FlashAttention 是一种 IO 感知（IO-aware）的精确注意力（exact attention）算法，由 Tri Dao 等人在 2022 年的论文中提出。它通过分块（tiling）减少 GPU 高带宽显存（HBM）与片上 SRAM 之间的读写次数，在不牺牲模型质量的前提下加速 Transformer 训练。官方实现由 Dao-AILab 维护，并已演进到 FlashAttention-2、FlashAttention-3 与 FlashAttention-4。
思维链蒸馏（Chain-of-Thought Distillation，CoTD）是一种把大语言模型的逐步推理能力迁移到小模型上的训练方法：由大模型生成推理链，再用这些推理链微调小模型。ACL 2023 的 SCoTD 论文表明，参数量小几个数量级的学生模型也能从思维链提示中获益；LREC-COLING 2024 的 MoDE-CoTD 论文则提出用解耦的 LoRA 专家来缓解多任务蒸馏中的能力冲突。
分页注意力（PagedAttention）是一种受操作系统虚拟内存与分页机制启发而提出的注意力算法，用于管理大语言模型推理中的键值缓存（KV cache）内存。该算法由 arXiv 论文《Efficient Memory Management for Large Language Model Serving with PagedAttention》提出，并在其之上构建了 vLLM 服务系统。
对比解码（Contrastive Decoding）是一种免训练的解码方法，它通过同时使用一个强模型（专家）和一个弱模型（业余者），在生成时最大化两者对数似然之差，并施加合理性约束，从而改善开放式文本生成与推理任务的表现。该方法由 Li 等人于 2022 年提出，O'Brien 与 Lewis 于 2023 年将其扩展到推理任务。
KV 缓存（KV Cache）是自回归生成模型在逐 token 解码时复用键值向量计算结果的一种机制。Hugging Face Transformers 官方文档指出，它通过避免重复计算来降低计算开销、提升生成速度，并提供了多种 Cache 类以在显存占用与生成速度之间取舍。
上下文并行（Context Parallelism，CP）是 NVIDIA 官方文档描述的一种模型并行方法：它把神经网络激活值沿序列维度切分到多张 GPU 上，使每张 GPU 只保存一段序列的激活与必要的 Key/Value 对，从而降低长序列训练的内存占用与计算开销。
推测解码是一种加速自回归模型推理的采样算法，由 arXiv 论文《Fast Inference from Transformers via Speculative Decoding》提出。它借助更高效的近似模型先并行猜出若干候选词元，再由大模型并行校验，从而在不改变输出分布的前提下减少串行解码步数。
持续预训练（Continual Pre-training）指在已有语言模型的基础上，用新的（通常是无标注的）语料继续做预训练，使模型适配新领域或新任务，同时尽量不丢失原有能力。它要解决的核心问题是：如何让一个通用模型在获得新知识的同时，不被「灾难性遗忘」抹掉旧知识。
近端策略优化（PPO）是 John Schulman 等人在 2017 年论文中提出的一族策略梯度强化学习方法。它通过与环境的交互采样数据，再用随机梯度上升优化一个「替代」目标函数，其目标函数允许对同一批数据做多个轮次的小批量更新。论文称它具备信赖域策略优化（TRPO）的部分优点，但实现更简单、更通用，经验上的样本复杂度也更好。
无分类器引导（Classifier-Free Guidance，CFG）是一种在条件扩散模型与流匹配模型中权衡样本质量与多样性的引导方法。它由 Ho 与 Salimans 在论文《Classifier-Free Diffusion Guidance》中提出，核心做法是联合训练条件与无条件两个扩散模型，再组合两者的分数估计，从而无需额外训练图像分类器即可实现引导。
组相对策略优化（GRPO）是 DeepSeekMath 论文提出的一种强化学习后训练方法，属于近端策略优化（PPO）的变体。它通过对同一提示采样一组回答、用组内相对得分估计优势，从而在提升数学推理能力的同时降低 PPO 的内存开销。Hugging Face TRL 提供了对应的 GRPO Trainer 实现。
分组查询注意力（Grouped-Query Attention，GQA）是 arXiv 论文《GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints》提出的一种注意力机制，它让多个查询头共享同一组键值头，键值头数量介于多头注意力与多查询注意力之间，从而在解码速度与模型质量之间取得折中。
直接偏好优化（DPO）是一种用于语言模型对齐的训练方法，由 arXiv 论文《Direct Preference Optimization: Your Language Model is Secretly a Reward Model》提出。它把 RLHF 中「先训练奖励模型、再用强化学习优化策略」的两阶段流程，改写成一个只需偏好数据即可直接优化的分类式损失，从而省去显式奖励模型与训练中的在线采样。
Loop engineering is a new paradigm in AI engineering, referring to building autonomous, looping systems around large models, thus upgrading AI from single-response tools to long-term autonomous agents. Analogous to the PDCA cycle in traditional engineering, the model is the execution...
Harness Engineering is a new paradigm for AI engineering, which refers to building a complete operating system around a large model, upgrading it from a one-time response tool to a long-term autonomous agent.
UCP (Universal Commerce Protocol) is an open-source standard protocol jointly developed by Google and numerous e-commerce giants and payment platforms, aiming to promote the development of a new generation of proxy commerce. UCP uses standardized interfaces and protocols...
Rue is an experimental programming language developed by Rust core contributor Steve Klabnik using Claude. Rue explores ways to provide a more user-friendly, memory-safe programming experience than Rust without relying on garbage collection.
AI agents, also known as artificial intelligence agents, are a core concept in the field of artificial intelligence. They refer to intelligent entities or computational entities that can perceive their environment, make autonomous decisions, and take actions to achieve specific goals...
Artificial General Intelligence (AGI) is a form of artificial intelligence that possesses broad cognitive abilities comparable to or even surpassing those of humans. It can perform tasks and environments as effectively as humans...
Deep learning frameworks are a collection of software libraries and tools used to design, train, and deploy deep learning models. They provide the infrastructure needed to build complex neural networks, including data...
Federated learning is a distributed machine learning approach that allows multiple clients to collaboratively train models while maintaining data locality and avoiding direct sharing of raw data. Privacy-preserving federated learning...
Context-aware computing is an intelligent computing paradigm that can identify and understand a user's real-time environment, state, and activities, providing personalized and highly adaptive services. Context-aware computing...
Exploration and exploitation are two core concepts. Exploration refers to an agent trying new or unfamiliar actions to discover better behavioral strategies, while exploitation refers to an agent using known best practices...
Question answering systems (QA) are a type of information retrieval system in the field of artificial intelligence. They can understand questions posed by users in natural language and provide accurate and direct answers. They can process text, voice, etc.