": "The pipeline retries or replays records after failures." }, { "source": "mlsys.online-feature-consistency", "relation": "mitigates", "target": "mlsys.train-serve-skew", "statement": "prevents", "conditions": "Online and offline stores are populated by separate jobs." }, { "source": "mlsys.online-feature-consistency", "relation": "trades-off", "target": "mlsys.feature-freshness", "statement": "trades against", "conditions": "Features are updated faster than they can be materialized consistently." }, { "source": "mlsys.ood-detection", "relation": "detects", "target": "mlsys.distribution-shift", "statement": "flags individual inputs affected by", "conditions": "Inputs fall outside the training distribution." }, { "source": "mlsys.ood-detection", "relation": "enables", "target": "mlsys.selective-prediction", "statement": "supplies abstention signals to", "conditions": "The system may defer uncertain inputs." }, { "source": "mlsys.model-feature-compatibility", "relation": "limits", "target": "mlsys.rollout", "statement": "constrains", "conditions": "Model, features, and consumers are released on different schedules." }, { "source": "mlsys.schema-evolution", "relation": "limits", "target": "mlsys.model-feature-compatibility", "statement": "can break", "conditions": "A schema change is not mirrored in the model's feature transformations." }, { "source": "mlsys.secrets-management", "relation": "implements", "target": "mlsys.least-privilege", "statement": "enforces", "conditions": "Credentials are scoped per workload and rotated." }, { "source": "mlsys.agent-credentials", "relation": "uses", "target": "mlsys.secrets-management", "statement": "draws on", "conditions": "Agents act with credentials issued by the platform." }, { "source": "mlsys.deletion-propagation", "relation": "requires", "target": "mlsys.machine-unlearning", "statement": "extends to models through", "conditions": "Deleted data has already been used for training." }, { "source": "mlsys.deletion-propagation", "relation": "uses", "target": "mlsys.lineage", "statement": "locates copies through", "conditions": "Derived datasets and features record their sources." }, { "source": "mlsys.caching-allocator", "relation": "causes", "target": "mlsys.memory-fragmentation", "statement": "can cause", "conditions": "Tensor sizes vary across steps or requests." }, { "source": "mlsys.memory-fragmentation", "relation": "increases", "target": "mlsys.memory-footprint", "statement": "inflates reserved", "conditions": "Freed blocks cannot be reused for new allocation sizes." }, { "source": "mlsys.perception-latency-budget", "relation": "part-of", "target": "mlsys.feedback-delay", "statement": "is the sensing share of", "conditions": "Perception runs inside the control loop." }, { "source": "mlsys.perception-latency-budget", "relation": "limits", "target": "mlsys.robot-perception", "statement": "bounds model size and resolution for", "conditions": "Perception runs on fixed onboard hardware." }, { "source": "mlsys.onboard-compute-partitioning", "relation": "uses", "target": "mlsys.mixed-criticality", "statement": "respects", "conditions": "Safety-critical and best-effort tasks share hardware." }, { "source": "mlsys.onboard-compute-partitioning", "relation": "specializes", "target": "mlsys.placement", "statement": "is execution placement on a robot", "conditions": "Workloads can run on several onboard processors or be offloaded." }, { "source": "mlsys.health-checks", "relation": "enables", "target": "mlsys.load-balancing", "statement": "tells which replicas may receive traffic from", "conditions": "Replicas can be unhealthy or still loading weights." }, { "source": "mlsys.health-checks", "relation": "mitigates", "target": "mlsys.cold-start", "statement": "keeps traffic away from replicas in", "conditions": "A replica reports ready only after its model is loaded and warmed." }, { "source": "mlsys.python-gil", "relation": "increases", "target": "mlsys.dispatch", "statement": "serializes and adds to", "conditions": "Several Python threads issue work concurrently." }, { "source": "mlsys.python-gil", "relation": "limits", "target": "mlsys.host-data-loader", "statement": "pushes preprocessing into worker processes in", "conditions": "Preprocessing runs in Python, so loaders use worker processes rather than threads." }, { "source": "mlsys.flash-endurance", "relation": "limits", "target": "mlsys.ota-model-update", "statement": "bounds the frequency of", "conditions": "Updates and logs are written to on-device flash." }, { "source": "mlsys.flash-endurance", "relation": "limits", "target": "mlsys.microcontroller", "statement": "limits the write budget of", "conditions": "The device stores weights and data in flash." }, { "source": "mlsys.intermittent-computing", "relation": "uses", "target": "mlsys.microcontroller", "statement": "typically runs on", "conditions": "On energy-harvesting deployments built around microcontroller-class processors." }, { "source": "mlsys.intermittent-computing", "relation": "requires", "target": "mlsys.checkpoint", "statement": "requires fine-grained", "conditions": "Power can fail mid-inference." }, { "source": "mlsys.mmap-weights", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "removes duplicate copies from", "conditions": "Several processes serve the same read-only weights." }, { "source": "mlsys.mmap-weights", "relation": "reduces", "target": "mlsys.cold-start", "statement": "shortens", "conditions": "Only the pages touched by the first requests must load." }, { "source": "mlsys.os-memory-pressure", "relation": "limits", "target": "mlsys.on-device-memory-budget", "statement": "enforces", "conditions": "The OS kills or pages out apps that exceed their memory limit." }, { "source": "mlsys.mmap-weights", "relation": "mitigates", "target": "mlsys.os-memory-pressure", "statement": "makes weight pages reclaimable without killing the process under", "conditions": "When file-backed pages can be dropped and reread instead of the app being terminated." }, { "source": "mlsys.managed-runtime-interop", "relation": "limits", "target": "mlsys.on-device-runtime", "statement": "adds copy and pause overhead to", "conditions": "The app calls native inference from a managed language." }, { "source": "mlsys.managed-runtime-interop", "relation": "increases", "target": "mlsys.tail-latency", "statement": "adds garbage-collection pauses to", "conditions": "Large buffers are allocated on the managed heap per request." }, { "source": "mlsys.zero-copy-dma", "relation": "reduces", "target": "mlsys.feedback-delay", "statement": "removes copy time from", "conditions": "Sensor data feeds a latency-critical loop." }, { "source": "mlsys.sensor-fusion", "relation": "mitigates", "target": "mlsys.sensor-interference", "statement": "can outvote corrupted readings from", "conditions": "Independent modalities observe the same scene." }, { "source": "mlsys.network-propagation-delay", "relation": "limits", "target": "mlsys.tail-latency", "statement": "sets a floor under", "conditions": "Requests cross regions or continents." }, { "source": "mlsys.network-propagation-delay", "relation": "part-of", "target": "mlsys.alpha-beta-model", "statement": "contributes to the latency term of", "conditions": "Messages travel long physical distances." }, { "source": "mlsys.cap-tradeoff", "relation": "limits", "target": "mlsys.consensus", "statement": "forces a minority partition to stop answering in", "conditions": "A partition separates replicas." }, { "source": "mlsys.cap-tradeoff", "relation": "determines", "target": "mlsys.feature-freshness", "statement": "decides whether stale values are served for", "conditions": "Feature or model replicas span regions that can partition." }, { "source": "mlsys.gossip-averaging", "relation": "implements", "target": "mlsys.federated-learning", "statement": "coordinates, without a central server,", "conditions": "When participants are separately controlled data holders that exchange model updates with peers rather than with a server." }, { "source": "mlsys.gossip-averaging", "relation": "alternative-to", "target": "mlsys.collectives", "statement": "replaces global collectives with neighbor exchange", "conditions": "Global synchronization is too slow or unreliable." }, { "source": "mlsys.interrupt-latency", "relation": "increases", "target": "mlsys.feedback-delay", "statement": "adds to", "conditions": "Sensor events are delivered through interrupts." }, { "source": "mlsys.transformer-variants", "relation": "specializes", "target": "mlsys.transformer", "statement": "are arrangements of", "conditions": "Models are built from Transformer blocks." }, { "source": "mlsys.transformer-variants", "relation": "determines", "target": "mlsys.prefill-decode", "statement": "decides whether serving has separate", "conditions": "The model generates output autoregressively." }, { "source": "mlsys.data-transfer-cost", "relation": "increases", "target": "mlsys.unit-cost", "statement": "adds to", "conditions": "Data, checkpoints, or responses cross billed network boundaries." }, { "source": "mlsys.data-transfer-cost", "relation": "part-of", "target": "mlsys.tco", "statement": "is a component of", "conditions": "Workloads span zones, regions, or providers." }, { "source": "mlsys.duty-cycling", "relation": "uses", "target": "mlsys.idle-dynamic-power", "statement": "exploits the gap between", "conditions": "Sleep power is far below active power." }, { "source": "mlsys.duty-cycling", "relation": "reduces", "target": "mlsys.energy", "statement": "cuts average", "conditions": "Sensing and inference are needed only intermittently." }, { "source": "mlsys.priority-inversion", "relation": "limits", "target": "mlsys.real-time-scheduling", "statement": "breaks the guarantees of", "conditions": "Tasks of different priorities share a lock or resource." }, { "source": "mlsys.priority-inversion", "relation": "increases", "target": "mlsys.feedback-delay", "statement": "can unboundedly extend", "conditions": "No priority inheritance protocol is in place." }, { "source": "mlsys.randomized-smoothing", "relation": "mitigates", "target": "mlsys.evasion", "statement": "provably bounds", "conditions": "Perturbations are within the certified radius." }, { "source": "mlsys.randomized-smoothing", "relation": "alternative-to", "target": "mlsys.adversarial-training", "statement": "trades inference cost for certified guarantees against", "conditions": "A provable bound is required rather than empirical robustness." }, { "source": "mlsys.zero-copy-dma", "relation": "uses", "target": "mlsys.sensor-interconnect", "statement": "moves data arriving over", "conditions": "The peripheral supports direct memory access." }, { "source": "mlsys.finite-buffer", "relation": "limits", "target": "mlsys.queueing", "statement": "caps, by dropping or blocking arrivals once full, the", "conditions": "When the buffer bound is reached under bursty input; blocked or dropped arrivals pay loss or upstream delay instead of waiting." }, { "source": "mlsys.backpressure", "relation": "mitigates", "target": "mlsys.finite-buffer", "statement": "prevents overflow of", "conditions": "Upstream producers can slow down when the buffer fills." }, { "source": "mlsys.ui-thread-inference", "relation": "uses", "target": "mlsys.on-device-runtime", "statement": "schedules calls into", "conditions": "The app has a main thread that must stay responsive." }, { "source": "mlsys.ui-thread-inference", "relation": "reduces", "target": "mlsys.tail-latency", "statement": "removes inference time from interactive", "conditions": "Inference takes longer than a frame interval." }, { "source": "mlsys.multi-head-latent-attention", "relation": "reduces", "target": "mlsys.kv-cache", "statement": "shrinks the per-token footprint of", "conditions": "When the latent and any decoupled positional components are smaller than the per-head keys and values they replace." }, { "source": "mlsys.multi-head-latent-attention", "relation": "alternative-to", "target": "mlsys.grouped-query-attention", "statement": "is an alternative KV-size lever to", "conditions": "When an architecture chooses how to reduce cached attention state at design time." }, { "source": "mlsys.multi-head-latent-attention", "relation": "specializes", "target": "mlsys.attention", "statement": "is a variant of", "conditions": "When keys and values are produced from a shared compressed representation." }, { "source": "mlsys.training-data-extraction", "relation": "causes", "target": "mlsys.privacy-leakage", "statement": "reveals training content as", "conditions": "When recovered training examples reveal information about people or records." }, { "source": "mlsys.dp-sgd", "relation": "mitigates", "target": "mlsys.training-data-extraction", "statement": "bounds the per-example memorization behind", "conditions": "When the privacy budget is small enough to limit each example's influence, at a utility cost." }, { "source": "mlsys.deduplication", "relation": "reduces", "target": "mlsys.training-data-extraction", "statement": "lowers extractable memorization by removing the repeated sequences behind", "conditions": "When near-duplicates are removed from the pretraining corpus." }, { "source": "mlsys.system-identification", "relation": "alternative-to", "target": "mlsys.domain-randomization", "statement": "narrows the reality gap by measurement instead of randomizing over", "conditions": "As a way to close the sim-to-real dynamics gap when parameters are observable in logged data; the two are often combined." }, { "source": "mlsys.system-identification", "relation": "enables", "target": "mlsys.model-predictive-control", "statement": "supplies the model required by", "conditions": "When MPC uses a physics-based rather than learned dynamics model." }, { "source": "mlsys.system-identification", "relation": "enables", "target": "mlsys.sim-to-real", "statement": "aligns simulated dynamics with hardware for", "conditions": "When identified plant and actuator parameters are imported into the training simulator." }, { "source": "mlsys.off-policy-evaluation", "relation": "alternative-to", "target": "mlsys.online-experiment", "statement": "estimates offline what is otherwise measured by", "conditions": "When logs record action propensities and the new policy's actions are covered by the logged ones." }, { "source": "mlsys.off-policy-evaluation", "relation": "requires", "target": "mlsys.inference-logging", "statement": "needs logged actions, propensities, and outcomes from", "conditions": "When the deployed policy is stochastic or its propensities are recorded." }, { "source": "mlsys.off-policy-evaluation", "relation": "measures", "target": "mlsys.decision-policy", "statement": "estimates the expected outcome of", "conditions": "When the logging policy gave nonzero probability to the candidate's actions and unmeasured confounding is bounded." }, { "source": "mlsys.pipeline-schedule", "relation": "determines", "target": "mlsys.pipeline-bubble", "statement": "sets the idle fraction known as", "conditions": "For a given number of stages and microbatches; fill-drain, one-forward-one-backward, interleaved, and zero-bubble orders give different bubbles." }, { "source": "mlsys.pipeline-schedule", "relation": "part-of", "target": "mlsys.pipeline-parallel", "statement": "is the execution order of", "conditions": "When a model is partitioned into pipeline stages." }, { "source": "mlsys.pipeline-schedule", "relation": "determines", "target": "mlsys.memory-footprint", "statement": "sets how many microbatches' activations each stage holds in", "conditions": "When stages keep forward activations until the matching backward pass runs." }, { "source": "mlsys.sequence-packing", "relation": "part-of", "target": "mlsys.input-pipeline", "statement": "forms the packed-batch construction stage of", "conditions": "When the training input pipeline assembles multiple examples into each pack." }, { "source": "mlsys.sequence-packing", "relation": "increases", "target": "mlsys.model-flops-utilization", "statement": "removes pad-token work to raise", "conditions": "When variable-length examples would otherwise be padded to the batch's longest sequence." }, { "source": "mlsys.document-chunking", "relation": "uses", "target": "mlsys.tokenization", "statement": "uses token counts to enforce chunk boundaries", "conditions": "For token-budgeted chunkers, whose sizes and overlaps are specified in model tokens." }, { "source": "mlsys.document-chunking", "relation": "determines", "target": "mlsys.dense-retrieval", "statement": "sets the unit embedded and returned by", "conditions": "When documents exceed encoder limits or one document vector would blur distinct topics." }, { "source": "mlsys.document-chunking", "relation": "part-of", "target": "mlsys.rag", "statement": "defines the indexed units of", "conditions": "When the RAG application builds its retrieval corpus from document chunks." }, { "source": "mlsys.resource-exhaustion-attack", "relation": "increases", "target": "mlsys.energy-per-inference", "statement": "inflates", "conditions": "When inputs are chosen to defeat sparsity, early exit, or bounded output length." }, { "source": "mlsys.admission", "relation": "mitigates", "target": "mlsys.resource-exhaustion-attack", "statement": "caps per-client work to blunt", "conditions": "When per-client quotas and token or step limits are enforced." }, { "source": "mlsys.resource-exhaustion-attack", "relation": "reduces", "target": "mlsys.slo-goodput", "statement": "takes completed useful work away from", "conditions": "When attacker-amplified requests contend for shared compute, KV-cache memory, or tool-execution slots." }, { "source": "mlsys.generated-content-watermarking", "relation": "alternative-to", "target": "mlsys.content-provenance", "statement": "marks the content itself instead of attaching signed metadata as", "conditions": "When the requirement is identifying output from a participating generator, rather than verifying signed origin and edit history." }, { "source": "mlsys.generated-content-watermarking", "relation": "uses", "target": "mlsys.sampling-controls", "statement": "biases token choice through", "conditions": "For token-sampling text watermarks applied during decoding." }, { "source": "mlsys.generated-content-watermarking", "relation": "enables", "target": "mlsys.algorithmic-audit", "statement": "supports generated-origin checks in", "conditions": "When an audit uses a compatible detector on marked outputs." }, { "source": "mlsys.model-merging", "relation": "alternative-to", "target": "mlsys.multi-adapter-serving", "statement": "folds several fine-tunes into one weight set instead of", "conditions": "When the fine-tunes need not stay separately addressable and the merged model keeps acceptable quality on each task." }, { "source": "mlsys.model-merging", "relation": "mitigates", "target": "mlsys.catastrophic-forgetting", "statement": "limits", "conditions": "Relative to the fine-tuned model alone, when interpolating with the pretrained base improves measured retention on the original tasks." }, { "source": "mlsys.optical-circuit-switching", "relation": "determines", "target": "mlsys.network-topology", "statement": "determines the active endpoint connectivity", "conditions": "When circuit configurations change which endpoints have direct optical paths." }, { "source": "mlsys.optical-circuit-switching", "relation": "enables", "target": "mlsys.all-to-all", "statement": "supports traffic-matched exchange schedules for", "conditions": "When the exchange pattern is stable long enough that reconfiguration time, typically milliseconds, is amortized." }, { "source": "mlsys.co-packaged-optics", "relation": "reduces", "target": "mlsys.data-movement-energy", "statement": "can reduce interface energy per delivered bit", "conditions": "Compared with pluggable optical modules, with optical conversion and laser power included in the accounting." }, { "source": "mlsys.co-packaged-optics", "relation": "uses", "target": "mlsys.advanced-accelerator-packaging", "statement": "integrates optical engines through", "conditions": "When optical engines share a package substrate or interposer with the switch or compute die." }, { "source": "mlsys.advanced-accelerator-packaging", "relation": "uses", "target": "mlsys.hbm", "statement": "uses closely integrated stacked memory", "conditions": "When a compute package connects processor dies to HBM stacks through dense package-level interfaces." }, { "source": "mlsys.advanced-accelerator-packaging", "relation": "determines", "target": "mlsys.thermal-dynamics", "statement": "determines package heat paths and thermal coupling", "conditions": "When die placement, stacking, and bonding geometry constrain heat removal." }, { "source": "mlsys.motion-forecasting", "relation": "requires", "target": "mlsys.multi-object-tracking", "statement": "takes agent histories from", "conditions": "When surrounding agents are tracked with persistent identities over time." }, { "source": "mlsys.motion-planning", "relation": "uses", "target": "mlsys.motion-forecasting", "statement": "plans against the predicted occupancy from", "conditions": "In environments shared with other moving agents such as people or vehicles." }, { "source": "mlsys.representational-bias-evaluation", "relation": "uses", "target": "mlsys.sliced-evaluation", "statement": "compares portrayals across groups through", "conditions": "When outputs for different groups are measured separately." }, { "source": "mlsys.representational-bias-evaluation", "relation": "uses", "target": "mlsys.llm-judge", "statement": "scores outputs at scale with", "conditions": "When scoring is done by a model grader rather than classifiers or likelihood comparisons, and the grader's own bias is measured." }, { "source": "mlsys.group-relative-policy-optimization", "relation": "specializes", "target": "mlsys.reinforcement-learning", "statement": "specializes policy optimization within", "conditions": "When group-relative advantages drive policy updates." }, { "source": "mlsys.group-relative-policy-optimization", "relation": "uses", "target": "mlsys.rl-rollout", "statement": "uses grouped sampled responses from", "conditions": "When several responses or trajectories are sampled for each prompt." }, { "source": "mlsys.group-relative-policy-optimization", "relation": "reduces", "target": "mlsys.training-state", "statement": "removes the critic model's parameters and optimizer state from", "conditions": "Compared with PPO-style actor-critic training that keeps a separate value model." }, { "source": "mlsys.inter-token-latency", "relation": "measures", "target": "mlsys.token-streaming", "statement": "measures gaps in incremental output delivery", "conditions": "When timestamps and the number of tokens carried by each streamed output are recorded." }, { "source": "mlsys.inter-token-latency", "relation": "determines", "target": "mlsys.slo-goodput", "statement": "decides whether streamed requests count toward", "conditions": "When the service objective bounds per-gap latency, with workload and other acceptance criteria fixed." }, { "source": "mlsys.hybrid-retrieval", "relation": "uses", "target": "mlsys.lexical-retrieval", "statement": "uses lexical candidate retrieval", "conditions": "When exact terms, identifiers, or keyword matches contribute a candidate list." }, { "source": "mlsys.hybrid-retrieval", "relation": "uses", "target": "mlsys.dense-retrieval", "statement": "uses embedding-based candidate retrieval", "conditions": "When semantic similarity contributes a candidate list that is fused with lexical results." }, { "source": "mlsys.fpga-acceleration", "relation": "implements", "target": "mlsys.domain-specific-architecture", "statement": "implements workload-specialized hardware", "conditions": "When the configured fabric realizes an ML accelerator datapath." }, { "source": "mlsys.fpga-acceleration", "relation": "uses", "target": "mlsys.dataflow", "statement": "uses explicit operand placement and streaming schedules", "conditions": "When operators are mapped onto configured arithmetic and memory resources." }, { "source": "mlsys.device-initiated-communication", "relation": "uses", "target": "mlsys.rdma", "statement": "uses direct remote-memory transport", "conditions": "For inter-node transfers over a device-initiated RDMA path." }, { "source": "mlsys.device-initiated-communication", "relation": "reduces", "target": "mlsys.dispatch", "statement": "reduces per-exchange host orchestration", "conditions": "When device-side communication replaces repeated host launches or host-issued communication operations." }, { "source": "mlsys.synthetic-data", "relation": "causes", "target": "mlsys.model-collapse", "statement": "can cause recursive distributional degradation", "conditions": "When successive generations train mainly on earlier models' outputs without retaining real data." }, { "source": "mlsys.model-collapse", "relation": "reduces", "target": "mlsys.generalization", "statement": "reduces performance on the original target distribution", "conditions": "When disappearing modes or tails remain relevant to evaluation and deployment." }, { "source": "mlsys.action-tokenization", "relation": "uses", "target": "mlsys.quantization", "statement": "uses discretization of continuous action information", "conditions": "When continuous commands or transformed action coefficients are mapped to a finite vocabulary." }, { "source": "mlsys.action-tokenization", "relation": "determines", "target": "mlsys.execution-time", "statement": "sets the sequential decoding work in", "conditions": "When actions are decoded autoregressively and architecture, hardware, and action horizon are held fixed." }, { "source": "mlsys.action-tokenization", "relation": "part-of", "target": "mlsys.vla-model", "statement": "is the action output representation of", "conditions": "For vision-language-action models that emit actions as tokens." }, { "source": "mlsys.wafer-scale-acceleration", "relation": "specializes", "target": "mlsys.domain-specific-architecture", "statement": "specializes accelerator organization at wafer scale", "conditions": "When the wafer-scale device is designed for ML computation." }, { "source": "mlsys.wafer-scale-acceleration", "relation": "uses", "target": "mlsys.dataflow", "statement": "uses explicit operand and intermediate placement", "conditions": "When work is mapped across distributed compute tiles and local memories." }, { "source": "mlsys.embedding-migration", "relation": "uses", "target": "mlsys.model-feature-compatibility", "statement": "keeps query and corpus encoders aligned through", "conditions": "When query encoders and stored corpus representations change across releases." }, { "source": "mlsys.embedding-migration", "relation": "uses", "target": "mlsys.feature-backfill", "statement": "re-encodes historical corpus records in shadow indexes using", "conditions": "Before switching live query traffic to the new embedding model." }, { "source": "mlsys.register-pressure", "relation": "limits", "target": "mlsys.occupancy", "statement": "caps the number of resident warps in", "conditions": "When per-thread register allocation times threads per block exceeds the register file." }, { "source": "mlsys.register-pressure", "relation": "limits", "target": "mlsys.tiling", "statement": "bounds the register-tile size usable in", "conditions": "When sizing thread-level accumulator tiles in matrix-multiply or attention kernels." }, { "source": "mlsys.software-pipelining", "relation": "uses", "target": "mlsys.async-data-movement", "statement": "issues the multi-stage copies it overlaps through", "conditions": "On hardware with asynchronous copy support." }, { "source": "mlsys.software-pipelining", "relation": "requires", "target": "mlsys.on-chip-sram", "statement": "allocates two or more stage buffers in", "conditions": "So loads for the next iteration do not overwrite operands the current iteration still reads." }, { "source": "mlsys.software-pipelining", "relation": "uses", "target": "mlsys.tiling", "statement": "overlaps the loads and compute of successive tiles from", "conditions": "When a kernel's loop iterates over tiles staged through on-chip memory." }, { "source": "mlsys.ecc-memory", "relation": "mitigates", "target": "mlsys.silent-data-corruption", "statement": "turns many would-be silent memory corruptions into corrected or reported faults, limiting", "conditions": "For errors within the code's correction and detection strength, in ECC-protected memory." }, { "source": "mlsys.ecc-memory", "relation": "increases", "target": "mlsys.availability", "statement": "keeps jobs running through correctable errors, raising", "conditions": "When correctable memory errors would otherwise crash jobs or corrupt their results." }, { "source": "mlsys.integer-only-inference", "relation": "requires", "target": "mlsys.quantization", "statement": "runs on values represented by", "conditions": "When every operator executes in integer arithmetic." }, { "source": "mlsys.on-device-runtime", "relation": "uses", "target": "mlsys.integer-only-inference", "statement": "executes fully integer operator graphs through", "conditions": "When the runtime targets an embedded device without a floating-point unit." }, { "source": "mlsys.edge-npu", "relation": "requires", "target": "mlsys.integer-only-inference", "statement": "accepts only fixed-point operator graphs lowered by", "conditions": "For NPUs that support only integer operators." }, { "source": "mlsys.execute-in-place", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "keeps frozen weights out of working SRAM in", "conditions": "When read-only parameters stay in memory-mapped flash during inference." }, { "source": "mlsys.microcontroller", "relation": "uses", "target": "mlsys.execute-in-place", "statement": "runs models whose weights exceed on-chip SRAM via", "conditions": "On microcontrollers whose flash is memory-mapped onto the processor bus." }, { "source": "mlsys.memory-bandwidth", "relation": "limits", "target": "mlsys.execute-in-place", "statement": "caps the weight-fetch rate of", "conditions": "Flash read bandwidth and wait states bound execute-in-place inference when operators reuse weights little." }, { "source": "mlsys.prompt-injection", "relation": "causes", "target": "mlsys.confused-deputy", "statement": "can steer a privileged agent into", "conditions": "When injected instructions induce a privileged agent to perform an action the untrusted party was not authorized to request." }, { "source": "mlsys.capability-based-security", "relation": "mitigates", "target": "mlsys.confused-deputy", "statement": "binds each action to narrow, task-scoped authority to prevent", "conditions": "When ambient user or service credentials are replaced with attenuated capabilities." }, { "source": "mlsys.agent-credentials", "relation": "mitigates", "target": "mlsys.confused-deputy", "statement": "scopes and brokers per-action tokens outside model context to limit", "conditions": "When tool calls require explicit user- and resource-scoped delegation." }, { "source": "mlsys.judge-bias", "relation": "limits", "target": "mlsys.llm-judge", "statement": "distorts rankings and win-rate estimates produced by", "conditions": "When candidate responses differ in length, formatting, or presentation order." }, { "source": "mlsys.human-evaluation", "relation": "detects", "target": "mlsys.judge-bias", "statement": "measures the size of", "conditions": "When human labels are collected on a calibration sample." }, { "source": "mlsys.guardrails", "relation": "causes", "target": "mlsys.over-refusal", "statement": "produces", "conditions": "When a safety check or refusal policy incorrectly blocks a benign request." }, { "source": "mlsys.over-refusal", "relation": "reduces", "target": "mlsys.slo-goodput", "statement": "rejects valid requests, reducing", "conditions": "When benign task completion is required for work to count as useful." }, { "source": "mlsys.sensor-fusion", "relation": "requires", "target": "mlsys.coordinate-frames", "statement": "projects heterogeneous sensor measurements into a common reference frame using", "conditions": "Before combining camera, lidar, radar, and inertial observations." }, { "source": "mlsys.sensing", "relation": "produces", "target": "mlsys.coordinate-frames", "statement": "yields the static mounting transforms of", "conditions": "When sensors are calibrated on the physical platform." }, { "source": "mlsys.coordinate-frames", "relation": "enables", "target": "mlsys.slam", "statement": "tracks the evolving transform between map, odometry, and base frames for", "conditions": "During continuous state estimation and loop closure." }, { "source": "mlsys.sensor-degradation", "relation": "limits", "target": "mlsys.operating-domain", "statement": "constrains the weather, lighting, and visibility conditions admitted in", "conditions": "Unless detected at runtime and mitigated by complementary modalities or fallback." }, { "source": "mlsys.sensor-fusion", "relation": "mitigates", "target": "mlsys.sensor-degradation", "statement": "combines modalities with distinct physical failure modes to mitigate", "conditions": "When complementary sensors retain useful measurements and fusion handles degraded inputs." }, { "source": "mlsys.data-curriculum", "relation": "uses", "target": "mlsys.data-mixture", "statement": "schedules over the course of training", "conditions": "When the curriculum changes source proportions between training phases." }, { "source": "mlsys.data-curriculum", "relation": "enables", "target": "mlsys.context-extension", "statement": "raises sequence length late in pretraining to support", "conditions": "When positional handling supports the longer lengths and evaluation verifies useful behavior at them." }, { "source": "mlsys.data-curriculum", "relation": "part-of", "target": "mlsys.pretraining", "statement": "structures the data schedule of", "conditions": "When pretraining runs in phases with different data distributions." }, { "source": "mlsys.streaming-dataset-format", "relation": "specializes", "target": "mlsys.dataset-storage-layout", "statement": "is the sequential-shard form of", "conditions": "When small records are packed into large files read sequentially from object storage or remote file systems." }, { "source": "mlsys.streaming-dataset-format", "relation": "uses", "target": "mlsys.object-storage", "statement": "streams contiguous byte ranges from", "conditions": "When shards live in object storage rather than on a local file system." }, { "source": "mlsys.streaming-dataset-format", "relation": "enables", "target": "mlsys.input-pipeline", "statement": "supports sharded, resumable reads for", "conditions": "When the loader also checkpoints shard assignment, record progress, and shuffle state." }, { "source": "mlsys.collective-algorithm-selection", "relation": "uses", "target": "mlsys.alpha-beta-model", "statement": "chooses the crossover message size between latency- and bandwidth-optimal schedules using", "conditions": "When tuning collective execution across intra-node and inter-node links." }, { "source": "mlsys.collective-algorithm-selection", "relation": "uses", "target": "mlsys.ring-allreduce", "statement": "selects, for large messages,", "conditions": "When message size is large relative to participant count." }, { "source": "mlsys.collective-algorithm-selection", "relation": "uses", "target": "mlsys.tree-allreduce", "statement": "selects, for small messages or many participants,", "conditions": "When per-step latency dominates transfer time." }, { "source": "mlsys.always-on-cascade", "relation": "uses", "target": "mlsys.duty-cycling", "statement": "keeps the main application processor asleep until triggered using", "conditions": "While the low-power island continuously monitors raw sensor streams." }, { "source": "mlsys.always-on-cascade", "relation": "uses", "target": "mlsys.model-cascade", "statement": "applies the staged escalation of", "conditions": "When the wake-up stage is itself a small model rather than a sensing circuit." }, { "source": "mlsys.always-on-cascade", "relation": "reduces", "target": "mlsys.energy-per-inference", "statement": "rejects background frames on low-power hardware to lower the average", "conditions": "Averaged per sensed frame over long idle intervals, not per main-model inference." }, { "source": "mlsys.fault-tree-analysis", "relation": "specializes", "target": "mlsys.hazard-analysis", "statement": "is a method of", "conditions": "When the analysis starts from a defined hazardous top event." }, { "source": "mlsys.fault-tree-analysis", "relation": "uses", "target": "mlsys.failure-model", "statement": "combines the basic fault events of", "conditions": "When defining the fault events and assumptions the tree combines." }, { "source": "mlsys.retry-backoff", "relation": "mitigates", "target": "mlsys.metastable-failure", "statement": "damps the retry amplification that sustains", "conditions": "When retries are jittered and capped by a retry budget." }, { "source": "mlsys.admission", "relation": "mitigates", "target": "mlsys.metastable-failure", "statement": "sheds the excess load that sustains", "conditions": "When early rejection lowers admitted load below the degraded system's recovery threshold long enough to break the feedback loop." }, { "source": "mlsys.fan-out-tail-amplification", "relation": "increases", "target": "mlsys.tail-latency", "statement": "raises", "conditions": "When completion waits for all subrequests and their slow events are not perfectly coincident." }, { "source": "mlsys.hedged-requests", "relation": "mitigates", "target": "mlsys.fan-out-tail-amplification", "statement": "cuts the slow-subrequest tail behind", "conditions": "When alternate replicas have independent delays, spare capacity exists, and redundant work is cancelled." }, { "source": "mlsys.group-fairness-metric", "relation": "measures", "target": "mlsys.proxy-discrimination", "statement": "measures the disparity produced by", "conditions": "When controlled comparisons or causal evidence isolate the part of the disparity due to proxy features." }, { "source": "mlsys.bias-mitigation", "relation": "mitigates", "target": "mlsys.proxy-discrimination", "statement": "reduces the disparity from", "conditions": "When mitigation constrains outcomes rather than only dropping the attribute." }, { "source": "mlsys.cross-embodiment-training", "relation": "uses", "target": "mlsys.teleoperation-data", "statement": "pools", "conditions": "When demonstrations come from several robot types." }, { "source": "mlsys.cross-embodiment-training", "relation": "uses", "target": "mlsys.data-mixture", "statement": "weights robot datasets through", "conditions": "When embodiments contribute very unequal amounts of data." }, { "source": "mlsys.dual-rate-hierarchical-policy", "relation": "alternative-to", "target": "mlsys.action-chunking", "statement": "decouples model rate from control rate differently from", "conditions": "When the large model is too slow to emit a new chunk before the current one runs out." }, { "source": "mlsys.dual-rate-hierarchical-policy", "relation": "uses", "target": "mlsys.vla-model", "statement": "queries, as its slow high-level layer,", "conditions": "When the slow layer is a vision-language-action model." }, { "source": "mlsys.trajectory-feasibility", "relation": "requires", "target": "mlsys.robot-kinematics", "statement": "checks joint limits and singularities through", "conditions": "When a task-space trajectory must be mapped into joint space." }, { "source": "mlsys.motion-planning", "relation": "uses", "target": "mlsys.robot-kinematics", "statement": "maps between task and configuration space with", "conditions": "When goals are given as end-effector poses." }, { "source": "mlsys.quantization-granularity", "relation": "part-of", "target": "mlsys.quantization", "statement": "is a design choice within", "conditions": "When choosing scales for weights, activations, or caches." }, { "source": "mlsys.quantization-granularity", "relation": "mitigates", "target": "mlsys.activation-outliers", "statement": "isolates the range inflation from", "conditions": "When groups split along the outlier channels, per channel or finer." }, { "source": "mlsys.block-scaled-format", "relation": "implements", "target": "mlsys.quantization-granularity", "statement": "fixes in hardware a fine-grained choice of", "conditions": "When the number format itself carries one shared scale per small block." }, { "source": "mlsys.agent-training-environment", "relation": "enables", "target": "mlsys.rl-rollout", "statement": "provides the world acted in during", "conditions": "When rollouts are multi-turn and call tools." }, { "source": "mlsys.agent-training-environment", "relation": "uses", "target": "mlsys.ephemeral-sandbox", "statement": "isolates each episode with", "conditions": "When episodes execute code or mutate files." }, { "source": "mlsys.model-directed-control-flow", "relation": "alternative-to", "target": "mlsys.workflow-orchestration", "statement": "replaces code-defined step selection in", "conditions": "When the steps cannot be enumerated in advance." }, { "source": "mlsys.model-directed-control-flow", "relation": "part-of", "target": "mlsys.agent-harness", "statement": "is the decision loop inside", "conditions": "When the harness lets model outputs choose the next action." }, { "source": "mlsys.agent-budget", "relation": "limits", "target": "mlsys.model-directed-control-flow", "statement": "bounds the runs of", "conditions": "When the harness enforces step, token, or cost limits on model-chosen actions." }, { "source": "mlsys.misuse-monitoring", "relation": "uses", "target": "mlsys.inference-logging", "statement": "analyzes", "conditions": "When logs keep enough request context and retention policy allows." }, { "source": "mlsys.misuse-monitoring", "relation": "detects", "target": "mlsys.jailbreak", "statement": "detects patterns of", "conditions": "When request and output traces show an actual policy bypass, with cross-session correlation supporting the diagnosis." }, { "source": "mlsys.masked-diffusion-decoding", "relation": "uses", "target": "mlsys.diffusion-model", "statement": "generates text with a discrete denoising model from", "conditions": "When a language model is trained to predict masked tokens over several refinement passes." }, { "source": "mlsys.masked-diffusion-decoding", "relation": "limits", "target": "mlsys.prefix-cache", "statement": "limits unchanged-prefix state reuse in", "conditions": "When refinement passes can revise earlier positions, so a fixed prefix's attention state is not final." }, { "source": "mlsys.residual-connection", "relation": "part-of", "target": "mlsys.transformer", "statement": "wraps each attention and feed-forward sublayer of", "conditions": "In standard pre-norm and post-norm transformer blocks." }, { "source": "mlsys.residual-connection", "relation": "mitigates", "target": "mlsys.training-instability", "statement": "helps keep deep stacks trainable to limit", "conditions": "When identity shortcuts and residual-branch scaling are used in networks deep enough that gradients through stacked layers would otherwise degrade." }, { "source": "mlsys.multi-task-learning", "relation": "alternative-to", "target": "mlsys.multi-model-serving", "statement": "consolidates several tasks into one shared-backbone model instead of hosting separate models in", "conditions": "When tasks share inputs and enough representation that one backbone serves them without unacceptable quality loss." }, { "source": "mlsys.multi-task-learning", "relation": "alternative-to", "target": "mlsys.transfer-learning", "statement": "learns tasks jointly rather than carrying learning across as in", "conditions": "When several related tasks are available at training time." }, { "source": "mlsys.recovery-objectives", "relation": "limits", "target": "mlsys.checkpoint-interval", "statement": "caps the longest acceptable", "conditions": "When recovery restarts from the last checkpoint and the RPO is shorter than the interval that failure rate and write cost would otherwise favor." }, { "source": "mlsys.recovery-objectives", "relation": "determines", "target": "mlsys.replication-failover", "statement": "sets the recovery speed required of", "conditions": "When the RTO is shorter than the time to restore from a checkpoint, so a standby must take over." }, { "source": "mlsys.incremental-checkpointing", "relation": "implements", "target": "mlsys.checkpoint", "statement": "produces, as a base plus a chain of changes,", "conditions": "When only a fraction of the state, such as touched embedding rows, changes between checkpoints." }, { "source": "mlsys.incremental-checkpointing", "relation": "reduces", "target": "mlsys.checkpoint-io", "statement": "cuts the bytes written by", "conditions": "When the changed fraction of state is small; restore must read the base and every later delta." }, { "source": "mlsys.multiple-comparisons", "relation": "mitigates", "target": "mlsys.selection-bias", "statement": "limits false declarations of a winner that feed", "conditions": "When many candidates, metrics, or slices are compared on the same evaluation data; it does not correct the winner's overstated effect size." }, { "source": "mlsys.online-experiment", "relation": "uses", "target": "mlsys.multiple-comparisons", "statement": "controls error rates across many decision metrics with", "conditions": "When an experiment's launch decision rests on several metrics or segments." }, { "source": "mlsys.model-assisted-labeling", "relation": "trades-off", "target": "mlsys.labels", "statement": "trades annotation throughput against the quality of", "conditions": "When annotators review model drafts rather than labeling from scratch, so anchoring can carry model errors into labels." }, { "source": "mlsys.model-assisted-labeling", "relation": "alternative-to", "target": "mlsys.weak-supervision", "statement": "keeps a per-item human review step that is not part of", "conditions": "When the label budget allows human confirmation of each draft label." }, { "source": "mlsys.forward-error-correction", "relation": "part-of", "target": "mlsys.scale-up-interconnect", "statement": "corrects bit errors on the links of", "conditions": "When high-speed SerDes links run a FEC code on every hop, adding decoding delay to each." }, { "source": "mlsys.rdma", "relation": "uses", "target": "mlsys.forward-error-correction", "statement": "avoids costly go-back-N retransmissions on fast links through", "conditions": "When the RDMA transport recovers from loss by go-back-N retransmission, so even rare uncorrected bit errors would stall throughput." }, { "source": "mlsys.object-storage", "relation": "uses", "target": "mlsys.erasure-coding", "statement": "protects stored objects with", "conditions": "When the store keeps parity fragments rather than full copies." }, { "source": "mlsys.erasure-coding", "relation": "alternative-to", "target": "mlsys.replication-failover", "statement": "provides storage durability with less overhead than the full-copy replication in", "conditions": "For durable stored data at comparable failure tolerance, not for request failover; reads after a loss pay reconstruction cost." }, { "source": "mlsys.bandwidth-delay-product", "relation": "determines", "target": "mlsys.congestion-control", "statement": "sets the window that", "conditions": "When a window-controlled flow must keep a long or fast path full while acknowledgments are in flight." }, { "source": "mlsys.network-propagation-delay", "relation": "determines", "target": "mlsys.bandwidth-delay-product", "statement": "sets the round-trip floor of", "conditions": "For a single path; queueing and processing delay add to the round-trip time." }, { "source": "mlsys.backfill-scheduling", "relation": "increases", "target": "mlsys.utilization", "statement": "fills idle resources to raise", "conditions": "When job runtime estimates are accurate enough that backfilled jobs finish before the reserved start." }, { "source": "mlsys.backfill-scheduling", "relation": "reduces", "target": "mlsys.queueing", "statement": "shortens the wait of small jobs in", "conditions": "When small jobs fit into gaps ahead of a large gang job's reserved start and their runtime estimates are accurate." }, { "source": "mlsys.traffic-classes", "relation": "mitigates", "target": "mlsys.performance-interference", "statement": "isolates competing flows to limit", "conditions": "When collective, storage, and control traffic share the same links and switches and are mapped to separate classes." }, { "source": "mlsys.lossless-fabric", "relation": "uses", "target": "mlsys.traffic-classes", "statement": "applies pause-based flow control per priority class through", "conditions": "When the fabric is made lossless with priority flow control, as in RoCE, rather than with credit-based link flow control." }, { "source": "mlsys.telemetry-sampling", "relation": "trades-off", "target": "mlsys.observability", "statement": "trades telemetry cost against the investigative reach of", "conditions": "When sampling drops the records, or aggregation drops the label detail, that would explain a rare failure." }, { "source": "mlsys.distributed-tracing", "relation": "uses", "target": "mlsys.telemetry-sampling", "statement": "keeps a fraction of request traces through", "conditions": "When recording every span for every request would exceed storage or processing budgets." }, { "source": "mlsys.zero-copy-dma", "relation": "requires", "target": "mlsys.dma-cache-maintenance", "statement": "stays correct on non-coherent hardware only with", "conditions": "When the CPU accesses cached buffers that a sensor or accelerator also accesses by DMA without hardware coherence." }, { "source": "mlsys.coherent-interconnect", "relation": "alternative-to", "target": "mlsys.dma-cache-maintenance", "statement": "replaces in hardware the software work of", "conditions": "When the link keeps device and host caches coherent, making explicit cache maintenance unnecessary." }, { "source": "mlsys.expert-offloading", "relation": "alternative-to", "target": "mlsys.cpu-gpu-hybrid-inference", "statement": "moves expert weights to the accelerator instead of computing on the host as in", "conditions": "When mixture-of-experts weights exceed accelerator memory and must stay partly in host memory." }, { "source": "mlsys.host-device-interconnect", "relation": "limits", "target": "mlsys.expert-offloading", "statement": "bounds the expert-loading rate of", "conditions": "When each batch needs experts that are not already resident in accelerator memory." }, { "source": "mlsys.expert-offloading", "relation": "uses", "target": "mlsys.expert-routing", "statement": "prefetches the experts predicted by", "conditions": "When routing for upcoming layers or batches can be predicted early enough to hide the transfer." }, { "source": "mlsys.variable-length-parallelism", "relation": "uses", "target": "mlsys.sequence-parallel", "statement": "varies the degree and layout of", "conditions": "When training sequence lengths vary widely within and across steps." }, { "source": "mlsys.variable-length-parallelism", "relation": "mitigates", "target": "mlsys.stragglers", "statement": "rebalances per-rank attention work to limit", "conditions": "When a few long documents would otherwise make some ranks finish each step much later than others." }, { "source": "mlsys.variable-length-parallelism", "relation": "uses", "target": "mlsys.sequence-packing", "statement": "chooses document-to-rank assignments in", "conditions": "When packed sequences are distributed across the ranks of a sequence-parallel group." }, { "source": "mlsys.fine-grained-accelerator-sharing", "relation": "specializes", "target": "mlsys.gpu-partitioning", "statement": "is the sub-kernel scheduling form of", "conditions": "When co-located workloads need tighter latency isolation than whole-kernel time slicing gives and more flexibility than fixed hardware partitions." }, { "source": "mlsys.fine-grained-accelerator-sharing", "relation": "mitigates", "target": "mlsys.performance-interference", "statement": "bounds the slowdown of latency-sensitive work from", "conditions": "When the scheduler can preempt or throttle best-effort kernels within a kernel's execution." }, { "source": "mlsys.intra-device-overlap", "relation": "alternative-to", "target": "mlsys.pd-disaggregation", "statement": "co-locates on one device the phases separated by", "conditions": "When a deployment chooses between running prefill and decode on the same devices or on separate pools." }, { "source": "mlsys.intra-device-overlap", "relation": "increases", "target": "mlsys.utilization", "statement": "keeps several hardware resources busy at once, raising", "conditions": "When the co-scheduled operations are bound by different resources and scheduling overhead stays small." }, { "source": "mlsys.intra-device-overlap", "relation": "uses", "target": "mlsys.chunked-prefill", "statement": "overlaps the prefill and decode work in the mixed batches formed by", "conditions": "When chunked prefill puts prefill chunks and decode steps in the same iteration, and a fused kernel runs them on the same compute units." }, { "source": "mlsys.graph-pipeline-parallelism", "relation": "specializes", "target": "mlsys.pipeline-parallel", "statement": "is the branching-graph form of", "conditions": "When the model's layer graph has branches with no data dependence between them." }, { "source": "mlsys.graph-pipeline-parallelism", "relation": "reduces", "target": "mlsys.pipeline-bubble", "statement": "shortens the pipeline's critical path, shrinking the fill and drain time of", "conditions": "Relative to a linearized schedule on the same devices and microbatch count, when the branches are balanced closely enough that parallel stages finish at similar times." }, { "source": "mlsys.rag-pipeline-scheduling", "relation": "reduces", "target": "mlsys.ttft", "statement": "overlaps retrieval with generation work to shorten", "conditions": "When retrieval sits on the request's critical path before prefill." }, { "source": "mlsys.rag-pipeline-scheduling", "relation": "reduces", "target": "mlsys.queueing", "statement": "avoids waiting for generation memory through joint configuration and scheduling, cutting", "conditions": "When a fixed RAG configuration would wait for accelerator memory but another configuration meeting the query's quality target fits." }, { "source": "mlsys.cross-stack-profiling", "relation": "specializes", "target": "mlsys.profiling", "statement": "is the cross-layer form of", "conditions": "When the profiler records events at more than one software layer and can link each device event to the host call that issued it." }, { "source": "mlsys.cross-stack-profiling", "relation": "detects", "target": "mlsys.dispatch", "statement": "exposes host-side gaps between device kernels caused by", "conditions": "When host and device events share a clock or are aligned by correlation identifiers." }, { "source": "mlsys.bottleneck-analysis", "relation": "uses", "target": "mlsys.cross-stack-profiling", "statement": "locates the limiting software layer with", "conditions": "When the bottleneck lies in host code or framework overhead rather than in a single kernel." }, { "source": "mlsys.profiling", "relation": "measures", "target": "mlsys.memory-footprint", "statement": "records per-tensor and per-phase allocations that make up", "conditions": "When the profiler tracks allocations over a training step rather than only peak usage." }, { "source": "mlsys.loss-scaling", "relation": "mitigates", "target": "mlsys.training-instability", "statement": "keeps small gradients representable to limit", "conditions": "When low-precision gradients would otherwise underflow to zero or overflow to infinity." }, { "source": "mlsys.profiling", "relation": "detects", "target": "mlsys.stragglers", "statement": "exposes the slow ranks behind", "conditions": "When per-rank step and collective times are recorded and compared." }, { "source": "mlsys.latency-percentiles", "relation": "measures", "target": "mlsys.ttft", "statement": "summarizes the distribution of", "conditions": "When time to first token is recorded per request rather than averaged." }, { "source": "mlsys.autoscaling", "relation": "reduces", "target": "mlsys.queueing", "statement": "adds capacity before requests wait, cutting", "conditions": "When new replicas become ready faster than the load rises; cold starts delay the effect." }, { "source": "mlsys.kv-cache-eviction", "relation": "reduces", "target": "mlsys.kv-cache", "statement": "bounds the resident size of", "conditions": "When the evicted entries receive little attention, so quality holds." }, { "source": "mlsys.kv-cache-offload", "relation": "reduces", "target": "mlsys.kv-cache", "statement": "moves cold blocks out of accelerator memory, shrinking the resident part of", "conditions": "When transfer bandwidth to host memory or storage can bring blocks back before they are needed." }, { "source": "mlsys.online-experiment", "relation": "detects", "target": "mlsys.selection-bias", "statement": "tests on live traffic a winner chosen under", "conditions": "When a model picked by offline comparison is exposed to a controlled share of real users." }, { "source": "mlsys.duty-cycling", "relation": "mitigates", "target": "mlsys.thermal-dynamics", "statement": "lowers average power to limit the heating behind", "conditions": "When average power, not peak power, sets the device's sustained temperature." }, { "source": "mlsys.cost-attribution", "relation": "measures", "target": "mlsys.cost-per-token", "statement": "assigns infrastructure spend to requests to compute", "conditions": "When usage is metered per request, model, or tenant." }, { "source": "mlsys.quantization", "relation": "reduces", "target": "mlsys.cost-per-token", "statement": "raises throughput per accelerator, cutting", "conditions": "When the lower precision keeps quality acceptable and the serving stack has kernels for it." }, { "source": "mlsys.sliced-evaluation", "relation": "detects", "target": "mlsys.over-refusal", "statement": "exposes, on a slice of borderline benign prompts,", "conditions": "When the evaluation set includes benign prompts that resemble harmful ones." }, { "source": "mlsys.preference-optimization", "relation": "mitigates", "target": "mlsys.over-refusal", "statement": "teaches helpful answers to borderline benign requests, limiting", "conditions": "When the preference data includes benign requests that resemble harmful ones, paired with helpful responses." }, { "source": "mlsys.admission", "relation": "limits", "target": "mlsys.kv-cache", "statement": "caps the number of concurrent sequences whose state must stay resident in", "conditions": "When admission is based on projected KV memory rather than request count alone." }, { "source": "mlsys.energy-measurement", "relation": "measures", "target": "mlsys.energy-per-inference", "statement": "records the energy behind", "conditions": "When device power is sampled over the same window as the inference workload." }, { "source": "mlsys.inter-token-latency", "relation": "measures", "target": "mlsys.tpot", "statement": "gives the per-gap distribution whose request average is", "conditions": "When streamed outputs are timestamped individually rather than only at request end." }, { "source": "mlsys.model-swapping", "relation": "part-of", "target": "mlsys.multi-model-serving", "statement": "is the weight-movement mechanism of", "conditions": "When more models are offered than fit in accelerator memory at once." }, { "source": "mlsys.host-device-interconnect", "relation": "limits", "target": "mlsys.model-swapping", "statement": "bounds how fast models can be loaded in", "conditions": "When a model's weights must move from host memory or storage before it can serve." }, { "source": "mlsys.model-swapping", "relation": "increases", "target": "mlsys.utilization", "statement": "lets otherwise idle accelerators serve other models, raising", "conditions": "When demand across models is uneven over time and the transfer latency fits the service objective." }, { "source": "mlsys.paged-kv", "relation": "specializes", "target": "mlsys.kv-memory-allocation", "statement": "is the block-based form of", "conditions": "When attention state is stored in fixed-size blocks mapped by a block table." }, { "source": "mlsys.kv-memory-allocation", "relation": "determines", "target": "mlsys.memory-fragmentation", "statement": "sets how much KV memory is lost to", "conditions": "When sequences of varying and growing length share one accelerator memory pool." }, { "source": "mlsys.kv-memory-allocation", "relation": "implements", "target": "mlsys.kv-cache", "statement": "provides the storage that holds", "conditions": "When a serving engine manages the cache's memory rather than the model framework." }, { "source": "mlsys.feed-forward-block", "relation": "part-of", "target": "mlsys.transformer", "statement": "is the position-wise sublayer paired with attention in each block of", "conditions": "In standard Transformer blocks, where each block alternates an attention sublayer and a feed-forward sublayer under residual connections." }, { "source": "mlsys.mixture-of-experts", "relation": "alternative-to", "target": "mlsys.feed-forward-block", "statement": "replaces the dense sublayer with routed experts as an alternative to", "conditions": "When a block's feed-forward sublayer is implemented as several routed experts instead of one dense network." }, { "source": "mlsys.feed-forward-block", "relation": "uses", "target": "mlsys.matrix-multiply", "statement": "executes its projections as", "conditions": "The projections are dense GEMMs over all tokens, so in dense Transformers they dominate parameter count and prefill compute; gated variants add a third projection." }, { "source": "mlsys.activation-sparsity", "relation": "uses", "target": "mlsys.feed-forward-block", "statement": "skips the predicted-inactive neurons of", "conditions": "When a hidden unit's activation is predicted near zero for the current token, so its weight rows need not be read or multiplied." }, { "source": "mlsys.arrival-rate", "relation": "increases", "target": "mlsys.utilization", "statement": "raises, as a fraction of service capacity, the", "conditions": "When service capacity is fixed and the system is stable, so utilization is the ratio of arrival rate to service capacity." }, { "source": "mlsys.dispatch", "relation": "determines", "target": "mlsys.bound-regime", "statement": "places small operations on the overhead-bound side of", "conditions": "When per-kernel work is so small that host launch and orchestration time exceeds device execution time, as in small-batch decode of small models." }, { "source": "mlsys.service-level-agreement", "relation": "uses", "target": "mlsys.sli", "statement": "is adjudicated on an agreed", "conditions": "When the contract names the indicator, its measurement window, and who measures it." }, { "source": "mlsys.requirements", "relation": "determines", "target": "mlsys.service-level-agreement", "statement": "set the externally promised terms of", "conditions": "When service commitments to customers are part of the stated deployment requirements." }, { "source": "mlsys.head-of-line-blocking", "relation": "increases", "target": "mlsys.tail-latency", "statement": "lengthens, through long items ahead of short ones, the", "conditions": "When service is first-come-first-served and service times vary widely, so one long request delays every shorter one behind it." }, { "source": "mlsys.continuous-batching", "relation": "mitigates", "target": "mlsys.head-of-line-blocking", "statement": "admits and retires requests each iteration to avoid", "conditions": "When finished sequences would otherwise wait for the longest sequence in a static batch before new work is admitted." }, { "source": "mlsys.chunked-prefill", "relation": "mitigates", "target": "mlsys.head-of-line-blocking", "statement": "bounds prefill work per iteration to limit", "conditions": "When a long prompt's prefill would otherwise hold the decode step of every co-batched request." }, { "source": "mlsys.lossless-fabric", "relation": "causes", "target": "mlsys.head-of-line-blocking", "statement": "can cause, through pause frames that stop a whole port or class,", "conditions": "When priority flow control pauses all traffic in a class because one downstream queue is congested." }, { "source": "mlsys.open-closed-loop-load", "relation": "determines", "target": "mlsys.arrival-rate", "statement": "sets whether responses throttle the", "conditions": "Closed-loop clients cap concurrency at their own count, so arrival rate falls as latency rises; open-loop arrivals do not respond to latency." }, { "source": "mlsys.open-closed-loop-load", "relation": "determines", "target": "mlsys.latency-throughput-curve", "statement": "changes, through response feedback and concurrency limits, the offered load behind", "conditions": "When generator concurrency, think time, and response feedback change offered arrivals, the measured curve must be interpreted under that load model; closed-loop tests can still reach saturation." }, { "source": "mlsys.inference-load-scenario", "relation": "uses", "target": "mlsys.open-closed-loop-load", "statement": "fixes as open or closed loop the", "conditions": "Server scenarios issue queries by a timed random process; single-stream and multi-stream scenarios wait for each response." }, { "source": "mlsys.transfer-learning", "relation": "reduces", "target": "mlsys.time-to-train", "statement": "shortens the target-task", "conditions": "When a suitable pretrained model exists and fine-tuning converges in far fewer steps than training from scratch." }, { "source": "mlsys.pretraining", "relation": "uses", "target": "mlsys.compute-optimal-allocation", "statement": "chooses model size and token count with", "conditions": "When a fixed training compute budget is allocated before the run and the fitted law covers the target scale." }, { "source": "mlsys.diffusion-feature-caching", "relation": "uses", "target": "mlsys.diffusion-model", "statement": "reuses intermediate features across the denoising steps of", "conditions": "When adjacent sampling steps produce similar features in some blocks, so those blocks can be skipped and their cached output reused." }, { "source": "mlsys.experiment-budget", "relation": "uses", "target": "mlsys.scaling-laws", "statement": "allocates small proxy runs to fit, and a large run to exploit,", "conditions": "When the budget is split between runs that fit a law at small scale and the final run whose size the law predicts; extrapolation beyond the fitted range is not guaranteed." }, { "source": "mlsys.inference-compute-estimate", "relation": "measures", "target": "mlsys.memory-footprint", "statement": "estimates the parameter and KV portions of", "conditions": "When parameter bytes follow from count and precision and KV bytes from layers, KV heads, head dimension, context, batch, and precision; runtime buffers and activations are excluded." }, { "source": "mlsys.dispatch", "relation": "increases", "target": "mlsys.serial-fraction", "statement": "contributes host-side, non-overlapped work to", "conditions": "When launch and orchestration time is not overlapped with device execution, so it stays constant as device throughput or worker count grows." }, { "source": "mlsys.all-reduce", "relation": "specializes", "target": "mlsys.collectives", "statement": "is a kind of", "conditions": "When every participant contributes an equal-shaped tensor and every participant receives the full reduced result." }, { "source": "mlsys.ring-allreduce", "relation": "implements", "target": "mlsys.all-reduce", "statement": "is an algorithm for", "conditions": "When participants can be arranged in a ring with comparable link bandwidth." }, { "source": "mlsys.tree-allreduce", "relation": "implements", "target": "mlsys.all-reduce", "statement": "is an algorithm for", "conditions": "When participants are arranged in trees or hierarchical groups so latency grows logarithmically." }, { "source": "mlsys.in-network-reduction", "relation": "implements", "target": "mlsys.all-reduce", "statement": "performs inside switches the reduction steps of", "conditions": "When switches support aggregation for the data type and reduction operation used and also distribute the reduced result to every participant." }, { "source": "mlsys.data-parallel", "relation": "uses", "target": "mlsys.all-reduce", "statement": "synchronizes gradients through", "conditions": "When replicas combine gradients before each optimizer step, as in synchronous data parallelism." }, { "source": "mlsys.reduce-scatter-all-gather", "relation": "specializes", "target": "mlsys.collectives", "statement": "are kinds of", "conditions": "When each participant holds or receives a distinct shard rather than the whole tensor." }, { "source": "mlsys.state-sharding", "relation": "uses", "target": "mlsys.reduce-scatter-all-gather", "statement": "all-gathers parameters before use and reduce-scatters gradients after, through", "conditions": "When parameters, gradients, or optimizer state are sharded across data-parallel workers, as in ZeRO stage 3 or FSDP." }, { "source": "mlsys.ring-allreduce", "relation": "uses", "target": "mlsys.reduce-scatter-all-gather", "statement": "is composed of", "conditions": "When all-reduce is decomposed into a reduce-scatter followed by an all-gather around the ring." }, { "source": "mlsys.tensor-parallel", "relation": "uses", "target": "mlsys.reduce-scatter-all-gather", "statement": "replaces per-layer all-reduce with", "conditions": "When tensor-parallel layers are combined with Megatron-style sequence parallelism so each all-reduce becomes a reduce-scatter and all-gather pair." }, { "source": "mlsys.torus-topology", "relation": "specializes", "target": "mlsys.network-topology", "statement": "is a kind of", "conditions": "When endpoints connect to neighbors directly, without a switch hierarchy." }, { "source": "mlsys.torus-topology", "relation": "alternative-to", "target": "mlsys.fat-tree", "statement": "is an alternative to", "conditions": "Both connect accelerator nodes at scale; a torus saves switch cost and suits ring collectives, while a non-oversubscribed Clos fat-tree with suitable routing gives uniform bisection bandwidth for arbitrary traffic." }, { "source": "mlsys.torus-topology", "relation": "uses", "target": "mlsys.optical-circuit-switching", "statement": "can be reconfigured and routed around failed nodes by", "conditions": "When torus links pass through an optical circuit switch so slices of a pod can be joined into different torus shapes, as in TPU pods." }, { "source": "mlsys.ring-allreduce", "relation": "uses", "target": "mlsys.torus-topology", "statement": "runs at full link bandwidth on the rings of", "conditions": "When each ring of the collective maps onto one torus dimension so every step uses a direct link." }, { "source": "mlsys.torus-topology", "relation": "limits", "target": "mlsys.all-to-all", "statement": "bounds, through multi-hop routing and lower bisection bandwidth,", "conditions": "When exchange traffic crosses many torus hops rather than staying among neighbors." }, { "source": "mlsys.network-oversubscription", "relation": "reduces", "target": "mlsys.bisection-bandwidth", "statement": "lowers, below the sum of endpoint links,", "conditions": "When traffic crosses the oversubscribed tier rather than staying under one leaf switch." }, { "source": "mlsys.network-oversubscription", "relation": "limits", "target": "mlsys.collectives", "statement": "caps the achievable rate of cross-tier", "conditions": "When synchronized collective flows from many endpoints exceed a leaf switch's uplink capacity at the same moment." }, { "source": "mlsys.network-oversubscription", "relation": "reduces", "target": "mlsys.tco", "statement": "lowers switch and cabling cost within", "conditions": "When fewer uplinks and spine switches are provisioned per endpoint and the workload tolerates the contention." }, { "source": "mlsys.fusion", "relation": "increases", "target": "mlsys.arithmetic-intensity", "statement": "raises, by keeping intermediates on chip instead of round-tripping them through device memory,", "conditions": "When the fused operations are memory-bound elementwise or reduction kernels whose intermediates would otherwise be written to and read back from device memory." }, { "source": "mlsys.quantization", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "shrinks the bytes of stored weights, activations, or caches in", "conditions": "When tensors are stored in the reduced format rather than only computed in it; dequantization buffers and scale factors add back a small share." }, { "source": "mlsys.host-memory", "relation": "part-of", "target": "mlsys.memory-hierarchy", "statement": "is the capacity tier above device memory in", "conditions": "When the accelerator is attached to a host that has its own DRAM." }, { "source": "mlsys.memory-offloading", "relation": "uses", "target": "mlsys.host-memory", "statement": "spills optimizer state, parameters, or activations into", "conditions": "When per-device memory cannot hold the full training state." }, { "source": "mlsys.kv-cache-offload", "relation": "uses", "target": "mlsys.host-memory", "statement": "keeps evicted KV blocks in", "conditions": "When blocks are likely to be reused and host memory is cheaper than device memory." }, { "source": "mlsys.host-data-loader", "relation": "uses", "target": "mlsys.host-memory", "statement": "stages decoded batches in page-locked regions of", "conditions": "When batches are copied to the device by DMA from pinned host buffers." }, { "source": "mlsys.schedule-separation", "relation": "uses", "target": "mlsys.tiling", "statement": "names, as a schedule primitive,", "conditions": "When tile sizes and loop order are schedule choices rather than part of the algorithm's definition." }, { "source": "mlsys.schedule-separation", "relation": "enables", "target": "mlsys.autotuning", "statement": "exposes the search space explored by", "conditions": "When schedule parameters are enumerable and the algorithm is fixed, so candidate implementations differ only in schedule." }, { "source": "mlsys.schedule-separation", "relation": "uses", "target": "mlsys.intermediate-representation", "statement": "applies schedule transformations to the loop-level", "conditions": "When the compiler lowers the algorithm to a loop nest that schedule primitives rewrite." }, { "source": "mlsys.schedule-separation", "relation": "uses", "target": "mlsys.software-pipelining", "statement": "selects, as a schedule primitive,", "conditions": "When pipelining stages are chosen by the schedule rather than written by hand in the kernel." }, { "source": "mlsys.kernel-synchronization", "relation": "uses", "target": "mlsys.thread-hierarchy", "statement": "coordinates threads at the scopes defined by", "conditions": "When threads of a block or warp share on-chip memory and must order their accesses." }, { "source": "mlsys.kernel-synchronization", "relation": "causes", "target": "mlsys.numerical-nondeterminism", "statement": "introduces order-dependent floating-point results through contended", "conditions": "When partial sums are combined with atomic adds whose completion order varies from run to run." }, { "source": "mlsys.kernel-synchronization", "relation": "limits", "target": "mlsys.compute-throughput", "statement": "serializes contended updates, bounding achieved", "conditions": "When many threads atomically update the same address, as in histogram, scatter-add, or gradient-aggregation kernels." }, { "source": "mlsys.sparse-tensor-kernels", "relation": "uses", "target": "mlsys.kernel-synchronization", "statement": "scatter-adds irregular contributions through", "conditions": "When several threads contribute to the same output element." }, { "source": "mlsys.link-reach", "relation": "limits", "target": "mlsys.scale-up-interconnect", "statement": "bounds the physical extent of", "conditions": "When scale-up links run over passive copper, so the domain must fit within copper reach, typically one or a few racks." }, { "source": "mlsys.link-reach", "relation": "increases", "target": "mlsys.rack-power-density", "statement": "forces accelerators into few racks, raising", "conditions": "When a scale-up domain must stay within copper reach rather than spreading across the floor." }, { "source": "mlsys.co-packaged-optics", "relation": "mitigates", "target": "mlsys.link-reach", "statement": "extends beyond copper, at lower energy per bit than pluggable modules,", "conditions": "When links must span more than copper reach allows." }, { "source": "mlsys.storage-access-pattern", "relation": "determines", "target": "mlsys.dataset-storage-layout", "statement": "motivates large sequential shards in", "conditions": "When training streams whole corpora so sequential bandwidth, not IOPS, is the binding resource." }, { "source": "mlsys.storage-access-pattern", "relation": "limits", "target": "mlsys.input-pipeline", "statement": "bounds the read stage of", "conditions": "When examples are fetched by many small random reads from media or stores with high per-request latency." }, { "source": "mlsys.storage-access-pattern", "relation": "determines", "target": "mlsys.storage-tiering", "statement": "decides which tier serves which data in", "conditions": "When random-access working sets, such as retrieval indexes or hot shards, are kept on NVMe or in memory while bulk corpora stream from capacity tiers." }, { "source": "mlsys.tensor-parallel", "relation": "part-of", "target": "mlsys.hybrid-parallel", "statement": "is one dimension of", "conditions": "When tensor-parallel groups are placed within the fastest scale-up domain and combined with data and pipeline parallelism across nodes." }, { "source": "mlsys.data-parallel", "relation": "part-of", "target": "mlsys.hybrid-parallel", "statement": "is the outermost dimension of", "conditions": "When replicas of a tensor- or pipeline-parallel model group are themselves data-parallel across nodes." }, { "source": "mlsys.matrix-unit", "relation": "increases", "target": "mlsys.compute-throughput", "statement": "supplies most of the peak low-precision", "conditions": "When the workload's operations are dense matrix multiplies in input and accumulation formats the unit supports." }, { "source": "mlsys.topology-aware-placement", "relation": "uses", "target": "mlsys.network-topology", "statement": "places communicating workers according to", "conditions": "When the scheduler holds a map of links, switches, and scale-up domains and the job declares which workers communicate most." }, { "source": "mlsys.quantization-aware-training", "relation": "enables", "target": "mlsys.integer-only-inference", "statement": "makes accurate, by simulating rounding and requantization during training,", "conditions": "When the simulated quantization matches the deployed integer arithmetic and requantization rules." }, { "source": "mlsys.embedding-cache", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "keeps only hot embedding rows resident, shrinking the accelerator-side", "conditions": "When footprint is measured in accelerator memory and lookup skew lets cold rows live in host memory or SSD; total stored model bytes are unchanged." }, { "source": "mlsys.dataset-shuffling", "relation": "part-of", "target": "mlsys.input-pipeline", "statement": "is the ordering stage of", "conditions": "When examples are read in a randomized order rather than in stored order." }, { "source": "mlsys.dataset-shuffling", "relation": "uses", "target": "mlsys.streaming-dataset-format", "statement": "shuffles shards, then records within a buffer, from", "conditions": "When the corpus is stored as sharded files so a two-level shuffle approximates a global permutation." }, { "source": "mlsys.collectives", "relation": "causes", "target": "mlsys.training-power-swings", "statement": "idles every accelerator in lockstep during exposed, producing", "conditions": "When collectives are synchronous and not overlapped, so the whole cluster alternates between full-power compute and low-power waiting at the same moments." }, { "source": "mlsys.peer-to-peer-transfer", "relation": "uses", "target": "mlsys.scale-up-interconnect", "statement": "moves data between accelerators over", "conditions": "When transfers are carried by dedicated accelerator-to-accelerator scale-up links rather than over the host link." }, { "source": "mlsys.overfitting", "relation": "reduces", "target": "mlsys.generalization", "statement": "widens the gap between training and held-out performance, lowering", "conditions": "When the model has capacity to fit training-set noise and training continues past the point where held-out performance peaks." }, { "source": "mlsys.regularization", "relation": "mitigates", "target": "mlsys.overfitting", "statement": "mitigates", "conditions": "When the penalty or procedure limits the model's ability to fit training-set noise." }, { "source": "mlsys.training-diagnostics", "relation": "detects", "target": "mlsys.overfitting", "statement": "reveals, through diverging training and validation curves,", "conditions": "When validation metrics are logged alongside training loss at step granularity." }, { "source": "mlsys.overfitting", "relation": "increases", "target": "mlsys.membership-inference", "statement": "raises the train-versus-nontrain confidence gap exploited by", "conditions": "When overfitting creates train-versus-nontrain confidence or loss differences that the attack exploits, raising its advantage." }, { "source": "mlsys.predictive-uncertainty", "relation": "enables", "target": "mlsys.selective-prediction", "statement": "supplies the confidence signal that decides abstention in", "conditions": "When the estimate ranks errors well enough that abstaining on the least certain inputs lowers risk at the chosen coverage." }, { "source": "mlsys.active-learning", "relation": "uses", "target": "mlsys.predictive-uncertainty", "statement": "selects examples to label from", "conditions": "For uncertainty-sampling acquisition functions." }, { "source": "mlsys.ood-detection", "relation": "uses", "target": "mlsys.predictive-uncertainty", "statement": "thresholds epistemic signals from", "conditions": "When uncertainty rises on inputs far from the training data; raw softmax confidence is often overconfident there." }, { "source": "mlsys.statistical-power", "relation": "determines", "target": "mlsys.online-experiment", "statement": "sets the traffic share and duration required by", "conditions": "When the launch decision must detect an effect of a stated minimum size at stated error rates; metric variance and the number of independent units fix the answer." }, { "source": "mlsys.paired-comparison", "relation": "increases", "target": "mlsys.statistical-power", "statement": "raises, by cancelling shared item variance,", "conditions": "When both systems are scored on the same items so per-item differences vary less than independent samples." }, { "source": "mlsys.statistical-uncertainty", "relation": "reduces", "target": "mlsys.statistical-power", "statement": "lowers, through outcome variance,", "conditions": "At a fixed sample size; a noisier metric needs more units to detect the same effect." }, { "source": "mlsys.retrieval-metrics", "relation": "specializes", "target": "mlsys.metrics", "statement": "is the ranked-list family of", "conditions": "When outputs are ranked candidate lists judged against relevance labels." }, { "source": "mlsys.retrieval-metrics", "relation": "measures", "target": "mlsys.retrieval", "statement": "scores the ranked results of", "conditions": "When relevance judgments exist for the query set; judged pools miss relevant unjudged records." }, { "source": "mlsys.retrieval-metrics", "relation": "measures", "target": "mlsys.ann", "statement": "quantifies, as recall against exact search, the approximation loss of", "conditions": "When ground-truth neighbors are computed by exhaustive search on the same corpus." }, { "source": "mlsys.rejection-sampling-fine-tuning", "relation": "specializes", "target": "mlsys.supervised-fine-tuning", "statement": "is the self-generated-data form of", "conditions": "When the training targets are the model's own accepted samples rather than human demonstrations." }, { "source": "mlsys.rejection-sampling-fine-tuning", "relation": "uses", "target": "mlsys.verifier", "statement": "filters candidates with", "conditions": "When acceptance is decided by a programmatic check, a reward model, or a grader; filter errors become training labels." }, { "source": "mlsys.rejection-sampling-fine-tuning", "relation": "alternative-to", "target": "mlsys.rlvr", "statement": "is an offline, SFT-based alternative to", "conditions": "Both learn from verifiable outcomes; rejection sampling updates only on accepted samples while RLVR takes policy gradients over all sampled outcomes." }, { "source": "mlsys.rejection-sampling-fine-tuning", "relation": "causes", "target": "mlsys.mode-collapse", "statement": "narrows sampled diversity across rounds toward", "conditions": "When successive rounds train only on the model's own high-scoring samples, concentrating training mass on a narrowing response family without fresh human or real data." }, { "source": "mlsys.early-stopping", "relation": "specializes", "target": "mlsys.regularization", "statement": "acts, by truncating training, as a form of", "conditions": "When validation performance peaks before the training loss plateaus." }, { "source": "mlsys.early-stopping", "relation": "requires", "target": "mlsys.evaluation-split", "statement": "monitors a validation partition from", "conditions": "The stopping decision must not use the final test partition, or the reported score is optimistically biased." }, { "source": "mlsys.experiment-budget", "relation": "uses", "target": "mlsys.early-stopping", "statement": "stops unpromising trials early under", "conditions": "When a trial is abandoned once its validation metric plateaus, freeing its remaining budget for other configurations." }, { "source": "mlsys.group-dro", "relation": "uses", "target": "mlsys.loss", "statement": "optimizes a worst-group aggregation of", "conditions": "When examples carry group labels and training minimizes the highest group loss rather than the mean." }, { "source": "mlsys.group-dro", "relation": "mitigates", "target": "mlsys.distribution-shift", "statement": "protects worst-group performance under subpopulation forms of", "conditions": "When the groups that dominate at deployment are among those declared at training." }, { "source": "mlsys.loss-masking", "relation": "uses", "target": "mlsys.chat-template", "statement": "locates assistant spans through", "conditions": "When non-target tokens are identified by role markers in the rendered sequence." }, { "source": "mlsys.chat-template", "relation": "causes", "target": "mlsys.train-serve-skew", "statement": "causes, when rendered differently in training and serving,", "conditions": "When the serving stack renders turns with different markers or whitespace from the training renderer." }, { "source": "mlsys.chat-template", "relation": "uses", "target": "mlsys.tokenization", "statement": "reserves special tokens in", "conditions": "When role markers are single reserved tokens rather than plain text." }, { "source": "mlsys.generation-stopping", "relation": "limits", "target": "mlsys.kv-cache", "statement": "caps per-request growth of", "conditions": "A server-enforced output-token cap limits additional retained token positions in engines whose autoregressive cache grows with generation; it does not bound all allocated cache memory." }, { "source": "mlsys.generation-stopping", "relation": "limits", "target": "mlsys.tail-latency", "statement": "limits the output-length contribution to", "conditions": "Decode-step time is bounded, while queueing, stalls, and retries are accounted for separately." }, { "source": "mlsys.generation-stopping", "relation": "mitigates", "target": "mlsys.resource-exhaustion-attack", "statement": "caps the output tokens one request can consume, blunting", "conditions": "A server-enforced output cap limits generated-token work per request; separate controls are needed for prompt processing, request rate, and aggregate client consumption." }, { "source": "mlsys.runbook", "relation": "part-of", "target": "mlsys.incident-response", "statement": "scripts the first steps of", "conditions": "When the failure has been seen before and its mitigation is known." }, { "source": "mlsys.alerting", "relation": "uses", "target": "mlsys.runbook", "statement": "links each page to a", "conditions": "When alerts are actionable and name the procedure to follow." }, { "source": "mlsys.postmortem", "relation": "produces", "target": "mlsys.runbook", "statement": "writes or revises a", "conditions": "When corrective actions include documenting the response for next time." }, { "source": "mlsys.runbook", "relation": "uses", "target": "mlsys.rollback", "statement": "often names, as the first mitigation,", "conditions": "For model-serving incidents where the previous version is known good." }, { "source": "mlsys.rl-training-colocation", "relation": "determines", "target": "mlsys.weight-sync", "statement": "sets the transfer path and cost of", "conditions": "Colocated designs reshard in device memory; disaggregated ones send weights across the network." }, { "source": "mlsys.rl-training-colocation", "relation": "uses", "target": "mlsys.dynamic-resharding", "statement": "switches between training and generation layouts with", "conditions": "When the colocated design runs both phases on one device set." }, { "source": "mlsys.rl-training-colocation", "relation": "increases", "target": "mlsys.policy-staleness", "statement": "raises, in its disaggregated asynchronous form,", "conditions": "When rollout workers keep generating while trainers update." }, { "source": "mlsys.sequential-revision", "relation": "alternative-to", "target": "mlsys.best-of-n", "statement": "spends compute serially rather than in parallel samples as in", "conditions": "When one task's inference budget is allocated between dependent revisions and independent candidates." }, { "source": "mlsys.sequential-revision", "relation": "uses", "target": "mlsys.verifier", "statement": "benefits from external feedback supplied by", "conditions": "When the critique comes from a checker rather than the model's own judgment, which is often unreliable." }, { "source": "mlsys.sequential-revision", "relation": "increases", "target": "mlsys.tail-latency", "statement": "adds serial passes to", "conditions": "Each pass must finish before the next begins." }, { "source": "mlsys.semi-supervised-learning", "relation": "uses", "target": "mlsys.supervised-learning", "statement": "trains on labeled and pseudo-labeled examples with", "conditions": "After unlabeled examples receive model-assigned targets." }, { "source": "mlsys.semi-supervised-learning", "relation": "alternative-to", "target": "mlsys.weak-supervision", "statement": "derives extra labels from the model itself rather than from heuristics as in", "conditions": "Both reduce manual labeling; semi-supervision needs a labeled seed and a model confident enough to self-label." }, { "source": "mlsys.semi-supervised-learning", "relation": "uses", "target": "mlsys.selective-prediction", "statement": "keeps only confident pseudo-labels through", "conditions": "When a confidence threshold decides which unlabeled examples enter training." }, { "source": "mlsys.contrastive-learning", "relation": "produces", "target": "mlsys.embeddings", "statement": "trains the encoders behind", "conditions": "When the learned representation is reused for retrieval, classification heads, or multimodal alignment." }, { "source": "mlsys.contrastive-learning", "relation": "enables", "target": "mlsys.dense-retrieval", "statement": "trains the bi-encoders used by", "conditions": "When query-document pairs supply positives and in-batch items supply negatives." }, { "source": "mlsys.batch-size", "relation": "limits", "target": "mlsys.contrastive-learning", "statement": "bounds the in-batch negatives available to", "conditions": "When negatives are drawn from the same batch rather than a memory bank." }, { "source": "mlsys.training-memory-estimate", "relation": "measures", "target": "mlsys.training-state", "statement": "sizes", "conditions": "For dense models where each component's precision and the optimizer's state count are known." }, { "source": "mlsys.training-memory-estimate", "relation": "uses", "target": "mlsys.precision", "statement": "takes byte widths from", "conditions": "Each state component may use a different format." }, { "source": "mlsys.training-memory-estimate", "relation": "measures", "target": "mlsys.memory-footprint", "statement": "estimates the training share of", "conditions": "Runtime buffers and allocator overhead are added separately." }, { "source": "mlsys.rl-rollout", "relation": "uses", "target": "mlsys.continuous-batching", "statement": "generates many sampled responses per prompt through", "conditions": "When rollouts are produced by a serving engine rather than the training framework, so engine throughput bounds samples per step." }, { "source": "mlsys.retraining", "relation": "produces", "target": "mlsys.model-artifact", "statement": "yields a replacement", "conditions": "After the retrained model passes validation and is versioned for release." }, { "source": "mlsys.hyperparameter-search", "relation": "uses", "target": "mlsys.experiment-budget", "statement": "allocates trials under", "conditions": "Whenever the number of configurations exceeds what can be trained to completion, so trials, repetitions, and fidelities compete for one budget." }, { "source": "mlsys.continuous-training", "relation": "uses", "target": "mlsys.data-validation", "statement": "validates incoming data before each run of", "conditions": "When the pipeline is automated end to end, so a schema or distribution failure must stop retraining rather than a person." }, { "source": "mlsys.model-gateway", "relation": "implements", "target": "mlsys.admission", "statement": "enforces per-client rate limits as", "conditions": "When all model calls pass through the gateway and limits are expressed per key, tenant, or application." }, { "source": "mlsys.alerting", "relation": "uses", "target": "mlsys.drift-detection", "statement": "fires on thresholds over statistics from", "conditions": "When drift statistics are alerted on directly; detected drift is a signal to diagnose, not proof of degraded quality." }, { "source": "mlsys.kv-cache-offload", "relation": "alternative-to", "target": "mlsys.kv-cache-eviction", "statement": "keeps cold entries recoverable instead of discarding them as in", "conditions": "Both bound resident KV memory; offload spends transfer bandwidth to preserve exactness, eviction spends quality to avoid transfers; they can be combined." }, { "source": "mlsys.pass-at-k", "relation": "uses", "target": "mlsys.test-oracle", "statement": "counts a sample as correct according to a", "conditions": "When the check is unit tests or a programmatic grader; the metric inherits the oracle's false positives." }, { "source": "mlsys.failure-attribution", "relation": "uses", "target": "mlsys.inference-logging", "statement": "reconstructs the failing request from", "conditions": "When logs retain inputs, model version, and outputs for the incident window." }, { "source": "mlsys.privacy-budget", "relation": "part-of", "target": "mlsys.differential-privacy", "statement": "is the accounting side of", "conditions": "When several releases or training steps on the same data must be composed into one guarantee." }, { "source": "mlsys.dp-sgd", "relation": "uses", "target": "mlsys.privacy-budget", "statement": "tracks its noise multiplier, sampling rate, and step count through", "conditions": "At every optimizer step; the accountant decides when training must stop." }, { "source": "mlsys.privacy-budget", "relation": "limits", "target": "mlsys.retraining", "statement": "caps how many times private data can be reused by", "conditions": "When each retraining run spends from a fixed budget rather than resetting it." }, { "source": "mlsys.bias-sources", "relation": "determines", "target": "mlsys.bias-mitigation", "statement": "selects which stage of intervention is appropriate in", "conditions": "When the dominant source is identified, since reweighting cannot fix a measurement problem and a constraint cannot fix a deployment mismatch." }, { "source": "mlsys.bias-sources", "relation": "causes", "target": "mlsys.proxy-discrimination", "statement": "can produce, through correlated features,", "conditions": "When historical or measurement bias is encoded in features correlated with protected attributes and those features drive decisions." }, { "source": "mlsys.carbon-accounting-method", "relation": "uses", "target": "mlsys.emission-scopes", "statement": "applies to the purchased-electricity category of", "conditions": "When purchased-electricity emissions are identified as Scope 2 before the attribution method is chosen." }, { "source": "mlsys.carbon-accounting-method", "relation": "measures", "target": "mlsys.operational-carbon", "statement": "reports, under declared factors and boundaries, the electricity share of", "conditions": "When contractual instruments cover a material share of consumption, so the market-based figure diverges from the location-based one." }, { "source": "mlsys.retry-backoff", "relation": "uses", "target": "mlsys.request-timeout", "statement": "treats as its failure signal the expiry of", "conditions": "When a retry is issued on timeout rather than on an explicit error; the retry budget and the deadline must be set together." }, { "source": "mlsys.request-timeout", "relation": "mitigates", "target": "mlsys.metastable-failure", "statement": "cancels abandoned work that would otherwise sustain", "conditions": "When deadlines are propagated so downstream services stop serving requests the client no longer waits for." }, { "source": "mlsys.circuit-breaker", "relation": "uses", "target": "mlsys.request-timeout", "statement": "counts expirations as errors in", "conditions": "When slow responses, not only explicit failures, trip the breaker." }, { "source": "mlsys.data-protection-regulation", "relation": "requires", "target": "mlsys.data-consent", "statement": "mandates a lawful basis and purpose limitation enforced through", "conditions": "When personal data of covered people is processed for training or inference." }, { "source": "mlsys.data-protection-regulation", "relation": "requires", "target": "mlsys.deletion-propagation", "statement": "creates erasure obligations fulfilled by", "conditions": "When a legally effective erasure obligation covers the retained personal data and its derived copies." }, { "source": "mlsys.data-protection-regulation", "relation": "determines", "target": "mlsys.privacy", "statement": "sets the legal floor for", "conditions": "Within the jurisdictions whose law applies to the people in the data." }, { "source": "mlsys.egress-control", "relation": "mitigates", "target": "mlsys.weight-exfiltration", "statement": "blocks the outbound channels used for", "conditions": "When weights can leave only through network paths the allowlist controls." }, { "source": "mlsys.confidential-computing", "relation": "mitigates", "target": "mlsys.weight-exfiltration", "statement": "keeps weights unreadable to the host operator, limiting", "conditions": "When the weights are decrypted only inside the attested environment." }, { "source": "mlsys.weight-exfiltration", "relation": "alternative-to", "target": "mlsys.model-extraction", "statement": "obtains the model directly instead of reconstructing it by queries as in", "conditions": "Both yield a usable copy; theft needs infrastructure access, extraction needs only the prediction interface." }, { "source": "mlsys.ai-bill-of-materials", "relation": "mitigates", "target": "mlsys.supply-chain-compromise", "statement": "lets consumers find known-bad components behind", "conditions": "When the manifest is complete and machine-readable and components flagged against advisories are rejected, replaced, or contained; the inventory alone reveals but does not prevent a compromise." }, { "source": "mlsys.ai-bill-of-materials", "relation": "uses", "target": "mlsys.lineage", "statement": "is compiled from", "conditions": "When data lineage records the sources and transformations the manifest lists." }, { "source": "mlsys.artifact-signing", "relation": "uses", "target": "mlsys.ai-bill-of-materials", "statement": "binds to the weights, through its signature, the contents of", "conditions": "When the signed payload includes the manifest rather than only the weight bytes." }, { "source": "mlsys.risk-tiered-regulation", "relation": "requires", "target": "mlsys.ai-incident-reporting", "statement": "obliges providers of high-risk systems to perform", "conditions": "For serious incidents under the governing regulation, within its reporting deadlines." }, { "source": "mlsys.ai-risk-management", "relation": "uses", "target": "mlsys.ai-incident-reporting", "statement": "feeds its measure and manage functions from", "conditions": "When reported incidents are reviewed against the risk register rather than only closed." }, { "source": "mlsys.ai-incident-reporting", "relation": "uses", "target": "mlsys.postmortem", "statement": "draws its causal account from", "conditions": "When the incident was handled by an internal incident process before disclosure." }, { "source": "mlsys.mechanistic-interpretability", "relation": "alternative-to", "target": "mlsys.feature-attribution", "statement": "explains internal computation rather than input contributions, unlike", "conditions": "Both aim at understanding a decision; attribution scores inputs, mechanistic methods inspect representations and need model access." }, { "source": "mlsys.mechanistic-interpretability", "relation": "uses", "target": "mlsys.linear-model", "statement": "fits linear probes from", "conditions": "For probing, where a linear classifier on activations tests whether a property is linearly represented." }, { "source": "mlsys.safety-certification", "relation": "uses", "target": "mlsys.safety-case", "statement": "assesses the argument and evidence in", "conditions": "Under standards that require a safety case; others prescribe process evidence directly." }, { "source": "mlsys.change-impact-analysis", "relation": "uses", "target": "mlsys.safety-certification", "statement": "scopes re-assessment against the claims certified by", "conditions": "When a model, data, or operating-domain change touches certified claims." }, { "source": "mlsys.hazard-analysis", "relation": "uses", "target": "mlsys.tolerable-risk", "statement": "ranks identified hazards against", "conditions": "When the analysis must decide which hazards receive controls rather than only list them." }, { "source": "mlsys.safety-case", "relation": "requires", "target": "mlsys.tolerable-risk", "statement": "argues acceptability against", "conditions": "The top claim of a safety case is that residual risk is acceptable, which is undefined without the criteria." }, { "source": "mlsys.fleet-hardware-screening", "relation": "detects", "target": "mlsys.silent-data-corruption", "statement": "finds the units responsible for", "conditions": "For faults that reproduce under the test workload; intermittent and workload-dependent faults need longer or in-production screening." }, { "source": "mlsys.fleet-hardware-screening", "relation": "uses", "target": "mlsys.test-oracle", "statement": "compares outputs against", "conditions": "When reference outputs are computed on trusted hardware or by a redundant run." }, { "source": "mlsys.fleet-hardware-screening", "relation": "trades-off", "target": "mlsys.utilization", "statement": "spends fleet time that would otherwise count toward", "conditions": "When screening runs on production capacity rather than on units already out of service." }, { "source": "mlsys.profiling", "relation": "uses", "target": "mlsys.hardware-performance-counters", "statement": "reads stall and throughput causes from", "conditions": "When the profiler collects counters in addition to timestamps; collection can perturb execution and on some accelerators serializes it." }, { "source": "mlsys.bottleneck-analysis", "relation": "uses", "target": "mlsys.hardware-performance-counters", "statement": "places a kernel on the roofline using measured bytes and operations from", "conditions": "When measured rather than estimated arithmetic intensity is needed." }, { "source": "mlsys.structured-pruning", "relation": "alternative-to", "target": "mlsys.nm-sparsity", "statement": "removes whole units for dense hardware instead of the fine per-group pattern of", "conditions": "Both aim at real speedup from pruning; structured pruning needs no sparse-engine support but usually costs more quality at equal parameter reduction, N:M needs a matrix engine with the matching sparse instruction." }, { "source": "mlsys.gray-failure", "relation": "limits", "target": "mlsys.health-checks", "statement": "passes undetected through", "conditions": "When the degraded node still answers liveness and readiness probes while failing the work callers send it." }, { "source": "mlsys.over-refusal", "relation": "trades-off", "target": "mlsys.jailbreak", "statement": "is the false-positive side of the operating point whose false-negative side is", "conditions": "When a single threshold or refusal policy decides both, so loosening it to admit benign requests also admits more successful bypasses." }, { "source": "mlsys.membership-inference", "relation": "measures", "target": "mlsys.machine-unlearning", "statement": "audits whether removed records still influence a model after", "conditions": "When attack advantage on the forgotten records, compared with a model retrained without them, is used as the empirical test of forgetting." }, { "source": "mlsys.guardrails", "relation": "increases", "target": "mlsys.tail-latency", "statement": "adds classifier passes on the request path to", "conditions": "When input and output checks run synchronously before the response is released; streaming output checks shift the cost to time-to-last-token." }, { "source": "mlsys.silent-data-corruption", "relation": "causes", "target": "mlsys.training-instability", "statement": "can inject the corrupted gradients or weights behind", "conditions": "When a miscomputed value propagates into optimizer state before any check catches it, so a loss spike or divergence appears with no software cause." }, { "source": "mlsys.stragglers", "relation": "reduces", "target": "mlsys.training-goodput", "statement": "idles every other rank, lowering", "conditions": "In synchronous training, where each step waits for the slowest rank; asynchronous or straggler-tolerant schemes break the dependence." }, { "source": "mlsys.human-oversight", "relation": "uses", "target": "mlsys.selective-prediction", "statement": "routes abstained cases to people through", "conditions": "When the system abstains on low-confidence inputs and a human review path exists with enough time and context to decide." }, { "source": "mlsys.agent-harness", "relation": "produces", "target": "mlsys.trajectory-log", "statement": "writes, as the sole author, the", "conditions": "When the harness, not the model, owns continuation and records each step before dispatching it." }, { "source": "mlsys.durable-execution", "relation": "uses", "target": "mlsys.trajectory-log", "statement": "recovers progress and reuses recorded results from the", "conditions": "When each step's intent is appended before the effect is issued, so a crash between the two can be reconciled." }, { "source": "mlsys.deterministic-replay", "relation": "uses", "target": "mlsys.trajectory-log", "statement": "substitutes recorded model outputs and tool results from the", "conditions": "When the log contains every input needed for the declared replay boundary." }, { "source": "mlsys.agent-evaluation", "relation": "uses", "target": "mlsys.trajectory-log", "statement": "scores intermediate effects and constraint compliance from the", "conditions": "When evaluation reads the record of what was dispatched rather than only the final answer." }, { "source": "mlsys.odometry", "relation": "specializes", "target": "mlsys.state-estimation", "statement": "is the relative-motion, drift-accumulating form of", "conditions": "Only relative motion is measured, so absolute pose is unobservable without an external reference." }, { "source": "mlsys.slam", "relation": "uses", "target": "mlsys.odometry", "statement": "takes its motion prior between keyframes from", "conditions": "When the front end propagates pose with odometry and the back end corrects drift at loop closure." }, { "source": "mlsys.recursive-filtering", "relation": "uses", "target": "mlsys.odometry", "statement": "propagates the predicted state in the prediction step with", "conditions": "When odometry supplies the control input to a Kalman-family or particle filter." }, { "source": "mlsys.odometry", "relation": "uses", "target": "mlsys.coordinate-frames", "statement": "accumulates pose in the odometry frame defined by", "conditions": "In systems that keep a continuous odometry frame beneath a map frame, so the map-to-odometry transform absorbs drift correction." }, { "source": "mlsys.capture-epoch", "relation": "determines", "target": "mlsys.observation-age", "statement": "fixes the start instant of", "conditions": "Observation age is measured from the capture epoch, not from the frame's arrival at the processor." }, { "source": "mlsys.capture-epoch", "relation": "limits", "target": "mlsys.sensor-fusion", "statement": "misaligns measurements taken at different instants within", "conditions": "When the platform or scene moves during exposure or a scan and the measurements are not de-skewed before fusion." }, { "source": "mlsys.recursive-filtering", "relation": "uses", "target": "mlsys.capture-epoch", "statement": "applies each measurement at its", "conditions": "When measurements arrive out of order or late, the filter must correct at the capture epoch rather than at arrival." }, { "source": "mlsys.timing-jitter", "relation": "limits", "target": "mlsys.feedback-control", "statement": "degrades, by varying the sample interval of,", "conditions": "When the controller's gains assume a fixed period and the actual release or completion time varies." }, { "source": "mlsys.real-time-scheduling", "relation": "reduces", "target": "mlsys.timing-jitter", "statement": "bounds, through fixed priorities and periodic release,", "conditions": "When the control task has the highest priority or is released by a timer rather than by a best-effort scheduler." }, { "source": "mlsys.xil-testing", "relation": "measures", "target": "mlsys.timing-jitter", "statement": "exposes, on the real compute and bus,", "conditions": "When the real software and hardware run in the loop; pure simulation hides it." }, { "source": "mlsys.thermal-dynamics", "relation": "increases", "target": "mlsys.timing-jitter", "statement": "adds, through frequency throttling,", "conditions": "When clock changes under thermal policy alter task completion times between periods." }, { "source": "mlsys.deterministic-field-network", "relation": "increases", "target": "mlsys.feedback-delay", "statement": "adds a bounded transport and cycle-time term to", "conditions": "When drives and sensors are reached over the bus rather than on-chip." }, { "source": "mlsys.robotics-middleware", "relation": "uses", "target": "mlsys.deterministic-field-network", "statement": "reaches drives and encoders through", "conditions": "When middleware nodes talk to actuators and sensors over a fieldbus rather than directly." }, { "source": "mlsys.deterministic-field-network", "relation": "limits", "target": "mlsys.real-time-scheduling", "statement": "bounds, by bus occupancy, the achievable period of", "conditions": "When the control task's period must fit the bus cycle and the number of devices on it." }, { "source": "mlsys.feedback-delay", "relation": "reduces", "target": "mlsys.loop-stability-margin", "statement": "spends, at a rate set by the loop's bandwidth,", "conditions": "Each unit of delay costs phase in proportion to the crossover frequency." }, { "source": "mlsys.loop-stability-margin", "relation": "limits", "target": "mlsys.feedback-control", "statement": "caps the achievable bandwidth and gain of", "conditions": "When the controller is tuned to keep a stated margin under the worst-case delay and plant variation." }, { "source": "mlsys.dynamics", "relation": "determines", "target": "mlsys.loop-stability-margin", "statement": "sets, together with the controller, the", "conditions": "The margin is a property of the plant-controller pair, not of the controller alone." }, { "source": "mlsys.impedance-force-control", "relation": "specializes", "target": "mlsys.feedback-control", "statement": "is the contact-regulating form of", "conditions": "The regulated variable is the force-motion relationship rather than position alone." }, { "source": "mlsys.impedance-force-control", "relation": "uses", "target": "mlsys.contact-compliance", "statement": "models the environment's stiffness and damping through", "conditions": "Stable contact depends on the combined compliance of robot, controller, and environment." }, { "source": "mlsys.feedback-delay", "relation": "limits", "target": "mlsys.impedance-force-control", "statement": "destabilizes stiff contact under", "conditions": "When commanded stiffness is high and the loop delay is long relative to the contact dynamics." }, { "source": "mlsys.safety-filter", "relation": "uses", "target": "mlsys.impedance-force-control", "statement": "bounds contact force at the actuator through", "conditions": "When the filter's safe action is expressed as a force or stiffness limit rather than a position correction." }, { "source": "mlsys.authorization", "relation": "uses", "target": "mlsys.tool-contract", "statement": "checks a proposed tool call against the authority and effects declared in the", "conditions": "When tool calls are authorized per action from the contract rather than by ambient session rights; schema conformance alone grants nothing." }, { "source": "mlsys.sensor-interference", "relation": "specializes", "target": "mlsys.sensor-degradation", "statement": "is the multi-emitter form of", "conditions": "When other active sensors share the scene; the hardware has not failed and the loss of fidelity depends on the operating environment." }, { "source": "mlsys.sensor-interconnect", "relation": "limits", "target": "mlsys.robot-perception", "statement": "bounds the resolution, frame rate, and sensor count available to", "conditions": "When raw sensor data streams into the onboard processor over MIPI, I2S, or SPI links rather than being pre-processed at the sensor." }, { "source": "mlsys.sensor-degradation", "relation": "limits", "target": "mlsys.robot-perception", "statement": "degrades the inputs to", "conditions": "When the perception model was trained and validated on clean conditions and no input-support monitor flags the degraded regime." }, { "source": "mlsys.approval-gate", "relation": "part-of", "target": "mlsys.agent-harness", "statement": "is the undispatched hold within", "conditions": "When the harness holds the proposed action, bound by digest and expiry, until an authorized decision, and re-runs admission after approval." }, { "source": "mlsys.agent-interruption", "relation": "requires", "target": "mlsys.agent-harness", "statement": "takes effect only at turn boundaries enforced by the", "conditions": "When a runtime owns continuation and dispatch, enforces cancellation at defined execution boundaries, and separately settles already issued external actions." }, { "source": "mlsys.diffusion-policy", "relation": "alternative-to", "target": "mlsys.action-tokenization", "statement": "generates continuous actions by denoising rather than through the discretization of", "conditions": "When choosing a robot policy's action head: tokens trade reconstruction error and decode steps against expressiveness; denoising trades inference iterations against multimodality." }, { "source": "mlsys.vla-model", "relation": "uses", "target": "mlsys.diffusion-policy", "statement": "can emit continuous actions through a", "conditions": "When the action head denoises or integrates a learned flow instead of decoding action tokens (pi0 and its successors)." }, { "source": "mlsys.vla-model", "relation": "uses", "target": "mlsys.cross-embodiment-training", "statement": "is commonly pretrained with", "conditions": "When the policy is pretrained on pooled multi-robot datasets before task fine-tuning." }, { "source": "mlsys.closed-loop-evaluation", "relation": "measures", "target": "mlsys.sim-to-real", "statement": "scores, by physical trial success rate,", "conditions": "When comparable closed-loop simulation and physical trials give an explicit transfer comparison under matched task and measurement conditions." }, { "source": "mlsys.multi-agent-orchestration", "relation": "uses", "target": "mlsys.capability-based-security", "statement": "attenuates delegated rights to each worker through", "conditions": "When a worker agent must receive narrower authority than its orchestrator." }, { "source": "mlsys.multi-agent-orchestration", "relation": "increases", "target": "mlsys.agentic-workload", "statement": "multiplies, by duplicating context across workers,", "conditions": "When each delegated agent carries its own context and makes its own dependent model calls, so total tokens exceed a single-agent run on the same task." }, { "source": "mlsys.context-selection", "relation": "uses", "target": "mlsys.agent-memory", "statement": "draws recalled records into the working set from", "conditions": "When the agent has persistent memory; the recalled record competes for the same token budget as instructions, tool descriptions, and observations." }, { "source": "mlsys.thermal-dynamics", "relation": "determines", "target": "mlsys.wcet", "statement": "fixes the clock and memory-timing assumptions under which a bound is a valid", "conditions": "When the bound was derived at full clock; the analysis must include the allowed throttled operating states or explicitly exclude them." }, { "source": "mlsys.power-delivery", "relation": "determines", "target": "mlsys.actuator-limits", "statement": "sets, when a shared rail sags, the torque available within", "conditions": "When compute and actuators share a supply and simultaneous demand or regeneration droops the rail below the drive's operating window." }, { "source": "mlsys.thermal-dynamics", "relation": "determines", "target": "mlsys.actuator-limits", "statement": "sets the sustained-duty component of", "conditions": "When motor winding temperature, not drive current, binds continuous torque; peak torque remains current-limited." }, { "source": "mlsys.sim-to-real", "relation": "uses", "target": "mlsys.parallel-simulation", "statement": "starts from policies trained in", "conditions": "When the policy's experience came from batched simulated environments rather than physical collection." }, { "source": "mlsys.idempotency", "relation": "mitigates", "target": "mlsys.in-doubt-action", "statement": "makes safe the retry of an", "conditions": "When the operation carries an identity the endpoint deduplicates." }, { "source": "mlsys.agent-interruption", "relation": "causes", "target": "mlsys.in-doubt-action", "statement": "leaves, when it cancels mid-dispatch, an", "conditions": "When a tool call was sent before the cancel and its result never arrived." }, { "source": "mlsys.in-doubt-action", "relation": "limits", "target": "mlsys.compensation", "statement": "cannot be compensated until it is known to have happened, so it bounds", "conditions": "A compensating action for an effect that never occurred is itself a new effect." }, { "source": "mlsys.model-bandwidth-utilization", "relation": "alternative-to", "target": "mlsys.model-flops-utilization", "statement": "is the memory-bound counterpart of", "conditions": "When one primary efficiency indicator is chosen for a stated bottleneck; both ratios can be reported for the same workload, and the bound regime says which one explains the ceiling." }, { "source": "mlsys.model-bandwidth-utilization", "relation": "uses", "target": "mlsys.tpot", "statement": "divides bytes per token by the measured", "conditions": "When achieved bandwidth is estimated from weight and KV-cache bytes over time per output token rather than from hardware counters." }, { "source": "mlsys.hot-spare-capacity", "relation": "increases", "target": "mlsys.training-goodput", "statement": "raises, by shortening node-replacement waits, the", "conditions": "When reduced recovery and rework losses exceed the reserve time charged to the same goodput denominator." }, { "source": "mlsys.hot-spare-capacity", "relation": "trades-off", "target": "mlsys.utilization", "statement": "holds idle capacity that lowers fleet", "conditions": "When fleet utilization counts the reserved idle spares in its capacity denominator." }, { "source": "mlsys.hot-spare-capacity", "relation": "alternative-to", "target": "mlsys.elastic-training", "statement": "replaces a failed worker instead of shrinking the job, unlike", "conditions": "When both replacement at the original world size and continuation at a reduced world size are feasible recovery choices." }, { "source": "mlsys.hot-spare-capacity", "relation": "uses", "target": "mlsys.mtbf", "statement": "uses the failure interval summarized by", "conditions": "When sizing converts component or job MTBF into a failure-arrival model and combines it with repair or replenishment times, failure correlations, and an acceptable spare-exhaustion probability." }, { "source": "mlsys.geo-distributed-training", "relation": "uses", "target": "mlsys.local-sgd", "statement": "cuts cross-site synchronization frequency with", "conditions": "When inter-site bandwidth cannot carry per-step gradient exchange." }, { "source": "mlsys.geo-distributed-training", "relation": "uses", "target": "mlsys.pipeline-parallel", "statement": "places stage boundaries on inter-site links using", "conditions": "When per-micro-batch activations are far smaller than the gradients a data-parallel cut would send across the WAN." }, { "source": "mlsys.network-propagation-delay", "relation": "limits", "target": "mlsys.geo-distributed-training", "statement": "sets a lower bound on each synchronous exchange of", "conditions": "When sites are hundreds of kilometers apart, so the round trip bounds every cross-site exchange regardless of bandwidth." }, { "source": "mlsys.geo-distributed-training", "relation": "uses", "target": "mlsys.hybrid-parallel", "statement": "extends to the datacenter tier the hierarchy mapping of", "conditions": "When the hardware hierarchy runs chip, node, rack, and site, and each dimension is matched to a tier's link speed." }, { "source": "mlsys.quarantined-model-pattern", "relation": "mitigates", "target": "mlsys.prompt-injection", "statement": "removes the channel from untrusted text to privileged actions used by", "conditions": "When the planner and runtime are trusted, untrusted content stays isolated in the quarantined model, and runtime policies validate derived arguments and data flows before any privileged effect." }, { "source": "mlsys.quarantined-model-pattern", "relation": "uses", "target": "mlsys.information-flow-control", "statement": "tags quarantined outputs and checks tool arguments with", "conditions": "When a value derived from untrusted data must be policy-checked before it reaches a side-effecting tool." }, { "source": "mlsys.sycophancy-evaluation", "relation": "detects", "target": "mlsys.reward-hacking", "statement": "surfaces agreement-seeking learned as a form of", "conditions": "When preference training rewarded user approval and the evaluation shows agreement rising while task correctness falls for the same behavior." }, { "source": "mlsys.sycophancy-evaluation", "relation": "specializes", "target": "mlsys.robustness-evaluation", "statement": "is a kind of", "conditions": "When the perturbation is conversational pressure rather than input corruption." }, { "source": "mlsys.sycophancy-evaluation", "relation": "uses", "target": "mlsys.llm-judge", "statement": "scores free-form multi-turn answers with", "conditions": "When responses are open-ended, so a flip cannot be read off a fixed label." }, { "source": "mlsys.communication-intensity", "relation": "limits", "target": "mlsys.parallel-efficiency", "statement": "together with link bandwidth bounds", "conditions": "When the exchange is not fully overlapped with computation." }, { "source": "mlsys.gradient-compression", "relation": "reduces", "target": "mlsys.communication-intensity", "statement": "lowers the bytes per step counted by", "conditions": "When compressed gradients replace full-precision exchange." }, { "source": "mlsys.hybrid-parallel", "relation": "determines", "target": "mlsys.communication-intensity", "statement": "sets, per parallelism dimension, the", "conditions": "When the parallel plan fixes which tensors cross which links on each step." }, { "source": "mlsys.technology-scaling", "relation": "causes", "target": "mlsys.dark-silicon", "statement": "produces, once constant power density ends,", "conditions": "When transistor count keeps growing but power density no longer falls with feature size." }, { "source": "mlsys.technology-scaling", "relation": "limits", "target": "mlsys.compute-throughput", "statement": "constrains, once power density stops falling, per-chip gains in", "conditions": "When transistor-density gains lack corresponding voltage and power improvements, so a fixed power budget prevents all added arithmetic units from operating at once at the intended clock rate." }, { "source": "mlsys.compression-composition", "relation": "uses", "target": "mlsys.pruning", "statement": "typically applies first", "conditions": "When sparsity is introduced before quantization so the quantizer calibrates on the surviving weights." }, { "source": "mlsys.compression-composition", "relation": "uses", "target": "mlsys.quantization", "statement": "applies to the already-pruned or distilled model", "conditions": "When the bit-width is chosen after structure has been removed." }, { "source": "mlsys.compression-composition", "relation": "uses", "target": "mlsys.distillation", "statement": "recovers quality lost by the other stages through", "conditions": "When a teacher is available to fine-tune the compressed student." }, { "source": "mlsys.compression-composition", "relation": "determines", "target": "mlsys.efficiency-frontier", "statement": "shapes, through the techniques chosen and their order, the attainable", "conditions": "When the frontier is measured over compressed variants rather than architectures." }, { "source": "mlsys.fidelity-adaptive-serving", "relation": "specializes", "target": "mlsys.graceful-degradation", "statement": "is a kind of", "conditions": "When the degraded mode is a cheaper model variant rather than a feature cut." }, { "source": "mlsys.fidelity-adaptive-serving", "relation": "uses", "target": "mlsys.weight-only-quantization", "statement": "obtains its cheaper variants from", "conditions": "When the fallback variant is a lower-precision copy of the same weights." }, { "source": "mlsys.compound-workflow-pipelining", "relation": "uses", "target": "mlsys.token-streaming", "statement": "connects stages through", "conditions": "When stages exchange streamed tokens rather than complete responses." }, { "source": "mlsys.compound-workflow-pipelining", "relation": "reduces", "target": "mlsys.execution-time", "statement": "shortens the end-to-end", "conditions": "When stage dependencies are partial, so the critical path is shorter than the serial sum of stages." }, { "source": "mlsys.rag-pipeline-scheduling", "relation": "uses", "target": "mlsys.compound-workflow-pipelining", "statement": "overlaps retrieval and generation through", "conditions": "When retrieval or other upstream stages expose stable partial results that downstream generation can process while upstream work continues." }, { "source": "mlsys.safety-filter", "relation": "uses", "target": "mlsys.reachability-analysis", "statement": "certifies candidate actions against sets computed by", "conditions": "When the filter's safe set is the complement of the backward reachable set of the failure region." }, { "source": "mlsys.stopping-envelope", "relation": "uses", "target": "mlsys.reachability-analysis", "statement": "is defined from forward reachable sets computed by", "conditions": "When the envelope is the set of states from which a stopping maneuver stays clear of everything reachable by others." }, { "source": "mlsys.reachability-analysis", "relation": "alternative-to", "target": "mlsys.control-barrier-function", "statement": "gives a set-based safety certificate as an alternative to", "conditions": "CBFs give local differentiable constraints; reachability gives the safe set itself at the cost of state-space computation." }, { "source": "mlsys.reachability-analysis", "relation": "uses", "target": "mlsys.dynamics", "statement": "propagates the", "conditions": "Always; the reachable set is defined by the dynamics model and the input and disturbance bounds." }, { "source": "mlsys.fairness-monitoring", "relation": "uses", "target": "mlsys.group-fairness-metric", "statement": "computes on live traffic the", "conditions": "When protected attributes or accepted proxies are available at monitoring time." }, { "source": "mlsys.fairness-monitoring", "relation": "specializes", "target": "mlsys.monitoring", "statement": "is a kind of", "conditions": "When the monitored signal is a disparity across groups rather than an aggregate." }, { "source": "mlsys.fairness-monitoring", "relation": "uses", "target": "mlsys.sliced-evaluation", "statement": "applies in production the per-group measurement of", "conditions": "When slices are protected groups and the evaluation is continuous rather than a release gate." }, { "source": "mlsys.forward-error-correction", "relation": "mitigates", "target": "mlsys.link-signal-integrity", "statement": "absorbs, within its correction capability, the raw errors of", "conditions": "When the error rate stays within the code's correction budget." }, { "source": "mlsys.link-signal-integrity", "relation": "causes", "target": "mlsys.stragglers", "statement": "produces retransmission and retraining stalls that appear as", "conditions": "When one link degrades under heat or marginal cabling while the others stay clean." }, { "source": "mlsys.link-signal-integrity", "relation": "causes", "target": "mlsys.gray-failure", "statement": "degrades a link that still passes link-up checks, a form of", "conditions": "When flaps are intermittent and health checks sample too coarsely to see them." }, { "source": "mlsys.activation-function", "relation": "part-of", "target": "mlsys.neural-network", "statement": "is the nonlinearity between the linear layers of", "conditions": "When the architecture applies pointwise or gated nonlinearities between affine layers, as nearly all do; without them stacked linear layers collapse to one." }, { "source": "mlsys.fusion", "relation": "uses", "target": "mlsys.activation-function", "statement": "folds into a matrix-multiply epilogue the", "conditions": "When the nonlinearity is elementwise, so it can be applied to tile outputs before they are written to memory." }, { "source": "mlsys.activation-function", "relation": "enables", "target": "mlsys.activation-sparsity", "statement": "produces the exact zeros exploited by", "conditions": "When the nonlinearity is ReLU-like; smooth activations such as GELU or SwiGLU need thresholding or prediction to find skippable units." }, { "source": "mlsys.attention-ffn-disaggregation", "relation": "increases", "target": "mlsys.arithmetic-intensity", "statement": "pools tokens from many attention instances to raise the expert-side", "conditions": "When the FFN or expert pool batches tokens across several attention replicas, so each weight read serves more tokens." }, { "source": "mlsys.attention-ffn-disaggregation", "relation": "uses", "target": "mlsys.communication-overlap", "statement": "hides inter-pool activation transfer with", "conditions": "When micro-batches are pipelined so one batch computes while another's activations are in flight." }, { "source": "mlsys.attention-ffn-disaggregation", "relation": "uses", "target": "mlsys.expert-parallel", "statement": "serves the expert pool with", "conditions": "For mixture-of-experts models whose experts are spread across the FFN pool." }, { "source": "mlsys.attention-ffn-disaggregation", "relation": "trades-off", "target": "mlsys.inter-token-latency", "statement": "adds a per-layer activation transfer that lengthens", "conditions": "When the activation exchange between pools is not hidden by ping-pong micro-batching." }, { "source": "mlsys.attention-ffn-disaggregation", "relation": "uses", "target": "mlsys.heterogeneous-accelerator-execution", "statement": "can place each pool on a different accelerator type through", "conditions": "When attention is deployed on bandwidth-rich devices and FFN on compute-rich devices." }, { "source": "mlsys.label-shift-estimation", "relation": "mitigates", "target": "mlsys.distribution-shift", "statement": "corrects, without new labels, the label-shift form of", "conditions": "When class priors change but class-conditional input distributions do not; it does not address covariate or concept drift." }, { "source": "mlsys.label-shift-estimation", "relation": "uses", "target": "mlsys.calibration", "statement": "adjusts posteriors by Bayes' rule and so depends on", "conditions": "When predicted probabilities are re-weighted by the ratio of new to old priors; miscalibrated scores give wrong corrections." }, { "source": "mlsys.label-shift-estimation", "relation": "alternative-to", "target": "mlsys.retraining", "statement": "is a label-free alternative, for prior changes, to", "conditions": "When the shift is confined to class priors and fresh labels are unavailable or delayed." }, { "source": "mlsys.slam", "relation": "produces", "target": "mlsys.map-representation", "statement": "builds and maintains the", "conditions": "When mapping runs online from the platform's own observations." }, { "source": "mlsys.motion-planning", "relation": "uses", "target": "mlsys.map-representation", "statement": "answers collision and clearance queries against the", "conditions": "The representation must support the query the planner issues at planning rate, such as distance-to-obstacle or occupancy lookup." }, { "source": "mlsys.on-device-memory-budget", "relation": "limits", "target": "mlsys.map-representation", "statement": "bounds the resolution and extent of the", "conditions": "When the map lives in onboard memory beside the perception and policy models." }, { "source": "mlsys.motion-forecasting", "relation": "uses", "target": "mlsys.map-representation", "statement": "conditions predicted trajectories on the lanes and free space in the", "conditions": "When a prior or online map supplies road or room structure." }, { "source": "mlsys.compute-cost-estimate", "relation": "uses", "target": "mlsys.training-compute-estimate", "statement": "takes its operation count from", "conditions": "When the job is dense-transformer training." }, { "source": "mlsys.compute-cost-estimate", "relation": "uses", "target": "mlsys.model-flops-utilization", "statement": "divides peak-rate hours by the expected", "conditions": "When required operations are converted into wall-clock accelerator-hours." }, { "source": "mlsys.compute-cost-estimate", "relation": "uses", "target": "mlsys.inference-compute-estimate", "statement": "takes per-token work from", "conditions": "When pricing a serving fleet rather than a training run." }, { "source": "mlsys.dp-sgd", "relation": "increases", "target": "mlsys.time-to-train", "statement": "adds per-example gradient computation and clipping to", "conditions": "When per-example gradient and clipping overhead is not offset by fewer optimization steps or faster kernels at the same declared quality target." }, { "source": "mlsys.retraining", "relation": "uses", "target": "mlsys.drift-detection", "statement": "is triggered by", "conditions": "When a detected shift is diagnosed as harmful rather than benign; drift alone is a signal to investigate, not proof of degraded quality." }, { "source": "mlsys.fat-tree", "relation": "uses", "target": "mlsys.network-oversubscription", "statement": "is commonly built, to save cost, with", "conditions": "When a fat-tree is built with fewer spine ports than leaf downlinks to save cost." }, { "source": "mlsys.dataset-shuffling", "relation": "limits", "target": "mlsys.experiment-reproducibility", "statement": "breaks, unless its seed and shard order are recorded,", "conditions": "When runs must visit examples in the same order to be compared." }, { "source": "mlsys.supervised-fine-tuning", "relation": "uses", "target": "mlsys.chat-template", "statement": "renders every example through", "conditions": "For conversational or tool-using fine-tuning data turned into tokens before loss computation." }, { "source": "mlsys.safety-integrity-level", "relation": "determines", "target": "mlsys.safety-certification", "statement": "sets the rigor of evidence demanded by", "conditions": "When the function's integrity level has been assigned by hazard analysis." }, { "source": "mlsys.loop-stability-margin", "relation": "uses", "target": "mlsys.system-identification", "statement": "takes its plant model from", "conditions": "When an identified plant response is combined with a specified controller to compute the margin." }, { "source": "mlsys.in-doubt-action", "relation": "limits", "target": "mlsys.durable-execution", "statement": "blocks, until settled, the resumption of", "conditions": "When a resumed workflow reaches a step whose effect may or may not have happened." }, { "source": "mlsys.model-bandwidth-utilization", "relation": "measures", "target": "mlsys.utilization", "statement": "reports, as an achieved fraction of aggregate peak bandwidth, the decode-phase", "conditions": "When the phase is memory-bound, so model-required bytes per step rather than FLOPs set the ceiling; physical traffic beyond model-required bytes is excluded." }, { "source": "mlsys.fairness-monitoring", "relation": "uses", "target": "mlsys.drift-detection", "statement": "supplements group metrics with subgroup tests from", "conditions": "When a subgroup's metric moves and the cause must be separated from model, label, or sampling changes." }, { "source": "mlsys.tco", "relation": "uses", "target": "mlsys.compute-cost-estimate", "statement": "allocates workload compute cost through", "conditions": "When workload costs are allocated consistently with the ownership boundary and without double-counting capital, energy, or other included expenses." }, { "source": "mlsys.agent-harness", "relation": "uses", "target": "mlsys.idempotency", "statement": "retries tool steps safely through", "conditions": "When a step is re-issued after a timeout or crash and the tool contract declares an idempotency key or deduplication." }, { "source": "mlsys.wcet", "relation": "uses", "target": "mlsys.profiling", "statement": "takes observed samples, as evidence but not as a bound, from", "conditions": "When a measurement-based bound is argued from execution traces plus analysis of unobserved states." }, { "source": "mlsys.interrupt-latency", "relation": "limits", "target": "mlsys.real-time-scheduling", "statement": "limits the response-time budget available to", "conditions": "An interrupt releases the scheduled task, and the response deadline is measured from the interrupt request; handler execution and task execution consume additional time." }, { "source": "mlsys.precision-recall", "relation": "specializes", "target": "mlsys.metrics", "statement": "is the two-error-type family of", "conditions": "For classification or detection tasks where false positives and false negatives carry different costs." }, { "source": "mlsys.decision-policy", "relation": "uses", "target": "mlsys.precision-recall", "statement": "sets its operating threshold from the curve of", "conditions": "When the policy thresholds a score and the two error rates must be traded explicitly." }, { "source": "mlsys.instruction-hierarchy", "relation": "mitigates", "target": "mlsys.prompt-injection", "statement": "lowers the success rate of", "conditions": "When the model has been trained to respect source precedence or uses a provenance-marking prompt scheme validated against the applicable injection attacks, with trusted role assignments controlled outside attacker-supplied content." }, { "source": "mlsys.instruction-hierarchy", "relation": "uses", "target": "mlsys.rlhf", "statement": "learns instruction precedence using", "conditions": "When human preference comparisons over instruction-following behavior train a reward model and reinforcement learning optimizes the policy against that reward." }, { "source": "mlsys.restart-free-recovery", "relation": "increases", "target": "mlsys.training-goodput", "statement": "raises, by avoiding full-job restart and reducing recovery losses, the", "conditions": "When, under the same accelerator-time accounting, saved restart and rollback losses exceed replication, redundant-computation, reconfiguration, and reserved-capacity costs for failures within the supported failure model." }, { "source": "mlsys.restart-free-recovery", "relation": "uses", "target": "mlsys.pipeline-parallel", "statement": "replicates or re-partitions the stages of", "conditions": "For pipeline-parallel jobs, where a failed stage's work can be absorbed by redundant stages or by re-partitioning survivors." }, { "source": "mlsys.tensor-parallel", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "shards a layer's parameters and activations to lower the per-device", "conditions": "Aggregate footprint is unchanged; per-device savings depend on which tensors are partitioned." }, { "source": "mlsys.sequence-parallel", "relation": "reduces", "target": "mlsys.memory-footprint", "statement": "partitions activations along the sequence to lower the per-device", "conditions": "Only activation memory shrinks; parameters and optimizer state are unaffected." }, { "source": "mlsys.batch-size", "relation": "determines", "target": "mlsys.bound-regime", "statement": "sets the arithmetic intensity, and so the binding ceiling, of weight-streaming kernels in", "conditions": "Holds for kernels whose bytes are dominated by weights reused across the batch, such as GEMM and decode attention." }, { "source": "mlsys.io-aware-attention", "relation": "increases", "target": "mlsys.locality", "statement": "keeps attention tiles resident on chip to improve", "conditions": "Requires on-chip storage large enough for the chosen tile and an exact tiled formulation." }, { "source": "mlsys.mixed-precision", "relation": "determines", "target": "mlsys.bound-regime", "statement": "lowers the bytes moved per operation for operands held in the reduced format, shifting a kernel's position in", "conditions": "When the reduced format is applied to the operands that dominate traffic; master weights and accumulators keep full width, and attainable arithmetic throughput changes with the format as well." }, { "source": "mlsys.lossless-fabric", "relation": "causes", "target": "mlsys.gray-failure", "statement": "can propagate pause frames into stalls that look healthy to some observers, a form of", "conditions": "Priority flow control enabled end to end; a single paused queue can block unrelated flows." }, { "source": "mlsys.llm-request-scheduling", "relation": "determines", "target": "mlsys.ttft", "statement": "sets, through its prefill-versus-decode priority, how long a new request waits for", "conditions": "Continuous-batching engines where prefill and decode share iterations." }, { "source": "mlsys.cache-aware-routing", "relation": "reduces", "target": "mlsys.ttft", "statement": "skips recomputing a cached prefix and so lowers", "conditions": "Requests share prefixes and the routed replica still holds the cached state." }, { "source": "mlsys.llm-request-scheduling", "relation": "limits", "target": "mlsys.kv-cache", "statement": "preempts in-flight requests by evicting or swapping blocks, bounding resident", "conditions": "Continuous-batching engines with preemption; recompute-on-resume costs prefill time." }, { "source": "mlsys.distributed-tracing", "relation": "measures", "target": "mlsys.tail-latency", "statement": "attributes a slow request's time to the spans behind", "conditions": "Request-scoped tracing enabled across every hop on the path, with sampling that keeps slow requests." }, { "source": "mlsys.graceful-degradation", "relation": "mitigates", "target": "mlsys.metastable-failure", "statement": "sheds work to a cheaper mode that breaks the sustaining loop of", "conditions": "The degraded mode must cost less than the overloaded path and be entered before saturation." }, { "source": "mlsys.sliced-evaluation", "relation": "detects", "target": "mlsys.distribution-shift", "statement": "localizes, by a drop concentrated in one slice, the population affected by", "conditions": "Slices are defined on production inputs and labels or proxies exist for the window." }, { "source": "mlsys.off-policy-evaluation", "relation": "detects", "target": "mlsys.proxy-objective", "statement": "exposes divergence between the optimized proxy and the estimated deployed outcome of", "conditions": "When logged propensities are known, the logging and candidate policies overlap, and the intended outcome is observed separately from the proxy." }, { "source": "mlsys.cost-sensitive-decision", "relation": "mitigates", "target": "mlsys.proxy-objective", "statement": "re-weights decisions by the real cost of each error type, correcting a misaligned", "conditions": "Error costs must be stated per outcome, not inferred from the proxy." }, { "source": "mlsys.thermal-dynamics", "relation": "causes", "target": "mlsys.measurement-variability", "statement": "makes short runs overstate sustained performance, a source of", "conditions": "Measurement windows shorter than the device's thermal time constant." }, { "source": "mlsys.quantization", "relation": "reduces", "target": "mlsys.energy-per-inference", "statement": "cuts bytes moved and energy per multiply-accumulate, lowering", "conditions": "The target hardware has low-precision datapaths; dequantization overhead must not dominate." }, { "source": "mlsys.prefix-cache", "relation": "reduces", "target": "mlsys.cost-per-token", "statement": "avoids recomputing shared prefixes, lowering", "conditions": "Requests share long prefixes such as system prompts; cache hits must be routed to the holding replica." }, { "source": "mlsys.llm-call-tracing", "relation": "detects", "target": "mlsys.over-refusal", "statement": "attributes a refusal to the input filter, the model, or the output filter, localizing", "conditions": "Filter decisions are recorded as spans alongside the model call, and the refused requests are labeled benign by an independent rubric." }, { "source": "mlsys.supervised-fine-tuning", "relation": "mitigates", "target": "mlsys.over-refusal", "statement": "trains compliance on curated borderline-benign requests, reducing", "conditions": "The curated set must be reviewed so it does not also teach compliance on harmful requests." }, { "source": "mlsys.agent-interruption", "relation": "limits", "target": "mlsys.model-directed-control-flow", "statement": "lets an operator override the next step chosen by", "conditions": "Interruption boundaries must sit at tool-call granularity so a pending action is not half-applied." }, { "source": "mlsys.egress-control", "relation": "mitigates", "target": "mlsys.prompt-injection", "statement": "removes the exfiltration channel exploited by", "conditions": "Allow-listed destinations; injected instructions can still act within the sandbox." }, { "source": "mlsys.checkpoint-io", "relation": "uses", "target": "mlsys.storage-tiering", "statement": "stages snapshots on a fast local tier before draining them through", "conditions": "The local tier must hold several snapshots and drain faster than the checkpoint interval." }, { "source": "mlsys.llm-call-tracing", "relation": "measures", "target": "mlsys.context-staleness", "statement": "records the passages and index timestamps needed to judge", "conditions": "When recorded index timestamps are compared with source versions or validity intervals; recency alone does not establish freshness." }, { "source": "mlsys.watchdog-timer", "relation": "detects", "target": "mlsys.feedback-delay", "statement": "reports a missed refresh deadline as a fault, revealing excess", "conditions": "The window is derived from the control period and the refresh is placed after actuation." }, { "source": "mlsys.mixed-criticality", "relation": "mitigates", "target": "mlsys.priority-inversion", "statement": "partitions time and resources so lower-criticality tasks cannot hold what high-criticality tasks need, preventing", "conditions": "Shared resources must have bounded access or be partitioned; priority inheritance still applies inside a partition." }, { "source": "mlsys.cross-stack-profiling", "relation": "measures", "target": "mlsys.input-pipeline", "statement": "attributes device idle gaps to the host-side stages of", "conditions": "Host and device timelines are correlated on one clock." }, { "source": "mlsys.evaluation-harness", "relation": "detects", "target": "mlsys.catastrophic-forgetting", "statement": "reruns a fixed base-capability suite on each checkpoint, revealing", "conditions": "The suite and prompts are held fixed across checkpoints." }, { "source": "mlsys.regularization", "relation": "mitigates", "target": "mlsys.catastrophic-forgetting", "statement": "penalizes drift from the base parameters or outputs, limiting", "conditions": "The penalty, such as a KL term to the base model, must be weighted against the target-task loss." }, { "source": "mlsys.peft", "relation": "mitigates", "target": "mlsys.catastrophic-forgetting", "statement": "freezes the base weights so adaptation cannot overwrite them, limiting", "conditions": "Holds when the adapter capacity is small relative to the base and base weights stay frozen." }, { "source": "mlsys.release-gate", "relation": "detects", "target": "mlsys.model-feature-compatibility", "statement": "checks the candidate's tokenizer, schema, and operator set against the serving stack, revealing broken", "conditions": "The gate runs against the same runtime image that will serve the model." }, { "source": "mlsys.backfill-scheduling", "relation": "uses", "target": "mlsys.gang-scheduling", "statement": "uses whole-job allocation reservations established by", "conditions": "When the backfill planner consults reserved start times and resource sets for gang-allocated jobs and admits other jobs only when their bounded resource use preserves those starts." }, { "source": "mlsys.agent-memory", "relation": "enables", "target": "mlsys.prompt-injection", "statement": "enables persistent", "conditions": "When attacker-controlled instructions can be stored in agent memory and later retrieved into a context where the model may treat them as authoritative." }, { "source": "mlsys.tensor-parallel", "relation": "reduces", "target": "mlsys.tpot", "statement": "pools the memory bandwidth of several devices to shorten", "conditions": "For the same model, precision, batch, and output workload, when tensor sharding reduces per-device decode memory traffic and the saved memory-access time exceeds the added collective, synchronization, and execution overhead." }, { "source": "mlsys.service-level-agreement", "relation": "limits", "target": "mlsys.slo", "statement": "constrains the acceptable targets for", "conditions": "When the internal objective covers the same service indicator, population, and measurement window as the contractual commitment and must provide a stricter target with room for intervention." }, { "source": "mlsys.compute-cost-estimate", "relation": "produces", "target": "mlsys.cost-per-token", "statement": "produces an estimate of", "conditions": "When estimated accelerator expenditure is divided by a declared processed-token or generated-token count and the reported cost boundary covers accelerator compute only." }, { "source": "mlsys.deterministic-field-network", "relation": "limits", "target": "mlsys.timing-jitter", "statement": "bounds the transport contribution to", "conditions": "When control messages use synchronized, fixed transmission slots with bounded interfering traffic; the bound concerns network-induced variation rather than all sensing, computation, and actuation jitter." }, { "source": "mlsys.self-supervised-learning", "relation": "uses", "target": "mlsys.contrastive-learning", "statement": "often trains through", "conditions": "When positives and negatives are constructed from the data itself, without example-level labels." }, { "source": "mlsys.data-parallel", "relation": "uses", "target": "mlsys.state-sharding", "statement": "partitions state across workers through", "conditions": "When each worker processes different examples but holds only a shard of parameters, gradients, or optimizer state." }, { "source": "mlsys.shared-memory-bank-conflict", "relation": "reduces", "target": "mlsys.memory-bandwidth", "statement": "reduces effective shared-memory", "conditions": "Different addresses in the same shared-memory bank require serialized requests; same-address broadcast is excluded." } ], "retired": [ { "id": "mlsys.llm-kernel-generation", "replaced_by": "mlsys.autotuning", "version": "0.2-draft", "reason": "Using ML to optimize systems is outside the taxonomy's scope; automated kernel search is covered by autotuning." } ], "guides": { "roles_note": "Role profiles name the subjects an interviewer or hiring manager in that role would expect a strong engineer to know (core) and to be able to work with (working). They describe knowledge, not seniority; depth comes from the level at which a question or task is posed.", "roles": [ { "id": "inference-engineer", "label": "ML inference and serving engineer", "summary": "Builds and operates the systems that answer model requests within latency, throughput, and cost targets.", "core": [ "serving-architecture", "latency-goodput", "request-scheduling", "llm-inference", "model-compression", "profiling-utilization" ], "working": [ "requirements-slos", "queueing-models", "runtime-dispatch", "memory-systems", "parallelism", "capacity-provisioning", "failure-semantics", "monitoring", "cost-economics", "conditional-computation", "efficient-sequence-models", "structured-generation", "compound-ai", "model-artifacts", "numerics", "tensor-operators", "system-benchmarking", "release-management" ] }, { "id": "training-infra-engineer", "label": "Training infrastructure engineer", "summary": "Makes large training jobs fit, scale, and survive failures across accelerators and nodes.", "core": [ "parallelism", "parallelism-planning", "collectives", "training-memory", "checkpoint-recovery", "profiling-utilization" ], "working": [ "hardware-faults", "node-interconnect", "network-fabrics", "storage-systems", "cluster-scheduling", "speedup-limits", "numerics", "training-optimization", "data-loading", "scaling-laws", "coordination", "datacenter-facilities", "fault-injection", "programming-models", "memory-systems", "roofline-locality", "system-benchmarking", "capacity-provisioning" ] }, { "id": "mlops-engineer", "label": "ML platform and MLOps engineer", "summary": "Builds the shared ML platform and runs the pipelines, releases, and monitoring that keep models reproducible, current, and correct in production.", "core": [ "pipelines-reproducibility", "release-management", "monitoring", "drift-retraining", "ml-platforms", "data-lineage", "model-artifacts" ], "working": [ "ml-testing", "incident-response", "serving-architecture", "data-quality", "data-leakage-skew", "experiment-management", "retirement", "cost-economics", "continual-learning", "requirements-slos", "data-pipelines", "failure-semantics", "access-isolation", "capacity-provisioning", "model-supply-chain", "evaluation-protocols" ] }, { "id": "post-training-engineer", "label": "Post-training engineer", "summary": "Turns pretrained models into instruction-following, reasoning, and tool-using models through fine-tuning, preference optimization, and reinforcement learning, including the rollout inference inside the training loop.", "core": [ "post-training", "fine-tuning", "training-memory", "llm-inference", "generative-agent-eval", "training-optimization", "data-curation" ], "working": [ "test-time-compute", "collectives", "numerics", "checkpoint-recovery", "experiment-management", "runtime-enforcement", "model-families", "learning-paradigms", "parallelism", "scaling-laws", "evaluation-protocols", "structured-generation", "data-collection", "ml-attacks", "parallelism-planning", "tool-interfaces" ] }, { "id": "ml-data-engineer", "label": "ML data engineer", "summary": "Builds the data collection, curation, validation, and delivery that training and serving depend on.", "core": [ "data-collection", "data-curation", "data-quality", "data-pipelines", "data-lineage", "data-leakage-skew" ], "working": [ "data-loading", "storage-systems", "pipelines-reproducibility", "drift-retraining", "privacy-dp", "evaluation-protocols", "governance-regulation" ] }, { "id": "edge-ml-engineer", "label": "Edge and embedded ML engineer", "summary": "Fits models onto phones, microcontrollers, and devices within memory, power, and latency limits.", "core": [ "on-device-deployment", "edge-embedded-hardware", "model-compression", "chip-power-thermal", "numerics", "profiling-utilization" ], "working": [ "hardware-aware-design", "memory-systems", "workload-characterization", "runtime-dispatch", "ml-compilers", "model-artifacts", "real-time-systems", "energy-carbon", "tensor-operators", "system-benchmarking", "federated-learning" ] }, { "id": "performance-engineer", "label": "AI accelerator and performance engineer", "summary": "Explains and removes the bottleneck between a model and the hardware it runs on.", "core": [ "roofline-locality", "accelerator-architecture", "memory-systems", "gpu-programming", "tensor-operators", "ml-compilers", "profiling-utilization", "runtime-dispatch" ], "working": [ "workload-characterization", "numerics", "node-interconnect", "collectives", "system-benchmarking", "speedup-limits", "chip-power-thermal", "programming-models", "architecture-blocks", "parallelism", "llm-inference", "model-compression", "hardware-aware-design" ] }, { "id": "ml-security-engineer", "label": "ML security engineer", "summary": "Defends models, data, and ML services against attack, theft, and leakage.", "core": [ "ml-attacks", "model-supply-chain", "privacy-dp", "access-isolation" ], "working": [ "runtime-enforcement", "robustness-slices", "model-artifacts", "tool-interfaces", "monitoring", "incident-response", "governance-regulation", "data-lineage", "federated-learning", "agent-harness", "generative-agent-eval", "ml-testing" ] }, { "id": "safety-eval-engineer", "label": "AI safety and evaluation engineer", "summary": "Measures what a model or agent can and cannot be trusted to do, specifies the limits it must stay within, and designs the evaluations that check them.", "core": [ "evaluation-protocols", "metrics-uncertainty", "robustness-slices", "generative-agent-eval", "ml-testing", "ml-attacks" ], "working": [ "runtime-enforcement", "verification-validation", "fairness-bias", "post-training", "hazards-operating-domain", "human-oversight", "governance-regulation", "objectives-decisions", "transparency-explainability", "monitoring", "safety-cases" ] }, { "id": "agent-platform-engineer", "label": "Agent platform engineer", "summary": "Builds the runtime that lets model-directed agents use tools and state safely, reliably, and within budget.", "core": [ "agent-harness", "tool-interfaces", "context-memory", "durable-execution", "access-isolation", "generative-agent-eval", "failure-semantics" ], "working": [ "multi-agent", "structured-generation", "llm-inference", "compound-ai", "runtime-enforcement", "monitoring", "retrieval-rag", "cost-economics", "ml-attacks", "coordination", "requirements-slos", "test-time-compute", "incident-response", "human-oversight" ] }, { "id": "robotics-ml-engineer", "label": "Robotics and autonomy ML engineer", "summary": "Puts learned perception and control on machines that act in the physical world under timing and safety limits.", "core": [ "perception-sensing", "state-estimation", "planning-control", "real-time-systems", "robot-learning", "runtime-enforcement" ], "working": [ "hazards-operating-domain", "safety-cases", "edge-embedded-hardware", "on-device-deployment", "model-compression", "profiling-utilization", "chip-power-thermal", "data-collection", "verification-validation" ] }, { "id": "applied-ml-engineer", "label": "Applied ML engineer", "summary": "Trains, adapts, evaluates, and ships models for a product, owning model quality from the data it learns from to the service that serves it.", "core": [ "model-families", "fine-tuning", "training-optimization", "evaluation-protocols", "metrics-uncertainty", "experiment-management", "objectives-decisions", "data-quality", "data-leakage-skew" ], "working": [ "ml-testing", "robustness-slices", "serving-architecture", "model-artifacts", "monitoring", "drift-retraining", "model-compression", "retrieval-rag", "data-collection", "release-management", "requirements-slos" ] }, { "id": "pretraining-engineer", "label": "Pretraining and model-scaling engineer", "summary": "Designs and runs the large training runs that produce foundation models, choosing data mixture, architecture, and compute allocation and keeping the run stable at scale.", "core": [ "scaling-laws", "data-curation", "training-optimization", "architecture-blocks", "numerics", "parallelism" ], "working": [ "training-memory", "parallelism-planning", "experiment-management", "evaluation-protocols", "conditional-computation", "efficient-sequence-models", "model-families", "learning-paradigms", "data-pipelines", "checkpoint-recovery", "profiling-utilization", "hardware-faults", "collectives" ] }, { "id": "ai-infrastructure-engineer", "label": "AI cluster and infrastructure engineer", "summary": "Builds and operates the shared accelerator fleet, from scheduling, networking, storage, and power to the capacity that training and serving teams consume.", "core": [ "cluster-scheduling", "network-fabrics", "storage-systems", "capacity-provisioning", "datacenter-facilities", "hardware-faults" ], "working": [ "node-interconnect", "collectives", "coordination", "failure-semantics", "fault-injection", "checkpoint-recovery", "monitoring", "incident-response", "access-isolation", "cost-economics", "energy-carbon", "embodied-lifecycle", "chip-power-thermal", "accelerator-architecture", "requirements-slos", "system-benchmarking" ] }, { "id": "ml-compiler-engineer", "label": "ML compiler and runtime engineer", "summary": "Lowers models to efficient kernels and runtime execution across accelerators.", "core": [ "ml-compilers", "programming-models", "tensor-operators", "runtime-dispatch", "gpu-programming" ], "working": [ "accelerator-architecture", "memory-systems", "numerics", "roofline-locality", "profiling-utilization", "system-benchmarking", "model-artifacts", "parallelism-planning" ] }, { "id": "recommendation-ranking-engineer", "label": "Recommendation and ranking ML engineer", "summary": "Builds retrieval and ranking models whose offline measurements, online decisions, and feature pipelines improve product outcomes under serving constraints.", "core": [ "model-families", "retrieval-rag", "objectives-decisions", "data-quality", "data-leakage-skew", "evaluation-protocols", "metrics-uncertainty" ], "working": [ "architecture-blocks", "training-optimization", "experiment-management", "data-collection", "data-curation", "data-pipelines", "data-loading", "parallelism", "serving-architecture", "latency-goodput", "monitoring", "drift-retraining", "fairness-bias", "privacy-dp" ] } ], "symptoms_note": "Each symptom lists the concepts most often behind it, what to check, and what to change. Every check and lever must share a relationship with one of the listed causes, so the guide cannot point at an idea the graph does not connect to the problem. Causes are starting points for diagnosis, not a complete fault tree.", "symptoms": [ { "id": "training-oom", "label": "Training runs out of accelerator memory", "signal": "Out-of-memory errors when model size, batch, or sequence length grows, or reserved memory well above allocated memory (fragmentation).", "causes": [ "mlsys.training-state", "mlsys.memory-footprint", "mlsys.memory-fragmentation", "mlsys.memory-capacity", "mlsys.batch-size", "mlsys.context", "mlsys.caching-allocator" ], "checks": [ "mlsys.profiling", "mlsys.training-memory-estimate" ], "levers": [ "mlsys.activation-recompute", "mlsys.state-sharding", "mlsys.mixed-precision", "mlsys.gradient-accumulation", "mlsys.memory-offloading", "mlsys.memory-efficient-optimizer", "mlsys.memory-planning", "mlsys.pipeline-parallel", "mlsys.peft", "mlsys.tensor-parallel", "mlsys.sequence-parallel" ] }, { "id": "low-utilization", "label": "Useful throughput is far below what the hardware can deliver", "signal": "Model FLOPs utilization or step throughput well below what the workload's roofline allows, from idle gaps between kernels or steps, exposed communication, or pipeline bubbles.", "causes": [ "mlsys.bound-regime", "mlsys.memory-bandwidth", "mlsys.dispatch", "mlsys.input-pipeline", "mlsys.collectives", "mlsys.pipeline-bubble", "mlsys.stragglers", "mlsys.python-gil", "mlsys.locality", "mlsys.batch-size" ], "checks": [ "mlsys.bottleneck-analysis", "mlsys.cross-stack-profiling", "mlsys.model-flops-utilization", "mlsys.profiling", "mlsys.roofline" ], "levers": [ "mlsys.fusion", "mlsys.graph-replay", "mlsys.device-streams", "mlsys.persistent-kernel", "mlsys.input-prefetching", "mlsys.communication-overlap", "mlsys.pipeline-schedule", "mlsys.sequence-packing", "mlsys.storage-tiering", "mlsys.topology-aware-placement", "mlsys.tiling", "mlsys.io-aware-attention", "mlsys.mixed-precision" ] }, { "id": "loss-spikes", "label": "Training loss spikes or diverges", "signal": "Sudden loss jumps, NaN or Inf values, or a run that stops improving after a change in scale or precision.", "causes": [ "mlsys.training-instability", "mlsys.numerical-stability", "mlsys.precision", "mlsys.activation-outliers", "mlsys.data-quality", "mlsys.silent-data-corruption", "mlsys.policy-staleness" ], "checks": [ "mlsys.training-diagnostics", "mlsys.abft", "mlsys.data-validation" ], "levers": [ "mlsys.gradient-clipping", "mlsys.learning-rate-schedule", "mlsys.normalization-layer", "mlsys.loss-scaling", "mlsys.accumulation-precision", "mlsys.weight-initialization", "mlsys.checkpoint", "mlsys.block-scaled-format", "mlsys.quality-filtering", "mlsys.fleet-hardware-screening" ] }, { "id": "collective-hang", "label": "A distributed job hangs or times out", "signal": "Ranks stall waiting in a collective, or the job times out with no error from the failing rank.", "causes": [ "mlsys.gray-failure", "mlsys.process-group-rendezvous", "mlsys.stragglers", "mlsys.collectives", "mlsys.lossless-fabric", "mlsys.link-signal-integrity" ], "checks": [ "mlsys.distributed-hang-diagnosis", "mlsys.distributed-tracing", "mlsys.profiling" ], "levers": [ "mlsys.elastic-training", "mlsys.congestion-control", "mlsys.forward-error-correction" ] }, { "id": "slow-steps", "label": "Step time varies and the job runs at the pace of its slowest worker", "signal": "Per-step time has a long tail and every step waits for its slowest rank, so the job runs at the pace of one worker even though most ranks finish early.", "causes": [ "mlsys.stragglers", "mlsys.performance-interference", "mlsys.gray-failure", "mlsys.input-pipeline", "mlsys.collectives", "mlsys.thermal-dynamics", "mlsys.network-topology" ], "checks": [ "mlsys.profiling", "mlsys.distributed-tracing", "mlsys.parallel-efficiency" ], "levers": [ "mlsys.variable-length-parallelism", "mlsys.expert-placement", "mlsys.traffic-classes", "mlsys.congestion-control", "mlsys.load-balancing-loss", "mlsys.input-prefetching", "mlsys.fine-grained-accelerator-sharing", "mlsys.topology-aware-placement", "mlsys.communication-overlap", "mlsys.gpu-partitioning", "mlsys.storage-tiering" ] }, { "id": "ttft-spike", "label": "Time to first token rises under load", "signal": "First tokens arrive late for some requests while decode speed looks normal.", "quantities": [ "mlsys.ttft" ], "causes": [ "mlsys.queueing", "mlsys.prefill-decode", "mlsys.arrival-burstiness", "mlsys.cold-start", "mlsys.batching", "mlsys.head-of-line-blocking" ], "checks": [ "mlsys.latency-percentiles", "mlsys.workload-trace", "mlsys.latency-throughput-curve" ], "levers": [ "mlsys.prefix-cache", "mlsys.pd-disaggregation", "mlsys.autoscaling", "mlsys.admission", "mlsys.load-balancing", "mlsys.non-prefix-kv-reuse", "mlsys.mmap-weights", "mlsys.llm-request-scheduling", "mlsys.cache-aware-routing", "mlsys.health-checks", "mlsys.continuous-batching", "mlsys.chunked-prefill", "mlsys.rag-pipeline-scheduling" ] }, { "id": "kv-exhaustion", "label": "Serving runs out of KV-cache memory", "signal": "Requests are preempted, recomputed, or rejected as context lengths or concurrency grow.", "quantities": [ "mlsys.kv-cache" ], "causes": [ "mlsys.context", "mlsys.memory-fragmentation", "mlsys.memory-footprint", "mlsys.agentic-workload", "mlsys.memory-capacity", "mlsys.kv-memory-allocation", "mlsys.batching" ], "checks": [ "mlsys.inference-compute-estimate", "mlsys.profiling" ], "levers": [ "mlsys.paged-kv", "mlsys.kv-cache-quantization", "mlsys.kv-cache-eviction", "mlsys.kv-cache-offload", "mlsys.prefix-cache", "mlsys.admission", "mlsys.reasoning-budget", "mlsys.grouped-query-attention", "mlsys.multi-head-latent-attention", "mlsys.sliding-window-attention", "mlsys.tensor-parallel", "mlsys.llm-request-scheduling", "mlsys.generation-stopping" ] }, { "id": "tail-latency", "label": "p99 latency is far above the median", "signal": "Most requests are fast, a few are much slower, and the gap widens with load or fan-out.", "quantities": [ "mlsys.tail-latency" ], "causes": [ "mlsys.queueing", "mlsys.fan-out-tail-amplification", "mlsys.arrival-burstiness", "mlsys.performance-interference", "mlsys.batching", "mlsys.cold-start", "mlsys.metastable-failure", "mlsys.managed-runtime-interop", "mlsys.utilization" ], "checks": [ "mlsys.latency-percentiles", "mlsys.workload-trace", "mlsys.distributed-tracing", "mlsys.latency-throughput-curve" ], "levers": [ "mlsys.hedged-requests", "mlsys.admission", "mlsys.backpressure", "mlsys.load-balancing", "mlsys.autoscaling", "mlsys.llm-request-scheduling", "mlsys.gpu-partitioning", "mlsys.retry-backoff", "mlsys.latency-budget", "mlsys.continuous-batching", "mlsys.capacity-planning", "mlsys.health-checks", "mlsys.graceful-degradation", "mlsys.request-timeout" ] }, { "id": "quality-drop", "label": "Production quality drops with no error or alert", "signal": "Business metrics or user feedback worsen while the service stays up and offline evaluation still passes.", "causes": [ "mlsys.distribution-shift", "mlsys.concept-drift", "mlsys.train-serve-skew", "mlsys.feedback-loop", "mlsys.schema-evolution", "mlsys.feature-freshness", "mlsys.data-quality", "mlsys.rollout", "mlsys.delayed-outcomes", "mlsys.data-dependency", "mlsys.model-feature-compatibility" ], "checks": [ "mlsys.data-validation", "mlsys.drift-detection", "mlsys.monitoring", "mlsys.ood-detection", "mlsys.shadow-deployment", "mlsys.integration-testing", "mlsys.lineage", "mlsys.calibration", "mlsys.sliced-evaluation" ], "levers": [ "mlsys.retraining", "mlsys.online-feature-consistency", "mlsys.feature-store", "mlsys.feature-backfill", "mlsys.data-contract", "mlsys.rollback", "mlsys.release-gate", "mlsys.online-learning", "mlsys.fleet-data-engine", "mlsys.label-shift-estimation" ] }, { "id": "offline-online-gap", "label": "The model wins offline and loses online", "signal": "A model that beats the baseline on held-out data does not improve, or hurts, the live metric.", "causes": [ "mlsys.train-serve-skew", "mlsys.leakage", "mlsys.proxy-objective", "mlsys.selection-bias", "mlsys.feedback-loop", "mlsys.distribution-shift", "mlsys.benchmark-contamination", "mlsys.experiment-interference", "mlsys.feature-freshness", "mlsys.delayed-outcomes" ], "checks": [ "mlsys.shadow-deployment", "mlsys.online-experiment", "mlsys.integration-testing", "mlsys.feature-attribution", "mlsys.drift-detection", "mlsys.off-policy-evaluation", "mlsys.calibration" ], "levers": [ "mlsys.point-in-time", "mlsys.online-feature-consistency", "mlsys.feature-store", "mlsys.evaluation-split", "mlsys.multiple-comparisons", "mlsys.retraining", "mlsys.cost-sensitive-decision" ] }, { "id": "device-throttling", "label": "On-device inference slows after a few minutes", "signal": "During sustained use, latency rises as the device heats and clocks drop, or the battery drains faster than the workload should need.", "causes": [ "mlsys.thermal-dynamics", "mlsys.tdp", "mlsys.energy-per-inference", "mlsys.heterogeneous-soc" ], "checks": [ "mlsys.energy-measurement", "mlsys.execution-time", "mlsys.measurement-variability" ], "levers": [ "mlsys.duty-cycling", "mlsys.dvfs", "mlsys.edge-npu", "mlsys.always-on-cascade", "mlsys.early-exit", "mlsys.device-cloud-partitioning", "mlsys.quantization" ] }, { "id": "serving-cost", "label": "Serving cost grows faster than traffic", "signal": "Cost per request or per token rises as usage grows, or capacity sits idle between peaks.", "quantities": [ "mlsys.cost-per-token", "mlsys.unit-cost" ], "causes": [ "mlsys.utilization", "mlsys.workload", "mlsys.idle-dynamic-power", "mlsys.cold-start", "mlsys.test-time-scaling", "mlsys.data-transfer-cost", "mlsys.agentic-workload" ], "checks": [ "mlsys.cost-attribution", "mlsys.profiling", "mlsys.workload-trace", "mlsys.compute-cost-estimate" ], "levers": [ "mlsys.batching", "mlsys.model-routing", "mlsys.model-cascade", "mlsys.semantic-cache", "mlsys.reasoning-budget", "mlsys.multi-model-serving", "mlsys.gpu-partitioning", "mlsys.autoscaling", "mlsys.quantization", "mlsys.spot-capacity", "mlsys.batch-inference", "mlsys.multi-adapter-serving", "mlsys.capacity-planning", "mlsys.model-swapping", "mlsys.serverless-inference", "mlsys.prefix-cache" ] }, { "id": "over-refusal-complaints", "label": "Users report that the assistant refuses harmless requests", "signal": "Evaluations or reviewed reports show benign requests rejected, with traces showing whether an input filter, the model, or an output filter refused.", "quantities": [ "mlsys.over-refusal" ], "causes": [ "mlsys.decision-policy", "mlsys.reward-hacking", "mlsys.guardrails", "mlsys.rlhf" ], "checks": [ "mlsys.sliced-evaluation", "mlsys.calibration", "mlsys.llm-call-tracing" ], "levers": [ "mlsys.cost-sensitive-decision", "mlsys.preference-optimization", "mlsys.reward-model", "mlsys.supervised-fine-tuning" ] }, { "id": "runaway-agent", "label": "An agent loops, overspends, or takes an unintended action", "signal": "Runs that exceed step or cost limits, repeat tool calls, or act on injected instructions.", "causes": [ "mlsys.model-directed-control-flow", "mlsys.coordination-failure", "mlsys.prompt-injection", "mlsys.confused-deputy", "mlsys.irreversible-effects", "mlsys.context-degradation", "mlsys.tool-contract" ], "checks": [ "mlsys.llm-call-tracing", "mlsys.agent-loop-detection" ], "levers": [ "mlsys.agent-budget", "mlsys.guardrails", "mlsys.ephemeral-sandbox", "mlsys.approval-gate", "mlsys.capability-based-security", "mlsys.least-privilege", "mlsys.agent-credentials", "mlsys.information-flow-control", "mlsys.compensation", "mlsys.idempotency", "mlsys.context-compaction", "mlsys.agent-interruption", "mlsys.egress-control" ] }, { "id": "slow-decode", "label": "Token generation is slow even when lightly loaded", "signal": "High time per output token that does not improve with spare capacity, or streaming that stutters as context or batch grows.", "quantities": [ "mlsys.tpot" ], "causes": [ "mlsys.memory-bandwidth", "mlsys.prefill-decode", "mlsys.dispatch", "mlsys.kv-cache", "mlsys.context", "mlsys.batching", "mlsys.expert-offloading" ], "checks": [ "mlsys.inter-token-latency", "mlsys.roofline", "mlsys.cross-stack-profiling", "mlsys.bound-regime", "mlsys.latency-throughput-curve", "mlsys.model-bandwidth-utilization" ], "levers": [ "mlsys.speculative-decoding", "mlsys.weight-only-quantization", "mlsys.kv-cache-quantization", "mlsys.grouped-query-attention", "mlsys.graph-replay", "mlsys.fusion", "mlsys.persistent-kernel", "mlsys.activation-sparsity", "mlsys.kv-cache-eviction", "mlsys.tensor-parallel" ] }, { "id": "checkpoint-stall", "label": "Training loses hours to checkpointing or recovery", "signal": "Periodic step-time gaps at each save, or long recomputation after a failure because saves are infrequent.", "causes": [ "mlsys.checkpoint-io", "mlsys.mtbf", "mlsys.training-state", "mlsys.consistent-cut", "mlsys.job-preemption" ], "checks": [ "mlsys.training-goodput" ], "levers": [ "mlsys.async-checkpointing", "mlsys.incremental-checkpointing", "mlsys.parallel-file-system", "mlsys.storage-tiering", "mlsys.state-sharding", "mlsys.elastic-training", "mlsys.checkpoint-interval" ] }, { "id": "rag-ungrounded", "label": "Retrieval-augmented answers are wrong or out of date", "signal": "Answers cite stale or irrelevant passages, miss recently updated documents, or state facts absent from the retrieved context.", "causes": [ "mlsys.hallucination", "mlsys.retrieval", "mlsys.dense-retrieval", "mlsys.context-staleness", "mlsys.context-degradation", "mlsys.embedding-migration", "mlsys.ann", "mlsys.context" ], "checks": [ "mlsys.llm-judge", "mlsys.index-freshness", "mlsys.llm-call-tracing", "mlsys.retrieval-metrics" ], "levers": [ "mlsys.hybrid-retrieval", "mlsys.reranking", "mlsys.document-chunking", "mlsys.context-selection", "mlsys.multi-hop-retrieval", "mlsys.context-compaction" ] }, { "id": "control-loop-miss", "label": "A robot or vehicle misses its control-loop deadline", "signal": "Perception, planning, or control finishes after its period deadline, or the state estimate used for actuation lags behind physical motion, when compute load rises.", "causes": [ "mlsys.feedback-delay", "mlsys.observation-age", "mlsys.wcet", "mlsys.priority-inversion", "mlsys.interrupt-latency", "mlsys.robotics-middleware", "mlsys.vla-model", "mlsys.thermal-dynamics", "mlsys.timing-jitter" ], "checks": [ "mlsys.xil-testing", "mlsys.time-synchronization", "mlsys.profiling", "mlsys.perception-latency-budget", "mlsys.watchdog-timer" ], "levers": [ "mlsys.real-time-scheduling", "mlsys.action-chunking", "mlsys.zero-copy-dma", "mlsys.mixed-criticality", "mlsys.dual-rate-hierarchical-policy" ] }, { "id": "input-starvation", "label": "Accelerators sit idle waiting for input data", "signal": "Device utilization drops between steps while host CPUs, loader workers, or storage reads are saturated, and step time falls when real input is replaced by synthetic batches.", "causes": [ "mlsys.input-pipeline", "mlsys.host-data-loader", "mlsys.python-gil", "mlsys.dataset-storage-layout", "mlsys.storage-access-pattern" ], "checks": [ "mlsys.utilization", "mlsys.cross-stack-profiling" ], "levers": [ "mlsys.input-prefetching", "mlsys.streaming-dataset-format", "mlsys.storage-tiering", "mlsys.parallel-file-system" ] }, { "id": "quantization-accuracy-loss", "label": "A quantized model loses accuracy", "signal": "After quantization, held-out quality falls more than the format should cost, often on particular inputs or layers, while the full-precision model is fine.", "causes": [ "mlsys.quantization", "mlsys.activation-outliers", "mlsys.quantization-granularity", "mlsys.post-training-quantization", "mlsys.precision", "mlsys.accumulation-precision" ], "checks": [ "mlsys.metrics" ], "levers": [ "mlsys.quantization-aware-training", "mlsys.block-scaled-format", "mlsys.weight-only-quantization" ] }, { "id": "finetune-regression", "label": "A fine-tuned model forgets what the base model could do", "signal": "The target task improves after fine-tuning while general benchmarks, safety behavior, or other tasks regress relative to the base checkpoint.", "causes": [ "mlsys.catastrophic-forgetting", "mlsys.supervised-fine-tuning", "mlsys.incremental-training", "mlsys.data-mixture", "mlsys.reward-hacking", "mlsys.chat-template" ], "checks": [ "mlsys.evaluation-harness" ], "levers": [ "mlsys.experience-replay", "mlsys.model-merging", "mlsys.regularization", "mlsys.peft", "mlsys.loss-masking" ] }, { "id": "deploy-incompatibility", "label": "A new model version fails at deployment", "signal": "The release errors on load or returns wrong shapes or garbage because an operator is unsupported by the runtime, the tokenizer or feature schema no longer matches, or a dependency differs from training.", "causes": [ "mlsys.model-feature-compatibility", "mlsys.model-signature", "mlsys.dependency-environment", "mlsys.model-conversion", "mlsys.execution-providers", "mlsys.chat-template" ], "checks": [ "mlsys.integration-testing", "mlsys.release-gate" ], "levers": [ "mlsys.containerization", "mlsys.infrastructure-as-code", "mlsys.model-exchange-format" ] }, { "id": "cluster-queue-wait", "label": "Jobs wait hours in the cluster queue while accelerators sit idle", "signal": "Queued jobs cannot start although the cluster shows free accelerators scattered across nodes, or a large job starves behind many small ones.", "causes": [ "mlsys.gang-scheduling", "mlsys.bin-packing", "mlsys.resource-quota", "mlsys.job-preemption" ], "checks": [ "mlsys.utilization" ], "levers": [ "mlsys.backfill-scheduling", "mlsys.elastic-training", "mlsys.ml-aware-scheduling", "mlsys.dominant-resource-fairness", "mlsys.checkpoint" ] }, { "id": "sim-to-real-gap", "label": "A policy works in simulation and fails on the robot", "signal": "Task success falls on hardware, or errors compound during execution, although simulation and held-out data looked good.", "causes": [ "mlsys.sim-to-real", "mlsys.compounding-error", "mlsys.contact-compliance", "mlsys.sensor-degradation", "mlsys.distribution-shift" ], "checks": [ "mlsys.closed-loop-evaluation", "mlsys.xil-testing" ], "levers": [ "mlsys.domain-randomization", "mlsys.system-identification", "mlsys.interactive-imitation", "mlsys.fleet-data-engine" ] } ], "paths_note": "Learning paths are authored reading orders through subjects for a stated goal. They are suggestions, not prerequisite claims; each step says why it comes where it does.", "learning_paths": [ { "id": "single-machine", "label": "ML systems on one machine", "audience": "A first course in ML systems, following Introduction to Machine Learning Systems.", "steps": [ { "subject": "requirements-slos", "why": "Start from what the system must achieve; every later choice is judged against it." }, { "subject": "data-quality", "why": "Validate the data before training on it." }, { "subject": "architecture-blocks", "why": "Know the layers a model is built from, because their shapes set its compute and memory." }, { "subject": "workload-characterization", "why": "Count the compute, memory, and data a model needs before choosing hardware." }, { "subject": "roofline-locality", "why": "Learn to tell whether a workload is limited by compute or by memory." }, { "subject": "memory-systems", "why": "See why bandwidth, not arithmetic, sets the limit for inference and small batches." }, { "subject": "numerics", "why": "Learn the number formats the hardware supports before trading precision for size." }, { "subject": "accelerator-architecture", "why": "See why accelerators are built the way they are." }, { "subject": "programming-models", "why": "Learn how a framework turns a model into accelerator work." }, { "subject": "data-loading", "why": "Feed training without starving the accelerator." }, { "subject": "training-optimization", "why": "Train a model and see how optimization choices affect speed and stability." }, { "subject": "training-memory", "why": "Fit parameters, optimizer state, and activations into device memory." }, { "subject": "metrics-uncertainty", "why": "Know what quality means before trading it away." }, { "subject": "profiling-utilization", "why": "Measure where time actually goes before optimizing." }, { "subject": "system-benchmarking", "why": "Compare systems under a stated workload and quality target, not a single run." }, { "subject": "model-compression", "why": "Shrink a model to fit a budget." }, { "subject": "serving-architecture", "why": "Serve it to users." }, { "subject": "monitoring", "why": "Keep it working after launch." } ] }, { "id": "fleet", "label": "Scaling to the fleet", "audience": "Engineers who know single-node ML systems and now run training or serving across many machines.", "steps": [ { "subject": "speedup-limits", "why": "Know what adding machines can and cannot buy." }, { "subject": "scaling-laws", "why": "Know what more compute and data buy before paying for a fleet." }, { "subject": "parallelism", "why": "Learn the ways to split a model and its data." }, { "subject": "collectives", "why": "Understand the communication that parallelism creates." }, { "subject": "node-interconnect", "why": "See the fast links inside a node or scale-up domain first." }, { "subject": "network-fabrics", "why": "Then the network that connects nodes." }, { "subject": "parallelism-planning", "why": "Choose and place a parallelism plan for real hardware." }, { "subject": "storage-systems", "why": "Feed data to the fleet and absorb concurrent checkpoint writes." }, { "subject": "failure-semantics", "why": "Specify failure assumptions, retries, and degradation before designing fleet recovery." }, { "subject": "hardware-faults", "why": "Expect the crashes, gray failures, and silent corruption that scale makes routine." }, { "subject": "checkpoint-recovery", "why": "Survive those failures without losing the run." }, { "subject": "cluster-scheduling", "why": "Share the fleet among many jobs." }, { "subject": "queueing-models", "why": "Relate load and variability to waiting time before sizing a service." }, { "subject": "request-scheduling", "why": "Batch, admit, and autoscale requests before sizing the service." }, { "subject": "capacity-provisioning", "why": "Size and share serving capacity across tenants." }, { "subject": "datacenter-facilities", "why": "See the power, cooling, and density limits the building imposes on the fleet." }, { "subject": "energy-carbon", "why": "Account for the power and carbon a fleet consumes." }, { "subject": "cost-economics", "why": "Turn utilization, energy, and hardware life into the cost per useful outcome the fleet exists to lower." } ] }, { "id": "generative-agentic", "label": "Generative and agentic systems", "audience": "Engineers building LLM services and agents, following Agentic Machine Learning Systems.", "steps": [ { "subject": "requirements-slos", "why": "Define task success, latency, spend, and action limits before choosing an architecture." }, { "subject": "llm-inference", "why": "Learn how token generation uses memory and compute." }, { "subject": "fine-tuning", "why": "Specify behavior by prompting, in-context examples, or adaptation before building around the model." }, { "subject": "structured-generation", "why": "Control what the model emits." }, { "subject": "test-time-compute", "why": "Spend extra inference to buy accuracy, and know what it costs." }, { "subject": "retrieval-rag", "why": "Ground answers in external data." }, { "subject": "compound-ai", "why": "Compose several models and tools into one system." }, { "subject": "tool-interfaces", "why": "Let a model act through tools." }, { "subject": "ml-attacks", "why": "Understand prompt injection and jailbreaks before deciding what an agent may reach." }, { "subject": "access-isolation", "why": "Limit what injected instructions and tools can reach." }, { "subject": "agent-harness", "why": "Run the loop that decides each next step." }, { "subject": "context-memory", "why": "Manage what the agent knows from step to step." }, { "subject": "failure-semantics", "why": "Learn retries, idempotency, and in-doubt actions before making runs durable." }, { "subject": "durable-execution", "why": "Make long runs survive failures without repeating effects." }, { "subject": "multi-agent", "why": "Delegate across agents only after a single agent's loop, state, and recovery are solid." }, { "subject": "monitoring", "why": "Trace model calls and tool use in production." }, { "subject": "generative-agent-eval", "why": "Measure whether the agent did the task, against the targets set at the start." }, { "subject": "runtime-enforcement", "why": "Bound what it is allowed to do in production." } ] }, { "id": "physical", "label": "Learned systems that act in the physical world", "audience": "Engineers putting ML on robots and vehicles, following Physical AI.", "steps": [ { "subject": "hazards-operating-domain", "why": "Define where the system may operate and what must never happen, before design choices." }, { "subject": "perception-sensing", "why": "Turn sensor data into an understanding of the scene." }, { "subject": "state-estimation", "why": "Track where the system and the things around it are." }, { "subject": "real-time-systems", "why": "Meet the deadlines that physical action imposes." }, { "subject": "planning-control", "why": "Choose and carry out actions." }, { "subject": "edge-embedded-hardware", "why": "Run the time-critical parts on the machine itself." }, { "subject": "on-device-deployment", "why": "Convert, package, and update models on the machines." }, { "subject": "robot-learning", "why": "Learn policies and close the gap between simulation and reality." }, { "subject": "runtime-enforcement", "why": "Bound what learned components may command." }, { "subject": "verification-validation", "why": "Gather closed-loop and scenario evidence that the system stays within its domain." }, { "subject": "safety-cases", "why": "Argue, with that evidence, that it is safe enough to deploy." } ] }, { "id": "production", "label": "Taking a model into production", "audience": "Model developers and platform engineers moving a trained model into a tested, monitored production service.", "steps": [ { "subject": "requirements-slos", "why": "Define the outcome, acceptance criteria, service targets, and budgets." }, { "subject": "objectives-decisions", "why": "Tie the model's output to the decision it drives and the proxy it optimizes." }, { "subject": "data-quality", "why": "Validate incoming data and keep feature definitions consistent." }, { "subject": "data-leakage-skew", "why": "Rule out leakage and training-serving skew before trusting offline results." }, { "subject": "evaluation-protocols", "why": "Design offline comparisons and online experiments that support the launch decision." }, { "subject": "data-lineage", "why": "Version and trace the data a run consumed so the run can be reproduced and audited." }, { "subject": "pipelines-reproducibility", "why": "Make the model-producing process repeatable." }, { "subject": "ml-testing", "why": "Test data, model, and service integration before users see it." }, { "subject": "model-artifacts", "why": "Export a portable, signed artifact with a declared signature before any server loads it." }, { "subject": "serving-architecture", "why": "Choose how predictions reach consumers." }, { "subject": "release-management", "why": "Roll out gradually with a working rollback." }, { "subject": "monitoring", "why": "Watch service health and quality with production signals." }, { "subject": "incident-response", "why": "Find and contain regressions when they happen." }, { "subject": "drift-retraining", "why": "Decide when new evidence warrants a replacement model." } ] }, { "id": "trustworthy", "label": "Trustworthy, secure, and assured ML systems", "audience": "Engineers and reviewers responsible for robustness, security, privacy, safety, and governance across ML systems.", "steps": [ { "subject": "requirements-slos", "why": "State the security, privacy, safety, fairness, and policy targets the system must satisfy." }, { "subject": "metrics-uncertainty", "why": "Quantify prediction uncertainty and calibration before making high-stakes decisions." }, { "subject": "robustness-slices", "why": "Test performance across subpopulations, distribution shifts, and worst-group slices." }, { "subject": "fairness-bias", "why": "Measure and mitigate disparate impact across protected groups." }, { "subject": "ml-attacks", "why": "Understand poisoning, evasion, extraction, and prompt injection threat models." }, { "subject": "privacy-dp", "why": "Protect training and inference data with minimization and differential privacy." }, { "subject": "model-supply-chain", "why": "Verify weights, dependencies, and provenance across the artifact supply chain." }, { "subject": "access-isolation", "why": "Enforce least privilege and sandboxing around models and tools." }, { "subject": "hazards-operating-domain", "why": "Bound the operating domain and identify system-level hazards." }, { "subject": "runtime-enforcement", "why": "Interpose guardrails, safety filters, and runtime assurance on live outputs." }, { "subject": "transparency-explainability", "why": "Document data and models and provide explanations suitable for audit." }, { "subject": "human-oversight", "why": "Keep human operators able to inspect, override, and contest automated decisions." }, { "subject": "verification-validation", "why": "Produce the scenario, closed-loop, and formal evidence a safety case will cite." }, { "subject": "safety-cases", "why": "Assemble structured arguments and evidence that the deployed system is acceptably safe." }, { "subject": "governance-regulation", "why": "Align lifecycle controls and documentation with organizational and legal regimes." } ] }, { "id": "on-device", "label": "Putting models on devices", "audience": "Engineers deploying models to phones, embedded boards, and microcontrollers under memory, energy, and latency limits.", "steps": [ { "subject": "requirements-slos", "why": "State the latency, memory, and energy budget the device imposes before touching the model." }, { "subject": "workload-characterization", "why": "Count the operations and bytes the model needs per inference." }, { "subject": "edge-embedded-hardware", "why": "Learn what the SoC, NPU, or microcontroller can actually hold and run." }, { "subject": "numerics", "why": "Know the number formats the device supports; they decide what quantization can do." }, { "subject": "profiling-utilization", "why": "Measure latency and memory on the device itself, not a laptop." }, { "subject": "model-compression", "why": "Quantize, prune, and distill until the model fits the budget." }, { "subject": "hardware-aware-design", "why": "Search architectures against measured device cost when compression is not enough." }, { "subject": "ml-compilers", "why": "Lower the model to the target's kernels and operator set." }, { "subject": "on-device-deployment", "why": "Convert, package, run, and update the model on the device fleet." }, { "subject": "chip-power-thermal", "why": "Keep sustained performance within the power and thermal envelope." }, { "subject": "energy-carbon", "why": "Measure energy per inference, the metric that binds on a battery." } ] } ] } }