NCP-GENL Concept Glossary — 52 Terms Defined

Scaled dot-product attentionNCPG-C1

Query, Key, and Value vectors combined as softmax(QKᵀ/√dₖ)·V, where dividing by √dₖ keeps dot products from growing large enough to saturate softmax.

Introduced in Scaled dot-product attention: Query, Key, Value, and the √dₖ scale

Multi-head attention, positional encoding, and layer normalizationNCPG-C2

Parallel attention heads over different representation subspaces (not a parameter-reduction trick), plus the positional signal and normalization that keep a transformer trainable.

Introduced in Multi-head attention, positional encoding, and layer normalization

Architecture familiesNCPG-C3

Encoder-only (bidirectional, MLM), decoder-only (causal, CLM), and encoder-decoder families matched to the training objective and task they fit.

Introduced in Architecture families: encoder-only, decoder-only, and encoder-decoder

Embedding extraction from encoder and decoder modelsNCPG-C4

Extracting and comparing dense vector embeddings from both encoder and decoder models via cosine similarity, requiring query and document embeddings from the same vector space.

Introduced in Embeddings: extraction from encoder and decoder models, and cosine similarity

Output sampling for decodersNCPG-C5

Greedy, beam search, temperature, top-k, and top-p decoding — temperature reshapes the whole distribution while top-k/top-p truncate it; beam search maximizes likelihood, not diversity.

Introduced in Output sampling for decoders: greedy, beam search, temperature, top-k, top-p

In-context learningNCPG-C6

Zero-shot, one-shot, and few-shot prompting with no gradient updates — the name "few-shot learning" invites confusion with fine-tuning despite training no weights.

Introduced in In-context learning: zero-shot, one-shot, and few-shot prompting

Chain-of-thought prompting and prompt templatesNCPG-C7

Step-by-step reasoning prompts that help multi-step tasks but cost tokens and latency, making them a poor fit for simple lookup tasks.

Introduced in Chain-of-thought prompting and prompt templates for small or specialized data

Causal language modeling (CLM)NCPG-C8

The left-to-right, next-token-prediction training objective that makes autoregressive generation possible and that prompt engineering ultimately steers.

Introduced in Causal language modeling: the training objective behind generation

Constrained decoding and validation wrappersNCPG-C9

Decoding-time controls that reduce malformed and hallucinated output without touching model weights or adding knowledge — pair with RAG for the latter.

Introduced in Output control: constrained decoding and validation wrappers

Choosing prompting vs. RAG vs. fine-tuningNCPG-C10

A decision rule under a stated constraint: dynamic or citable knowledge points to RAG, a new skill or style baked in permanently points to fine-tuning, and both should be exhausted first when data or compute is limited.

Introduced in Choosing prompting vs. RAG vs. fine-tuning under a stated constraint

Dataset cleaning and curationNCPG-C11

Deduplication and class-imbalance handling using tools like NeMo Curator and RAPIDS cuDF, with scalers and encoders fit on the training split only.

Introduced in Cleaning and curating a dataset: dedup, imbalance, and NeMo Curator

Dataset organization and formattingNCPG-C12

Correct formats for a given task — JSONL for instruction data, prompt/response pairs for SFT, chunked passages for RAG — with clean, non-leaking splits.

Introduced in Organizing and formatting datasets for pretraining, fine-tuning, and RAG

BPE vs. WordPiece subword tokenizationNCPG-C13

Two subword merge rules: BPE merges the most frequent adjacent pair; WordPiece merges the pair that most increases corpus likelihood. Neither is character- or word-level.

Introduced in Subword tokenization: BPE vs. WordPiece

The vocabulary-size tradeoffNCPG-C14

A bigger vocabulary shortens sequences but enlarges the embedding and softmax tables; because tokenization changes token counts, perplexity is only comparable across models sharing one tokenizer.

Introduced in The vocabulary-size tradeoff and why it caps perplexity comparisons

The five-step EDA checklist before fine-tuningNCPG-C15

Distribution, length, vocabulary, label, and quality checks that catch imbalance, leakage, and truncation risk before they quietly wreck results.

Introduced in Exploratory data analysis before fine-tuning: the five-step checklist

Quantization: PTQ vs. QAT vs. GPTQNCPG-C16

PTQ calibrates scales from observed activations with no retraining; QAT retrains with fake-quant nodes for better low-precision accuracy; GPTQ is a distinct one-shot, post-training, weight-only method using Hessian information — none of them improve accuracy, only protect it while cutting memory and latency.

Introduced in Quantization: PTQ vs QAT vs GPTQ

Knowledge distillation: the DistilBERT trioNCPG-C17

DistilBERT's three measured properties — roughly 40% smaller, 60% faster, and about 97% of BERT's performance retained — distinguishing "40% smaller" from "40% of the size."

Introduced in Knowledge distillation: the DistilBERT trio

Pruning and structured 2:4 sparsityNCPG-C18

Unstructured pruning zeroes individual weights but may not speed up dense hardware; structured 2:4 sparsity maps directly to Tensor Core acceleration and pairs with INT8 in TensorRT.

Introduced in Pruning and structured 2:4 sparsity

KV cachingNCPG-C19

Storing per-token keys and values so decoding does not recompute attention over the whole prefix each step — the primary latency lever for autoregressive inference, spending memory to buy speed.

Introduced in KV caching as the primary latency lever

Streaming attention and TensorRT runtime optimizationNCPG-C20

Sliding-window attention bounding memory for long sequences, and TensorRT compiling, fusing kernels, and auto-tuning for a target GPU — optimization, distinct from Triton serving.

Introduced in Streaming attention and TensorRT runtime optimization

Encoder foundation models and masked language modelingNCPG-C21

Training encoder-based foundation models with MLM, alongside the quantization, distillation, and pruning levers that shrink them for deployment.

Introduced in Encoder foundation models and masked language modeling

Parameter-efficient fine-tuning: LoRA, adapters, P-tuningNCPG-C22

LoRA freezing base weights and training small low-rank matrices that merge back in with no added inference latency, unlike bottleneck adapters — PEFT never shrinks the base model.

Introduced in Parameter-efficient fine-tuning: LoRA, adapters, and P-tuning

Alignment: SFT, RLHF, DPO, and GRPONCPG-C23

Classic RLHF trains a separate reward model and PPO critic; DPO drops the reward model and optimizes preference pairs directly; GRPO keeps a reward signal but drops the critic, using group-relative scores instead.

Introduced in Alignment: SFT, RLHF, DPO, and GRPO

Contrastive loss for embeddingsNCPG-C24

A loss pulling similar pairs together and pushing dissimilar pairs apart in embedding space — the basis for retrieval and semantic-search models feeding RAG.

Introduced in Contrastive loss for embeddings

Early stopping and fine-tuning impact assessmentNCPG-C25

Capturing the best-generalizing checkpoint before the validation optimum is passed, and measuring impact against a stated before/after baseline.

Introduced in Early stopping and fine-tuning impact assessment

When to fine-tune vs. prompt or RAGNCPG-C26

Fine-tuning for a new skill, style, or behavior baked in permanently; prompting or RAG when knowledge is dynamic, must be cited, or data and compute are limited.

Introduced in When to fine-tune vs. prompt or RAG

PerplexityNCPG-C27

A metric where lower is better, undefined for masked LMs like BERT, and comparable only across models sharing the same tokenizer.

Introduced in Perplexity: what it measures and where it does not apply

BLEU vs. ROUGE vs. METEORNCPG-C28

BLEU is precision-oriented (built for translation); ROUGE is recall-oriented (built for summarization); METEOR adds stem and synonym awareness beyond raw n-gram overlap.

Introduced in BLEU vs. ROUGE vs. METEOR

LLM-as-a-judge and error analysisNCPG-C29

Combining automatic scores with judge or human review and systematic error-mode categorization to catch fluent-but-wrong answers that surface-overlap metrics miss.

Introduced in LLM-as-a-judge, human-in-the-loop review, and error analysis

Benchmarking and scalable evaluation frameworksNCPG-C30

Standardized metrics enabling cross-platform comparisons, with tools like NeMo Evaluator packaging benchmark harnesses, LLM-as-a-judge scoring, and RAG/agent metrics into one service.

Introduced in Benchmarking and scalable evaluation frameworks

RAG evaluation metricsNCPG-C31

Faithfulness (grounding), answer relevancy, context precision, and context recall — splitting quality across the retrieval/generation boundary so scoring only the final answer does not hide which stage is weak.

Introduced in RAG evaluation: faithfulness, answer relevancy, context precision, and context recall

The parallelism familiesNCPG-C32

Data, tensor, pipeline, sequence, context, and expert parallelism combining to scale training from billions to trillions of parameters, each with its own precondition.

Introduced in The parallelism families: data, tensor, pipeline, sequence, context, and expert

Tensor Parallelism vs. Pipeline ParallelismNCPG-C33

Tensor parallelism splits within a single layer (intra-layer); pipeline parallelism splits across consecutive layers (inter-layer) — the domain's most common distractor pair.

Introduced in Tensor Parallelism vs Pipeline Parallelism: the #1 distractor pair

Memory sharding: FSDP and ZeRONCPG-C34

Sharding optimizer state — and optionally parameters and gradients — across data-parallel ranks; a memory technique layered on data parallelism, not a separate parallelism axis.

Introduced in Memory sharding: FSDP and ZeRO

Mixed precision and Tensor CoresNCPG-C35

FP16/BF16 compute on Tensor Cores as the default fast path on modern GPUs, where FP16 risks gradient underflow without loss scaling and BF16's wider exponent range often avoids needing it.

Introduced in Mixed precision and Tensor Cores

Gradient accumulation and effective batch sizeNCPG-C36

Per-device batch size times accumulation steps, letting a large effective batch fit under a memory cap without reducing total compute.

Introduced in Gradient accumulation and effective batch size

Profiling with NsightNCPG-C37

Locating kernel occupancy, memory-bound versus compute-bound kernels, and CUDA memory issues before changing batch size, precision, or parallelism configuration.

Introduced in Profiling and troubleshooting with Nsight

Dynamic vs. sequence batchingNCPG-C38

Dynamic batching combines independent requests at runtime for stateless models; sequence batching routes a stateful sequence's requests to the same model instance — applying the former to the latter is the standing trap.

Introduced in Dynamic batching vs. sequence batching in Dynamo-Triton

NVIDIA NIM as a prepackaged microserviceNCPG-C39

NIM LLM 2.0's "one container, one backend" design with vLLM as the inference engine, shipping validated containers with OpenAI-compatible endpoints — a higher-level microservice above the general Triton server.

Introduced in NVIDIA NIM as an LLM-serving container: model formats and the TensorRT-LLM backend

Concurrent model execution and instance groupsNCPG-C40

Running several copies of the same model, or several different models, in parallel on one system including a single GPU — complementary to, not the same mechanism as, dynamic batching.

Introduced in Concurrent model execution and instance groups

Containerization, Kubernetes, and MIGNCPG-C41

Docker packaging model plus runtime for reproducible deployment, Kubernetes scaling and health-checking it, and Multi-Instance GPU partitioning a single GPU for multi-tenant serving.

Introduced in Containerization, Kubernetes, and Multi-Instance GPU

Model-type compute tradeoffsNCPG-C42

Decoder generation latency scaling with output length because decoding is sequential — unlike a single-pass encoder, bigger batches alone cannot fix it; KV caching is the main mitigation.

Introduced in Model-type compute tradeoffs: encoder vs. decoder vs. encoder-decoder

Reliability and performance metricsNCPG-C43

Latency percentiles (p95, p99), throughput, and error rate, chosen because reporting only average latency hides the tail problems users actually feel.

Introduced in Reliability and performance metrics: latency percentiles, throughput, error rate

Logging and root-cause analysisNCPG-C44

Capturing request/response logs and alerting on latency spikes, error surges, and quality drops — turning an anomaly into a diagnosed root cause.

Introduced in Logging, anomalies, and root-cause analysis

Data drift and concept driftNCPG-C45

A model that passed evaluation at launch drifting in production as real-world inputs shift away from the training distribution — launch evaluation is not a standing guarantee.

Introduced in Drift: data drift and concept drift

Regression benchmarking against prior versionsNCPG-C46

Continuously comparing a deployment against previous versions on a fixed suite, treating evaluation as ongoing rather than a one-time launch gate.

Introduced in Offline eval-set regression: benchmarking a deployment against prior versions

Automated tuning, retraining, and versioningNCPG-C47

Automated (not manual) retraining and tuning, with versioning providing traceability and safe rollback.

Introduced in Automated model retraining, tuning, and versioning

NeMo Guardrails' three rail categoriesNCPG-C48

Topical, safety/content, and security (anti-injection) rails governing both inputs and outputs — three distinct jobs, not one blanket filter.

Introduced in Guardrails: topical, safety/content, and security rails

Measuring guardrail effectivenessNCPG-C49

Evaluating a guardrail's coverage alongside its latency and false-positive-rate cost, not coverage alone.

Introduced in Measuring guardrail effectiveness and its latency/false-positive tradeoffs

Disaggregated bias and fairness auditingNCPG-C50

Per-group evaluation as the only way to detect disparate impact a high overall accuracy number can mask — mitigation happens at the data and training level, not inside a guardrail.

Introduced in Bias and fairness auditing: disaggregated evaluation

Hallucination mitigationNCPG-C51

RAG grounding, constrained decoding, and trustworthiness checks that reduce but do not eliminate hallucination, since a model can still misuse or ignore given context.

Introduced in Hallucination mitigation: RAG grounding, constrained decoding, trustworthiness checks

Responsible-AI monitoring and complianceNCPG-C52

Ongoing compliance monitoring for toxicity and PII leakage, extending production monitoring to ethical and regulatory signals across the whole lifecycle.

Introduced in Responsible-AI monitoring and compliance