| 中文 | 英文 | 缩写 |
|---|---|---|
| $n$元语言模型 | $n$-gram Language Model | |
| 四比特正态浮点 | 4-bit NormalFloat | NF4 |
| 八位浮点 | 8-bit Floating Point | FP8 |
| 绝对位置嵌入 | Absolute Positional Embedding | |
| 拒答 | Abstention | |
| 准确率 | Accuracy | |
| 声学词元 | Acoustic Token | |
| 动作价值 | Action Value | |
| 激活重计算 | Activation Recomputation | |
| 激活感知权重量化 | Activation-Aware Weight Quantization | AWQ |
| 激活加权低秩近似 | Activation-weighted Low-rank Approximation | |
| 行动者—评论家 | Actor--Critic | |
| 适配器注册表 | Adapter Registry | |
| 适配器路由 | Adapter Routing | |
| 自适应低秩适配 | Adaptive Low-Rank Adaptation | AdaLoRA |
| 自适应矩估计 | Adaptive Moment Estimation | Adam |
| 准入控制 | Admission Control | |
| 优势函数 | Advantage Function | |
| 仿射量化 | Affine Quantization | |
| 仿射变换 | Affine Transformation | |
| 智能体 | Agent | |
| 智能体执行与控制系统 | Agent Harness | |
| 老化 | Aging | |
| 全收集 | All-Gather | |
| 全归约 | All-Reduce | |
| 全互换 | All-to-All | |
| 全互换通信 | All-to-All | |
| 美国数学邀请赛 | American Invitational Mathematics Examination | AIME |
| 答案规范化 | Answer Canonicalization | |
| 近似最近邻搜索 | Approximate Nearest Neighbor Search | ANN |
| 算术强度 | Arithmetic Intensity | |
| 制品完整性 | Artifact Integrity | |
| 保障论证 | Assurance Case | |
| 非对称距离计算 | Asymmetric Distance Computation | ADC |
| 注意力机制 | Attention | |
| 注意力蒸馏 | Attention Distillation | |
| 注意力掩码 | Attention Mask | |
| 自回归 | Autoregression | |
| 自回归空白填充 | Autoregressive Blank Infilling | |
| 无辅助损失负载均衡 | Auxiliary-Loss-Free Load Balancing | |
| 平均精确率 | Average Precision | AP |
| 后门 | Backdoor | |
| 背压 | Backpressure | |
| 束搜索 | Beam Search | |
| 序列开始符 | Beginning of Sequence | BOS |
| 旧策略 | Behavior Policy | |
| 贝尔曼方程 | Bellman Equation | |
| 基准测试 | Benchmark | |
| 最佳匹配评分 | Best Matching 25 | BM25 |
| 多候选择优 | Best-of-$K$ Selection | Best-of-$K$ |
| 多候选择优 | Best-of-N | |
| 基于 Transformer 的双向编码表示 | Bidirectional Encoder Representations from Transformers | BERT |
| 双线性插值 | Bilinear Interpolation | |
| 双语评估替代指标 | Bilingual Evaluation Understudy | BLEU |
| 割带宽 | Bisection Bandwidth | |
| 偏置微调 | BitFit | |
| 蓝绿部署 | Blue–Green Deployment | |
| 自助重采样 | Bootstrap | |
| 自举 | Bootstrapping | |
| 瓶颈适配器 | Bottleneck Adapter | |
| 脑浮点16 | Brain Floating Point 16 | BF16 |
| 缓存淘汰 | Cache Eviction | |
| 金丝雀发布 | Canary Release | |
| 容量因子 | Capacity Factor | |
| 灾难性遗忘 | Catastrophic Forgetting | |
| 因果注意力 | Causal Attention | |
| 因果编码器—解码器 | Causal Encoder-Decoder | CED |
| 因果掩码 | Causal Mask | |
| 中央处理器 | Central Processing Unit | CPU |
| 思维链 | Chain of Thought | CoT |
| 思维链提示 | Chain-of-Thought Prompting | CoT Prompting |
| 变更数据捕获 | Change Data Capture | CDC |
| 会话模板 | Chat Template | |
| 检查点 | Checkpoint | |
| 中文大规模多任务语言理解 | Chinese Massive Multitask Language Understanding | CMMLU |
| Chrome 开发者工具协议 | Chrome DevTools Protocol | CDP |
| 分块预填充 | Chunked Prefill | |
| 分类标记 | Classification Token | CLS |
| 无分类器引导 | Classifier-Free Guidance | CFG |
| 干净样本预测 | Clean Sample Prediction | |
| 裁剪 | Clipping | |
| 闭合环负载 | Closed-Loop Load | |
| 簇重采样 | Cluster Bootstrap | |
| 合并访存 | Coalesced Memory Access | |
| 科恩一致性系数 | Cohen's Kappa | $\kappa$ |
| 集合通信 | Collective Communication | |
| 合并 | Combine | |
| 比较并交换 | Compare-and-Swap | CAS |
| 补偿事务 | Compensating Transaction | |
| 压缩稀疏注意力第二代 | Compressed Sparse Attention 2 | CSA2 |
| 计算机操作 | Computer Use | |
| 条件计算 | Conditional Computation | |
| 条件独立 | Conditional Independence | |
| 条件记忆 | Conditional Memory | |
| 置信区间 | Confidence Interval | CI |
| 混淆矩阵 | Confusion Matrix | |
| 连接器 | Connector | |
| 一致性 | Consistency | |
| 包含率 | Containment | |
| 上下文压缩 | Context Compression | |
| 上下文并行 | Context Parallelism | CP |
| 上下文窗口 | Context Window | |
| 上下文校准 | Contextual Calibration | |
| 继续预训练 | Continued Pretraining | CPT |
| 连续批处理 | Continuous Batching | |
| 对比语言图像预训练 | Contrastive Language-Image Pre-training | CLIP |
| 对比学习 | Contrastive Learning | |
| 控制面 | Control Plane | |
| 协调遗漏 | Coordinated Omission | |
| 写时复制 | Copy-on-Write | CoW |
| 余弦衰减 | Cosine Decay | |
| 余弦相似度 | Cosine Similarity | |
| 反事实评估 | Counterfactual Evaluation | |
| 信用分配 | Credit Assignment | |
| 交叉编码器 | Cross Encoder | |
| 交叉注意力 | Cross-Attention | |
| 交叉注意力融合 | Cross-attention Fusion | |
| 交叉编码器 | Cross-encoder | |
| 交叉熵 | Cross-Entropy | CE |
| 跨模态对齐 | Cross-modal Alignment | |
| 课程学习 | Curriculum Learning | |
| 数据增强 | Data Augmentation | |
| 数据契约 | Data Contract | |
| 数据工程 | Data Engineering | |
| 数据新鲜度 | Data Freshness | |
| 数据血缘 | Data Lineage | |
| 数据防泄漏 | Data Loss Prevention | DLP |
| 数据面 | Data Plane | |
| 数据投毒 | Data Poisoning | |
| 截止期限 | Deadline | |
| 增量解码 | Decode | |
| 仅解码器 | Decoder-only | |
| 解耦权重衰减 | Decoupled Weight Decay | |
| 纵深防御 | Defense in Depth | |
| 删除血缘 | Deletion Lineage | |
| 去噪扩散隐式模型 | Denoising Diffusion Implicit Model | DDIM |
| 去噪扩散概率模型 | Denoising Diffusion Probabilistic Model | DDPM |
| 稠密专家混合 | Dense Mixture of Experts | |
| 稠密模型 | Dense Model | |
| 稠密检索 | Dense Retrieval | |
| 反量化 | Dequantization | |
| 设计效应 | Design Effect | |
| 确定性变换 | Deterministic Transformation | |
| 设备亲和性 | Device Affinity | |
| 差分隐私 | Differential Privacy | DP |
| 扩散模型 | Diffusion Model | |
| 扩散 Transformer | Diffusion Transformer | DiT |
| 数字签名 | Digital Signature | |
| 直接内存访问 | Direct Memory Access | DMA |
| 直接偏好优化 | Direct Preference Optimization | DPO |
| 直接提示注入 | Direct Prompt Injection | |
| 折损累计增益 | Discounted Cumulative Gain | DCG |
| 分发 | Dispatch | |
| 分布式数据并行 | Distributed Data Parallel | DDP |
| 领域本体 | Domain Ontology | |
| 双重量化 | Double Quantization | |
| 排空 | Draining | |
| 随机失活 | Dropout | |
| 双编码器 | Dual Encoder | |
| 持久性 | Durability | |
| 持久执行 | Durable Execution | |
| 动态填充 | Dynamic Padding | |
| 动态量化 | Dynamic Quantization | |
| 动态分辨率 | Dynamic Resolution | |
| 嵌入 | Embedding | |
| 涌现能力 | Emergent Abilities | |
| 编码器—解码器 | Encoder--Decoder | |
| 仅编码器 | Encoder-only | |
| 序列结束符 | End of Sequence | EOS |
| 企业资源计划 | Enterprise Resource Planning | ERP |
| 熵 | Entropy | |
| 情景记忆 | Episodic Memory | |
| 待估目标量 | Estimand | |
| 欧氏距离 | Euclidean Distance | |
| 评测污染 | Evaluation Contamination | |
| 评估污染 | Evaluation Contamination | |
| 评测运行框架 | Evaluation Harness | |
| 证据组装 | Evidence Assembly | |
| 证据忠实度 | Evidence Faithfulness | |
| 证据下界 | Evidence Lower Bound | ELBO |
| 精确去重 | Exact Deduplication | |
| 完全匹配 | Exact Match | EM |
| 精确最近邻搜索 | Exact Nearest Neighbor Search | |
| 恰好一次 | Exactly-Once | |
| 示例库 | Example Bank | |
| 可执行验证器 | Executable Verifier | |
| 执行图回放 | Execution Graph Replay | |
| 期望校准误差 | Expected Calibration Error | ECE |
| 专家并行 | Expert Parallelism | EP |
| 显式拥塞通知 | Explicit Congestion Notification | ECN |
| 暴露偏差 | Exposure Bias | |
| F1 分数 | F1 Score | |
| 故障域 | Failure Domain | |
| 伪量化 | Fake Quantization | |
| 漏放 | False Negative | |
| 假负例 | False Negative | |
| 误拦 | False Positive | |
| 逐位置前馈网络 | Feed-Forward Network | FFN |
| 隔离令牌 | Fencing Token | |
| 少样本提示 | Few-Shot Prompting | |
| 中间填充 | Fill-in-the-Middle | FIM |
| 细粒度专家 | Fine-Grained Expert | |
| 先到先服务 | First-Come, First-Served | FCFS |
| 浮点操作数 | Floating-point Operations | FLOPs |
| 流匹配 | Flow Matching | |
| 前向边缘分布 | Forward Marginal Distribution | |
| 前向过程 | Forward Process | |
| 新鲜度 | Freshness | |
| 全参数微调 | Full Fine-Tuning | FFT |
| 完全分片数据并行 | Fully Sharded Data Parallel | FSDP |
| 融合反量化 | Fused Dequantization | |
| 成组调度 | Gang Scheduling | |
| 门控线性单元 | Gated Linear Unit | GLU |
| 门控网络 | Gating Network | |
| 高斯误差线性单元 | Gaussian Error Linear Unit | GELU |
| 高斯误差线性门控单元 | GELU-Gated Linear Unit | GEGLU |
| 广义优势估计 | Generalized Advantage Estimation | GAE |
| 生成式预训练 Transformer | Generative Pre-trained Transformer | GPT |
| 全局平均池化 | Global Average Pooling | GAP |
| 有效吞吐 | Goodput | |
| 生成式模型训练后量化 | GPTQ | |
| 梯度累积 | Gradient Accumulation | |
| 梯度范数裁剪 | Gradient Norm Clipping | |
| 研究生级难检索问答 | Graduate-Level Google-Proof Question Answering | GPQA |
| 图编译 | Graph Compilation | |
| 图形处理器 | Graphics Processing Unit | GPU |
| 贪心解码 | Greedy Decoding | |
| 动作定位 | Grounding | |
| 组相对策略优化 | Group Relative Policy Optimization | GRPO |
| 分组划分 | Group Split | |
| 分组量化 | Group-Wise Quantization | |
| 分组查询注意力 | Grouped-Query Attention | GQA |
| 半精度浮点 | Half Precision Floating Point | FP16 |
| 幻觉 | Hallucination | |
| 硬标签 | Hard Label | |
| 难负例 | Hard Negative | |
| 队首阻塞 | Head-of-Line Blocking | HOL Blocking |
| 海森矩阵 | Hessian Matrix | |
| 分层可导航小世界图 | Hierarchical Navigable Small World | HNSW |
| 高带宽内存 | High Bandwidth Memory | HBM |
| 命中率 | Hit Rate | Hit@K |
| 人类最后的考试 | Humanity's Last Exam | HLE |
| 混合检索 | Hybrid Retrieval | |
| 幂等性 | Idempotency | |
| 幂等键 | Idempotency Key | |
| 不可变制品 | Immutable Artifact | |
| 重要性采样 | Importance Sampling | |
| 批内负例 | In-Batch Negative | |
| 上下文学习 | In-Context Learning | ICL |
| 事件响应 | Incident Response | |
| 增量解码 | Incremental Decoding | |
| 索引快照 | Index Snapshot | |
| 间接提示注入 | Indirect Prompt Injection | |
| 归纳偏置 | Inductive Bias | |
| InfiniBand 网络 | InfiniBand | IB |
| 信息熵 | Information Entropy | |
| 信息检索 | Information Retrieval | IR |
| 内激活抑制与放大适配 | Infused Adapter by Inhibiting and Amplifying Inner Activations | IA$^3$ |
| 每秒输入输出操作数 | Input/Output Operations Per Second | IOPS |
| 指令微调 | Instruction Tuning | |
| 指令遵循评估 | Instruction-Following Evaluation | IFEval |
| 探索不足 | Insufficient Exploration | |
| 意向处理分析 | Intention-to-Treat Analysis | ITT |
| 词元间时延 | Inter-Token Latency | ITL |
| 中间表示蒸馏 | Intermediate Representation Distillation | |
| 逆文档频率 | Inverse Document Frequency | IDF |
| 倒排文件向量索引 | Inverted File Index | IVF |
| 倒排索引 | Inverted Index | |
| 迭代检索 | Iterative Retrieval | |
| Jaccard 相似度 | Jaccard Similarity | |
| 雅可比矩阵 | Jacobian Matrix | |
| 越狱 | Jailbreak | |
| 键 | Key | K |
| 键值缓存 | Key--Value Cache | KV Cache |
| 键值缓存 | Key-Value Cache | KV Cache |
| 键值缓存量化 | Key-Value Cache Quantization | |
| 知识蒸馏 | Knowledge Distillation | KD |
| 库尔贝克—莱布勒散度 | Kullback--Leibler Divergence | KL |
| 标签基数 | Label Cardinality | |
| 潜空间扩散 | Latent Diffusion | |
| 潜变量 | Latent Variable | |
| 层归一化 | Layer Normalization | LayerNorm |
| 学习率预热 | Learning Rate Warmup | |
| 学习排序 | Learning to Rank | LTR |
| 最小权限 | Least Privilege | |
| 留一基线 | Leave-one-out Baseline | LOO Baseline |
| 长度分桶 | Length Bucketing | |
| 线性探测 | Linear Probing | |
| 存活探针 | Liveness Probe | |
| 模型评审 | LLM-as-a-Judge | |
| 局部敏感哈希 | Locality-Sensitive Hashing | LSH |
| 未归一化分数 | Logits | |
| 长短期记忆 | Long Short-Term Memory | LSTM |
| 长期记忆 | Long-Term Memory | |
| 损失掩码 | Loss Mask | |
| 损失缩放 | Loss Scaling | |
| 中间信息利用退化 | Lost in the Middle | |
| 低精度训练 | Low-Precision Training | |
| 低秩适配 | Low-Rank Adaptation | LoRA |
| 低秩分解 | Low-rank Factorization | |
| 宏平均 | Macro Average | |
| 幅值剪枝 | Magnitude Pruning | |
| 发布清单 | Manifest | |
| 制造执行系统 | Manufacturing Execution System | MES |
| 边际收益 | Marginal Utility | |
| 马尔可夫决策过程 | Markov Decision Process | MDP |
| 掩码语言建模 | Masked Language Modeling | MLM |
| 大规模多学科多模态理解 | Massive Multi-discipline Multimodal Understanding | MMMU |
| 大规模多任务语言理解 | Massive Multitask Language Understanding | MMLU |
| 矩阵吸收 | Matrix Absorption | |
| 最大内积搜索 | Maximum Inner Product Search | MIPS |
| 最大似然估计 | Maximum Likelihood Estimation | MLE |
| 平均倒数排名 | Mean Reciprocal Rank | MRR |
| 成员推断 | Membership Inference | |
| 元数据 | Metadata | |
| 元数据过滤 | Metadata Filtering | |
| 微平均 | Micro Average | |
| 最小哈希 | MinHash | |
| 混合精度 | Mixed Precision | |
| 混合专家模型 | Mixture of Experts | MoE |
| 混合增强 | Mixup | |
| 模态 | Modality | |
| 模型适配器 | Model Adapter | |
| 模型制品 | Model Artifact | |
| 级联调用 | Model Cascade | |
| 模型压缩 | Model Compression | |
| 模型上下文协议 | Model Context Protocol | MCP |
| 模型提取 | Model Extraction | |
| 模型路由 | Model Routing | |
| 动量 | Momentum | |
| 蒙特卡洛估计 | Monte Carlo Estimation | MC |
| 多智能体系统 | Multi-Agent System | MAS |
| 多头注意力 | Multi-Head Attention | MHA |
| 多头潜在注意力 | Multi-Head Latent Attention | MLA |
| 多头潜在注意力 | Multi-head Latent Attention | MLA |
| 多查询注意力 | Multi-Query Attention | MQA |
| 多轮共指消解 | Multi-Round Coreference Resolution | MRCR |
| 多租户隔离 | Multi-tenant Isolation | |
| 多词元预测 | Multi-Token Prediction | MTP |
| N-gram 语言模型 | N-gram Language Model | |
| 命名实体识别 | Named Entity Recognition | NER |
| 负缓存 | Negative Cache | |
| 下一句预测 | Next Sentence Prediction | NSP |
| 噪声预测 | Noise Prediction | |
| 非一致内存访问 | Non-Uniform Memory Access | NUMA |
| 非易失性存储器快速接口 | Non-Volatile Memory Express | NVMe |
| 归一化折损累计增益 | Normalized Discounted Cumulative Gain | nDCG |
| 核采样 | Nucleus Sampling | Top-p |
| 可观测性 | Observability | |
| 观测 | Observation | |
| 占用率 | Occupancy | |
| 离策略学习 | Off-Policy Learning | |
| 卸载 | Offload | |
| 在策略学习 | On-Policy Learning | |
| 一前向一反向调度 | One-Forward-One-Backward | 1F1B |
| 开放权重 | Open Weights | |
| 开放环负载 | Open-Loop Load | |
| 算子融合 | Operator Fusion | |
| 结果奖励模型 | Outcome Reward Model | ORM |
| 结果监督 | Outcome Supervision | |
| 超售比 | Oversubscription Ratio | |
| 连续提示调优 | P-Tuning | |
| 深层提示调优 | P-Tuning v2 | |
| 分页键值缓存 | Paged KV Cache | |
| 分页优化器 | Paged Optimizer | |
| 参数高效微调 | Parameter-Efficient Fine-Tuning | PEFT |
| 帕累托前沿 | Pareto Frontier | |
| 解析失败 | Parse Failure | |
| 部分可观测马尔可夫决策过程 | Partially Observable Markov Decision Process | POMDP |
| 前 $k$ 次采样通过率 | Pass at k | pass@k |
| 补丁 | Patch | |
| 补丁嵌入 | Patch Embedding | |
| 补丁化 | Patchification | |
| 逐通道量化 | Per-Channel Quantization | |
| 逐张量量化 | Per-Tensor Quantization | |
| 高速串行计算机扩展总线 | Peripheral Component Interconnect Express | PCIe |
| 排列等变性 | Permutation Equivariance | |
| 排列不变性 | Permutation Invariance | |
| 置换检验 | Permutation Test | |
| 困惑度 | Perplexity | PPL |
| 个人可识别信息 | Personally Identifiable Information | PII |
| 页锁定内存 | Pinned Memory | |
| 流水并行 | Pipeline Parallelism | PP |
| 规划执行架构 | Plan-and-Execute Architecture | |
| 策略滞后 | Policy Lag | |
| 位置插值 | Position Interpolation | PI |
| 后置层归一化 | Post-Layer Normalization | Post-LN |
| 训练后量化 | Post-Training Quantization | PTQ |
| 倒排列表 | Posting List | |
| 电能使用效率 | Power Usage Effectiveness | PUE |
| 前置层归一化 | Pre-Layer Normalization | Pre-LN |
| 精确率 | Precision | |
| 精确率—召回率曲线 | Precision--Recall Curve | PR |
| 抢占 | Preemption | |
| 偏好对 | Preference Pair | |
| 预填充 | Prefill | |
| 预填充与解码分离 | Prefill/Decode Disaggregation | P/D 分离 |
| 预填充—解码分离 | Prefill–Decode Disaggregation | P/D 分离 |
| 前缀缓存 | Prefix Cache | |
| 前缀解码器 | Prefix Decoder | |
| 前缀融合 | Prefix Fusion | |
| 前缀微调 | Prefix Tuning | |
| 预训练 | Pretraining | |
| 基于优先级的流量控制 | Priority-based Flow Control | PFC |
| 隐私 | Privacy | |
| 概率校准 | Probability Calibration | |
| 概率密度函数 | Probability Density Function | |
| 概率质量函数 | Probability Mass Function | PMF |
| 程序性记忆 | Procedural Memory | |
| 过程奖励模型 | Process Reward Model | PRM |
| 过程监督 | Process Supervision | |
| 乘积量化 | Product Quantization | PQ |
| 投影器 | Projector | |
| 提示注入 | Prompt Injection | |
| 提示微调 | Prompt Tuning | |
| 来源跨度 | Provenance Span | |
| 近端策略优化 | Proximal Policy Optimization | PPO |
| 剪枝 | Pruning | |
| 量化校准 | Quantization Calibration | |
| 量化尺度 | Quantization Scale | |
| 量化感知训练 | Quantization-Aware Training | QAT |
| 量化低秩适配 | Quantized Low-Rank Adaptation | QLoRA |
| 查询 | Query | Q |
| 查询分解 | Query Decomposition | |
| 查询改写 | Query Rewriting | |
| 查询变换器 | Querying Transformer | Q-Former |
| 序列级 RAG | RAG-Sequence | |
| 词元级 RAG | RAG-Token | |
| 随机变量 | Random Variable | |
| 随机对照实验 | Randomized Controlled Experiment | |
| 进程序号 | Rank | |
| 融合以太网远程直接内存访问 | RDMA over Converged Ethernet | RoCE |
| 重排 | Re-ranking | |
| 反应式架构 | Reactive Architecture | |
| 就绪探针 | Readiness Probe | |
| 推理与行动交替 | Reasoning and Acting | ReAct |
| 推理蒸馏 | Reasoning Distillation | |
| 推理强度 | Reasoning Effort | |
| 召回率 | Recall | |
| 面向召回的摘要评估指标 | Recall-Oriented Understudy for Gisting Evaluation | ROUGE |
| 受试者工作特征曲线 | Receiver Operating Characteristic | ROC |
| 倒数排名融合 | Reciprocal Rank Fusion | RRF |
| 核对 | Reconciliation | |
| 恢复训练 | Recovery Training | |
| 整流线性单元 | Rectified Linear Unit | ReLU |
| 循环神经网络 | Recurrent Neural Network | RNN |
| 红队评估 | Red Teaming | |
| 归约散播 | Reduce-Scatter | |
| 引用计数 | Reference Counting | |
| 参考策略 | Reference Policy | |
| 强化学习 | Reinforcement Learning | RL |
| 基于人类反馈的强化学习 | Reinforcement Learning from Human Feedback | RLHF |
| 可验证奖励强化学习 | Reinforcement Learning with Verifiable Rewards | RLVR |
| 拒绝采样微调 | Rejection Sampling Fine-Tuning | RFT |
| 可靠性 | Reliability | |
| 远程直接内存访问 | Remote Direct Memory Access | RDMA |
| 回放数据 | Replay Data | |
| 再量化 | Requantization | |
| 重排序 | Reranking | |
| 重采样器 | Resampler | |
| 检索增强生成 | Retrieval-Augmented Generation | RAG |
| 反向条件后验 | Reverse Conditional Posterior | |
| 反向过程 | Reverse Process | |
| 奖励投机 | Reward Hacking | |
| 环形缓存 | Ring Buffer | |
| 屋顶线模型 | Roofline Model | |
| 均方根归一化 | Root Mean Square Normalization | RMSNorm |
| 旋转位置编码 | Rotary Position Embedding | RoPE |
| 最近舍入 | Round to Nearest | RTN |
| 路由器 | Router | |
| 评分量表 | Rubric | |
| 行为安全 | Safety | |
| 采样率 | Sampling Rate | |
| 样本单位 | Sampling Unit | |
| 沙箱验证 | Sandbox Verification | |
| 饱和误差 | Saturation Error | |
| 缩放点积注意力 | Scaled Dot-Product Attention | |
| 规模定律 | Scaling Law | |
| 结构模式 | Schema | |
| 数据模式 | Schema | |
| 模式适配器 | Schema Adapter | |
| 得分函数 | Score Function | |
| 得分函数估计 | Score-Function Estimation | |
| 安全防护 | Security | |
| 自注意力 | Self-Attention | |
| 自一致性 | Self-Consistency | |
| 自监督学习 | Self-Supervised Learning | |
| 语义切分 | Semantic Chunking | |
| 语义记忆 | Semantic Memory | |
| 半结构化稀疏 | Semi-structured Sparsity | |
| 序列打包 | Sequence Packing | |
| 序列装箱 | Sequence Packing | |
| 序列并行 | Sequence Parallelism | SP |
| 序列级蒸馏 | Sequence-level Distillation | |
| 服务端发送事件 | Server-Sent Events | SSE |
| 服务降级 | Service Degradation | |
| 影子流量 | Shadow Traffic | |
| 形状分桶 | Shape Bucketing | |
| 分片 | Shard | |
| 共享专家 | Shared Expert | |
| 连续片段 | Shingle | |
| 短时傅里叶变换 | Short-Time Fourier Transform | STFT |
| 副作用 | Side Effect | |
| Sigmoid线性单元 | Sigmoid Linear Unit | SiLU |
| 信噪比 | Signal-to-Noise Ratio | SNR |
| 简化噪声损失 | Simplified Noise Loss | |
| 单指令多线程 | Single Instruction, Multiple Threads | SIMT |
| 奇异值分解 | Singular Value Decomposition | SVD |
| 正弦位置编码 | Sinusoidal Positional Encoding | |
| 技能 | Skill | |
| 切片评估 | Slice Evaluation | |
| 滑动窗口注意力 | Sliding-Window Attention | |
| 平滑量化 | SmoothQuant | |
| 快照清单 | Snapshot Manifest | |
| 软标签 | Soft Label | |
| 软专家混合 | Soft Mixture of Experts | Soft MoE |
| 软提示 | Soft Prompt | |
| 归一化指数函数 | Softmax | |
| 软件物料清单 | Software Bill of Materials | SBOM |
| 稀疏注意力 | Sparse Attention | |
| 稀疏专家混合 | Sparse Mixture of Experts | |
| 稀疏专家模型 | Sparse Mixture-of-Experts Model | MoE |
| 稀疏检索 | Sparse Retrieval | |
| 稀疏奖励 | Sparse Reward | |
| 稀疏率 | Sparsity | |
| 推测解码 | Speculative Decoding | |
| 启动探针 | Startup Probe | |
| 启动风暴 | Startup Storm | |
| 状态机 | State Machine | |
| 状态价值 | State Value | |
| 静态批处理 | Static Batching | |
| 静态量化 | Static Quantization | |
| 直通估计器 | Straight-Through Estimator | STE |
| 流式读取 | Streaming | |
| 流式因果性 | Streaming Causality | |
| 流式多处理器 | Streaming Multiprocessor | SM |
| 强扩展 | Strong Scaling | |
| 结构化剪枝 | Structured Pruning | |
| 监督微调 | Supervised Fine-Tuning | SFT |
| 监督学习 | Supervised Learning | |
| 监督者架构 | Supervisor Architecture | |
| 支持集 | Support | |
| 替代目标 | Surrogate Objective | |
| Swish门控线性单元 | Swish-Gated Linear Unit | SwiGLU |
| 对称量化 | Symmetric Quantization | |
| 合成数据 | Synthetic Data | |
| 合成指令数据 | Synthetic Instruction Data | |
| 任务单元 | Task Unit | |
| 教师强制 | Teacher Forcing | |
| 温度缩放 | Temperature Scaling | |
| 温度混合 | Temperature-Based Mixing | |
| 时间对齐 | Temporal Alignment | |
| 时序差分 | Temporal Difference | TD |
| 张量核心 | Tensor Core | |
| 张量并行 | Tensor Parallelism | TP |
| 词项 | Term | |
| 测试污染 | Test Contamination | |
| 测试时计算 | Test-Time Compute | |
| 测试时计算 | Test-time Compute | |
| 威胁模型 | Threat Model | |
| 每输出词元时间 | Time per Output Token | TPOT |
| 平均输出词元时间 | Time per Output Token | TPOT |
| 首词元时延 | Time to First Token | TTFT |
| 首词元延迟 | Time to First Token | TTFT |
| 存活时间 | Time to Live | TTL |
| 检查与使用时差 | Time-of-Check to Time-of-Use | TOCTOU |
| 词元 F1 | Token F1 | |
| 词元物化 | Token Materialization | |
| 词元级蒸馏 | Token-level Distillation | |
| 删除标记 | Tombstone | |
| 工具调用 | Tool Calling | |
| 前 $k$ 项采样 | Top-k Sampling | |
| 总拥有成本 | Total Cost of Ownership | TCO |
| 训练块 | Training Block | |
| 轨迹 | Trajectory | |
| 轨迹回放 | Trajectory Replay | |
| 事务发件箱 | Transactional Outbox | |
| 平移等变性 | Translation Equivariance | |
| 信任边界 | Trust Boundary | |
| 二维位置插值 | Two-Dimensional Positional Interpolation | |
| 类型化记录 | Typed Record | |
| U形网络 | U-Net | |
| 非结构化剪枝 | Unstructured Pruning | |
| 无监督学习 | Unsupervised Learning | |
| 值 | Value | V |
| 信息价值 | Value of Information | VOI |
| 变分自编码器 | Variational Autoencoder | VAE |
| 向量雅可比积 | Vector--Jacobian Product | VJP |
| 速度参数化 | Velocity Parameterization | |
| 标签词 | Verbalizer | |
| 验证器 | Verifier | |
| 视觉 Transformer | Vision Transformer | ViT |
| 视觉语言模型 | Vision-Language Model | VLM |
| 视觉指令微调 | Visual Instruction Tuning | |
| 线程束 | Warp | |
| 弱扩展 | Weak Scaling | |
| WebDriver 双向协议 | WebDriver BiDi | |
| 权重绑定 | Weight Tying | |
| 权重分解低秩适配 | Weight-Decomposed Low-Rank Adaptation | DoRA |
| 仅权重量化 | Weight-Only Quantization | |
| 威尔逊区间 | Wilson Interval | |
| 工作流 | Workflow | |
| 工作记忆 | Working Memory | |
| 零点 | Zero Point | |
| 零冗余优化器 | Zero Redundancy Optimizer | ZeRO |
| 零样本提示 | Zero-Shot Prompting | |
| 零方差组 | Zero-variance Group |
大语言模型：从理论到实践